Air fryer control method and system based on internet of things data collection

Abstract

This invention discloses an air fryer control method and system based on IoT data acquisition, comprising the following steps: S1, synchronously acquiring images of food inside the fryer and multi-dimensional sensor data of the environment through an IoT gateway; S2, preprocessing the data and aligning timestamps to generate a standardized spatiotemporal input dataset; S3, extracting and segmenting features from the images, and encoding and outputting the microscopic spectral features of the food; S4, retrieving and matching a knowledge graph, and generating a cooking state feature set by combining the DeepONet model thermodynamic mapping; S5, inputting an improved N-BEATS model, introducing a physical state deviation correction mechanism, performing deep nonlinear decomposition and physical deviation correction, estimating the critical evaporation moment, and outputting a heat intervention command; S6, adjusting and cyclically detecting according to the command, and generating a completion signal to stop heating. This invention achieves precise perception and dynamic control of the cooking process, effectively improving cooking quality and consistency.

Description

Technical Field

[0001] This invention relates to the field of smart home appliance control and Internet of Things (IoT) application technology, and in particular to an air fryer control method and system based on IoT data acquisition. Background Technology

[0002] IoT gateways based on unified device coding and multimodal deep learning technology have been widely used in smart home appliances, healthy cooking, and industrial control in recent years, becoming a key development direction for promoting the intelligent transformation of traditional equipment. However, in practical applications, smart cooking scenarios face many challenges such as nonlinear changes in the state of ingredients, complex heat conduction processes, and large fluctuations in environmental interference. The deployment effectiveness of existing control methods is still constrained by many factors.

[0003] Most current intelligent cooking control methods rely on single-modal sensor data input, making it difficult to fully utilize multi-source information such as food image texture and multi-dimensional environmental parameters of the cavity. This results in a lack of comprehensive modeling of food doneness and internal thermal state. Some systems only use simple PID algorithms or fixed-curve temperature control strategies to evaluate model output, ignoring the combined influence of multiple factors such as dehydration rate, core heat penetration depth, oil precipitation state, and physical deviation correction, thus limiting the adaptive optimization capability of temperature control strategies. Furthermore, the cooking decision-making logic lacks a physically-based explanatory path, making it difficult to provide users or the system with clear thermodynamic evolution evidence, affecting the credibility and usability of cooking results.

[0004] Furthermore, existing time-series prediction networks for cooking control are mostly designed with pure data-driven feature extraction, failing to dynamically adjust network weights by incorporating physical laws such as Fourier's law of heat conduction. This results in the model consistently performing poorly when faced with changes in the type or weight of ingredients, and is unable to effectively correct its course. It is difficult to adapt to the continuous changes and evolution of complex cooking scenarios, seriously affecting the practical value and stability of the model in real kitchen scenarios. Summary of the Invention

[0005] One objective of this invention is to propose an air fryer control method and system based on IoT data acquisition. This invention fully integrates key steps such as multimodal data preprocessing, pixel-level image feature extraction, DeepONet model thermodynamic mapping, improved N-BEATS model prediction, and adaptive heat intervention. It constructs an intelligent cooking control process with spatiotemporal data standardization, microscopic spectral feature encoding, operator learning mechanism for physical solution, and physical deviation correction, achieving accurate mapping and dynamic state correction of food thermodynamic evolution under complex cooking environments. Specifically, this invention employs the DeepONet model to rapidly generate dehydration rate and core heat penetration depth through collaborative solution of branch and backbone sub-networks; and utilizes the improved N-BEATS model to introduce a physical state deviation correction mechanism, comparing the theoretical penetration depth with the predicted depth to dynamically adjust network weights. This invention possesses advantages such as multi-dimensional fusion of state modeling, accurate thermodynamic evolution mapping, strong adaptive physical deviation correction, and flexible heat intervention commands. It can significantly improve the control accuracy, anti-interference ability, and consistency of food cooking degree in the cooking process, thereby effectively solving problems such as difficulty in solving the heat conduction process, large model prediction deviations, and lack of physical constraints in decision-making strategies in existing methods.

[0006] An air fryer control method based on Internet of Things (IoT) data acquisition according to an embodiment of the present invention includes the following steps: S1. Synchronously collect images of food and multi-dimensional environmental sensor data inside the air fryer through an IoT gateway; S2. Preprocess the food images and align the timestamps of the environmental multidimensional sensor data with the timestamps of the video frames to generate a standardized spatiotemporal input dataset. S3. Extract pixel-level features from food images in the standardized spatiotemporal input dataset, segment and identify the main food region, and deeply encode the texture density, edge contour evolution, and pixel distribution information of oil precipitation spots on the food surface, outputting the food microspectral feature code. S4. Based on the microscopic spectral feature encoding of the ingredients, search and match in the preset cooking knowledge graph to obtain the optimal ripeness data of the corresponding ingredients attributes, and input the ingredients microscopic spectral feature encoding into the DeepONet model to perform nonlinear mapping on the thermodynamic evolution process, generate the dehydration rate and core heat penetration depth of the ingredients at the current moment, and output them as the current cooking state feature set along with the historical temperature control curve. S5. Input the current cooking state feature set into the improved N-BEATS model for deep nonlinear decomposition, introduce a physical state deviation correction mechanism, dynamically correct the previous cooking state feature set, reverse-engineer the critical evaporation time of the internal moisture of the ingredients, and output an adaptive heat intervention command. S6. Based on the adaptive heat intervention command, perform real-time adjustment, acquire updated food images, and after cyclic detection, generate a cooking completion signal and stop heating.

[0007] Optionally, S1 specifically includes: calling a preset synchronous acquisition interface through an IoT gateway to trigger an image acquisition device and multiple sets of environmental sensors deployed inside the air fryer cavity, capturing images of the food inside the air fryer at the current moment, and simultaneously reading the data stream transmitted by the environmental sensors to obtain multi-dimensional environmental sensing data inside the air fryer. The multi-dimensional environmental sensing data specifically includes cavity center temperature, humidity sensor readings, current power percentage of the heating element, and weight change data of the food tray.

[0008] Optionally, S2 specifically includes: S21. Perform noise reduction preprocessing, read the grayscale values ​​of the pixels in the food image, calculate the average grayscale value of the preset neighborhood and replace the original value, statistically analyze the grayscale histogram of the food image, calculate the average brightness of the grayscale histogram, and calculate the exposure adjustment coefficient based on the difference between the average brightness and the preset target brightness. Multiply the grayscale value of each pixel in the food image by the exposure adjustment coefficient. S22. Read the camera's internal parameters and distortion coefficients, calculate the distortion offset of each pixel in the food image based on the distortion coefficients, and extract the timestamp of each data point in the environmental multidimensional sensing data and the timestamp of the video frame corresponding to the food image after reverse mapping of the pixel coordinates. S23. Find the first most recent environmental multidimensional sensor data timestamp before the video frame timestamp and the second most recent environmental multidimensional sensor data timestamp. Use linear interpolation to calculate the interpolated data at the video frame timestamp based on the data values ​​corresponding to the two environmental multidimensional sensor data timestamps. Combine the interpolated data with the video frame to generate a standardized spatiotemporal input dataset.

[0009] Optionally, S3 specifically includes: S31. Input the food images in the standardized spatiotemporal input dataset into the image segmentation network, classify each pixel, mark the pixels belonging to the main body of the food as the foreground region, calculate the local change frequency of the gray value of the pixels in the foreground region, count the number of pixels with a gray value change frequency higher than the preset frequency, calculate the ratio of the number of pixels to the total number of pixels in the foreground region, and quantify it as the texture density of the food surface. S32. Extract pixel coordinates along the boundary of the foreground region, construct edge contour curves based on pixel coordinates, calculate the change in the enclosed area between the edge contour curve at the current time and the edge contour curve at the previous time, and mark it as edge contour evolution. S33. Traverse the grayscale value of each pixel in the foreground area, find the pixels with grayscale values ​​higher than the preset brightness threshold, count the distribution density of the brightness threshold pixels in the foreground area, and map them to the pixel distribution information of the oil precipitation spot. S34. Combine texture density, edge contour evolution and pixel distribution information into a high-dimensional feature vector, and use a preset fully connected layer to perform deep encoding on the high-dimensional feature vector to generate and output the food micro-spectral feature encoding.

[0010] Optionally, S4 specifically includes: S41. Input the microscopic spectral feature encoding of the ingredients as a query vector into the preset cooking knowledge graph, traverse all the ingredient entity nodes stored in the cooking knowledge graph, calculate the cosine similarity between the query vector and the feature vector of each ingredient entity node, select the ingredient entity node with the largest value as the target matching node, read the attribute label data associated with the target matching node, and parse it into the best ripeness data of the corresponding ingredient attribute. S42. Construct the branch sub-networks contained in the DeepONet model, input the optimal maturity data into the input layer of the branch sub-network, encode the optimal maturity data using a multi-layer preset fully connected neural network, multiply it with the preset weight matrix and superimpose the bias vector, process it through the tanh activation function, output a latent code vector of a set dimension, and use it as the boundary condition parameter to control the boundary of the heat conduction equation. S43. Construct the backbone sub-network of the DeepONet model. The microscopic spectral features of the food ingredients are encoded as input features and input to the backbone sub-network of the pre-sized convolutional neural network structure containing four residual blocks. Each residual block performs two convolution operations on the input features and adds them to the input features. The added features are then subjected to global average pooling to output the spatiotemporal features of the solution domain. S44. The DeepONet model reads the boundary condition parameters and the spatiotemporal features of the solution domain, uses the boundary condition parameters as function coefficients, and uses the spatiotemporal features of the solution domain as function independent variables. It performs a dot product operation to fuse the two, generates a fused feature tensor, constructs the heat conduction partial differential equation formula, and calculates the first derivative of the fused feature tensor with respect to the time coordinate to obtain the temperature change rate. S45. Calculate the second derivative of the fused feature tensor with respect to spatial coordinates to obtain the curvature of the temperature spatial distribution. Subtract the product of the temperature spatial distribution curvature and the preset thermal diffusivity from the rate of temperature change, and calculate the difference between the two as the physical residual. Use the mean square error to convert the physical residual into a loss value, and use the gradient descent algorithm to adjust the network weights to make the loss value converge. S46. Input the fused feature tensor after updating the weights into the DeepONet model for forward inference calculation, and output the predicted temperature distribution matrix to approximate the solution function of the heat conduction partial differential equation, thereby obtaining a nonlinear mapping of the thermodynamic evolution process inside the food. S47. Based on the results of nonlinear mapping, the DeepONet model performs differential operations on the fusion feature tensor along the time dimension, calculates the derivative of moisture change over time, converts the derivative value into mass units, generates the dehydration rate of the food at the current moment, and performs integral operations on the fusion feature tensor along the spatial dimension. The depth where the cumulative temperature is higher than the preset protein denaturation temperature is the core heat penetration depth. The dehydration rate and the core heat penetration depth are the current state data. S48. Read the historical cooking records stored in the system memory, extract the cavity center temperature data within a preset time period before the current moment, arrange the cavity center temperature data in chronological order, connect them to generate a historical temperature control curve, and perform vector splicing of the dehydration rate and core heat penetration depth of the ingredients at the current moment with the historical temperature control curve to generate a current cooking state feature set and output it.

[0011] Optionally, S5 specifically includes: S51. Input the current cooking state feature set into the improved N-BEATS model containing a double-layer stacked structure. The first layer stacked structure performs a one-dimensional convolution operation with a preset kernel on the historical temperature control curve to extract the trend term. The second layer stacked structure performs a preset fully connected operation on the current state data to extract the seasonal term. The trend term and the seasonal term are added point by point to obtain a deep nonlinear decomposition of the historical temperature control curve and the current dehydration rate and core heat penetration depth state data. S52. Construct a physical constraint loss function at the output of the improved N-BEATS model, read the cavity center temperature and the current power percentage of the heating tube from the environmental multidimensional sensing data, multiply the rated power of the heating tube by the current power percentage to obtain the actual output power, divide it by the heat dissipation area of ​​the heating tube to obtain the heat flux density, and divide the heat flux density by the temperature difference between the cavity center temperature and the food surface temperature to obtain the comprehensive heat transfer coefficient. S53. Multiply the overall heat transfer coefficient by the difference between the cavity center temperature and the food surface temperature to obtain the input heat flux. Use the Fourier heat conduction formula to divide the input heat flux by the food thermal conductivity to obtain the theoretical heat flux density along the food thickness direction. Use the theoretical heat flux density to divide by the product of the food specific heat capacity and density to calculate the theoretical temperature penetration depth. Mark the theoretical temperature penetration depth as the actual penetration depth. S54. Extract the core heat penetration depth from the current cooking state feature set, calculate the difference between the actual penetration depth and the core heat penetration depth, and define it as the physical state deviation value. Construct a deviation correction matrix with the same dimension as the hidden layer of the improved N-BEATS model, and add each element in the deviation correction matrix to the corresponding element in the hidden layer weight matrix to complete the dynamic correction of the network weights of the improved N-BEATS model. S55. The improved N-BEATS model with corrected weights is used to perform forward computation on the input features. The multiplication operation between the current cooking state feature set and the weight matrix is ​​performed in the preset fully connected layer and the bias term is added. The corrected feature vector is output after passing through the ReLU activation function, and the original values ​​in the current cooking state feature set are replaced by the corrected feature vector. S56. Perform an inverse transformation operation on the dynamically corrected feature vector, project the dynamically corrected feature vector back into the data space, calculate the first difference value of adjacent data points in the projected sequence, filter out data points whose absolute value of the first difference value is less than the preset change threshold, record the timestamp corresponding to the filtered data points, and mark it as the critical evaporation time of the internal moisture of the food in the reverse inference. S57. Read the current time of the system clock and calculate the difference between the current time and the critical evaporation time. Multiply the time difference by the preset heat dissipation coefficient to obtain the heat deficit. Divide the heat deficit by the rated power of the heating tube to obtain the compensation heating time. Convert the compensation heating time into a control duty cycle signal to generate an adaptive heat intervention command to prevent the food from drying out excessively.

[0012] Optionally, the construction of the bias correction matrix with the same hidden layer dimension as the improved N-BEATS model specifically includes: reading the number of rows and columns of the hidden layer weights of the improved N-BEATS model, creating an all-zero matrix as the initialization matrix, multiplying the physical state bias value by a preset learning rate coefficient to obtain a scaling factor, replacing each element value in the initialization matrix with the scaling factor, and constructing the bias correction matrix with the same hidden layer dimension as the improved N-BEATS model.

[0013] Optionally, S6 specifically includes: adjusting the heating tube power according to the adaptive heat intervention command, controlling the equipment operation status in real time during the cooking process, continuously monitoring the changes in the state of the ingredients after the control, acquiring updated images of the ingredients through an image acquisition device, and inputting them into an image segmentation network for cyclic detection. When the coverage rate of scorch marks on the surface of the ingredients reaches a preset threshold, a cooking completion signal is generated and the air fryer is controlled to stop heating.

[0014] An air fryer control system based on Internet of Things (IoT) data acquisition according to an embodiment of the present invention includes: The IoT data acquisition and synchronization module is used to collect food images and multi-dimensional environmental sensor data. The multimodal data preprocessing module is used to preprocess food images and align the timestamps of environmental multidimensional sensor data with the timestamps of video frames to generate a standardized spatiotemporal input dataset. The food micro-feature extraction module is used to extract pixel-level features from food images in a standardized spatiotemporal input dataset, segment and identify the main food region and encode it deeply, and output the food micro-spectral feature code. The thermodynamic evolution mapping module is used to search and match in the preset cooking knowledge graph to obtain the best cooking data, and input it into the DeepONet model along with the food micro-spectral feature encoding. Through the operator learning mechanism, the thermodynamic evolution process inside the food is nonlinearly mapped to generate the current cooking state feature set. The physical deviation correction and prediction module is used to input the current cooking state feature set into the improved N-BEATS model, perform deep nonlinear decomposition, introduce a physical state deviation correction mechanism, dynamically correct the current cooking state feature set, reverse-engineer the critical evaporation time, and output adaptive heat intervention instructions. The adaptive control and execution module is used to make real-time adjustments based on the adaptive heat intervention command, monitor changes in the state of the food, acquire updated food images and perform cyclic detection. When the coverage of scorch marks on the surface of the food reaches a preset threshold, a cooking completion signal is generated and the air fryer is controlled to stop heating.

[0015] The beneficial effects of this invention are: This invention addresses the issues of ambiguous food state perception and nonlinear heat conduction in air fryer cooking by deploying IoT gateways and multimodal sensors. It employs image segmentation algorithms to extract microscopic spectral features and aligns them with environmental data timestamps to generate a standardized spatiotemporal input dataset. Feature encoding and optimal ripeness data from a knowledge graph are input into a DeepONet model. Branch subnetworks and a backbone subnetwork are used to process boundary conditions and spatiotemporal features respectively, and an operator learning mechanism outputs the dehydration rate and core heat penetration depth. In the predictive control stage, thermodynamic state features are input into an improved N-BEATS model, introducing a physical state deviation correction mechanism. A deviation correction matrix is ​​constructed using the difference between the actual penetration depth and the predicted depth, dynamically adjusting network weights to inversely deduce the critical moisture evaporation moment and output adaptive heat intervention commands. Furthermore, heating power is adjusted in real-time according to the intervention commands, and scorch coverage is monitored cyclically to determine the cooking endpoint. Ultimately, this achieves multimodal data fusion, accurate solution of the thermodynamic physical field, and adaptive closed-loop control based on physical constraints in the air fryer cooking process, effectively improving the accuracy of food ripeness control, the rationality of the heat intervention strategy, and the robustness of the cooking model. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an air fryer control method based on Internet of Things data acquisition proposed in this invention; Figure 2 This is a structural diagram of an air fryer control system based on Internet of Things (IoT) data acquisition proposed in this invention. Figure 3 This is a flowchart of the thermodynamic evolution mapping and microscopic spectral feature encoding based on the DeepONet model proposed in this invention; Figure 4 This is a flowchart of the improved N-BEATS model for time-series prediction and adaptive thermal intervention command generation, which incorporates a physical state deviation correction mechanism as proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-4 An air fryer control method based on Internet of Things (IoT) data acquisition includes the following steps: S1. Real-time images of food and multi-dimensional environmental sensor data inside the air fryer are collected synchronously through the Internet of Things gateway. The multi-dimensional environmental sensor data includes the temperature of the cavity center, humidity sensor readings, current power percentage of the heating element, and weight change data of the food tray. S2. Perform noise reduction, exposure compensation and distortion correction preprocessing on real-time food images, and use linear interpolation to align the timestamps of environmental multidimensional sensor data with the timestamps of video frames to generate a standardized spatiotemporal input dataset containing multimodal information. S3. Extract pixel-level features from food images in the standardized spatiotemporal input dataset, segment and identify the main food region, and deeply encode the texture density, edge contour evolution, and pixel distribution information of oil precipitation spots on the food surface, outputting the food microspectral feature code. S4. Based on the microscopic spectral feature encoding of the ingredients, search and match in the preset cooking knowledge graph to obtain the optimal ripeness data of the corresponding ingredient attributes. Input the optimal ripeness data and the microscopic spectral feature encoding of the ingredients into the DeepONet model. Use the branch sub-network to use the optimal ripeness data as the boundary condition parameter, and use the backbone sub-network to use the microscopic spectral feature encoding of the ingredients as the spatiotemporal features of the solution domain. Through the operator learning mechanism, perform nonlinear mapping on the thermodynamic evolution process inside the ingredients to generate the dehydration rate and core heat penetration depth of the ingredients at the current moment. Output the current cooking state feature set along with the historical temperature control curve. S5. Input the current cooking state feature set into the improved N-BEATS model for deep nonlinear decomposition, introduce a physical state deviation correction mechanism, use the deviation value between the actual penetration depth and the core heat penetration depth to dynamically correct the current cooking state feature set, reverse the deduction of the critical evaporation time of the internal moisture of the food, and output an adaptive heat intervention command to prevent the food from drying out excessively. S6. Based on the adaptive heat intervention command, perform real-time adjustment, acquire updated food images, and after cyclic detection, generate a cooking completion signal and stop heating.

[0019] This invention significantly improves the precision of intelligent cooking and the consistency of finished products. By synchronously collecting food images and multi-dimensional sensor data through an IoT gateway and aligning them with timestamps, it achieves comprehensive digital monitoring of the cooking process. Utilizing a deep integration of the DeepONet and N-BEATS models, combined with a physical state deviation correction mechanism, it can nonlinearly map the thermodynamic evolution process in real time, accurately predicting the critical moment of moisture evaporation. This system not only solves the problems of traditional control methods being susceptible to environmental interference, leading to burnt surfaces or undercooked interiors, but also effectively prevents over-drying, ensuring a tender texture. Through real-time adjustment of adaptive heat intervention commands, it achieves precise matching between the cooking endpoint and optimal doneness, significantly reducing the defect rate and greatly improving the intelligent control level and user experience of the air fryer.

[0020] In this embodiment, S1 specifically includes: calling a preset synchronous acquisition interface through an IoT gateway to trigger an image acquisition device and multiple sets of environmental sensors deployed inside the air fryer cavity, capturing images of the food inside the air fryer at the current moment, and simultaneously reading the data stream transmitted by the environmental sensors to obtain multi-dimensional environmental sensing data inside the air fryer. The multi-dimensional environmental sensing data specifically includes cavity center temperature, humidity sensor readings, current power percentage of the heating element, and weight change data of the food tray.

[0021] In this embodiment, S2 specifically includes: S21. Perform noise reduction preprocessing, read the grayscale values ​​of the pixels in the food image, calculate the average grayscale value of the preset neighborhood and replace the original value, statistically analyze the grayscale histogram of the food image, calculate the average brightness of the grayscale histogram, and calculate the exposure adjustment coefficient based on the difference between the average brightness and the preset target brightness. Multiply the grayscale value of each pixel in the food image by the exposure adjustment coefficient. S22. Read the camera's internal parameters and distortion coefficients, calculate the distortion offset of each pixel in the food image based on the distortion coefficients, and extract the timestamp of each data point in the environmental multidimensional sensing data and the timestamp of the video frame corresponding to the food image after reverse mapping of the pixel coordinates. S23. Find the first most recent environmental multidimensional sensor data timestamp before the video frame timestamp and the second most recent environmental multidimensional sensor data timestamp. Use linear interpolation to calculate the interpolated data at the video frame timestamp based on the data values ​​corresponding to the two environmental multidimensional sensor data timestamps. Combine the interpolated data with the video frame to generate a standardized spatiotemporal input dataset containing multimodal information.

[0022] In this embodiment, S3 specifically includes: S31. Input the food images in the standardized spatiotemporal input dataset into the image segmentation network, classify each pixel, mark the pixels belonging to the main body of the food as the foreground region, calculate the local change frequency of the gray value of the pixels in the foreground region, count the number of pixels with a gray value change frequency higher than the preset frequency, calculate the ratio of the number of pixels to the total number of pixels in the foreground region, and quantify it as the texture density of the food surface. S32. Extract pixel coordinates along the boundary of the foreground region, construct edge contour curves based on pixel coordinates, calculate the change in the enclosed area between the edge contour curve at the current time and the edge contour curve at the previous time, and mark it as edge contour evolution. S33. Traverse the grayscale value of each pixel in the foreground area, find the pixels with grayscale values ​​higher than the preset brightness threshold, count the distribution density of the brightness threshold pixels in the foreground area, and map them to the pixel distribution information of the oil precipitation spot. S34. Combine texture density, edge contour evolution and pixel distribution information into a high-dimensional feature vector, and use a preset fully connected layer to perform deep encoding on the high-dimensional feature vector to generate and output the food micro-spectral feature encoding.

[0023] In this embodiment, S4 specifically includes: S41. Input the microscopic spectral feature encoding of the ingredients as a query vector into the preset cooking knowledge graph, traverse all the ingredient entity nodes stored in the cooking knowledge graph, calculate the cosine similarity between the query vector and the feature vector of each ingredient entity node, select the ingredient entity node with the largest value as the target matching node, read the attribute label data associated with the target matching node, and parse it into the best ripeness data of the corresponding ingredient attribute. S42. Construct the branch sub-networks contained in the DeepONet model, input the optimal maturity data into the input layer of the branch sub-network, encode the optimal maturity data using a multi-layer preset fully connected neural network, multiply it with the preset weight matrix and superimpose the bias vector, process it through the tanh activation function, output a latent code vector of a set dimension, and use it as the boundary condition parameter to control the boundary of the heat conduction equation. S43. Construct the backbone sub-network of the DeepONet model. The microscopic spectral features of the food are encoded as input features and input to the backbone sub-network of the pre-sized convolutional neural network structure containing four residual blocks. Each residual block performs two convolution operations on the input features and adds them to the input features. The added features are then subjected to global average pooling to output the spatiotemporal features of the solution domain describing the geometry and heating position of the food. S44. The DeepONet model reads the boundary condition parameters and the spatiotemporal features of the solution domain, uses the boundary condition parameters as function coefficients, and uses the spatiotemporal features of the solution domain as function independent variables. It performs a dot product operation to fuse the two, generates a fused feature tensor, constructs the heat conduction partial differential equation formula, and calculates the first derivative of the fused feature tensor with respect to the time coordinate to obtain the temperature change rate. S45. Calculate the second derivative of the fused feature tensor with respect to spatial coordinates to obtain the curvature of the temperature spatial distribution. Subtract the product of the temperature spatial distribution curvature and the preset thermal diffusivity from the rate of temperature change, and calculate the difference between the two as the physical residual. Use the mean square error to convert the physical residual into a loss value, and use the gradient descent algorithm to adjust the network weights to make the loss value converge. S46. Input the fused feature tensor after updating the weights into the DeepONet model for forward inference calculation, and output the predicted temperature distribution matrix to approximate the solution function of the heat conduction partial differential equation, thereby obtaining a nonlinear mapping of the thermodynamic evolution process inside the food. S47. Based on the results of nonlinear mapping, the DeepONet model performs differential operations on the fusion feature tensor along the time dimension, calculates the derivative of moisture change over time, converts the derivative value into mass units, generates the dehydration rate of the food at the current moment, and performs integral operations on the fusion feature tensor along the spatial dimension. The depth where the cumulative temperature is higher than the preset protein denaturation temperature is the core heat penetration depth. The dehydration rate and the core heat penetration depth are the current state data. S48. Read the historical cooking records stored in the system memory, extract the cavity center temperature data within a preset time period before the current moment, arrange the cavity center temperature data in chronological order, connect them to generate a historical temperature control curve, and perform vector splicing of the dehydration rate and core heat penetration depth of the ingredients at the current moment with the historical temperature control curve to generate a current cooking state feature set and output it.

[0024] This invention combines culinary knowledge graph retrieval with the DeepONet operator learning mechanism to achieve precise nonlinear mapping and state quantification of the internal thermodynamic evolution of ingredients. First, the cosine similarity between ingredient features and graph nodes is calculated to accurately match target ingredients and obtain optimal ripeness data, which is then encoded into boundary condition parameters using a branch sub-network. Simultaneously, the spatiotemporal features of the ingredient's geometry and heated location are extracted through the backbone sub-network, and these two are fused to construct a partial differential equation for heat conduction. Physical residuals are then calculated to optimize network weights. This method effectively addresses the problem of traditional models' difficulty in accurately predicting unsteady heat conduction processes, significantly improving the accuracy of derivation of key states such as dehydration rate and core heat penetration depth. By approximating the solution function of the real physical process, reliable data support is provided for subsequent adaptive heat intervention, ensuring high-precision control of the cooking process.

[0025] In this embodiment, S5 specifically includes: S51. Input the current cooking state feature set into the improved N-BEATS model containing a double-layer stacked structure. The first layer stacked structure performs a one-dimensional convolution operation with a preset kernel on the historical temperature control curve to extract the trend term. The second layer stacked structure performs a preset fully connected operation on the current state data to extract the seasonal term. The trend term and the seasonal term are added point by point to obtain a deep nonlinear decomposition of the historical temperature control curve and the current dehydration rate and core heat penetration depth state data. S52. Construct a physical constraint loss function at the output of the improved N-BEATS model, read the cavity center temperature and the current power percentage of the heating tube from the environmental multidimensional sensing data, multiply the rated power of the heating tube by the current power percentage to obtain the actual output power, divide it by the heat dissipation area of ​​the heating tube to obtain the heat flux density, and divide the heat flux density by the temperature difference between the cavity center temperature and the food surface temperature to obtain the comprehensive heat transfer coefficient. S53. Multiply the overall heat transfer coefficient by the difference between the cavity center temperature and the food surface temperature to obtain the input heat flux. Use the Fourier heat conduction formula to divide the input heat flux by the food thermal conductivity to obtain the theoretical heat flux density along the food thickness direction. Use the theoretical heat flux density to divide by the product of the food specific heat capacity and density to calculate the theoretical temperature penetration depth. Mark the theoretical temperature penetration depth as the actual penetration depth. S54. Extract the core heat penetration depth from the current cooking state feature set, calculate the difference between the actual penetration depth and the core heat penetration depth, and define it as the physical state deviation value. Construct a deviation correction matrix with the same dimension as the hidden layer of the improved N-BEATS model, and add each element in the deviation correction matrix to the corresponding element in the hidden layer weight matrix to complete the dynamic correction of the network weights of the improved N-BEATS model. S55. The improved N-BEATS model with corrected weights is used to perform forward computation on the input features. The multiplication operation between the current cooking state feature set and the weight matrix is ​​performed in the preset fully connected layer and the bias term is added. The corrected feature vector is output after passing through the ReLU activation function, and the original values ​​in the current cooking state feature set are replaced by the corrected feature vector. S56. Perform an inverse transformation operation on the dynamically corrected feature vector, project the dynamically corrected feature vector back into the data space, calculate the first difference value of adjacent data points in the projected sequence, filter out data points whose absolute value of the first difference value is less than the preset change threshold, record the timestamp corresponding to the filtered data points, and mark it as the critical evaporation time of the internal moisture of the food in the reverse inference. S57. Read the current time of the system clock and calculate the difference between the current time and the critical evaporation time. Multiply the time difference by the preset heat dissipation coefficient to obtain the heat deficit. Divide the heat deficit by the rated power of the heating tube to obtain the compensation heating time. Convert the compensation heating time into a control duty cycle signal to generate an adaptive heat intervention command to prevent the food from drying out excessively.

[0026] This invention introduces an improved N-BEATS model combined with a physical constraint mechanism to achieve deep nonlinear decomposition and dynamic deviation correction of cooking state characteristics. A double-layer stacked structure is used to extract the trend term from historical temperature control curves and the seasonal term from current state data, achieving multi-dimensional state decoupling. Theoretical heat flux density and actual penetration depth are calculated using heating power and temperature parameters to construct a physical constraint loss function. The deviation between the actual and predicted depths is mapped to a weight correction matrix, updating network parameters in real time. This invention effectively overcomes the lack of physical interpretability in data-driven models, accurately capturing the critical evaporation moment of moisture inside ingredients. By dynamically correcting feature vectors and inversely transforming them to deduce key time points, heat intervention commands are adaptively generated, ensuring intelligent control of the cooking process and the final tenderness of the food.

[0027] In this embodiment, the construction of a bias correction matrix with the same hidden layer dimension as the improved N-BEATS model specifically includes: reading the number of rows and columns of the hidden layer weights of the improved N-BEATS model, creating an all-zero matrix as an initialization matrix, multiplying the physical state bias value by a preset learning rate coefficient to obtain a scaling factor, replacing each element value in the initialization matrix with the scaling factor, and constructing a bias correction matrix with the same hidden layer dimension as the improved N-BEATS model.

[0028] In this embodiment, S6 specifically includes: adjusting the heating tube power according to the adaptive heat intervention command, controlling the equipment operation status in real time during the cooking process, continuously monitoring the changes in the food status after the control, acquiring the updated food image through the image acquisition device, and inputting it into the image segmentation network for cyclic detection. When the scorch coverage rate on the food surface reaches a preset threshold, a cooking completion signal is generated and the air fryer is controlled to stop heating.

[0029] An air fryer control system based on Internet of Things (IoT) data acquisition includes: The IoT data acquisition and synchronization module is used to collect food images and multi-dimensional sensor data of the environment; The multimodal data preprocessing module is used to preprocess food images and align the timestamps of environmental multidimensional sensor data with the timestamps of video frames to generate a standardized spatiotemporal input dataset. The food micro-feature extraction module is used to extract pixel-level features from food images in a standardized spatiotemporal input dataset, segment and identify the main food region and encode it deeply, and output the food micro-spectral feature code. The thermodynamic evolution mapping module is used to search and match in the preset cooking knowledge graph to obtain the best cooking data, and input it into the DeepONet model along with the food micro-spectral feature encoding. Through the operator learning mechanism, the thermodynamic evolution process inside the food is nonlinearly mapped to generate the current cooking state feature set. The physical deviation correction and prediction module is used to input the current cooking state feature set into the improved N-BEATS model, perform deep nonlinear decomposition, introduce a physical state deviation correction mechanism, dynamically correct the current cooking state feature set, reverse-engineer the critical evaporation time, and output adaptive heat intervention instructions. The adaptive control and execution module is used to make real-time adjustments based on the adaptive heat intervention command, monitor changes in the state of the food, acquire updated food images and perform cyclic detection. When the coverage of scorch marks on the surface of the food reaches a preset threshold, a cooking completion signal is generated and the air fryer is controlled to stop heating.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to the intelligent central kitchen of a well-known chain restaurant brand in a certain city. The workshop processes more than 5 tons of various ingredients daily and is equipped with 20 large-capacity industrial-grade intelligent air fryers. Due to differences in the biological characteristics of the raw materials and external interferences such as temperature, humidity, and voltage fluctuations in the workshop environment, the traditional fixed temperature control mode based on PID algorithm is difficult to accurately control the cooking endpoint. Typical problems include burnt or undercooked surfaces of potato products and substandard core temperatures, resulting in high scrap rates and inconsistent taste. The platform receives massive amounts of heterogeneous data daily, including high-definition images of the inside of the air fryers, multi-channel NTC temperature arrays, and humidity values. Traditional systems rely on preset time-temperature curves, lacking awareness of the actual state of the ingredients and unable to cope with interferences such as heat loss upon opening the door and changes in the thickness of the food stack, easily leading to thermal imbalance and batch defects.

[0031] In practical deployment, multi-source sensor data undergoes spatiotemporal alignment and preprocessing. An improved DeiT network extracts surface texture density, color gradient, and edge evolution features of the food ingredients and encodes them as microscopic spectral vectors. Environmental data is then fused to construct a standardized spatiotemporal input dataset. This dataset is input into the DeepONet model. Sub-networks process the optimal ripeness boundary conditions from knowledge graph matching, while the backbone network processes geometric morphology and spatiotemporal texture features. An operator learning mechanism rapidly solves the partial differential equation of heat conduction, outputting the dehydration rate and core heat penetration depth in real time. To eliminate data-driven model bias, an improved N-BEATS model is introduced. Fourier's law is used to calculate the theoretical penetration depth, which is compared with the DeepONet prediction results to calculate the physical state deviation value. This value is used to construct a dynamic deviation correction matrix, adjusting the weights of the N-BEATS model's hidden layers in real time. This accurately captures the critical moment of moisture evaporation, inversely extrapolates the heat deficit and compensation time, and ultimately generates adaptive heat intervention commands to adjust the heating element duty cycle in real time, ensuring optimal cooking. Table 1 below shows a detailed performance comparison of the method of the present invention and the traditional fixed temperature control method in cooking tasks for three typical types of ingredients: Table 1. Performance Indicators Comparison between the Invention and Traditional Methods

[0032] According to the comparative data shown in Table 1, the intelligent cooking control method proposed in this invention shows significant performance advantages over traditional methods in all key indicators.

[0033] Regarding the core temperature compliance rate, this invention maintains a rate above 96% across all three types of ingredients, significantly higher than the average of approximately 80% achieved by traditional methods. For example, in the scenario of "marinating chicken wings," traditional methods, limited by individual differences, achieve a compliance rate of only 76.8%, while this invention, through precise sensing, raises the compliance rate to 96.5%, effectively solving the problem of meat being "burnt on the outside and raw on the inside" during cooking.

[0034] In terms of cooking timeliness, the present invention controls the prediction error of the end of dehydration to within 10 seconds on average, while the error of traditional methods is as high as 45 seconds or more. This significantly shortens the average cooking time per batch, for example, the cooking efficiency of French fries is significantly improved.

[0035] Furthermore, this invention significantly reduces the surface scorch mark coverage deviation from over ±15% using traditional methods to within ±3%, and lowers the defect rate from around 10% to below 2%, greatly improving the consistency of the finished product. Simultaneously, unit energy consumption is reduced by approximately 6% to 8%, and the taste score remains consistently above 9 points. Overall, this invention, through a physically enhanced deep learning model, achieves high-precision control and energy efficiency optimization of the cooking process, demonstrating significant application value.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for controlling an air fryer based on Internet of Things (IoT) data acquisition, characterized in that, Includes the following steps: S1. Synchronously collect images of food and multi-dimensional environmental sensor data inside the air fryer through an IoT gateway; S2. Preprocess the food images and align the timestamps of the environmental multidimensional sensor data with the timestamps of the video frames to generate a standardized spatiotemporal input dataset. S3. Extract pixel-level features from food images in the standardized spatiotemporal input dataset, segment and identify the main food region, and deeply encode the texture density, edge contour evolution, and pixel distribution information of oil precipitation spots on the food surface, outputting the food microspectral feature code. S4. Based on the microscopic spectral feature encoding of the ingredients, search and match in the preset cooking knowledge graph to obtain the optimal ripeness data of the corresponding ingredients attributes, and input the ingredients microscopic spectral feature encoding into the DeepONet model to perform nonlinear mapping on the thermodynamic evolution process, generate the dehydration rate and core heat penetration depth of the ingredients at the current moment, and output them as the current cooking state feature set along with the historical temperature control curve. S5. Input the current cooking state feature set into the improved N-BEATS model for deep nonlinear decomposition, introduce a physical state deviation correction mechanism, dynamically correct the current cooking state feature set, reverse-engineer the critical evaporation time of the internal moisture of the ingredients, and output an adaptive heat intervention command. S6. Based on the adaptive heat intervention command, perform real-time adjustment, acquire updated food images, and after cyclic detection, generate a cooking completion signal and stop heating.

2. The air fryer control method based on Internet of Things data acquisition according to claim 1, characterized in that, S1 specifically includes: calling a preset synchronous acquisition interface through an IoT gateway to trigger an image acquisition device and multiple sets of environmental sensors deployed inside the air fryer cavity, capturing images of the food inside the air fryer at the current moment, and simultaneously reading the data stream transmitted by the environmental sensors to obtain multi-dimensional environmental sensing data inside the air fryer. The multi-dimensional environmental sensing data specifically includes the cavity center temperature, humidity sensor readings, current power percentage of the heating element, and weight change data of the food tray.

3. The air fryer control method based on Internet of Things data acquisition according to claim 1, characterized in that, S2 specifically includes: S21. Perform noise reduction preprocessing, read the grayscale values ​​of the pixels in the food image, calculate the average grayscale value of the preset neighborhood and replace the original value, statistically analyze the grayscale histogram of the food image, calculate the average brightness of the grayscale histogram, and calculate the exposure adjustment coefficient based on the difference between the average brightness and the preset target brightness. Multiply the grayscale value of each pixel in the food image by the exposure adjustment coefficient. S22. Read the camera's internal parameters and distortion coefficients, calculate the distortion offset of each pixel in the food image based on the distortion coefficients, and extract the timestamp of each data point in the environmental multidimensional sensing data and the timestamp of the video frame corresponding to the food image after reverse mapping of the pixel coordinates. S23. Find the first most recent environmental multidimensional sensor data timestamp before the video frame timestamp and the second most recent environmental multidimensional sensor data timestamp. Use linear interpolation to calculate the interpolated data at the video frame timestamp based on the data values ​​corresponding to the two environmental multidimensional sensor data timestamps. Combine the interpolated data with the video frame to generate a standardized spatiotemporal input dataset.

4. The air fryer control method based on Internet of Things data acquisition according to claim 1, characterized in that, S3 specifically includes: S31. Input the food images in the standardized spatiotemporal input dataset into the image segmentation network, classify each pixel, mark the pixels belonging to the main body of the food as the foreground region, calculate the local change frequency of the gray value of the pixels in the foreground region, count the number of pixels with a gray value change frequency higher than the preset frequency, calculate the ratio of the number of pixels to the total number of pixels in the foreground region, and quantify it as the texture density of the food surface. S32. Extract pixel coordinates along the boundary of the foreground region, construct edge contour curves based on pixel coordinates, calculate the change in the enclosed area between the edge contour curve at the current time and the edge contour curve at the previous time, and mark it as edge contour evolution. S33. Traverse the grayscale value of each pixel in the foreground area, find the pixels with grayscale values ​​higher than the preset brightness threshold, count the distribution density of the brightness threshold pixels in the foreground area, and map them to the pixel distribution information of the oil precipitation spot. S34. Combine texture density, edge contour evolution and pixel distribution information into a high-dimensional feature vector, and use a preset fully connected layer to perform deep encoding on the high-dimensional feature vector to generate and output the food micro-spectral feature encoding.

5. The air fryer control method based on Internet of Things data acquisition according to claim 1, characterized in that, S4 specifically includes: S41. Input the microscopic spectral feature encoding of the ingredients as a query vector into the preset cooking knowledge graph, traverse all the ingredient entity nodes stored in the cooking knowledge graph, calculate the cosine similarity between the query vector and the feature vector of each ingredient entity node, select the ingredient entity node with the largest value as the target matching node, read the attribute label data associated with the target matching node, and parse it into the best ripeness data of the corresponding ingredient attribute. S42. Construct the branch sub-networks contained in the DeepONet model, input the optimal maturity data into the input layer of the branch sub-network, encode the optimal maturity data using a multi-layer preset fully connected neural network, multiply it with the preset weight matrix and superimpose the bias vector, process it through the tanh activation function, output a latent code vector of a set dimension, and use it as the boundary condition parameter to control the boundary of the heat conduction equation. S43. Construct the backbone sub-network of the DeepONet model. The microscopic spectral features of the food ingredients are encoded as input features and input to the backbone sub-network of the pre-sized convolutional neural network structure containing four residual blocks. Each residual block performs two convolution operations on the input features and adds them to the input features. The added features are then subjected to global average pooling to output the spatiotemporal features of the solution domain. S44. The DeepONet model reads the boundary condition parameters and the spatiotemporal features of the solution domain, uses the boundary condition parameters as function coefficients, and uses the spatiotemporal features of the solution domain as function independent variables. It performs a dot product operation to fuse the two, generates a fused feature tensor, constructs the heat conduction partial differential equation formula, and calculates the first derivative of the fused feature tensor with respect to the time coordinate to obtain the temperature change rate. S45. Calculate the second derivative of the fused feature tensor with respect to spatial coordinates to obtain the curvature of the temperature spatial distribution. Subtract the product of the temperature spatial distribution curvature and the preset thermal diffusivity from the rate of temperature change, and calculate the difference between the two as the physical residual. Use the mean square error to convert the physical residual into a loss value, and use the gradient descent algorithm to adjust the network weights to make the loss value converge. S46. Input the fused feature tensor after updating the weights into the DeepONet model for forward inference calculation, and output the predicted temperature distribution matrix to approximate the solution function of the heat conduction partial differential equation, thereby obtaining a nonlinear mapping of the thermodynamic evolution process inside the food. S47. Based on the results of nonlinear mapping, the DeepONet model performs differential operations on the fusion feature tensor along the time dimension, calculates the derivative of moisture change over time, converts the derivative value into mass units, generates the dehydration rate of the food at the current moment, and performs integral operations on the fusion feature tensor along the spatial dimension. The depth where the cumulative temperature is higher than the preset protein denaturation temperature is the core heat penetration depth. The dehydration rate and the core heat penetration depth are the current state data. S48. Read the historical cooking records stored in the system memory, extract the cavity center temperature data within a preset time period before the current moment, arrange the cavity center temperature data in chronological order, connect them to generate a historical temperature control curve, and perform vector splicing of the dehydration rate and core heat penetration depth of the ingredients at the current moment with the historical temperature control curve to generate a current cooking state feature set and output it.

6. The air fryer control method based on Internet of Things data acquisition according to claim 1, characterized in that, S5 specifically includes: S51. Input the current cooking state feature set into the improved N-BEATS model containing a double-layer stacked structure. The first layer stacked structure performs a one-dimensional convolution operation with a preset kernel on the historical temperature control curve to extract the trend term. The second layer stacked structure performs a preset fully connected operation on the current state data to extract the seasonal term. The trend term and the seasonal term are added point by point to obtain a deep nonlinear decomposition of the historical temperature control curve and the current dehydration rate and core heat penetration depth state data. S52. Construct a physical constraint loss function at the output of the improved N-BEATS model, read the cavity center temperature and the current power percentage of the heating tube from the environmental multidimensional sensing data, multiply the rated power of the heating tube by the current power percentage to obtain the actual output power, divide it by the heat dissipation area of ​​the heating tube to obtain the heat flux density, and divide the heat flux density by the temperature difference between the cavity center temperature and the food surface temperature to obtain the comprehensive heat transfer coefficient. S53. Multiply the overall heat transfer coefficient by the difference between the cavity center temperature and the food surface temperature to obtain the input heat flux. Use the Fourier heat conduction formula to divide the input heat flux by the food thermal conductivity to obtain the theoretical heat flux density along the food thickness direction. Use the theoretical heat flux density to divide by the product of the food specific heat capacity and density to calculate the theoretical temperature penetration depth. Mark the theoretical temperature penetration depth as the actual penetration depth. S54. Extract the core heat penetration depth from the current cooking state feature set, calculate the difference between the actual penetration depth and the core heat penetration depth, and define it as the physical state deviation value. Construct a deviation correction matrix with the same dimension as the hidden layer of the improved N-BEATS model, and add each element in the deviation correction matrix to the corresponding element in the hidden layer weight matrix to complete the dynamic correction of the network weights of the improved N-BEATS model. S55. The improved N-BEATS model with corrected weights is used to perform forward computation on the input features. The multiplication operation between the current cooking state feature set and the weight matrix is ​​performed in the preset fully connected layer and the bias term is added. The corrected feature vector is output after passing through the ReLU activation function, and the original values ​​in the current cooking state feature set are replaced by the corrected feature vector. S56. Perform an inverse transformation operation on the dynamically corrected feature vector, project the dynamically corrected feature vector back into the data space, calculate the first difference value of adjacent data points in the projected sequence, filter out data points whose absolute value of the first difference value is less than the preset change threshold, record the timestamp corresponding to the filtered data points, and mark it as the critical evaporation time of the internal moisture of the food in the reverse inference. S57. Read the current time of the system clock and calculate the difference between the current time and the critical evaporation time. Multiply the time difference by the preset heat dissipation coefficient to obtain the heat deficit. Divide the heat deficit by the rated power of the heating tube to obtain the compensation heating time. Convert the compensation heating time into a control duty cycle signal to generate an adaptive heat intervention command to prevent the food from drying out excessively.

7. The air fryer control method based on Internet of Things data acquisition according to claim 6, characterized in that, The construction of the bias correction matrix with the same hidden layer dimension as the improved N-BEATS model specifically includes: reading the number of rows and columns of the hidden layer weights of the improved N-BEATS model, creating an all-zero matrix as the initialization matrix, multiplying the physical state bias value by a preset learning rate coefficient to obtain the scaling factor, replacing each element value in the initialization matrix with the scaling factor, and constructing the bias correction matrix with the same hidden layer dimension as the improved N-BEATS model.

8. The air fryer control method based on Internet of Things data acquisition according to claim 1, characterized in that, S6 specifically includes: adjusting the heating tube power according to the adaptive heat intervention command, controlling the equipment operation status in real time during the cooking process, continuously monitoring the changes in the state of the food after the control, acquiring the updated food image through the image acquisition device, and inputting it into the image segmentation network for cyclic detection. When the coverage rate of scorch marks on the surface of the food reaches a preset threshold, a cooking completion signal is generated and the air fryer is controlled to stop heating.

9. An air fryer control system based on Internet of Things (IoT) data acquisition, comprising the air fryer control method based on IoT data acquisition as described in any one of claims 1 to 8, characterized in that, include: The IoT data acquisition and synchronization module is used to collect food images and multi-dimensional sensor data of the environment; The multimodal data preprocessing module is used to preprocess food images and align the timestamps of environmental multidimensional sensor data with the timestamps of video frames to generate a standardized spatiotemporal input dataset. The food micro-feature extraction module is used to extract pixel-level features from food images in a standardized spatiotemporal input dataset, segment and identify the main food region and encode it deeply, and output the food micro-spectral feature code. The thermodynamic evolution mapping module is used to search and match in the preset cooking knowledge graph to obtain the best cooking data, and input it into the DeepONet model along with the food micro-spectral feature encoding. Through the operator learning mechanism, the thermodynamic evolution process inside the food is nonlinearly mapped to generate the current cooking state feature set. The physical deviation correction and prediction module is used to input the current cooking state feature set into the improved N-BEATS model, perform deep nonlinear decomposition, introduce a physical state deviation correction mechanism, dynamically correct the current cooking state feature set, reverse-engineer the critical evaporation time, and output adaptive heat intervention instructions. The adaptive control and execution module is used to make real-time adjustments based on the adaptive heat intervention command, monitor changes in the state of the food, acquire updated food images and perform cyclic detection. When the coverage of scorch marks on the surface of the food reaches a preset threshold, a cooking completion signal is generated and the air fryer is controlled to stop heating.

Images