A range hood control method and system based on Internet of Things

Abstract

The present invention provides a range hood control method and system based on the Internet of Things, which relates to the technical field of artificial intelligence and the Internet of Things, including acquiring multi-source hierarchical environmental characteristic data of heat, oil smoke and airflow by constructing a hierarchical data acquisition network, calculating the suction force required for each area and generating an adaptive fan control curve after edge computing and cloud server processing, and then generating an initial control parameter group including a main fan speed value, an auxiliary fan speed value and an air guide plate angle value; utilizing a cloud-edge collaborative architecture, constructing a vortex flow prediction model based on vorticity tensor and velocity divergence, predicting the airflow motion trajectory and calculating the real-time airflow deviation, dynamically adjusting the initial control parameter group to generate a target control parameter group, and finally continuously optimizing the prediction model through a deep reinforcement learning network.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and the Internet of Things, and in particular to a range hood control method and system based on the Internet of Things. Background Art

[0002] With the rapid development of Internet of Things technology and smart home appliances, the intelligent control of range hoods has become a key research direction in the industry. Traditional range hoods mainly rely on preset fixed gears for control and cannot be accurately adjusted according to actual cooking conditions. In recent years, researchers have proposed an intelligent control solution based on sensor networks, which can automatically adjust the fan speed by collecting environmental parameters such as oil fume concentration and temperature. However, due to the complexity and dynamic changes of the kitchen environment, the data collected by a single sensor is difficult to fully reflect the flow field state in the smoke guide cavity, which limits the control accuracy and effect of the range hood.

[0003] There are several deficiencies in the existing technology. The traditional single-layer sensor arrangement scheme cannot accurately capture the three-dimensional flow field characteristics in the smoke guide cavity, resulting in large errors in the prediction of the smoke diffusion trajectory; most of the existing control algorithms use a simple linear mapping relationship, which fails to fully consider the coupling between the thermal field, the smoke field and the airflow field, and is difficult to cope with the variable flow field under complex cooking conditions; there is a lack of effective prediction models and real-time optimization mechanisms, and the control system has poor adaptability to environmental changes, which is prone to problems such as smoke escape and energy waste.

[0004] In summary, a multi-source hierarchical environmental perception network is constructed to achieve accurate acquisition of flow field characteristics in the smoke guide cavity; an intelligent computing framework based on cloud-edge collaboration is established to improve the system's adaptability to complex cooking conditions; and a vortex flow prediction model and a deep reinforcement learning algorithm are used to achieve dynamic optimization of range hood control parameters, thereby improving oil fume purification efficiency and reducing energy consumption. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a range hood control method and system based on the Internet of Things, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a range hood control method based on the Internet of Things, comprising:

[0008] A hierarchical data acquisition network is constructed based on the Internet of Things communication protocol. The thermal distribution map and cooking fire power value are obtained through the bottom sensor array, the particle size distribution and concentration gradient of the oil fume particles are obtained through the middle sensor array, and the airflow velocity distribution matrix of the smoke guide cavity cross section is obtained through the upper sensor array. Multi-source hierarchical environmental feature data is determined, and the multi-source hierarchical environmental feature data is subjected to spatiotemporal fusion processing through the edge computing node to generate a cooking condition feature vector including thermal features, oil fume features and airflow features.

[0009] The cooking condition characteristic vector is uploaded to the cloud server, the rising speed of the hot air flow is calculated based on the thermal characteristics, the oil fume diffusion radius is calculated based on the oil fume characteristics, the airflow dead zone position and the eddy zone position in the smoke guide cavity are identified based on the airflow characteristics, and the flow field characteristic parameters are determined; a piecewise nonlinear optimization algorithm is executed using a distributed computing framework, and the required suction force of each area is calculated according to the flow field characteristic parameters, and an adaptive fan control curve is generated; based on the adaptive fan control curve, an initial airflow deviation is calculated, and an initial control parameter group is generated, wherein the initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value, and an initial air guide plate angle value;

[0010] Based on the cloud-edge collaborative architecture, the vorticity tensor of the airflow field in the smoke guide cavity is calculated using thermal characteristics and airflow characteristics, and the velocity divergence is calculated using oil smoke characteristics and airflow characteristics; a vortex flow prediction model is constructed on the edge side based on the vortex flow prediction model; the airflow motion trajectory is segmentedly predicted by the cloud server based on the vortex flow prediction model, and the real-time airflow deviation is calculated; the initial control parameter group is dynamically adjusted based on the real-time airflow deviation to generate a target control parameter group; the real-time airflow deviation and the target control parameter group are uploaded to the cloud-based deep reinforcement learning network to continuously optimize the vortex flow prediction model.

[0011] In an optional embodiment,

[0012] A hierarchical data acquisition network is constructed based on the IoT communication protocol. The thermal distribution map and cooking fire power value are obtained through the bottom sensor array. The particle size distribution and concentration gradient of the oil smoke particles are obtained through the middle sensor array. The airflow velocity distribution matrix of the smoke guide cavity cross section is obtained through the upper sensor array.

[0013] A hierarchical data acquisition network is constructed based on the Internet of Things communication protocol, and a multi-source sensor array is divided into a bottom sensor array, a middle sensor array and an upper sensor array. The bottom sensor array includes an infrared sensor and a flame spectrum sensor, the middle sensor array includes a scattering spectrum sensor and an absorption spectrum sensor, and the upper sensor array includes an annular array airflow sensor, which is composed of a plurality of hot wire wind speed sensors uniformly distributed along the circumference.

[0014] According to a preset first cycle, the temperature data matrix of the infrared sensor array is read, and the temperature data matrix is ​​expanded by a bicubic interpolation algorithm to generate a thermal distribution map; according to a preset second cycle, the spectrum waveform of the flame spectrum sensor is collected, the spectrum waveform is fast Fourier transformed to obtain a spectrum map, the main frequency corresponding to the maximum amplitude and the secondary frequency corresponding to the second maximum amplitude in the spectrum map are extracted, the main frequency and the secondary frequency are substituted into a pre-trained back propagation neural network, the thermal distribution map is used as an auxiliary input, and the cooking fire power value at the current moment is output;

[0015] Control the scattering spectrum sensor to emit a detection beam within a preset angle range and at a preset angle interval, collect light intensity data at multiple scattering angles, substitute the light intensity data into an inverse iteration algorithm, perform iterative calculations using the Mie scattering coefficient as an iteration parameter, and output the particle size distribution of the oil fume particles when the root mean square error between two adjacent iterations is less than a preset error threshold; control the absorption spectrum sensor to determine measurement points according to a preset height interval, measure transmittance at different heights, obtain transmittance data at multiple measurement points, calculate the oil fume concentration at each measurement point based on the Beer-Lambert law, use a cubic spline interpolation method to fit, obtain a continuous variation curve of the oil fume concentration with height, and calculate the first-order derivative of the continuous variation curve to obtain the oil fume concentration gradient;

[0016] The airflow measurement points are determined by the positions of multiple hot-wire wind speed sensors, and the airflow velocity values ​​are collected at each airflow measurement point at preset time intervals. The airflow velocity values ​​are Gaussian filtered to remove measurement noise to obtain filtered airflow velocity values, which are decomposed into radial components and tangential components. A polar coordinate grid is established on the cross-section of the smoke guide cavity, and the radial components and tangential components are extended to the grid points through radial basis function interpolation to generate an airflow velocity distribution matrix, and the curl field and divergence field of the airflow velocity distribution matrix are calculated.

[0017] In an optional embodiment,

[0018] Determining multi-source layered environmental feature data, performing spatiotemporal fusion processing on the multi-source layered environmental feature data through an edge computing node, and generating a cooking condition feature vector including thermal features, oil smoke features, and airflow features includes:

[0019] Based on the thermal distribution map, cooking fire power value, particle size distribution of oil fume particles, oil fume concentration gradient, and the curl field and divergence field of the airflow velocity distribution matrix, multi-source hierarchical environmental characteristic data are constructed;

[0020] Performing discrete wavelet transform on the thermal distribution map to obtain approximate components and detail components of different frequency bands, calculating the normalized energy of the detail components at each scale, and extracting the frequency band whose energy proportion exceeds a preset proportion threshold as the thermal distribution characteristic frequency band; performing continuous wavelet transform on the cooking fire power value, using Mexican hat wavelet to obtain time-frequency domain features, extracting the energy spectrum peak value and the corresponding characteristic scale in the time-frequency domain as the dynamic characteristics of firepower change, and combining the thermal distribution characteristic frequency band and the dynamic characteristics of firepower change to form a thermal characteristic;

[0021] Performing wavelet packet decomposition on the particle size distribution of the oil fume particles, adaptively selecting an optimal decomposition basis according to the Shannon entropy criterion, and extracting local features of the particle size distribution from the optimal decomposition basis nodes; performing nonlinear principal component analysis on the oil fume concentration gradient, mapping the oil fume concentration gradient to a high-dimensional feature space using a kernel function, extracting principal components as nonlinear features of concentration distribution, and combining the local features of the particle size distribution and the nonlinear features of the concentration distribution to form oil fume features;

[0022] Performing orthogonal empirical mode decomposition on the airflow velocity distribution data to obtain eigenmode functions of different scales, and calculating the instantaneous frequency characteristics and amplitude characteristics of the eigenmode functions; performing a two-dimensional Hilbert transform based on the curl field and divergence field of the airflow velocity distribution matrix, extracting local instantaneous phase characteristics and amplitude envelope characteristics, and combining the instantaneous frequency characteristics, the amplitude characteristics, the local instantaneous phase characteristics and the amplitude envelope characteristics to form airflow characteristics;

[0023] A multi-layer spatiotemporal feature fusion network is constructed through edge computing nodes. The multi-layer spatiotemporal feature fusion network includes a feature extraction layer and a feature fusion layer. The feature extraction layer includes a thermal feature sub-network, a fume feature sub-network, and an airflow feature sub-network. Each sub-network adopts an attention mechanism to adaptively weight the input thermal features, fume features, and airflow features to obtain weighted feature data;

[0024] In the feature fusion layer, a gated recurrent unit is used to extract time series features from the weighted feature data, and the association weights between different features are calculated through a cross-modal attention mechanism. Each feature data is mapped to a unified feature space according to the association weights to obtain unified feature data, and a graph convolutional network is used to model the unified feature data to obtain spatiotemporal dependent feature data;

[0025] The spatiotemporal dependency feature data is input into a fully connected layer for nonlinear mapping to generate a cooking condition feature vector.

[0026] In an optional embodiment,

[0027] The cooking condition characteristic vector is uploaded to the cloud server, the rising speed of the hot air flow is calculated based on the thermal characteristics, the fume diffusion radius is calculated based on the fume characteristics, and the airflow dead zone position and eddy zone position in the smoke guide cavity are identified based on the airflow characteristics. Determining the flow field characteristic parameters includes:

[0028] Upload the cooking condition feature vector to the cloud server;

[0029] The cloud server constructs a thermal gradient field based on the characteristic frequency band of thermal distribution, establishes a thermal buoyancy model in combination with the dynamic characteristics of firepower changes, takes the local temperature difference as the driving force, inputs the thermal gradient field and the thermal buoyancy model into the ideal gas state equation, and calculates the initial thermal airflow rise velocity field; generates a correction coefficient based on the temperature stratification effect, multiplies the correction coefficient by the initial thermal airflow rise velocity field, and obtains the corrected thermal airflow rise velocity field;

[0030] The cloud server constructs a multi-component diffusion model based on the local characteristics of the particle size distribution, determines the initial oil fume cluster diffusion coefficient according to the nonlinear characteristics of the concentration distribution, inputs the initial oil fume cluster diffusion coefficient into the Gaussian diffusion model, and calculates the multi-component diffusion flux; based on the multi-component diffusion flux, calculates the collision frequency function, multiplies the collision frequency function by the initial oil fume cluster diffusion coefficient, and obtains the corrected oil fume cluster diffusion coefficient; substitutes the corrected oil fume cluster diffusion coefficient into the diffusion equation, and solves to obtain the oil fume diffusion radius;

[0031] The cloud server inputs the instantaneous frequency characteristics and amplitude characteristics into the topological structure model to obtain the flow field topological structure, identifies the characteristic streamlines based on the local instantaneous phase characteristics and amplitude envelope characteristics, and inputs the characteristic streamlines into the critical point theoretical model to determine the saddle point position and the focal point position; calculates the Laplace operator of the velocity field based on the flow field topological structure, and compares the Laplace operator with the preset vorticity threshold to determine the airflow dead zone position in the smoke guide cavity; inputs the flow field topological structure into the Q criterion model and the λ2 criterion model to determine the vortex zone position in the smoke guide cavity.

[0032] In an optional embodiment,

[0033] A distributed computing framework is used to execute a piecewise nonlinear optimization algorithm. According to the flow field characteristic parameters, the required suction force of each area is calculated to generate an adaptive fan control curve. Based on the adaptive fan control curve, the initial airflow deviation is calculated, and an initial control parameter group is generated. The initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value, and an initial air guide plate angle value, including:

[0034] The cloud server uses the corrected hot air flow rising velocity field, oil smoke diffusion radius, airflow dead zone position and eddy zone position as flow field characteristic parameters, and inputs them into the flow field division module. Based on the velocity field gradient threshold, the flow field is divided into multiple sub-areas, and the sub-area boundary grid nodes and internal grid nodes are obtained.

[0035] The velocity field variance is calculated based on the velocity components of the internal grid nodes, and the velocity field variance is divided by the preset velocity field reference variance to obtain the airflow organization uniformity index; the modulus of the concentration field gradient is calculated based on the concentration values ​​of the internal grid nodes, and the modulus of the concentration field gradient is divided by the preset concentration gradient reference value to obtain the oil fume collection efficiency index; the fan input power is calculated, and the input power is divided by the preset power reference value to obtain the energy consumption index;

[0036] The first weighted term is obtained by multiplying the airflow organization uniformity index by the first preset weight coefficient, the second weighted term is obtained by multiplying the oil fume collection efficiency index by the second preset weight coefficient, and the third weighted term is obtained by multiplying the energy consumption index by the third preset weight coefficient; based on the first weighted term, the second weighted term and the third weighted term, a comprehensive objective function is obtained;

[0037] Based on the sub-region boundary grid nodes, boundary constraints are constructed, including velocity continuity constraints and pressure continuity constraints of the boundary grid nodes of adjacent sub-regions; the boundary constraints are combined with the comprehensive objective function to construct the Lagrangian function, and the quadratic penalty term is introduced into the Lagrangian function to obtain the augmented Lagrangian function;

[0038] Fix the initial value of the Lagrange multiplier, substitute the augmented Lagrangian function into the gradient descent algorithm, solve the local optimal solution of each sub-region, and obtain the sub-region control variable; based on the sub-region control variable, calculate the residual of the boundary constraint condition, multiply the residual by the step size factor, and update the Lagrange multiplier value;

[0039] Calculate the norm of the difference between the sub-region control variables of two adjacent iterations, stop the iteration when the norm is less than the preset convergence threshold, and convert the final sub-region control variables into the time-varying data points of the suction force in each region;

[0040] The time-varying data points of the extraction force are selected as nodes, the cubic spline basis function is constructed, and the cubic spline coefficients are solved by the least square method; the cubic spline coefficients are substituted into the cubic spline basis function to obtain the piecewise function expression;

[0041] The function value, first-order derivative value and second-order derivative value of the piecewise function expression at the node are calculated, and the continuity of the function value, first-order derivative value and second-order derivative value of adjacent piecewise functions is verified respectively; when the continuity verification passes, the piecewise function expression is determined as the adaptive fan control curve.

[0042] Based on the adaptive fan control curve, the cloud server compares the actual flow field with the ideal flow field through the flow field similarity principle, calculates the angle deviation and modulus deviation of the velocity vector field, and obtains the initial airflow deviation; the initial airflow deviation is input into the fuzzy control rule to generate an initial control parameter group, which includes the initial main fan speed value, the initial auxiliary fan speed value and the initial air guide plate angle value.

[0043] In an optional embodiment,

[0044] Based on the cloud-edge collaborative architecture, the vorticity tensor of the airflow field in the smoke guide cavity is calculated using the thermal characteristics and airflow characteristics, and the velocity divergence is calculated using the smoke characteristics and airflow characteristics; based on the vorticity tensor and the velocity divergence, a vortex flow prediction model is constructed on the edge side, including:

[0045] The characteristic frequency band of thermal distribution and the dynamic characteristics of firepower change in the thermal characteristics are input into the regularized spectrum analysis model to extract the dominant frequency components and amplitude components of the thermal field; the instantaneous frequency characteristics and amplitude characteristics in the airflow characteristics are input into the Hilbert-Huang transform model, and the intrinsic mode function of the airflow field is obtained by decomposition; based on the dominant frequency components and the intrinsic mode function, the thermal-airflow coupling tensor is constructed, and the vorticity tensor of the airflow field in the smoke guide cavity is calculated using the third-order compact difference format;

[0046] The local characteristics of particle size distribution and nonlinear characteristics of concentration distribution in the oil smoke characteristics are input into the continuity equation, and the velocity divergence of the airflow field in the smoke guide cavity is calculated by combining the local instantaneous phase characteristics in the airflow characteristics.

[0047] In the edge computing unit, a vortex particle method model is established based on the vorticity tensor, and the vorticity tensor is discretized into a set of vortex particles with intensity weights and kernel functions; the pressure Poisson's equation is constructed based on the velocity divergence, and the pressure field is solved by the multi-grid method; and a vortex flow prediction model is constructed according to the vortex particle set, the velocity divergence and the pressure field.

[0048] In an optional embodiment,

[0049] Based on the vortex flow prediction model, the cloud server predicts the airflow motion trajectory in segments and calculates the real-time airflow deviation; dynamically adjusts the initial control parameter group based on the real-time airflow deviation to generate a target control parameter group; uploads the real-time airflow deviation and the target control parameter group to the cloud deep reinforcement learning network, and continuously optimizes the vortex flow prediction model, including:

[0050] The vortex flow prediction model is transmitted to the cloud server, and a piecewise polynomial is used in the cloud server to describe the airflow motion trajectory; based on the vorticity tensor and pressure field at the current moment, the predicted velocity field at the next moment is calculated; the measured velocity field is obtained, the predicted velocity field is compared with the measured velocity field, and the direction deviation and amplitude deviation of the velocity vector are calculated to obtain the real-time airflow deviation degree;

[0051] A feedback correction matrix is ​​constructed according to the real-time airflow deviation, and the feedback correction matrix is ​​subjected to matrix multiplication operation with the initial control parameter group to obtain the corrected control parameters; based on the corrected control parameters, the flow field response characteristics are calculated, and the recursive least squares method is used to establish a dynamic mapping relationship between the corrected control parameters and the flow field response characteristics to generate the target control parameter group;

[0052] A dual-time deep neural network is constructed in the cloud server. The real-time airflow deviation is used as the state input, and the target control parameter group is used as the action input. Through temporal difference learning, the update strategy network parameters and the update value network parameters are obtained. The update strategy network parameters and the update value network parameters are passed to the vortex flow prediction model in the edge computing unit to achieve continuous optimization of the vortex flow prediction model.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] Provided is a range hood control system based on the Internet of Things, comprising:

[0055] The first unit is used to build a hierarchical data acquisition network based on the Internet of Things communication protocol, obtain the thermal distribution map and the cooking fire power value through the bottom sensor array, obtain the particle size distribution and concentration gradient of the oil fume particles through the middle sensor array, obtain the airflow velocity distribution matrix of the cross section of the smoke guide cavity through the upper sensor array, determine the multi-source hierarchical environmental feature data, perform spatiotemporal fusion processing on the multi-source hierarchical environmental feature data through the edge computing node, and generate a cooking condition feature vector including thermal features, oil fume features and airflow features;

[0056] The second unit is used to upload the cooking condition characteristic vector to the cloud server, calculate the rising speed of the hot air flow based on the thermal characteristics, calculate the oil fume diffusion radius based on the oil fume characteristics, identify the airflow dead zone position and the eddy zone position in the smoke guide cavity based on the airflow characteristics, and determine the flow field characteristic parameters; use a distributed computing framework to execute a piecewise nonlinear optimization algorithm, calculate the suction force required for each area according to the flow field characteristic parameters, and generate an adaptive fan control curve; calculate the initial airflow deviation based on the adaptive fan control curve, and generate an initial control parameter group, wherein the initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value, and an initial air guide plate angle value;

[0057] The third unit is used to calculate the vorticity tensor of the airflow field in the smoke guide cavity based on the cloud-edge collaborative architecture, using thermal characteristics and airflow characteristics, and calculate the velocity divergence using oil smoke characteristics and airflow characteristics; based on the vorticity tensor and the velocity divergence, a vortex flow prediction model is constructed on the edge side; based on the vortex flow prediction model, the airflow motion trajectory is segmentedly predicted by the cloud server, and the real-time airflow deviation is calculated; based on the real-time airflow deviation, the initial control parameter group is dynamically adjusted to generate a target control parameter group; the real-time airflow deviation and the target control parameter group are uploaded to the cloud-based deep reinforcement learning network to continuously optimize the vortex flow prediction model.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] A fourth aspect of the embodiments of the present invention is:

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] In an embodiment of the present invention, multi-dimensional environmental data such as heat, oil smoke and airflow are obtained through a multi-source layered sensor array, and combined with the powerful computing power of edge computing and cloud servers, precise control of the range hood fan and air guide plate is achieved, effectively avoiding the lag and inaccuracy of traditional range hood control methods; based on the cooking condition characteristic vector and flow field characteristic parameters, a piecewise nonlinear optimization algorithm is used to dynamically generate an adaptive fan control curve and a control parameter group, which can intelligently adjust the operating state of the range hood according to the real-time cooking situation to achieve optimal adaptability to different cooking scenes; the airflow motion trajectory is segmentedly predicted by the vortex flow prediction model, and the control parameters are dynamically adjusted based on the real-time airflow deviation, thereby achieving precise capture of oil smoke, avoiding energy waste, and improving the operating efficiency of the range hood. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of a range hood control method based on the Internet of Things according to an embodiment of the present invention;

[0067] Figure 2 The present invention is a schematic structural diagram of a range hood control system based on the Internet of Things. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of a range hood control method based on the Internet of Things according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] S101. Construct a hierarchical data acquisition network based on the Internet of Things communication protocol, obtain the thermal distribution map and cooking fire power value through the bottom sensor array, obtain the particle size distribution and concentration gradient of the oil fume particles through the middle sensor array, obtain the airflow velocity distribution matrix of the smoke guide cavity cross section through the upper sensor array, determine the multi-source hierarchical environmental feature data, perform spatiotemporal fusion processing on the multi-source hierarchical environmental feature data through the edge computing node, and generate a cooking condition feature vector including thermal features, oil fume features and airflow features;

[0072] In this embodiment, through the hierarchical data collection of the bottom, middle and upper sensor arrays, multi-dimensional environmental characteristic data such as thermal distribution, particle size and concentration of oil fume particles, and airflow velocity in the cooking process can be obtained comprehensively and accurately; the edge computing nodes are used to perform spatiotemporal fusion processing on multi-source data to achieve collaborative optimization between different data sources, which can more accurately describe the working condition characteristics of the cooking process; the generated cooking condition characteristic vector can provide real-time and intelligent cooking environment monitoring and control, and provide strong support for applications such as intelligent optimization of the cooking process and oil fume emission control; through a comprehensive grasp of thermal, oil fume and airflow characteristics, the energy utilization efficiency in the cooking process can be optimized, oil fume pollution can be reduced, and the safety and operation accuracy of cooking equipment can be improved.

[0073] S102. Upload the cooking condition characteristic vector to the cloud server, calculate the rising speed of the hot air flow based on the thermal characteristics, calculate the oil fume diffusion radius based on the oil fume characteristics, identify the airflow dead zone position and the eddy zone position in the smoke guide cavity based on the airflow characteristics, and determine the flow field characteristic parameters; use a distributed computing framework to execute a piecewise nonlinear optimization algorithm, calculate the required suction force of each area according to the flow field characteristic parameters, and generate an adaptive fan control curve; calculate the initial airflow deviation based on the adaptive fan control curve, and generate an initial control parameter group, the initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value and an initial air guide plate angle value;

[0074] In this embodiment, through analysis based on thermal, oil fume and airflow characteristics, the rising speed of the hot air flow, the oil fume diffusion radius, and the positions of the airflow dead zone and vortex zone can be accurately calculated, so as to fully grasp the airflow distribution and flow field characteristics during the cooking process; using a distributed computing framework and a nonlinear optimization algorithm, the fan control curve is adaptively generated according to the flow field characteristic parameters to ensure that the fan operation efficiency is maximized while avoiding energy waste; by calculating the suction force required for each area and adjusting the fan operating parameters, precise air extraction control can be achieved, the oil fume emission effect can be improved, and oil fume pollution can be reduced. At the same time, the energy consumption of the fan is optimized, and the efficiency of the overall cooking environment is improved; by generating an initial control parameter group and adjusting it in real time according to the deviation, the operating state of the fan system can be dynamically optimized, so as to achieve optimal air flow and oil fume emission control under different cooking conditions.

[0075] S103. Based on the cloud-edge collaborative architecture, the vorticity tensor of the airflow field in the smoke guide cavity is calculated using thermal characteristics and airflow characteristics, and the velocity divergence is calculated using oil smoke characteristics and airflow characteristics; based on the vorticity tensor and the velocity divergence, a vortex flow prediction model is constructed on the edge side; based on the vortex flow prediction model, the airflow motion trajectory is segmentedly predicted by the cloud server, and the real-time airflow deviation is calculated; based on the real-time airflow deviation, the initial control parameter group is dynamically adjusted to generate a target control parameter group; the real-time airflow deviation and the target control parameter group are uploaded to the cloud-based deep reinforcement learning network to continuously optimize the vortex flow prediction model.

[0076] In this embodiment, through the calculation of vorticity tensor and velocity divergence, combined with the vortex flow prediction model constructed on the edge side, accurate prediction of complex airflow movement in the smoke guide cavity is achieved; based on the real-time airflow deviation, the initial control parameter group is dynamically adjusted to generate a target control parameter group to ensure that the system can quickly adapt and optimize performance under different working conditions; by utilizing the collaborative architecture of cloud servers and edge computing, efficient segmented airflow prediction and parameter adjustment are achieved, which greatly improves the data processing speed and the responsiveness of the prediction model; through the continuous optimization of the cloud-based deep reinforcement learning network, the accuracy and adaptability of the vortex flow prediction model are continuously improved, so that the system can maintain efficient and precise airflow control in long-term operation; the optimized airflow control parameter group significantly improves the air purification efficiency of cooking equipment, reduces energy waste, and further improves the air quality of the cooking environment.

[0077] In an optional implementation, a hierarchical data acquisition network is constructed based on the Internet of Things communication protocol, and the thermal distribution map and cooking fire power value are obtained through the bottom sensor array, the particle size distribution and concentration gradient of the oil smoke particles are obtained through the middle sensor array, and the airflow velocity distribution matrix of the cross section of the smoke guide cavity is obtained through the upper sensor array, including:

[0078] A hierarchical data acquisition network is constructed based on the Internet of Things communication protocol, and a multi-source sensor array is divided into a bottom sensor array, a middle sensor array and an upper sensor array. The bottom sensor array includes an infrared sensor and a flame spectrum sensor, the middle sensor array includes a scattering spectrum sensor and an absorption spectrum sensor, and the upper sensor array includes an annular array airflow sensor, which is composed of a plurality of hot wire wind speed sensors uniformly distributed along the circumference.

[0079] According to a preset first cycle, the temperature data matrix of the infrared sensor array is read, and the temperature data matrix is ​​expanded by a bicubic interpolation algorithm to generate a thermal distribution map; according to a preset second cycle, the spectrum waveform of the flame spectrum sensor is collected, the spectrum waveform is fast Fourier transformed to obtain a spectrum map, the main frequency corresponding to the maximum amplitude and the secondary frequency corresponding to the second maximum amplitude in the spectrum map are extracted, the main frequency and the secondary frequency are substituted into a pre-trained back propagation neural network, the thermal distribution map is used as an auxiliary input, and the cooking fire power value at the current moment is output;

[0080] Control the scattering spectrum sensor to emit a detection beam within a preset angle range and at a preset angle interval, collect light intensity data at multiple scattering angles, substitute the light intensity data into an inverse iteration algorithm, perform iterative calculations using the Mie scattering coefficient as an iteration parameter, and output the particle size distribution of the oil fume particles when the root mean square error between two adjacent iterations is less than a preset error threshold; control the absorption spectrum sensor to determine measurement points according to a preset height interval, measure transmittance at different heights, obtain transmittance data at multiple measurement points, calculate the oil fume concentration at each measurement point based on the Beer-Lambert law, use a cubic spline interpolation method to fit, obtain a continuous variation curve of the oil fume concentration with height, and calculate the first-order derivative of the continuous variation curve to obtain the oil fume concentration gradient;

[0081] The airflow measurement points are determined by the positions of multiple hot-wire wind speed sensors, and the airflow velocity values ​​are collected at each airflow measurement point at preset time intervals. The airflow velocity values ​​are Gaussian filtered to remove measurement noise to obtain filtered airflow velocity values, which are decomposed into radial components and tangential components. A polar coordinate grid is established on the cross-section of the smoke guide cavity, and the radial components and tangential components are extended to the grid points through radial basis function interpolation to generate an airflow velocity distribution matrix, and the curl field and divergence field of the airflow velocity distribution matrix are calculated.

[0082] The bicubic interpolation algorithm specifically refers to a mathematical interpolation method based on 16 surrounding points, which is commonly used in image processing and data interpolation. Unlike linear interpolation, bicubic interpolation not only considers the value of the data point, but also the derivative of the data point, so it can provide a smoother interpolation effect. It is often used to estimate continuous values ​​from fewer known data points.

[0083] The inverse iterative algorithm specifically refers to a numerical method for solving matrix eigenvalues, which is usually used to solve linear algebra problems, such as eigenvalue calculation. The estimated eigenvalues ​​and eigenvectors are updated through repeated iterations until convergence.

[0084] The Mie scattering coefficient specifically refers to a physical quantity that describes the scattering of light in non-uniform materials (such as oil smoke particles). Mie scattering theory is mainly used to explain the relationship between particle size and scattered light intensity, and is usually used in atmospheric physics, meteorology and environmental monitoring.

[0085] The Beer-Lambert law specifically refers to a physical law used to describe the relationship between transmittance and substance concentration. Its mathematical form is: transmittance is exponentially related to the absorbance of the sample, the path length and the substance concentration. This law is widely used in chemical analysis, especially in spectral analysis to determine the concentration of a solution.

[0086] The curl field specifically refers to a vector field in which the curl represents a small rotation around the point. It is a vector that describes the degree of rotation of the flow around the point. The curl field is of great significance in fluid dynamics and can reflect the rotation of air or liquid flow.

[0087] The divergence field specifically refers to the quantity that describes the rate of change of volume at a point in a vector field, indicating the degree of diffusion or compression of the fluid or gas at that point. In fluid mechanics, a divergence of zero indicates that the flow of the fluid is sourceless and sinkless (i.e., mass conservation). The divergence field is very important in the analysis of air and liquid flows.

[0088] In an optional specific implementation, first, a data acquisition network is built. An IoT communication protocol that supports low power consumption and long-distance transmission, such as LoRaWAN or NB-IoT, is used to connect the bottom, middle and upper sensor arrays to the central processing unit. The bottom layer deploys an infrared sensor array and a flame spectrum sensor, the middle layer deploys a scattering spectrum sensor and an absorption spectrum sensor, and the upper layer deploys a circular array airflow sensor, which is composed of a plurality of hot wire wind speed sensors uniformly distributed along the circumference. The data of all sensors are transmitted to the central processing unit for processing through the IoT communication protocol.

[0089] Next, obtain the thermal distribution map and cooking fire power value. Read the temperature data matrix of the infrared sensor array once per second, for example, a 5x5 matrix, each element represents the temperature value of a point, for example: [[200, 210, 220, 215, 205], [205, 215, 225, 220, 210], [210, 220, 230, 225, 215], [205, 215, 225, 220, 210], [200, 210, 220, 215, 205]]. Use the bicubic interpolation algorithm to expand this 5x5 matrix to a 10x10 matrix to generate a more detailed thermal distribution map. At the same time, collect the data of the flame spectrum sensor at a frequency of once every two seconds to obtain the spectrum waveform. Perform fast Fourier transform on the collected spectrum waveform to obtain the spectrum map. Assume that the main frequency corresponding to the maximum amplitude in the spectrum graph is 100Hz, and the secondary frequency corresponding to the second maximum amplitude is 50Hz. These two frequency values ​​and the thermal distribution map are input into a pre-trained back propagation neural network, which has been trained using a large amount of flame spectrum data and corresponding power values. The neural network outputs the cooking fire power value at the current moment, for example, 1.5kW.

[0090] Then, the particle size distribution and concentration gradient of the oil fume particles are obtained. The scattering spectrum sensor is controlled to emit a detection beam within the range of 0 to 90 degrees at intervals of 5 degrees. The light intensity data at different scattering angles are collected. Assume that the light intensity at an angle of 30 degrees is 100 lux, the light intensity at an angle of 35 degrees is 90 lux, and so on. Substitute these light intensity data into the reverse iteration algorithm, and use the Mie scattering coefficient as the iteration parameter for iterative calculation. Set the error threshold to 0.01. When the root mean square error of two adjacent iterations is less than 0.01, the particle size distribution of the oil fume particles is output. For example, most of the particle diameters are between 0.5 microns and 1 micron. Control the absorption spectrum sensor to measure the transmittance at different heights from the stove with a height interval of 1 cm. Assume that the transmittance at a height of 10 cm is 90%, the transmittance at a height of 11 cm is 85%, and so on. Based on the Beer-Lambert law, the oil fume concentration at each measurement point is calculated. The cubic spline interpolation method is used to fit these concentration values ​​to obtain a continuous change curve of oil fume concentration with height. Calculate the slope of this curve to get the oil smoke concentration gradient. For example, the concentration gradient at a height of 10 cm decreases by 5% per cm.

[0091] Finally, the airflow velocity distribution matrix of the smoke guide cavity cross section is obtained. The airflow measurement points are determined by the positions of multiple hot wire wind speed sensors. The airflow velocity values ​​are collected at each measurement point at a time interval of 0.1 seconds. The collected airflow velocity values ​​are Gaussian filtered to remove measurement noise. The filtered airflow velocity values ​​are decomposed into radial components and tangential components. A polar coordinate grid is established in the cross section of the smoke guide cavity. The radial basis function interpolation method is used to extend the radial and tangential components to the grid points to generate an airflow velocity distribution matrix. The curl field and divergence field of the matrix are calculated to analyze the rotation and convergence of the airflow.

[0092] In this embodiment, by real-time monitoring of the thermal distribution diagram and the cooking fire power value, the fire power can be accurately controlled to avoid energy waste and improve cooking efficiency; by monitoring the particle size distribution and concentration gradient of the oil fume particles, the design and operation of the oil fume purifier can be optimized, the oil fume can be effectively removed, and the kitchen environment can be improved; by obtaining the air flow velocity distribution matrix of the smoke guide cavity cross-section, the design of the smoke exhaust duct can be optimized, the smoke exhaust efficiency can be improved, and the oil fume residue can be reduced.

[0093] In an optional implementation, determining multi-source hierarchical environmental feature data, performing spatiotemporal fusion processing on the multi-source hierarchical environmental feature data through an edge computing node, and generating a cooking condition feature vector including thermal features, oil smoke features, and airflow features includes:

[0094] Based on the thermal distribution map, cooking fire power value, particle size distribution of oil fume particles, oil fume concentration gradient, and the curl field and divergence field of the airflow velocity distribution matrix, multi-source hierarchical environmental characteristic data are constructed;

[0095] Performing discrete wavelet transform on the thermal distribution map to obtain approximate components and detail components of different frequency bands, calculating the normalized energy of the detail components at each scale, and extracting the frequency band whose energy proportion exceeds a preset proportion threshold as the thermal distribution characteristic frequency band; performing continuous wavelet transform on the cooking fire power value, using Mexican hat wavelet to obtain time-frequency domain features, extracting the energy spectrum peak value and the corresponding characteristic scale in the time-frequency domain as the dynamic characteristics of firepower change, and combining the thermal distribution characteristic frequency band and the dynamic characteristics of firepower change to form a thermal characteristic;

[0096] Performing wavelet packet decomposition on the particle size distribution of the oil fume particles, adaptively selecting an optimal decomposition basis according to the Shannon entropy criterion, and extracting local features of the particle size distribution from the optimal decomposition basis nodes; performing nonlinear principal component analysis on the oil fume concentration gradient, mapping the oil fume concentration gradient to a high-dimensional feature space using a kernel function, extracting principal components as nonlinear features of concentration distribution, and combining the local features of the particle size distribution and the nonlinear features of the concentration distribution to form oil fume features;

[0097] Performing orthogonal empirical mode decomposition on the airflow velocity distribution data to obtain eigenmode functions of different scales, and calculating the instantaneous frequency characteristics and amplitude characteristics of the eigenmode functions; performing a two-dimensional Hilbert transform based on the curl field and divergence field of the airflow velocity distribution matrix, extracting local instantaneous phase characteristics and amplitude envelope characteristics, and combining the instantaneous frequency characteristics, the amplitude characteristics, the local instantaneous phase characteristics and the amplitude envelope characteristics to form airflow characteristics;

[0098] A multi-layer spatiotemporal feature fusion network is constructed through edge computing nodes. The multi-layer spatiotemporal feature fusion network includes a feature extraction layer and a feature fusion layer. The feature extraction layer includes a thermal feature sub-network, a fume feature sub-network, and an airflow feature sub-network. Each sub-network adopts an attention mechanism to adaptively weight the input thermal features, fume features, and airflow features to obtain weighted feature data;

[0099] In the feature fusion layer, a gated recurrent unit is used to extract time series features from the weighted feature data, and the association weights between different features are calculated through a cross-modal attention mechanism. Each feature data is mapped to a unified feature space according to the association weights to obtain unified feature data, and a graph convolutional network is used to model the unified feature data to obtain spatiotemporal dependent feature data;

[0100] The spatiotemporal dependency feature data is input into a fully connected layer for nonlinear mapping to generate a cooking condition feature vector.

[0101] In an optional specific implementation, first, multi-source hierarchical environmental characteristic data are collected through various sensors, including a thermal distribution map, a cooking fire power value, a particle size distribution of oil fume particles, an oil fume concentration gradient, and a velocity matrix reflecting the airflow velocity distribution. For example, an infrared thermal imager can be used to obtain a thermal distribution map of the kitchen space, a power meter can be used to measure the cooking fire power value, a particle sensor can be used to measure the particle size distribution and concentration gradient of oil fume particles, and an ultrasonic anemometer can be used to measure the airflow velocity distribution matrix.

[0102] Next, feature extraction is performed on the collected multi-source data. For thermal characteristics, discrete wavelet transform is performed on the thermal distribution map, decomposed into components of different frequencies, and the energy proportion of each frequency component is calculated. The frequency components whose energy proportion exceeds the preset threshold (for example, 80%) are selected as the thermal distribution characteristic frequency band. At the same time, continuous wavelet transform is performed on the cooking fire power value, its time-frequency domain characteristics are analyzed, and the energy spectrum peak and the corresponding characteristic scale are extracted as the dynamic characteristics of firepower change. The thermal distribution characteristic frequency band and the dynamic characteristics of firepower change are combined to form thermal characteristics. For example, assuming that the main energy of the thermal distribution map is concentrated in the low-frequency part, the low-frequency component is extracted as the thermal distribution characteristic frequency band; assuming that the cooking fire power value has an energy peak at a certain time point, the characteristic scale corresponding to the time point is extracted.

[0103] For the characteristics of oil smoke, the particle size distribution of oil smoke particles is decomposed by wavelet packets, the optimal decomposition basis is selected according to the Shannon entropy criterion, and the local characteristics of particle size distribution are extracted from the optimal decomposition basis nodes. At the same time, the oil smoke concentration gradient is subjected to nonlinear principal component analysis, which is mapped to a high-dimensional feature space, and the principal components are extracted as nonlinear characteristics of concentration distribution. The local characteristics of particle size distribution and the nonlinear characteristics of concentration distribution are combined to form the characteristics of oil smoke. For example, assuming that the particle size of oil smoke particles is mainly concentrated in a certain range, the local characteristics of particle size distribution within this range are extracted; assuming that the oil smoke concentration gradient has a nonlinear change law, its main characteristics are extracted through nonlinear principal component analysis.

[0104] For airflow characteristics, orthogonal empirical mode decomposition is performed on the airflow velocity distribution data to obtain the intrinsic mode functions of different scales, and their instantaneous frequency characteristics and amplitude characteristics are calculated. At the same time, based on the curl field and divergence field of the airflow velocity distribution matrix, a two-dimensional Hilbert transform is performed to extract the local instantaneous phase characteristics and amplitude envelope characteristics. The instantaneous frequency characteristics, amplitude characteristics, local instantaneous phase characteristics and amplitude envelope characteristics are combined to form airflow characteristics. For example, assuming that the airflow velocity has periodic fluctuations, the corresponding instantaneous frequency and amplitude characteristics are extracted; assuming that there is a local vortex in the airflow, the phase and amplitude envelope characteristics are extracted through the two-dimensional Hilbert transform of the curl field and divergence field.

[0105] Subsequently, a multi-layer spatiotemporal feature fusion network is constructed on the edge computing node. The network includes a feature extraction layer and a feature fusion layer. The feature extraction layer includes a thermal feature subnetwork, a fume feature subnetwork, and an airflow feature subnetwork, which adaptively weight the input thermal features, fume features, and airflow features, respectively. The feature fusion layer uses a gated recurrent unit to extract temporal features, and calculates the association weights between different features through a cross-modal attention mechanism, mapping each feature data to a unified feature space. Then, a graph convolutional network is used to model the unified feature data to obtain spatiotemporal dependent feature data. Finally, the spatiotemporal dependent feature data is input into the fully connected layer for nonlinear mapping to generate a cooking condition feature vector.

[0106] In this embodiment, by fusing multi-source hierarchical environmental feature data, various states in the cooking process can be more comprehensively characterized, thereby improving the accuracy of cooking condition identification; data processing through edge computing nodes can reduce data transmission volume and delay, reduce calculation complexity, and improve real-time performance; the modular network structure design makes the system easy to expand, and new sensors and feature extraction modules can be easily added to adapt to different cooking scenarios and needs.

[0107] In an optional embodiment, the cooking condition characteristic vector is uploaded to a cloud server, the rising speed of the hot air flow is calculated based on the thermal characteristics, the fume diffusion radius is calculated based on the fume characteristics, and the airflow dead zone position and the eddy zone position in the smoke guide cavity are identified based on the airflow characteristics. Determining the flow field characteristic parameters includes:

[0108] Upload the cooking condition feature vector to the cloud server;

[0109] The cloud server constructs a thermal gradient field based on the characteristic frequency band of thermal distribution, establishes a thermal buoyancy model in combination with the dynamic characteristics of firepower changes, takes the local temperature difference as the driving force, inputs the thermal gradient field and the thermal buoyancy model into the ideal gas state equation, and calculates the initial thermal airflow rise velocity field; generates a correction coefficient based on the temperature stratification effect, multiplies the correction coefficient by the initial thermal airflow rise velocity field, and obtains the corrected thermal airflow rise velocity field;

[0110] The cloud server constructs a multi-component diffusion model based on the local characteristics of the particle size distribution, determines the initial oil fume cluster diffusion coefficient according to the nonlinear characteristics of the concentration distribution, inputs the initial oil fume cluster diffusion coefficient into the Gaussian diffusion model, and calculates the multi-component diffusion flux; based on the multi-component diffusion flux, calculates the collision frequency function, multiplies the collision frequency function by the initial oil fume cluster diffusion coefficient, and obtains the corrected oil fume cluster diffusion coefficient; substitutes the corrected oil fume cluster diffusion coefficient into the diffusion equation, and solves to obtain the oil fume diffusion radius;

[0111] The cloud server inputs the instantaneous frequency characteristics and amplitude characteristics into the topological structure model to obtain the flow field topological structure, identifies the characteristic streamlines based on the local instantaneous phase characteristics and amplitude envelope characteristics, and inputs the characteristic streamlines into the critical point theoretical model to determine the saddle point position and the focal point position; calculates the Laplace operator of the velocity field based on the flow field topological structure, and compares the Laplace operator with the preset vorticity threshold to determine the airflow dead zone position in the smoke guide cavity; inputs the flow field topological structure into the Q criterion model and the λ2 criterion model to determine the vortex zone position in the smoke guide cavity.

[0112] The collision frequency function specifically describes the frequency of collisions of particles (such as oil smoke particles) per unit time. This function is often used to describe the collision behavior of gas molecules, particulate matter, etc., and is associated with characteristic parameters such as diffusion coefficient.

[0113] The characteristic streamlines specifically refer to lines used to represent the trajectory of fluid motion in fluid flow. Streamlines are tangents to the fluid velocity vector, which help visualize the direction and structure of fluid flow. In airflow, characteristic streamlines usually indicate the mainstream direction or significant features of the airflow.

[0114] The critical point theory model is specifically used to describe the existence and characteristics of critical points (such as saddle points and focal points) in the flow field under specific conditions in fluid dynamics. The critical point refers to the point where physical quantities such as velocity or pressure in the fluid flow change dramatically, and is often used to analyze special structures and movements in the airflow.

[0115] The saddle point specifically refers to a point in a flow field or potential energy field that is a maximum in some directions and a minimum in other directions. In airflow analysis, a saddle point often indicates the particularity of a flow field, such as the center of a vortex or a turning point in airflow distribution.

[0116] The focus specifically refers to a point in a flow field or potential energy field, near which the flow of a fluid tends to converge. The focus usually appears in flow field analysis in fluid dynamics, indicating a convergence or acceleration area of ​​a fluid or airflow.

[0117] The vorticity specifically refers to a quantity in fluid dynamics that describes the intensity of rotation or vortex in a fluid. Vorticity is the curl of the velocity field, which can indicate the degree of rotation of the fluid at a certain point and is an important physical quantity for analyzing fluid vortex and rotational flow.

[0118] The Q criterion specifically refers to a criterion for identifying vortex regions in a flow field. The Q criterion determines whether there is a vortex by calculating the strain and rotational components of the flow field, and is commonly used in fluid dynamics and airflow simulation.

[0119] The λ2 criterion specifically refers to a flow field analysis tool for identifying vortex regions. It defines vortex regions based on the eigenvalues ​​of the velocity gradient tensor. The λ2 criterion can accurately identify vortices and dead zones in the flow field and is widely used in fluid dynamics and computational fluid dynamics (CFD) research.

[0120] In an optional specific implementation, first, cooking condition characteristic data is collected. The temperature, fume concentration, airflow speed and other data of the cooking process are collected in real time using sensors, and the data are converted into feature vectors. For example, data can be collected once per second, and the data within 10 seconds can be integrated into a feature vector, which contains data of multiple dimensions such as temperature, fume concentration, and airflow speed.

[0121] Then, the cooking condition feature vector is uploaded to the cloud server, which receives data from various sensors and stores and processes them.

[0122] Next, the cloud server calculates the rising speed of the thermal airflow. The server analyzes the frequency band of the thermal distribution characteristics and constructs the thermal gradient field. For example, the temperature data is decomposed according to the frequency to obtain the temperature change at different frequencies, so as to construct the thermal gradient field. At the same time, a thermal buoyancy model is established in combination with the dynamic characteristics of the firepower change. For example, a thermal buoyancy model is established according to the power change of the stove. The thermal gradient field and the thermal buoyancy model are input into the equivalent calculation method of the ideal gas state equation to calculate the initial thermal airflow rising speed field. For example, the initial thermal airflow rising speed at different positions is calculated based on parameters such as temperature difference and gas density. Taking into account the influence of the temperature stratification effect, a correction coefficient is generated. For example, a correction coefficient is generated based on the temperature difference at different heights. The correction coefficient is multiplied by the initial thermal airflow rising speed field to obtain a more accurate corrected thermal airflow rising speed field. For example, the initial speed is multiplied by the correction coefficient to obtain the final rising speed.

[0123] At the same time, the cloud server calculates the diffusion radius of the oil fume. The server analyzes the local characteristics of the oil fume particle size distribution and constructs a multi-component diffusion model. For example, a multi-component diffusion model is constructed based on the concentration distribution of oil fume particles of different particle sizes. The diffusion coefficient of the initial oil fume cluster is determined based on the nonlinear characteristics of the concentration distribution. For example, the initial diffusion coefficient is determined based on the change of the oil fume concentration with distance. The initial oil fume cluster diffusion coefficient is input into the equivalent calculation method of the Gaussian diffusion model to calculate the multi-component diffusion flux. For example, the diffusion flux in different directions is calculated based on the diffusion coefficient and the concentration gradient. Based on the multi-component diffusion flux, the collision frequency function is calculated. For example, the collision frequency function is calculated based on the collision probability between different oil fume particles. The collision frequency function is multiplied by the initial oil fume cluster diffusion coefficient to obtain the modified oil fume cluster diffusion coefficient. For example, the initial diffusion coefficient is multiplied by the collision frequency function to obtain the modified diffusion coefficient. The modified oil fume cluster diffusion coefficient is substituted into the equivalent calculation method of the diffusion equation to solve the oil fume diffusion radius. For example, the diffusion radius of the oil fume is calculated based on the modified diffusion coefficient and time. Assuming that the initial oil smoke diffusion coefficient is 0.1 and the collision frequency function is 0.9, the modified oil smoke diffusion coefficient is 0.09. After substituting it into the diffusion equation, the oil smoke diffusion radius is calculated to be 0.5 meters.

[0124] In addition, the cloud server identifies the position of the airflow dead zone and the eddy zone in the smoke guide cavity. The instantaneous frequency characteristics and amplitude characteristics are input into the topological structure model to obtain the flow field topological structure. For example, the frequency and amplitude information of the airflow velocity are input into the model to obtain the topological structure of the flow field. Characteristic streamlines are identified based on local instantaneous phase characteristics and amplitude envelope characteristics. For example, characteristic streamlines are identified according to the phase and amplitude changes of the airflow velocity. The characteristic streamlines are input into the critical point theory model to determine the saddle point position and the focal point position. For example, the positions of the saddle point and the focal point are determined according to the morphology of the characteristic streamlines. The Laplace operator of the velocity field is calculated based on the flow field topological structure. For example, the second-order derivative of the velocity field is calculated according to the topological structure of the flow field. The calculation result is compared with the preset vortex threshold to determine the position of the airflow dead zone in the smoke guide cavity. For example, if the calculation result is less than the threshold, the area is considered to be an airflow dead zone. The flow field topological structure is input into the Q criterion model and the λ2 criterion model to determine the position of the eddy zone in the smoke guide cavity. For example, the location of the eddy current zone can be determined based on the topological structure and velocity distribution of the flow field.

[0125] In this embodiment, by accurately calculating the rising speed of the hot air flow and the oil fume diffusion radius, the oil fume can be more effectively guided into the purification device, thereby improving the oil fume purification efficiency; by identifying the position of the airflow dead zone and the eddy zone in the smoke guide cavity, the design of the smoke guide cavity can be optimized, the generation of airflow dead zones and eddies can be reduced, and the oil fume purification efficiency can be further improved; by performing data processing and calculation on a cloud server, intelligent control of the oil fume purification process can be achieved, and the purification parameters can be automatically adjusted according to different cooking conditions to achieve the best purification effect.

[0126] In an optional implementation, a distributed computing framework is used to execute a piecewise nonlinear optimization algorithm, and the required suction force of each area is calculated according to the flow field characteristic parameters to generate an adaptive fan control curve; based on the adaptive fan control curve, the initial airflow deviation is calculated, and an initial control parameter group is generated, and the initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value, and an initial guide plate angle value, including:

[0127] The cloud server uses the corrected hot air flow rising velocity field, oil smoke diffusion radius, airflow dead zone position and eddy zone position as flow field characteristic parameters, and inputs them into the flow field division module. Based on the velocity field gradient threshold, the flow field is divided into multiple sub-areas, and the sub-area boundary grid nodes and internal grid nodes are obtained.

[0128] The velocity field variance is calculated based on the velocity components of the internal grid nodes, and the velocity field variance is divided by the preset velocity field reference variance to obtain the airflow organization uniformity index; the modulus of the concentration field gradient is calculated based on the concentration values ​​of the internal grid nodes, and the modulus of the concentration field gradient is divided by the preset concentration gradient reference value to obtain the oil fume collection efficiency index; the fan input power is calculated, and the input power is divided by the preset power reference value to obtain the energy consumption index;

[0129] The first weighted term is obtained by multiplying the airflow organization uniformity index by the first preset weight coefficient, the second weighted term is obtained by multiplying the oil fume collection efficiency index by the second preset weight coefficient, and the third weighted term is obtained by multiplying the energy consumption index by the third preset weight coefficient; based on the first weighted term, the second weighted term and the third weighted term, a comprehensive objective function is obtained;

[0130] Based on the sub-region boundary grid nodes, boundary constraints are constructed, including velocity continuity constraints and pressure continuity constraints of the boundary grid nodes of adjacent sub-regions; the boundary constraints are combined with the comprehensive objective function to construct the Lagrangian function, and the quadratic penalty term is introduced into the Lagrangian function to obtain the augmented Lagrangian function;

[0131] Fix the initial value of the Lagrange multiplier, substitute the augmented Lagrangian function into the gradient descent algorithm, solve the local optimal solution of each sub-region, and obtain the sub-region control variable; based on the sub-region control variable, calculate the residual of the boundary constraint condition, multiply the residual by the step size factor, and update the Lagrange multiplier value;

[0132] Calculate the norm of the difference between the sub-region control variables of two adjacent iterations, stop the iteration when the norm is less than the preset convergence threshold, and convert the final sub-region control variables into the time-varying data points of the suction force in each region;

[0133] The time-varying data points of the extraction force are selected as nodes, the cubic spline basis function is constructed, and the cubic spline coefficients are solved by the least square method; the cubic spline coefficients are substituted into the cubic spline basis function to obtain the piecewise function expression;

[0134] Calculate the function value, first-order derivative value and second-order derivative value of the piecewise function expression at the node, and verify the continuity of the function value, first-order derivative value and second-order derivative value of the adjacent piecewise functions respectively; when the continuity verification is passed, determine the piecewise function expression as the adaptive fan control curve;

[0135] Based on the adaptive fan control curve, the cloud server compares the actual flow field with the ideal flow field through the flow field similarity principle, calculates the angle deviation and modulus deviation of the velocity vector field, and obtains the initial airflow deviation; the initial airflow deviation is input into the fuzzy control rule to generate an initial control parameter group, which includes the initial main fan speed value, the initial auxiliary fan speed value and the initial air guide plate angle value.

[0136] In an optional specific implementation, the flow field is optimized in piecewise nonlinear manner to generate an adaptive fan control curve, and initial control parameters are calculated based on the curve to achieve precise control of the fan and air guide plates, ultimately achieving a highly efficient and energy-saving ventilation effect.

[0137] First, the cloud server receives real-time data from sensors, including flow field characteristic parameters such as the hot air flow rising velocity field, oil smoke diffusion radius, airflow dead zone position and vortex zone position.

[0138] Then, the flow field division module divides the entire flow field into several sub-areas according to the rate of change of the velocity field. The division process is based on a preset velocity field gradient threshold. When the velocity difference between adjacent positions exceeds the threshold, the two positions are considered to belong to different sub-areas. For example, the velocity field gradient threshold is set to 0.5m / s / m. If the velocity difference between two adjacent positions is 0.6m / s / m, the two positions are divided into different sub-areas. After division, the coordinate information of the boundary grid nodes and internal grid nodes of each sub-area is recorded.

[0139] Next, each sub-area is evaluated and three indicators are calculated: airflow organization uniformity index, oil fume collection efficiency index and energy consumption index. The airflow organization uniformity index is measured by calculating the degree of change of the velocity component of the grid nodes inside the sub-area. The greater the degree of change, the greater the index value. The oil fume collection efficiency index is measured by calculating the oil fume concentration gradient of the grid nodes inside the sub-area. The greater the gradient, the greater the index value. The energy consumption index is measured by calculating the input power of the fan. The greater the power, the greater the index value. Assume that the variance of the velocity component of the grid nodes inside a sub-area is 0.1 (m / s) 2 , the preset velocity field reference variance is 0.05 (m / s) 2 , then the airflow uniformity index is 0.1 / 0.05=2. Assuming the oil smoke concentration gradient modulus is 0.2mg / m 4 / m, the preset concentration gradient reference value is 0.1mg / m 4 / m, the oil fume collection efficiency index is 0.2 / 0.1=2. Assuming the fan input power is 100W and the preset power reference value is 50W, the energy consumption index is 100 / 50=2.

[0140] After that, the three indicators are multiplied by the corresponding preset weight coefficients, and the results are added to obtain the comprehensive objective function. For example, assuming that the weight coefficient of the airflow organization uniformity index is 0.4, the weight coefficient of the oil fume collection efficiency index is 0.5, and the weight coefficient of the energy consumption index is 0.1, the value of the comprehensive objective function is 2×0.4+2×0.5+2×0.1=2.

[0141] Then, the augmented Lagrangian function is constructed. This function contains the comprehensive objective function, boundary constraints and penalty terms. The boundary constraints ensure the continuity of velocity and pressure of the boundary grid nodes of adjacent sub-regions. The penalty term is used to penalize the violation of boundary constraints.

[0142] Next, the gradient descent algorithm is used to iteratively solve the local optimal solution of each sub-region to obtain the sub-region control variable, that is, the required suction force of each sub-region. During the iteration process, the Lagrange multiplier is continuously updated until the norm of the difference between the sub-region control variables of two adjacent iterations is less than the preset convergence threshold. Assuming that the preset convergence threshold is 0.01, the iteration is stopped when the norm of the difference between the sub-region control variables of two iterations is less than 0.01.

[0143] After converting the final sub-region control variables into the time-varying data points of the extraction force in each region, the cubic spline interpolation method is used to fit these discrete data points into a smooth curve, namely the adaptive fan control curve.

[0144] Finally, the cloud server calculates the initial airflow deviation by comparing the actual flow field with the ideal flow field based on the adaptive fan control curve. The initial airflow deviation is input into the fuzzy control rule library to generate the initial control parameter group, including the initial main fan speed value, the initial auxiliary fan speed value and the initial air guide angle value.

[0145] In this embodiment, by finely controlling the flow field, the suction force is more concentrated on the oil fume generating area, thereby improving the oil fume collection efficiency and reducing the diffusion of oil fume; it can effectively avoid the generation of airflow dead zones and eddies, making the airflow distribution more uniform and creating a more comfortable kitchen environment; by adaptively controlling the fan speed and the angle of the air guide plate, unnecessary energy waste is avoided and the goal of energy saving and consumption reduction is achieved.

[0146] In an optional implementation, based on the cloud-edge collaborative architecture, the vorticity tensor of the airflow field in the smoke guide cavity is calculated using the thermal characteristics and the airflow characteristics, and the velocity divergence is calculated using the oil smoke characteristics and the airflow characteristics; based on the vorticity tensor and the velocity divergence, constructing a vortex flow prediction model on the edge side includes:

[0147] The characteristic frequency band of thermal distribution and the dynamic characteristics of firepower change in the thermal characteristics are input into the regularized spectrum analysis model to extract the dominant frequency components and amplitude components of the thermal field; the instantaneous frequency characteristics and amplitude characteristics in the airflow characteristics are input into the Hilbert-Huang transform model, and the intrinsic mode function of the airflow field is obtained by decomposition; based on the dominant frequency components and the intrinsic mode function, the thermal-airflow coupling tensor is constructed, and the vorticity tensor of the airflow field in the smoke guide cavity is calculated using the third-order compact difference format;

[0148] The local characteristics of particle size distribution and nonlinear characteristics of concentration distribution in the oil smoke characteristics are input into the continuity equation, and the velocity divergence of the airflow field in the smoke guide cavity is calculated by combining the local instantaneous phase characteristics in the airflow characteristics.

[0149] In the edge computing unit, a vortex particle method model is established based on the vorticity tensor, and the vorticity tensor is discretized into a set of vortex particles with intensity weights and kernel functions; the pressure Poisson's equation is constructed based on the velocity divergence, and the pressure field is solved by the multi-grid method; and a vortex flow prediction model is constructed according to the vortex particle set, the velocity divergence and the pressure field.

[0150] The compact difference format is a finite difference method for numerical calculation, usually used for high-precision partial differential equation solutions. Compared with traditional difference formats, the compact difference format can improve the accuracy and efficiency of calculations, and is particularly suitable for vorticity and flow field characteristics in computational fluid dynamics.

[0151] The vorticity tensor is a mathematical tool for describing the vortex characteristics in a fluid. It is a generalization of vorticity (the degree of rotation of a fluid) and is usually used to represent vortex structures in multidimensional flow fields. In airflow simulation, the vorticity tensor can effectively describe the rotation behavior of the fluid and thus help predict vortex flow.

[0152] The velocity divergence is a quantity in fluid dynamics that describes the rate of change of velocity in a flow field, and reflects the degree of expansion or compression of the fluid in the flow field. When the divergence is zero, the fluid has no volume change; when it is positive, the fluid diffuses at a certain point; when it is negative, the fluid converges at that point. In fluid control and airflow prediction, the calculation of velocity divergence helps determine the compression or expansion characteristics of the airflow.

[0153] The vortex particle method is a numerical simulation method for simulating vortex flow in a fluid. The method discretizes the vorticity tensor into a number of vortex particles, each of which represents a vortex point in the fluid. The vortex particle method can effectively simulate vortex behavior in complex fluids and is widely used in problems such as airflow simulation and vortex prediction.

[0154] The vortex particle set is a set composed of multiple vorticity particles, each vortex particle represents a vortex in the flow field. The vortex particle set can discretely represent the rotating area in the flow field and is used to simulate the vortex behavior and the complex dynamics of the airflow.

[0155] In an optional specific implementation, based on the cloud-edge collaborative architecture, the vorticity tensor is calculated using thermal characteristics and airflow characteristics, the velocity divergence is calculated using oil smoke characteristics and airflow characteristics, and a vortex flow prediction model is constructed on the edge side.

[0156] First, collect the thermal characteristics and airflow characteristics in the smoke guide cavity. Thermal characteristics include thermal distribution characteristic frequency bands, such as using an infrared thermal imager to capture temperature information at different locations, obtain the spatial distribution of the temperature field, and convert it into frequency band data. For example, the frequency band range measured in the case is 0.5Hz to 2Hz; and dynamic characteristics of firepower changes, such as quantifying firepower changes through the pixel change rate of the flame image, and converting it into time series data. For example, the firepower change frequency in the case is 1Hz. Airflow characteristics include instantaneous frequency characteristics and amplitude characteristics. For example, a hot wire anemometer is used to measure the airflow velocity, and the instantaneous frequency and amplitude are obtained through fast Fourier transform. The frequency range measured in the case is 2Hz to 10Hz, and the amplitude range is 0.1m / s to 1m / s.

[0157] Then, the collected thermal characteristics are input into the regularized spectral analysis model. The characteristic frequency band of the thermal distribution is input into the model for spectral analysis to extract the dominant frequency component of the thermal field. For example, the dominant frequency extracted in the case is 1Hz. At the same time, the dynamic characteristics of firepower changes are input into the model to extract the amplitude component. For example, the amplitude component extracted in the case is 0.8. The collected airflow characteristics are input into the Hilbert-Huang transform model. The instantaneous frequency characteristics and amplitude characteristics are input into the model for decomposition to obtain the intrinsic mode function of the airflow field. For example, three intrinsic mode functions are decomposed in the case, representing airflow fluctuations of different scales.

[0158] Next, the thermal-airflow coupling tensor is constructed. Based on the dominant frequency component (1Hz) and amplitude component (0.8) of the thermal field, as well as the intrinsic mode functions (3) of the airflow field, these data are fused to construct a three-dimensional tensor that characterizes the coupling relationship between thermal and airflow. The vorticity tensor of the airflow field in the smoke guide cavity is calculated using the third-order compact difference format. For example, by performing a differential operation on the coupling tensor, the components of the vorticity tensor are obtained, which describes the rotation characteristics of the airflow field.

[0159] At the same time, the characteristics of oil smoke in the smoke guide cavity are collected. The characteristics of oil smoke include local characteristics of particle size distribution and nonlinear characteristics of concentration distribution. For example, the particle size distribution of oil smoke particles is measured using a particle counter. The particle size measured in the case is mainly concentrated between 0.5μm and 2μm; the concentration of oil smoke is measured using a photoelectric sensor. The concentration range measured in the case is 10mg / m³ to 50mg / m³. Combined with the local instantaneous phase characteristics in the previously collected airflow characteristics, for example, a hot wire anemometer is used to measure the airflow velocity and calculate its phase information. The measured phase range in the case is 0 to 2π. This information is input into the continuity equation to calculate the velocity divergence of the airflow field in the smoke guide cavity. For example, by calculating the oil smoke characteristics and the airflow phase information, the value of the velocity divergence is obtained, which describes the expansion or contraction characteristics of the airflow field.

[0160] In the edge computing unit, a vortex particle method model is established based on the calculated vorticity tensor. The vorticity tensor is discretized into a set of vortex particles with intensity weights and kernel functions. For example, the vorticity tensor is decomposed into several vortex particles, each of which has a position, intensity, and influence range. The pressure Poisson equation is constructed based on the calculated velocity divergence. The pressure field is solved using the multi-grid method. For example, the pressure values ​​at each position in the smoke guide cavity are obtained by iteratively solving the pressure Poisson equation. A vortex flow prediction model is constructed based on the vortex particle set, velocity divergence, and pressure field. This model can predict the changes in the airflow field in the smoke guide cavity over a period of time in the future.

[0161] In this embodiment, by combining thermal characteristics, airflow characteristics and oil smoke characteristics, and adopting advanced signal processing and numerical calculation methods, the complex dynamic characteristics of the airflow field in the smoke guide cavity can be captured more accurately, thereby improving the accuracy of vortex flow prediction; by adopting a cloud-edge collaborative architecture, complex computing tasks are offloaded to the edge computing unit, reducing the amount of data transmission and cloud computing load, thereby reducing computing costs and energy consumption; the deployment of edge computing units makes data processing and model prediction closer to the data source, reduces data transmission delays, thereby enhancing the real-time performance of vortex flow prediction, and can better meet the needs of real-time control and optimization.

[0162] In an optional embodiment, the cloud server performs segmented prediction of the airflow motion trajectory based on the vortex flow prediction model, and calculates the real-time airflow deviation; dynamically adjusts the initial control parameter group based on the real-time airflow deviation to generate a target control parameter group; uploads the real-time airflow deviation and the target control parameter group to the cloud deep reinforcement learning network, and continuously optimizes the vortex flow prediction model, including:

[0163] The vortex flow prediction model is transmitted to the cloud server, and a piecewise polynomial is used in the cloud server to describe the airflow motion trajectory; based on the vorticity tensor and pressure field at the current moment, the predicted velocity field at the next moment is calculated; the measured velocity field is obtained, the predicted velocity field is compared with the measured velocity field, and the direction deviation and amplitude deviation of the velocity vector are calculated to obtain the real-time airflow deviation degree;

[0164] A feedback correction matrix is ​​constructed according to the real-time airflow deviation, and the feedback correction matrix is ​​subjected to matrix multiplication operation with the initial control parameter group to obtain the corrected control parameters; based on the corrected control parameters, the flow field response characteristics are calculated, and the recursive least squares method is used to establish a dynamic mapping relationship between the corrected control parameters and the flow field response characteristics to generate the target control parameter group;

[0165] A dual-time deep neural network is constructed in the cloud server. The real-time airflow deviation is used as the state input, and the target control parameter group is used as the action input. Through temporal difference learning, the update strategy network parameters and the update value network parameters are obtained. The update strategy network parameters and the update value network parameters are passed to the vortex flow prediction model in the edge computing unit to achieve continuous optimization of the vortex flow prediction model.

[0166] In an optional specific implementation, first, a vortex flow prediction model is established. The model is based on physical laws and experimental data and can predict airflow motion under specific boundary conditions and initial states. The core of the model is to use piecewise polynomials to describe the airflow motion trajectory, decomposing the complex curve trajectory into several simple polynomial curves for easy calculation and analysis.

[0167] Then, the established vortex flow prediction model is transmitted to the cloud server. In the cloud server, the predicted velocity field at the next moment is calculated using the vorticity tensor (a physical quantity that describes the rotation of the fluid) and the pressure field data at the current moment. The vorticity tensor and pressure field can be obtained by real-time measurement by sensors or inferred from previous prediction results.

[0168] Next, the measured velocity field is obtained by collecting the airflow velocity data at each position in real time through the sensor network arranged in the flow field to obtain the measured velocity field.

[0169] The predicted velocity field is compared with the measured velocity field to calculate the direction deviation and amplitude deviation of the velocity vector. For example, assuming that the predicted velocity is 10 meters per second to the east and the measured velocity is 8 meters per second at 15 degrees north of the east, the direction deviation is 15 degrees and the amplitude deviation is 2 meters per second. The direction deviation and amplitude deviation are integrated to obtain the real-time airflow deviation, which can quantify the accuracy of the prediction model.

[0170] A feedback correction matrix is ​​constructed based on the real-time airflow deviation. The element values ​​of the matrix are related to the size and direction of the deviation. For example, the larger the deviation, the larger the corresponding matrix element value. Assuming that the real-time airflow deviation is 0.1, a 2×2 feedback correction matrix can be constructed, and its element values ​​are 0.05, 0.02, 0.03, and 0.04 respectively.

[0171] Perform matrix multiplication operation on the feedback correction matrix and the initial control parameter group to obtain the corrected control parameters. The initial control parameter group is a set of pre-set parameters used to control the motion state of the airflow, such as the injection speed and injection angle of the airflow. Assuming that the initial control parameter group is [10, 5], multiply the feedback correction matrix and the initial control parameter group to obtain the corrected control parameter group [10.35, 5.26].

[0172] Based on the modified control parameters, the flow field response characteristics are calculated. The modified control parameters are input into the flow field control system, and the actual motion state of the airflow, such as speed, pressure, temperature, etc., is observed to obtain the flow field response characteristics. Assume that the modified control parameters increase the airflow speed by 1 m / s.

[0173] The recursive least squares method is used to establish a dynamic mapping relationship between the modified control parameters and the flow field response characteristics to generate a target control parameter group. The recursive least squares method is an online learning algorithm that can continuously update model parameters based on new data. By analyzing the relationship between the modified control parameters and the flow field response characteristics, a mapping function can be obtained to predict the flow field response under specific control parameters. Based on this mapping function, a target control parameter group that can achieve the expected airflow state can be generated. Assuming that the target airflow velocity is 12 m / s, the corresponding target control parameter group can be calculated by the mapping function [12.5, 6.1].

[0174] A dual temporal deep neural network is built in the cloud server. The network consists of two sub-networks: a policy network and a value network. The policy network is used to select the best action based on the current state, and the value network is used to evaluate the value of the current state.

[0175] The real-time airflow deviation is taken as the state input, and the target control parameter group is taken as the action input, which are input into the dual temporal deep neural network.

[0176] Through temporal difference learning, the updated policy network parameters and updated value network parameters are obtained. Temporal difference learning is a reinforcement learning algorithm that finds the optimal control strategy through continuous trial and learning. Assume that through temporal difference learning, the parameters of the policy network are updated to [0.2, 0.5, 0.3], and the parameters of the value network are updated to [0.1, 0.4, 0.5].

[0177] The update strategy network parameters and the update value network parameters are passed to the vortex flow prediction model in the edge computing unit to achieve continuous optimization of the vortex flow prediction model. The edge computing unit is located on a device close to the sensor and can process data and control the airflow in real time. By continuously updating the model parameters, the accuracy of the prediction model and the efficiency of the control system can be improved.

[0178] In this embodiment, by real-time monitoring of airflow deviation and dynamic adjustment of control parameters, the prediction error can be effectively reduced and precise control of the airflow can be achieved; the control strategy based on deep reinforcement learning can be automatically adjusted according to environmental changes, improving the system's ability to cope with complex and uncertain situations; the application of edge computing units can reduce data transmission delays and improve the real-time response speed of the system.

[0179] Figure 2 FIG. 1 is a schematic diagram of a control system for a range hood based on the Internet of Things according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0180] The first unit is used to build a hierarchical data acquisition network based on the Internet of Things communication protocol, obtain the thermal distribution map and the cooking fire power value through the bottom sensor array, obtain the particle size distribution and concentration gradient of the oil fume particles through the middle sensor array, obtain the airflow velocity distribution matrix of the cross section of the smoke guide cavity through the upper sensor array, determine the multi-source hierarchical environmental feature data, perform spatiotemporal fusion processing on the multi-source hierarchical environmental feature data through the edge computing node, and generate a cooking condition feature vector including thermal features, oil fume features and airflow features;

[0181] The second unit is used to upload the cooking condition characteristic vector to the cloud server, calculate the rising speed of the hot air flow based on the thermal characteristics, calculate the oil fume diffusion radius based on the oil fume characteristics, identify the airflow dead zone position and the eddy zone position in the smoke guide cavity based on the airflow characteristics, and determine the flow field characteristic parameters; use a distributed computing framework to execute a piecewise nonlinear optimization algorithm, calculate the suction force required for each area according to the flow field characteristic parameters, and generate an adaptive fan control curve; calculate the initial airflow deviation based on the adaptive fan control curve, and generate an initial control parameter group, wherein the initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value, and an initial air guide plate angle value;

[0182] The third unit is used to calculate the vorticity tensor of the airflow field in the smoke guide cavity based on the cloud-edge collaborative architecture, using thermal characteristics and airflow characteristics, and calculate the velocity divergence using oil smoke characteristics and airflow characteristics; based on the vorticity tensor and the velocity divergence, a vortex flow prediction model is constructed on the edge side; based on the vortex flow prediction model, the airflow motion trajectory is segmentedly predicted by the cloud server, and the real-time airflow deviation is calculated; based on the real-time airflow deviation, the initial control parameter group is dynamically adjusted to generate a target control parameter group; the real-time airflow deviation and the target control parameter group are uploaded to the cloud-based deep reinforcement learning network to continuously optimize the vortex flow prediction model.

[0183] According to a third aspect of the embodiments of the present invention,

[0184] An electronic device is provided, comprising:

[0185] processor;

[0186] a memory for storing processor-executable instructions;

[0187] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0188] A fourth aspect of the embodiments of the present invention is:

[0189] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0190] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A range hood control method based on the Internet of Things, characterized in that: include: A hierarchical data acquisition network is constructed based on the Internet of Things communication protocol. The thermal distribution map and cooking fire power value are obtained through the bottom sensor array, the particle size distribution and concentration gradient of the oil fume particles are obtained through the middle sensor array, and the airflow velocity distribution matrix of the smoke guide cavity cross section is obtained through the upper sensor array. Multi-source hierarchical environmental feature data is determined, and the multi-source hierarchical environmental feature data is subjected to spatiotemporal fusion processing through the edge computing node to generate a cooking condition feature vector including thermal features, oil fume features and airflow features. The cooking condition characteristic vector is uploaded to the cloud server, the rising speed of the hot air flow is calculated based on the thermal characteristics, the oil fume diffusion radius is calculated based on the oil fume characteristics, the airflow dead zone position and the eddy zone position in the smoke guide cavity are identified based on the airflow characteristics, and the flow field characteristic parameters are determined; a piecewise nonlinear optimization algorithm is executed using a distributed computing framework, and the required suction force of each area is calculated according to the flow field characteristic parameters, and an adaptive fan control curve is generated; based on the adaptive fan control curve, an initial airflow deviation is calculated, and an initial control parameter group is generated, wherein the initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value, and an initial air guide plate angle value; Based on the cloud-edge collaborative architecture, the vorticity tensor of the airflow field in the smoke guide cavity is calculated using thermal characteristics and airflow characteristics, and the velocity divergence is calculated using oil smoke characteristics and airflow characteristics; a vortex flow prediction model is constructed on the edge side based on the vortex flow prediction model; the airflow motion trajectory is segmentedly predicted by the cloud server based on the vortex flow prediction model, and the real-time airflow deviation is calculated; the initial control parameter group is dynamically adjusted based on the real-time airflow deviation to generate a target control parameter group; the real-time airflow deviation and the target control parameter group are uploaded to the cloud-based deep reinforcement learning network to continuously optimize the vortex flow prediction model.

2. The method according to claim 1, characterized in that A hierarchical data acquisition network is constructed based on the IoT communication protocol. The thermal distribution map and cooking fire power value are obtained through the bottom sensor array. The particle size distribution and concentration gradient of the oil smoke particles are obtained through the middle sensor array. The airflow velocity distribution matrix of the smoke guide cavity cross section is obtained through the upper sensor array. A hierarchical data acquisition network is constructed based on the Internet of Things communication protocol, and a multi-source sensor array is divided into a bottom sensor array, a middle sensor array and an upper sensor array. The bottom sensor array includes an infrared sensor and a flame spectrum sensor, the middle sensor array includes a scattering spectrum sensor and an absorption spectrum sensor, and the upper sensor array includes an annular array airflow sensor, which is composed of a plurality of hot wire wind speed sensors uniformly distributed along the circumference. According to a preset first cycle, the temperature data matrix of the infrared sensor array is read, and the temperature data matrix is ​​expanded by a bicubic interpolation algorithm to generate a thermal distribution map; according to a preset second cycle, the spectrum waveform of the flame spectrum sensor is collected, the spectrum waveform is fast Fourier transformed to obtain a spectrum map, the main frequency corresponding to the maximum amplitude and the secondary frequency corresponding to the second maximum amplitude in the spectrum map are extracted, the main frequency and the secondary frequency are substituted into a pre-trained back propagation neural network, the thermal distribution map is used as an auxiliary input, and the cooking fire power value at the current moment is output; Control the scattering spectrum sensor to emit a detection beam within a preset angle range and at a preset angle interval, collect light intensity data at multiple scattering angles, substitute the light intensity data into an inverse iteration algorithm, perform iterative calculations using the Mie scattering coefficient as an iteration parameter, and output the particle size distribution of the oil fume particles when the root mean square error between two adjacent iterations is less than a preset error threshold; control the absorption spectrum sensor to determine measurement points according to a preset height interval, measure transmittance at different heights, obtain transmittance data at multiple measurement points, calculate the oil fume concentration at each measurement point based on the Beer-Lambert law, use a cubic spline interpolation method to fit, obtain a continuous variation curve of the oil fume concentration with height, and calculate the first-order derivative of the continuous variation curve to obtain the oil fume concentration gradient; The airflow measurement points are determined by the positions of multiple hot-wire wind speed sensors, and the airflow velocity values ​​are collected at each airflow measurement point at preset time intervals. The airflow velocity values ​​are Gaussian filtered to remove measurement noise to obtain filtered airflow velocity values, which are decomposed into radial components and tangential components. A polar coordinate grid is established on the cross-section of the smoke guide cavity, and the radial components and tangential components are extended to the grid points through radial basis function interpolation to generate an airflow velocity distribution matrix, and the curl field and divergence field of the airflow velocity distribution matrix are calculated.

3. The method according to claim 2, characterized in that Determining multi-source layered environmental feature data, performing spatiotemporal fusion processing on the multi-source layered environmental feature data through an edge computing node, and generating a cooking condition feature vector including thermal features, oil smoke features, and airflow features includes: Based on the thermal distribution map, cooking fire power value, particle size distribution of oil fume particles, oil fume concentration gradient, and the curl field and divergence field of the airflow velocity distribution matrix, multi-source hierarchical environmental characteristic data are constructed; Performing discrete wavelet transform on the thermal distribution map to obtain approximate components and detail components of different frequency bands, calculating the normalized energy of the detail components at each scale, and extracting the frequency band whose energy proportion exceeds a preset proportion threshold as the thermal distribution characteristic frequency band; performing continuous wavelet transform on the cooking fire power value, using Mexican hat wavelet to obtain time-frequency domain features, extracting the energy spectrum peak value and the corresponding characteristic scale in the time-frequency domain as the dynamic characteristics of firepower change, and combining the thermal distribution characteristic frequency band and the dynamic characteristics of firepower change to form a thermal characteristic; Performing wavelet packet decomposition on the particle size distribution of the oil fume particles, adaptively selecting an optimal decomposition basis according to the Shannon entropy criterion, and extracting local features of the particle size distribution from the optimal decomposition basis nodes; performing nonlinear principal component analysis on the oil fume concentration gradient, mapping the oil fume concentration gradient to a high-dimensional feature space using a kernel function, extracting principal components as nonlinear features of concentration distribution, and combining the local features of the particle size distribution and the nonlinear features of the concentration distribution to form oil fume features; Performing orthogonal empirical mode decomposition on the airflow velocity distribution data to obtain eigenmode functions of different scales, and calculating the instantaneous frequency characteristics and amplitude characteristics of the eigenmode functions; performing a two-dimensional Hilbert transform based on the curl field and divergence field of the airflow velocity distribution matrix, extracting local instantaneous phase characteristics and amplitude envelope characteristics, and combining the instantaneous frequency characteristics, the amplitude characteristics, the local instantaneous phase characteristics and the amplitude envelope characteristics to form airflow characteristics; A multi-layer spatiotemporal feature fusion network is constructed through edge computing nodes. The multi-layer spatiotemporal feature fusion network includes a feature extraction layer and a feature fusion layer. The feature extraction layer includes a thermal feature sub-network, a fume feature sub-network, and an airflow feature sub-network. Each sub-network adopts an attention mechanism to adaptively weight the input thermal features, fume features, and airflow features to obtain weighted feature data; In the feature fusion layer, a gated recurrent unit is used to extract time series features from the weighted feature data, and the association weights between different features are calculated through a cross-modal attention mechanism. Each feature data is mapped to a unified feature space according to the association weights to obtain unified feature data, and a graph convolutional network is used to model the unified feature data to obtain spatiotemporal dependent feature data; The spatiotemporal dependency feature data is input into a fully connected layer for nonlinear mapping to generate a cooking condition feature vector.

4. The method according to claim 1, characterized in that: The cooking condition characteristic vector is uploaded to the cloud server, the rising speed of the hot air flow is calculated based on the thermal characteristics, the fume diffusion radius is calculated based on the fume characteristics, and the airflow dead zone position and eddy zone position in the smoke guide cavity are identified based on the airflow characteristics. Determining the flow field characteristic parameters includes: Upload the cooking condition feature vector to the cloud server; The cloud server constructs a thermal gradient field based on the characteristic frequency band of thermal distribution, establishes a thermal buoyancy model in combination with the dynamic characteristics of firepower changes, takes the local temperature difference as the driving force, inputs the thermal gradient field and the thermal buoyancy model into the ideal gas state equation, and calculates the initial thermal airflow rise velocity field; generates a correction coefficient based on the temperature stratification effect, multiplies the correction coefficient by the initial thermal airflow rise velocity field, and obtains the corrected thermal airflow rise velocity field; The cloud server constructs a multi-component diffusion model based on the local characteristics of the particle size distribution, determines the initial oil fume cluster diffusion coefficient according to the nonlinear characteristics of the concentration distribution, inputs the initial oil fume cluster diffusion coefficient into the Gaussian diffusion model, and calculates the multi-component diffusion flux; based on the multi-component diffusion flux, calculates the collision frequency function, multiplies the collision frequency function by the initial oil fume cluster diffusion coefficient, and obtains the corrected oil fume cluster diffusion coefficient; substitutes the corrected oil fume cluster diffusion coefficient into the diffusion equation, and solves to obtain the oil fume diffusion radius; The cloud server inputs the instantaneous frequency characteristics and amplitude characteristics into the topological structure model to obtain the flow field topological structure, identifies the characteristic streamlines based on the local instantaneous phase characteristics and amplitude envelope characteristics, and inputs the characteristic streamlines into the critical point theoretical model to determine the saddle point position and the focal point position; calculates the Laplace operator of the velocity field based on the flow field topological structure, and compares the Laplace operator with the preset vorticity threshold to determine the airflow dead zone position in the smoke guide cavity; inputs the flow field topological structure into the Q criterion model and the λ2 criterion model to determine the vortex zone position in the smoke guide cavity.

5. The method according to claim 4, characterized in that A distributed computing framework is used to execute a piecewise nonlinear optimization algorithm. According to the flow field characteristic parameters, the required suction force of each area is calculated to generate an adaptive fan control curve. Based on the adaptive fan control curve, the initial airflow deviation is calculated, and an initial control parameter group is generated. The initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value, and an initial air guide plate angle value, including: The cloud server uses the corrected hot air flow rising velocity field, oil smoke diffusion radius, airflow dead zone position and eddy zone position as flow field characteristic parameters, and inputs them into the flow field division module. Based on the velocity field gradient threshold, the flow field is divided into multiple sub-areas, and the sub-area boundary grid nodes and internal grid nodes are obtained. The velocity field variance is calculated based on the velocity components of the internal grid nodes, and the velocity field variance is divided by the preset velocity field reference variance to obtain the airflow organization uniformity index; the modulus of the concentration field gradient is calculated based on the concentration values ​​of the internal grid nodes, and the modulus of the concentration field gradient is divided by the preset concentration gradient reference value to obtain the oil fume collection efficiency index; the fan input power is calculated, and the input power is divided by the preset power reference value to obtain the energy consumption index; The first weighted term is obtained by multiplying the airflow organization uniformity index by the first preset weight coefficient, the second weighted term is obtained by multiplying the oil fume collection efficiency index by the second preset weight coefficient, and the third weighted term is obtained by multiplying the energy consumption index by the third preset weight coefficient; based on the first weighted term, the second weighted term and the third weighted term, a comprehensive objective function is obtained; Based on the sub-region boundary grid nodes, boundary constraints are constructed, including velocity continuity constraints and pressure continuity constraints of the boundary grid nodes of adjacent sub-regions; the boundary constraints are combined with the comprehensive objective function to construct the Lagrangian function, and the quadratic penalty term is introduced into the Lagrangian function to obtain the augmented Lagrangian function; Fix the initial value of the Lagrange multiplier, substitute the augmented Lagrangian function into the gradient descent algorithm, solve the local optimal solution of each sub-region, and obtain the sub-region control variable; based on the sub-region control variable, calculate the residual of the boundary constraint condition, multiply the residual by the step size factor, and update the Lagrange multiplier value; Calculate the norm of the difference between the sub-region control variables of two adjacent iterations, stop the iteration when the norm is less than the preset convergence threshold, and convert the final sub-region control variables into the time-varying data points of the suction force in each region; The time-varying data points of the extraction force are selected as nodes, the cubic spline basis function is constructed, and the cubic spline coefficients are solved by the least square method; the cubic spline coefficients are substituted into the cubic spline basis function to obtain the piecewise function expression; Calculate the function value, first-order derivative value and second-order derivative value of the piecewise function expression at the node, and verify the continuity of the function value, first-order derivative value and second-order derivative value of the adjacent piecewise functions respectively; when the continuity verification is passed, determine the piecewise function expression as the adaptive fan control curve; Based on the adaptive fan control curve, the cloud server compares the actual flow field with the ideal flow field through the flow field similarity principle, calculates the angle deviation and modulus deviation of the velocity vector field, and obtains the initial airflow deviation; the initial airflow deviation is input into the fuzzy control rule to generate an initial control parameter group, which includes the initial main fan speed value, the initial auxiliary fan speed value and the initial air guide plate angle value.

6. The method according to claim 1, characterized in that Based on the cloud-edge collaborative architecture, the vorticity tensor of the airflow field in the smoke guide cavity is calculated using the thermal characteristics and airflow characteristics, and the velocity divergence is calculated using the smoke characteristics and airflow characteristics; based on the vorticity tensor and the velocity divergence, a vortex flow prediction model is constructed on the edge side, including: The characteristic frequency band of thermal distribution and the dynamic characteristics of firepower change in the thermal characteristics are input into the regularized spectrum analysis model to extract the dominant frequency components and amplitude components of the thermal field; the instantaneous frequency characteristics and amplitude characteristics in the airflow characteristics are input into the Hilbert-Huang transform model, and the intrinsic mode function of the airflow field is obtained by decomposition; based on the dominant frequency components and the intrinsic mode function, the thermal-airflow coupling tensor is constructed, and the vorticity tensor of the airflow field in the smoke guide cavity is calculated using the third-order compact difference format; The local characteristics of particle size distribution and nonlinear characteristics of concentration distribution in the oil smoke characteristics are input into the continuity equation, and the velocity divergence of the airflow field in the smoke guide cavity is calculated by combining the local instantaneous phase characteristics in the airflow characteristics. In the edge computing unit, a vortex particle method model is established based on the vorticity tensor, and the vorticity tensor is discretized into a set of vortex particles with intensity weights and kernel functions; the pressure Poisson's equation is constructed based on the velocity divergence, and the pressure field is solved by the multi-grid method; and a vortex flow prediction model is constructed according to the vortex particle set, the velocity divergence and the pressure field.

7. The method according to claim 6, characterized in that The cloud server performs segmented prediction of the airflow trajectory based on the vortex flow prediction model and calculates the real-time airflow deviation; dynamically adjusts the initial control parameter group based on the real-time airflow deviation to generate a target control parameter group; Uploading the real-time airflow deviation and target control parameter group to the cloud-based deep reinforcement learning network to continuously optimize the vortex flow prediction model includes: The vortex flow prediction model is transmitted to the cloud server, and a piecewise polynomial is used in the cloud server to describe the airflow motion trajectory; based on the vorticity tensor and pressure field at the current moment, the predicted velocity field at the next moment is calculated; the measured velocity field is obtained, the predicted velocity field is compared with the measured velocity field, and the direction deviation and amplitude deviation of the velocity vector are calculated to obtain the real-time airflow deviation degree; A feedback correction matrix is ​​constructed according to the real-time airflow deviation, and the feedback correction matrix is ​​subjected to matrix multiplication operation with the initial control parameter group to obtain the corrected control parameters; based on the corrected control parameters, the flow field response characteristics are calculated, and the recursive least squares method is used to establish a dynamic mapping relationship between the corrected control parameters and the flow field response characteristics to generate the target control parameter group; A dual-time deep neural network is constructed in the cloud server. The real-time airflow deviation is used as the state input, and the target control parameter group is used as the action input. Through temporal difference learning, the update strategy network parameters and the update value network parameters are obtained. The update strategy network parameters and the update value network parameters are passed to the vortex flow prediction model in the edge computing unit to achieve continuous optimization of the vortex flow prediction model.

8. A range hood control system based on the Internet of Things, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to build a hierarchical data acquisition network based on the Internet of Things communication protocol, obtain the thermal distribution map and the cooking fire power value through the bottom sensor array, obtain the particle size distribution and concentration gradient of the oil fume particles through the middle sensor array, obtain the airflow velocity distribution matrix of the cross section of the smoke guide cavity through the upper sensor array, determine the multi-source hierarchical environmental feature data, perform spatiotemporal fusion processing on the multi-source hierarchical environmental feature data through the edge computing node, and generate a cooking condition feature vector including thermal features, oil fume features and airflow features; The second unit is used to upload the cooking condition characteristic vector to the cloud server, calculate the rising speed of the hot air flow based on the thermal characteristics, calculate the oil fume diffusion radius based on the oil fume characteristics, identify the airflow dead zone position and the eddy zone position in the smoke guide cavity based on the airflow characteristics, and determine the flow field characteristic parameters; use a distributed computing framework to execute a piecewise nonlinear optimization algorithm, calculate the suction force required for each area according to the flow field characteristic parameters, and generate an adaptive fan control curve; calculate the initial airflow deviation based on the adaptive fan control curve, and generate an initial control parameter group, wherein the initial control parameter group includes an initial main fan speed value, an initial auxiliary fan speed value, and an initial air guide plate angle value; The third unit is used to calculate the vorticity tensor of the airflow field in the smoke guide cavity based on the cloud-edge collaborative architecture, using thermal characteristics and airflow characteristics, and calculate the velocity divergence using oil smoke characteristics and airflow characteristics; based on the vorticity tensor and the velocity divergence, a vortex flow prediction model is constructed on the edge side; based on the vortex flow prediction model, the airflow motion trajectory is segmentedly predicted by the cloud server, and the real-time airflow deviation is calculated; based on the real-time airflow deviation, the initial control parameter group is dynamically adjusted to generate a target control parameter group; the real-time airflow deviation and the target control parameter group are uploaded to the cloud-based deep reinforcement learning network to continuously optimize the vortex flow prediction model.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images