Multi-mode control method and system for intelligent variable-frequency cooking equipment

Abstract

The invention provides a multi-mode control method and system for intelligent variable-frequency cooking equipment, and relates to the technical field of intelligent control, and the method comprises the steps: obtaining a food material image and weight data through an image collection device, extracting feature information through a deep learning processor based on a pre-trained food material feature recognition model, and generating a feature data packet containing food material type, weight and volume information. The controller matches initial cooking mode parameters according to the data packet, collects real-time operation data such as temperature, pressure, heating power and stirring motor rotating speed in the cavity, and calculates power-temperature and motor rotating speed-torque response curves. And the controller generates a frequency conversion control instruction according to the response curve and a preset multi-mode switching judgment rule, and controls the heating unit and the stirring motor to perform cooperative adjustment. In the cooking process, real-time operation data are continuously collected, multi-mode self-adaptive switching is carried out, and the frequency conversion control instruction is optimized until cooking is completed.

Description

Technical Field

[0001] The present invention relates to intelligent control technology, and in particular to a multi-mode control method and system for intelligent variable frequency cooking equipment. Background Art

[0002] At present, intelligent cooking equipment is developing towards automation, multi-function and personalization. However, existing intelligent cooking equipment has some defects and shortcomings in the cooking process: The cooking mode is single and cannot be finely controlled according to the characteristics of ingredients and user preferences. The preset cooking mode is often relatively fixed and difficult to adapt to the characteristics of different ingredients, such as size, thickness, freshness, etc., resulting in unsatisfactory cooking results.

[0003] Lack of real-time dynamic adjustment, low control accuracy of the cooking process. Most existing equipment uses simple feedback control, which cannot make real-time adjustments according to changes in food status and cooking environment, such as heating power, stirring speed, etc., making it difficult to accurately control the cooking process.

[0004] The switching of cooking modes is abrupt and lacks smooth transition. In the process of making dishes that require switching between different cooking modes, existing equipment often switches parameters directly and lacks smooth transition control, which easily causes temperature fluctuations or sudden changes in stirring speed, affecting the cooking effect. Summary of the invention

[0005] The embodiments of the present invention provide a multi-mode control method and system for intelligent variable frequency cooking equipment, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, Provides a multi-mode control method for intelligent variable frequency cooking equipment, including: The image of the food is acquired through an image acquisition device, and the weight data of the food is collected at the same time. The image acquisition device transmits the food image to a deep learning processor. The deep learning processor extracts feature information from the food image based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data, wherein the food feature data packet includes food type information, weight information and volume information; the food feature recognition model is obtained by transferring the knowledge of the teacher network to the student network, wherein the multi-scale feature map of the teacher network is mapped to the feature space of the student network through a feature alignment module, and the sample difficulty is determined based on the predicted probability distribution entropy value, feature point density value and annotation difficulty value of the sample, and a hierarchical training strategy is used to optimize the model; Based on the food feature data packet, the controller matches the corresponding initial cooking mode parameters from a preset cooking mode database, and collects real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data, and the controller calculates a power-temperature dynamic response curve and a motor speed-torque response curve according to the real-time operation data; The controller generates a variable frequency control instruction based on the power-temperature dynamic response curve and the motor speed-torque response curve in combination with a preset multi-mode switching judgment rule, and coordinates the heating unit and the stirring motor unit of the cooking device through the variable frequency control module; the controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the variable frequency control instruction based on the switched mode until cooking is completed.

[0007] In an optional embodiment, The steps of acquiring food images and food weight data through an image acquisition device, wherein the image acquisition device transmits the food images to a deep learning processor, wherein the deep learning processor extracts feature information from the food images based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data include: The image of the food is acquired by an image acquisition device disposed on the inner side of the cover of the cooking device, wherein the image acquisition device comprises a CMOS sensor, a waterproof glass and a heat-insulating layer, wherein the CMOS sensor is disposed within the protection of the waterproof glass and the heat-insulating layer, and the image acquisition device uses an adaptive exposure algorithm to adjust the exposure time and gain of the acquired food image; The weight data of the food is obtained by a strain gauge weighing sensor arranged at the base of the cooking device, wherein the output end of the strain gauge weighing sensor is connected to a low-pass filter and a temperature compensation circuit, wherein the low-pass filter is used to eliminate vibration interference, and the temperature compensation circuit is used to eliminate the influence of ambient temperature on the weighing accuracy; The food image is input into a deep learning processor, and the deep learning processor processes the food image based on a pre-trained food feature recognition model, wherein the food feature recognition model adopts a ResNet50 network structure provided with a spatial pyramid pooling module, and the deep learning processor extracts features from the input food image through a feature extraction network of the food feature recognition model to obtain an image feature vector; the image feature vector is input into a feature dimension reduction module, and the feature dimension reduction module performs dimension reduction processing on the image feature vector using a principal component analysis method to obtain a reduced dimension feature vector; The reduced-dimensional feature vector and the food weight data are input into a feature fusion network, which includes a plurality of fully connected layers connected in series. The feature fusion network performs feature fusion on the reduced-dimensional feature vector and the normalized food weight data to generate a food feature data packet including food type information, weight information, volume information and appearance feature parameters, wherein the food volume information is calculated based on an image segmentation result and a pre-calibrated correspondence between pixels and actual sizes.

[0008] In an optional embodiment, The steps of training the food feature recognition model include: Collecting food images and cooking process images and annotating them to obtain a training data set, wherein the annotated content includes food type information, weight information, volume information and feature point location information; using the training data set to train a feature extraction network; processing the output features of the feature extraction network through a spatial feature processing branch and a temporal feature processing branch and weighted fusion to obtain a fusion feature; The fusion feature is input into a multi-task learning network, the multi-task learning network includes a food classification module, a feature point positioning module and a weight and volume prediction module, the food classification module is trained by a cross entropy loss function, the feature point positioning module is trained by a Euclidean distance loss function, and the weight and volume prediction module is trained by a mean square error loss function to obtain a teacher network; knowledge transfer is performed on the teacher network to obtain a student network, including: mapping the multi-scale feature map of the teacher network to the feature space of the student network through a feature alignment module, constructing a spatial attention matrix and a channel attention matrix of the multi-scale feature map of the teacher network and transferring them to the student network, and transferring the prediction result of the teacher network to the student network through a soft label method, wherein the student network adopts a depth-separable convolutional layer structure, and a channel rearrangement module is set between adjacent network layers; The sample difficulty is determined according to the prediction result of the student network. The sample difficulty is obtained by weighting the entropy value of the predicted probability distribution of the sample by the student network, the density value of the feature points in the sample and the annotation difficulty value of the training data set. The training samples are stratified based on the sample difficulty. During the training process, the sampling ratio of samples of different difficulties is adjusted according to the training rounds; data enhancement processing and clarity detection, feature point detection and signal-to-noise ratio detection are performed on the sampled samples; the enhanced samples that pass the detection are input into the student network for iterative training until the loss value of the student network converges.

[0009] In an optional embodiment, Based on the food feature data packet, the controller matches the corresponding initial cooking mode parameters from a preset cooking mode database, and collects real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data. The controller calculates the power-temperature dynamic response curve and the motor speed-torque response curve according to the real-time operation data, and the steps include: Constructing a cooking mode database, the cooking mode database comprising food attribute information and corresponding cooking parameter combinations, the food attribute information comprising weight characteristic value, volume characteristic value, water content characteristic value, shape regularity, surface smoothness and hardness characteristic value; constructing a to-be-cooked food feature vector according to the food feature data packet, calculating the feature importance of the food attribute information in the to-be-cooked food feature vector according to different cooking modes, calculating a feature weight coefficient based on the feature importance, and the feature weight coefficient is obtained by mapping and normalizing the feature importance through an exponential function; The feature weight coefficient is used to weight the feature vector of the to-be-cooked food, and the weighted threshold cosine similarity algorithm is used to calculate the similarity between the weighted feature vector of the to-be-cooked food and the food attribute information in the cooking mode database, wherein the weighted threshold cosine similarity algorithm sets an adaptive threshold constraint for each feature value difference, and the adaptive threshold is obtained through data statistics in the cooking mode database; the top three matching ingredients with the highest similarity are selected, and the cooking parameter combinations corresponding to the matching ingredients are weightedly fused based on the feature weight coefficient to obtain initial cooking parameters, wherein the initial cooking parameters include a heating power parameter and a stirring speed parameter; The temperature data, pressure data, heating power data and stirring motor speed data in the cavity of the cooking equipment are collected, and the initial cooking parameters are adaptively corrected according to the deviation between the reference cooking effect in the cooking mode database and the actual cooking effect; a second-order transfer function model is established based on the temperature data and the heating power data to obtain a power-temperature dynamic response curve, and a torque model considering the moment of inertia and viscous damping is established based on the stirring motor speed data and the motor current data to obtain a motor speed-torque response curve.

[0010] In an optional embodiment, The controller generates a variable frequency control instruction based on the power-temperature dynamic response curve and the motor speed-torque response curve in combination with a preset multi-mode switching judgment rule, and the step of cooperatively adjusting the heating unit and the stirring motor unit of the cooking device through the variable frequency control module includes: Construct a system state vector including temperature, rotation speed, temperature change rate and rotation speed change rate, design a Kalman state observer based on a state transfer matrix and an observation matrix to estimate the system state vector, and optimize the gain matrix of the Kalman state observer through a system noise covariance matrix; design a state coupling matrix based on the estimated system state vector, and obtain the coupling coefficient of the state coupling matrix through particle swarm optimization; Based on the estimated system state vector, the working mode is identified in combination with a preset multi-mode switching determination rule, wherein the multi-mode switching determination rule includes a temperature steady-state error threshold, a speed fluctuation range threshold, a temperature change rate threshold, and a speed change rate threshold, and the control mode to be currently adopted is determined according to the comparison result between each component of the system state vector and the corresponding threshold; A coupling control law is designed according to the state coupling matrix and the currently adopted control mode, wherein the coupling control law associates the heating power adjustment amount with the speed adjustment amount through the coupling coefficient, wherein the heating power adjustment amount is composed of a proportional integral term of the temperature deviation and a coupling term of the estimated system state vector, and the speed adjustment amount is composed of a proportional differential term of the speed deviation and a coupling term of the estimated system state vector, and the coupling coefficient is updated in real time by designing an adaptive mechanism based on a weighted product of an estimated system state vector error and an actual system state vector error; The output of the coupling control law is converted into a variable frequency control instruction, wherein the heating power variable frequency control instruction is composed of a proportional integral term of a reference frequency and a heating power adjustment amount, and the stirring motor variable frequency control instruction is composed of a proportional integral term of a reference frequency and a speed adjustment amount; the variable frequency control instruction is converted into a control signal through pulse width modulation to coordinately control the heating unit and the stirring motor unit of the cooking equipment.

[0011] In an optional embodiment, The coupling control law associates the heating power adjustment amount with the speed adjustment amount through the coupling coefficient, wherein the heating power adjustment amount is composed of a proportional integral term of the temperature deviation and a coupling term of the estimated system state vector, and the speed adjustment amount is composed of a proportional differential term of the speed deviation and a coupling term of the estimated system state vector, and the coupling coefficient is updated in real time by designing an adaptive mechanism based on a weighted product of an estimated system state vector error and an actual system state vector error, including: Obtain an instantaneous state error vector including a temperature error, a rotation speed error, a temperature change rate error, and a rotation speed change rate error, obtain a weighted state error vector based on an exponentially weighted sliding average of the instantaneous state error vector and a historical state error vector, wherein the exponentially weighted sliding average performs attenuation processing on historical error information through a forgetting factor; based on the instantaneous state error vector, calculate a difference between state error vectors at two adjacent sampling moments to obtain a state error change amount, and divide the state error change amount by a sampling period to obtain a state error change rate; Constructing an error covariance matrix considering the cross-coupling terms of temperature and speed, the error covariance matrix including the outer product term of the weighted state error vector, and a coupling strength term consisting of the product of the temperature error and the speed change rate error, and the product of the speed error and the temperature change rate error, calculated based on the state error change rate; Designing a dynamic adaptive learning rate based on the state error change rate, wherein the dynamic adaptive learning rate decays as the state error change rate increases through an exponential function, and a minimum learning rate threshold is set; A dynamic adjustment strategy for an adaptive weight matrix is ​​designed based on the relative amplitudes of the temperature error, the speed error, the temperature change rate error and the speed change rate error, wherein the diagonal elements of the adaptive weight matrix are dynamically adjusted according to the ratio of the corresponding state error amplitude to the maximum error amplitude; A coupling coefficient update equation is designed, wherein the coupling coefficient update equation includes a product term of the error covariance matrix and the adaptive weight matrix multiplied by the dynamic adaptive learning rate, and a coupling coefficient history update increment term, wherein the coupling coefficient history update increment term is used to accelerate convergence and suppress local oscillations; upper and lower limit constraints are imposed on the coupling coefficient calculated by the coupling coefficient update equation through a projection algorithm, and the constrained coupling coefficient is substituted into the coupling control law to coordinately control the heating unit and the stirring motor unit.

[0012] In an optional embodiment, The controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the frequency conversion control instructions based on the switched mode until the cooking is completed, the steps comprising: Preprocessing the collected real-time operation data to construct a real-time operation parameter vector, calculating the parameter error vector between the real-time operation parameter vector and the expected operation parameter vector, and calculating the parameter change rate vector; extracting a feature index set based on the parameter error vector and the parameter change rate vector, the feature index set including temperature rise rate, temperature fluctuation amplitude, pressure change rate, pressure fluctuation amplitude, power utilization rate, power fluctuation coefficient and speed fluctuation, constructing the feature index set into an operation state feature vector and performing dimensionality reduction; A multi-layer fuzzy neural network is used for working mode recognition, wherein the input layer of the multi-layer fuzzy neural network receives the running state feature vector after dimension reduction, the first hidden layer uses the Gaussian membership function to fuzzify the running state feature vector, the second hidden layer performs reasoning operation based on fuzzy rules, and the output layer gives the degree of attribution of each mode, the cluster center of the Gaussian membership function is obtained by offline training of the K-means algorithm, and the standard deviation of the Gaussian membership function is estimated by the sample variance; a fuzzy rule base including a heating dominant mode, a stirring dominant mode and a coupling control mode is constructed, and the rule weights in the fuzzy rule base are optimized online by a back propagation algorithm, and the optimization goal is to minimize the pattern recognition error; The current optimal working mode is determined according to the degree of belonging of each mode, and a mode soft switching strategy is designed based on a negative exponential decay function. The mode soft switching strategy adopts a time decay switching function to perform weighted fusion on the control quantities before and after the switching. The switching rate parameter of the time decay switching function is adaptively adjusted according to the norm of the parameter error vector; Based on the working mode after switching, a performance index function including a tracking error term, a control energy consumption term and a control stability term is constructed, the weight coefficient of the performance index function is associated with the mode attribution degree of the multi-layer fuzzy neural network output, and is dynamically adjusted according to the current working mode; based on the performance index function, a gradient update equation of the PID parameters is constructed, and a variable frequency control instruction is generated according to the updated PID parameters.

[0013] According to a second aspect of the embodiments of the present invention, Provide intelligent variable frequency cooking equipment multi-mode control system, including: The first unit is used to obtain food images through an image acquisition device and collect food weight data at the same time. The image acquisition device transmits the food images to a deep learning processor. The deep learning processor extracts feature information from the food images based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data, wherein the food feature data packet includes food type information, weight information and volume information; the food feature recognition model is obtained by transferring the teacher network knowledge to the student network, wherein the multi-scale feature map of the teacher network is mapped to the student network feature space through a feature alignment module, and the sample difficulty is determined based on the predicted probability distribution entropy value, feature point density value and annotation difficulty value of the sample, and a hierarchical training strategy is used to optimize the model; The second unit is used for matching the corresponding initial cooking mode parameters from a preset cooking mode database based on the food feature data packet, and collecting real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data, and the controller calculates a power-temperature dynamic response curve and a motor speed-torque response curve according to the real-time operation data; The third unit is used for the controller to generate a variable frequency control instruction according to the power-temperature dynamic response curve and the motor speed-torque response curve in combination with a preset multi-mode switching judgment rule, and coordinately adjust the heating unit and the stirring motor unit of the cooking equipment through the variable frequency control module; the controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the variable frequency control instruction based on the switched mode until cooking is completed.

[0014] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the embodiments of the present invention is: 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.

[0016] The present invention automatically identifies the type, weight and volume of ingredients through image recognition and weight data, and matches the corresponding initial cooking mode parameters, thereby realizing automated cooking without the need for users to manually set complex cooking parameters, thus simplifying the operating process.

[0017] The present invention utilizes the power-temperature dynamic response curve and the motor speed-torque response curve, combines multi-mode switching judgment rules and frequency conversion control technology, and coordinates the heating unit and the stirring motor to achieve precise temperature and speed control, thereby improving cooking efficiency and effects.

[0018] The present invention continuously collects real-time operating data during the cooking process, and performs multi-mode adaptive switching and frequency conversion control instruction optimization according to the data, making the cooking process more intelligent and capable of dynamic adjustment according to the state of ingredients and the cooking environment to ensure the best cooking effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of a multi-mode control method of an intelligent variable frequency cooking device according to an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a multi-mode control system of an intelligent variable frequency cooking device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] 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.

[0021] 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.

[0022] Figure 1 FIG. 1 is a flow chart of a multi-mode control method for an intelligent variable frequency cooking device according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Acquire food images through an image acquisition device, and collect food weight data at the same time, the image acquisition device transmits the food images to a deep learning processor, the deep learning processor extracts feature information from the food images based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data, wherein the food feature data packet includes food type information, weight information and volume information; the food feature recognition model is obtained by transferring the teacher network knowledge to the student network, wherein the multi-scale feature map of the teacher network is mapped to the student network feature space through a feature alignment module, and the sample difficulty is determined based on the predicted probability distribution entropy value, feature point density value and annotation difficulty value of the sample, and a hierarchical training strategy is used to optimize the model; S2. Based on the food feature data packet, the controller matches the corresponding initial cooking mode parameters from a preset cooking mode database, and collects real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data, and the controller calculates a power-temperature dynamic response curve and a motor speed-torque response curve according to the real-time operation data; S3. The controller generates a variable frequency control instruction based on the power-temperature dynamic response curve and the motor speed-torque response curve, combined with a preset multi-mode switching judgment rule, and coordinates the heating unit and the stirring motor unit of the cooking equipment through the variable frequency control module; the controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the variable frequency control instruction based on the switched mode until cooking is completed.

[0023] In an optional embodiment, The steps of acquiring food images and food weight data through an image acquisition device, wherein the image acquisition device transmits the food images to a deep learning processor, wherein the deep learning processor extracts feature information from the food images based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data include: The image of the food is acquired by an image acquisition device disposed on the inner side of the cover of the cooking device, wherein the image acquisition device comprises a CMOS sensor, a waterproof glass and a heat-insulating layer, wherein the CMOS sensor is disposed within the protection of the waterproof glass and the heat-insulating layer, and the image acquisition device uses an adaptive exposure algorithm to adjust the exposure time and gain of the acquired food image; The weight data of the food is obtained by a strain gauge weighing sensor arranged at the base of the cooking device, wherein the output end of the strain gauge weighing sensor is connected to a low-pass filter and a temperature compensation circuit, wherein the low-pass filter is used to eliminate vibration interference, and the temperature compensation circuit is used to eliminate the influence of ambient temperature on the weighing accuracy; The food image is input into a deep learning processor, and the deep learning processor processes the food image based on a pre-trained food feature recognition model, wherein the food feature recognition model adopts a ResNet50 network structure provided with a spatial pyramid pooling module, and the deep learning processor extracts features from the input food image through a feature extraction network of the food feature recognition model to obtain an image feature vector; the image feature vector is input into a feature dimension reduction module, and the feature dimension reduction module performs dimension reduction processing on the image feature vector using a principal component analysis method to obtain a reduced dimension feature vector; The reduced-dimensional feature vector and the food weight data are input into a feature fusion network, which includes a plurality of fully connected layers connected in series. The feature fusion network performs feature fusion on the reduced-dimensional feature vector and the normalized food weight data to generate a food feature data packet including food type information, weight information, volume information and appearance feature parameters, wherein the food volume information is calculated based on an image segmentation result and a pre-calibrated correspondence between pixels and actual sizes.

[0024] Exemplarily, first, an image of the food is acquired. An image acquisition device is installed on the inner side of the upper cover of the cooking device, and the device consists of a CMOS sensor, a waterproof glass, and a heat-insulating layer. The CMOS sensor is located under the protection of the waterproof glass and the heat-insulating layer, which can effectively prevent water vapor and high temperature from damaging the sensor. The image acquisition device adopts an adaptive exposure algorithm to automatically adjust the exposure time and gain according to the brightness and color of the food to ensure that the acquired image is clear and the exposure is appropriate. For example, when darker-colored food is placed, the exposure time will be appropriately extended and the gain will be increased; when lighter-colored food is placed, the exposure time will be shortened and the gain will be reduced. This ensures that high-quality food images can be obtained under different lighting conditions.

[0025] Next, obtain the weight data of the ingredients. A high-precision strain gauge weighing sensor is installed at the base of the cooking equipment to measure the weight of the ingredients. The output of the weighing sensor is connected to a low-pass filter and a temperature compensation circuit. The low-pass filter can effectively eliminate the influence of environmental vibration on the weighing result. For example, even if slight vibration occurs during the cooking process, it will not affect the accuracy of the weight data. The temperature compensation circuit can eliminate the influence of ambient temperature changes on the weighing accuracy, ensuring that accurate weight data can be obtained in different temperature environments. For example, when the ambient temperature rises from 20 degrees Celsius to 30 degrees Celsius, the temperature compensation circuit will correct the weighing result according to the preset compensation parameters to ensure measurement accuracy. Assume that the measured weight of the ingredients is 500 grams.

[0026] Then, extract the features of the food image. Input the acquired food image into the deep learning processor. The deep learning processor uses a pre-trained food feature recognition model to process the image. The model uses the ResNet50 network structure and adds a spatial pyramid pooling module, which can effectively extract the feature information of the food. Even if the position, size and angle of the food in the image are different, it can be accurately identified. The deep learning processor extracts features from the input food image through the feature extraction network of the model to obtain a high-dimensional image feature vector. For example, the vector may contain 2048 values, each of which represents a feature of the food image.

[0027] In order to reduce the computational complexity, the extracted high-dimensional image feature vector is input into the feature dimension reduction module. The feature dimension reduction module uses the principal component analysis method to reduce the dimension of the image feature vector and obtain a low-dimensional reduced dimension feature vector. For example, a 2048-dimensional vector can be reduced to 256 dimensions, retaining the main feature information while reducing the amount of calculation.

[0028] Finally, generate the food feature data packet. The reduced feature vector and food weight data are input into the feature fusion network. The feature fusion network consists of multiple fully connected layers in series, which fuses the reduced feature vector and the normalized food weight data to generate the final food feature data packet. The food feature data packet contains: food type information, such as tomatoes; weight information, such as 500 grams; volume information, such as calculated based on the image segmentation results and the pre-calibrated pixel-to-actual size correspondence, assuming 300 cubic centimeters; and appearance feature parameters, such as color, shape, etc.

[0029] The present invention adopts deep learning models and image recognition technology, which can accurately identify various ingredients and avoid the errors and inefficiencies caused by manual identification. It can not only identify the type and weight of ingredients, but also obtain information such as the volume and appearance characteristics of the ingredients, providing more comprehensive data support for subsequent cooking processes. By integrating functions such as image acquisition, weight measurement and deep learning processing, the cooking equipment is made more intelligent and the user experience is improved.

[0030] In an optional embodiment, The steps of training the food feature recognition model include: Collecting food images and cooking process images and annotating them to obtain a training data set, wherein the annotated content includes food type information, weight information, volume information and feature point location information; using the training data set to train a feature extraction network; processing the output features of the feature extraction network through a spatial feature processing branch and a temporal feature processing branch and weighted fusion to obtain a fusion feature; The fusion feature is input into a multi-task learning network, the multi-task learning network includes a food classification module, a feature point positioning module and a weight and volume prediction module, the food classification module is trained by a cross entropy loss function, the feature point positioning module is trained by a Euclidean distance loss function, and the weight and volume prediction module is trained by a mean square error loss function to obtain a teacher network; knowledge transfer is performed on the teacher network to obtain a student network, including: mapping the multi-scale feature map of the teacher network to the feature space of the student network through a feature alignment module, constructing a spatial attention matrix and a channel attention matrix of the multi-scale feature map of the teacher network and transferring them to the student network, and transferring the prediction result of the teacher network to the student network through a soft label method, wherein the student network adopts a depth-separable convolutional layer structure, and a channel rearrangement module is set between adjacent network layers; The sample difficulty is determined according to the prediction result of the student network. The sample difficulty is obtained by weighting the entropy value of the predicted probability distribution of the sample by the student network, the density value of the feature points in the sample and the annotation difficulty value of the training data set. The training samples are stratified based on the sample difficulty. During the training process, the sampling ratio of samples of different difficulties is adjusted according to the training rounds; data enhancement processing and clarity detection, feature point detection and signal-to-noise ratio detection are performed on the sampled samples; the enhanced samples that pass the detection are input into the student network for iterative training until the loss value of the student network converges.

[0031] Exemplarily, first, a training data set is constructed. Images of various ingredients and images of the cooking process are collected, such as tomatoes, potatoes, carrots, beef, chicken, etc., as well as images of them in different processing states such as dicing, shredding, and dicing. At the same time, images of these ingredients in different cooking stages such as stir-frying and stewing in a pot are also collected. The collected images are annotated, and the annotation content includes: ingredient type, such as "tomato"; weight, such as "200 grams"; volume, such as "150 cubic centimeters"; and feature point location information, such as the location of the tomato stem and the key points of the texture direction of the beef. At least 1,000 images of each ingredient in different states are collected to ensure the diversity of the data set.

[0032] Next, use the constructed training data set to train the feature extraction network. You can choose classic convolutional neural networks such as ResNet and VGG as the skeleton network of the feature extraction network, and adjust the network structure according to the actual situation. Input the image into the feature extraction network, and the network will automatically learn to extract the features in the image. For example, for a tomato image, the network may extract features such as color, shape, and texture.

[0033] Then, the extracted features are processed. The features output by the feature extraction network are input into the spatial feature processing branch and the temporal feature processing branch respectively. The spatial feature processing branch focuses on the spatial information in the image, such as the shape and texture of the food. The temporal feature processing branch focuses on the temporal information in the image sequence, such as the changes of the food during the cooking process. The spatial feature processing branch can use the spatial attention mechanism to capture the relationship between different regions in the image. The temporal feature processing branch can use a recurrent neural network, such as LSTM, to capture the temporal dependency in the image sequence. The outputs of the spatial feature processing branch and the temporal feature processing branch are weighted fused to obtain fused features. The fusion weight can be obtained by learning, for example, using a simple fully connected layer to predict the weight.

[0034] The fused features are input into the multi-task learning network. The network contains three modules: ingredient classification module, feature point location module and weight volume prediction module. The ingredient classification module is trained using the cross entropy loss function. For example, if a tomato is identified as an apple, a large loss will be incurred. The feature point location module is trained using the Euclidean distance loss function. For example, the greater the difference between the predicted feature point position and the actual position, the greater the loss. The weight volume prediction module is trained using the mean square error loss function. For example, the greater the difference between the predicted weight and the actual weight, the greater the loss. Through the joint training of these three modules, the teacher network is obtained.

[0035] The knowledge of the teacher network is transferred to obtain the student network. The student network adopts a lighter network structure, such as using a depth-separable convolutional layer. The multi-scale feature map of the teacher network is mapped to the feature space of the student network through the feature alignment module, and the spatial attention matrix and channel attention matrix of the teacher network are transferred to the student network. At the same time, the prediction results of the teacher network are transferred to the student network in the form of soft labels. For example, if the teacher network predicts that the probability of a certain ingredient being a tomato is 80% and the probability of an apple is 20%, this probability distribution is used as a soft label to guide the learning of the student network. A channel rearrangement module is set between adjacent network layers of the student network to improve the efficiency of information flow.

[0036] The sample difficulty is determined based on the prediction results of the student network. The sample difficulty is obtained by weighting the entropy value of the predicted probability distribution of the student network for the sample, the density value of the feature points in the sample, and the annotation difficulty value of the training data set. For example, if the prediction result of the student network for a certain sample is very uncertain, the sample difficulty is considered to be high. The training samples are stratified based on the sample difficulty, for example, the samples are divided into three levels: simple, medium, and difficult. During the training process, the sampling ratio of samples of different difficulty levels is adjusted according to the training rounds. For example, in the early stage of training, more simple samples are sampled; as the training progresses, the sampling ratio of medium and difficult samples is gradually increased.

[0037] The sampled samples are subjected to data enhancement processing, such as random cropping, rotation, scaling, etc., as well as clarity detection, feature point detection, and signal-to-noise ratio detection. The enhanced samples that have passed the detection are input into the student network for iterative training until the loss value of the student network converges. For example, when the loss value no longer decreases or decreases very little, stop training.

[0038] Through multi-task learning and knowledge transfer, the model of the present invention can more effectively learn the various characteristics of ingredients, thereby improving the recognition accuracy of ingredient types, weights, volumes and feature point positions; the student network adopts a lightweight network structure, which can reduce the computational complexity and storage space of the model, making it easier to deploy on resource-constrained platforms such as mobile devices; the hierarchical training strategy based on sample difficulty can make the model more focused on learning difficult samples, thereby improving training efficiency.

[0039] In an optional embodiment, Based on the food feature data packet, the controller matches the corresponding initial cooking mode parameters from a preset cooking mode database, and collects real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data. The controller calculates the power-temperature dynamic response curve and the motor speed-torque response curve according to the real-time operation data, and the steps include: Constructing a cooking mode database, the cooking mode database comprising food attribute information and corresponding cooking parameter combinations, the food attribute information comprising weight characteristic value, volume characteristic value, water content characteristic value, shape regularity, surface smoothness and hardness characteristic value; constructing a to-be-cooked food feature vector according to the food feature data packet, calculating the feature importance of the food attribute information in the to-be-cooked food feature vector according to different cooking modes, calculating a feature weight coefficient based on the feature importance, and the feature weight coefficient is obtained by mapping and normalizing the feature importance through an exponential function; The feature weight coefficient is used to weight the feature vector of the to-be-cooked food, and the weighted threshold cosine similarity algorithm is used to calculate the similarity between the weighted feature vector of the to-be-cooked food and the food attribute information in the cooking mode database, wherein the weighted threshold cosine similarity algorithm sets an adaptive threshold constraint for each feature value difference, and the adaptive threshold is obtained through data statistics in the cooking mode database; the top three matching ingredients with the highest similarity are selected, and the cooking parameter combinations corresponding to the matching ingredients are weightedly fused based on the feature weight coefficient to obtain initial cooking parameters, wherein the initial cooking parameters include a heating power parameter and a stirring speed parameter; The temperature data, pressure data, heating power data and stirring motor speed data in the cavity of the cooking equipment are collected, and the initial cooking parameters are adaptively corrected according to the deviation between the reference cooking effect in the cooking mode database and the actual cooking effect; a second-order transfer function model is established based on the temperature data and the heating power data to obtain a power-temperature dynamic response curve, and a torque model considering the moment of inertia and viscous damping is established based on the stirring motor speed data and the motor current data to obtain a motor speed-torque response curve.

[0040] Exemplarily, first, characteristic data of the ingredients to be cooked are obtained, including weight, volume, water content, shape regularity, surface smoothness, and hardness, etc. For example, there is a piece of beef brisket to be cooked, which weighs 1 kg, has a volume of 0.001 cubic meters, and a water content of 70%; the shape regularity is 0.8, and the closer the value is to 1, the more regular the shape; the surface smoothness is 0.6, and the closer the value is to 1, the smoother the surface; the hardness is 0.7, and the closer the value is to 1, the higher the hardness.

[0041] Then, a cooking mode database is constructed. The database contains the attribute information of various ingredients and the corresponding cooking parameter combinations. For example, the database contains information on various ingredients such as beef, chicken, and fish, as well as corresponding cooking parameters such as heating power and stirring speed. For beef, the database records the optimal cooking parameter combinations under different weights, volumes, water contents, shape regularity, surface smoothness, and hardness.

[0042] Next, a feature vector is constructed based on the feature data of the ingredients to be cooked. For example, the feature vector of the beef brisket can be expressed as [1, 0.001, 70, 0.8, 0.6, 0.7].

[0043] According to different cooking modes, the importance of each feature in the feature vector of the ingredients to be cooked is calculated. For example, for stewed beef, weight, water content and hardness are more important, while volume, shape regularity and surface smoothness are relatively less important.

[0044] The feature weight coefficient is calculated based on the feature importance. For example, an exponential function can be used to map the feature importance, and then the mapped value is normalized to obtain the final feature weight coefficient. Assume that in the beef stew mode, the weight coefficients of weight, volume, water content, shape regularity, surface smoothness and hardness are 0.3, 0.1, 0.25, 0.05, 0.05, and 0.25 respectively.

[0045] The feature weight coefficient is used to weight the feature vector of the ingredient to be cooked. For example, the feature vector of beef brisket [1, 0.001, 70, 0.8, 0.6, 0.7] is multiplied by the weight coefficient [0.3, 0.1, 0.25, 0.05, 0.05, 0.25] to obtain the weighted feature vector.

[0046] The weighted threshold cosine similarity algorithm is used to calculate the similarity between the weighted feature vector of the ingredients to be cooked and the ingredient attribute information in the cooking mode database. The algorithm sets an adaptive threshold constraint on each feature value difference, and the adaptive threshold is obtained through data statistics in the cooking mode database. For example, by counting the weight data of beef in the database, the reasonable range of the weight difference can be determined and set as the adaptive threshold.

[0047] Select the top three matching ingredients with the highest similarity. For example, if the three ingredients with the highest similarity to beef brisket are beef brisket A, beef brisket B, and beef shank, then select the cooking parameter combinations corresponding to these three ingredients.

[0048] The cooking parameter combinations corresponding to the matching ingredients are weighted and fused based on the feature weight coefficients to obtain the initial cooking parameters. For example, the heating power and stirring speed parameters of beef brisket A, beef brisket B, and beef tendon are multiplied by the corresponding weight coefficients, and then the results are added to obtain the initial heating power and stirring speed parameters.

[0049] Collect temperature data, pressure data, heating power data, and stirring motor speed data in the cooking equipment cavity. For example, collect these data once every certain period of time and record them.

[0050] The initial cooking parameters are adaptively corrected according to the deviation between the reference cooking effect in the cooking mode database and the actual cooking effect. For example, if the actual temperature rise speed is slower than the reference speed, the heating power is appropriately increased.

[0051] Based on the temperature data and heating power data, a second-order transfer function model is established to obtain the power-temperature dynamic response curve. By analyzing the relationship between temperature and heating power, a curve describing the dynamic response of the two can be obtained.

[0052] Based on the stirring motor speed data and motor current data, a torque model considering rotational inertia and viscous damping is established to obtain the motor speed-torque response curve. By analyzing the relationship between motor speed and current, a curve describing the dynamic response of the two can be obtained, taking into account the influence of rotational inertia and viscous damping.

[0053] The present invention can shorten cooking time and improve cooking efficiency by intelligently matching and dynamically adjusting cooking parameters; can make food achieve the best taste and nutritional value by accurately controlling the cooking process; and can automate the cooking process, reduce manual intervention, and simplify the operation process, so that even users with no cooking experience can easily cook delicious dishes.

[0054] In an optional embodiment, The controller generates a variable frequency control instruction based on the power-temperature dynamic response curve and the motor speed-torque response curve in combination with a preset multi-mode switching judgment rule, and the step of cooperatively adjusting the heating unit and the stirring motor unit of the cooking device through the variable frequency control module includes: Construct a system state vector including temperature, rotation speed, temperature change rate and rotation speed change rate, design a Kalman state observer based on a state transfer matrix and an observation matrix to estimate the system state vector, and optimize the gain matrix of the Kalman state observer through a system noise covariance matrix; design a state coupling matrix based on the estimated system state vector, and obtain the coupling coefficient of the state coupling matrix through particle swarm optimization; Based on the estimated system state vector, the working mode is identified in combination with a preset multi-mode switching determination rule, wherein the multi-mode switching determination rule includes a temperature steady-state error threshold, a speed fluctuation range threshold, a temperature change rate threshold, and a speed change rate threshold, and the control mode to be currently adopted is determined according to the comparison result between each component of the system state vector and the corresponding threshold; A coupling control law is designed according to the state coupling matrix and the currently adopted control mode, wherein the coupling control law associates the heating power adjustment amount with the speed adjustment amount through the coupling coefficient, wherein the heating power adjustment amount is composed of a proportional integral term of the temperature deviation and a coupling term of the estimated system state vector, and the speed adjustment amount is composed of a proportional differential term of the speed deviation and a coupling term of the estimated system state vector, and the coupling coefficient is updated in real time by designing an adaptive mechanism based on a weighted product of an estimated system state vector error and an actual system state vector error; The output of the coupling control law is converted into a variable frequency control instruction, wherein the heating power variable frequency control instruction is composed of a proportional integral term of a reference frequency and a heating power adjustment amount, and the stirring motor variable frequency control instruction is composed of a proportional integral term of a reference frequency and a speed adjustment amount; the variable frequency control instruction is converted into a control signal through pulse width modulation to coordinately control the heating unit and the stirring motor unit of the cooking equipment.

[0055] For example, first, a hardware platform of a cooking device is built, including a heating unit, a stirring motor unit, a temperature sensor, a speed sensor, and a control unit. The temperature sensor is a PT100 thermal resistor, and the speed sensor is a Hall effect sensor.

[0056] Next, the system state vector is constructed. The data of the temperature sensor and the speed sensor are collected, and the temperature change rate and the speed change rate are calculated. For example, the temperature and speed data are collected every 0.1 seconds, and then the difference between the two sampling values ​​is divided by the sampling time interval to obtain the temperature change rate and the speed change rate. The current temperature, speed, temperature change rate and speed change rate form a system state vector.

[0057] Then, design the Kalman state observer. Determine the state transfer matrix and the observation matrix according to the system characteristics. For example, the two matrices can be determined according to the physical models of the heating unit and the stirring motor and the sensor characteristics. Set the initial value of the system noise covariance matrix and adjust it through experimental data to optimize the gain matrix of the Kalman state observer so that the estimated system state vector is closer to the true value. For example, the system noise covariance matrix can be adjusted by comparing the error between the estimated temperature and the measured temperature to minimize the error.

[0058] After that, the state coupling matrix is ​​designed. The state coupling matrix is ​​used to describe the mutual influence between temperature and speed. Its coupling coefficient can be optimized by particle swarm algorithm. For example, the population size of particle swarm algorithm is set to 50 and the number of iterations is set to 100. The optimal coupling coefficient is found by minimizing the error between the estimated system state vector and the actual system state vector.

[0059] According to the estimated system state vector, the working mode is identified in combination with the preset multi-mode switching judgment rules. For example, the temperature steady-state error threshold is set to ±1°C, the speed fluctuation range threshold is ±50 rpm, the temperature change rate threshold is ±0.5°C / second, and the speed change rate threshold is ±10 rpm / second. If the deviation between the estimated temperature and the set temperature is within ±1°C, and the estimated speed fluctuation is within ±50 rpm, and the estimated temperature change rate is within ±0.5°C / second, and the estimated speed change rate is within ±10 rpm / second, then the system is judged to be in steady-state mode. The judgment rules for other modes can be set similarly.

[0060] The coupling control law is designed according to the state coupling matrix and the currently adopted control mode. For example, in the steady-state mode, the heating power regulation is mainly determined by the proportional integral term of the temperature deviation, and the speed regulation is mainly determined by the proportional differential term of the speed deviation, and the two are related by the coupling coefficient. The coupling coefficient is updated in real time by designing an adaptive mechanism based on the weighted product of the estimated system state vector error and the actual system state vector error to ensure the control effect. For example, the weighted coefficients can be set to 0.8 and 0.2, corresponding to the estimated error and the actual error, respectively.

[0061] The output of the coupling control law is converted into a variable frequency control instruction. For example, the base frequency is set to 50Hz, the proportional integral coefficients of the heating power adjustment amount are 0.5 and 0.1, and the proportional integral coefficients of the speed adjustment amount are 0.2 and 0.05. Then, the variable frequency control instruction is converted into a control signal through pulse width modulation to drive the heating unit and the stirring motor unit.

[0062] The present invention estimates the system state through the Kalman state observer, and combines the multi-mode switching judgment rule and the coupling control law to more accurately control the cooking temperature and stirring speed, thereby improving the consistency and stability of the cooking effect; by updating the coupling coefficient in real time through the adaptive mechanism, the control system can be more adaptable to different ingredients and cooking environments, thereby enhancing the robustness of the system; by collaboratively controlling the heating and stirring processes, energy consumption can be reduced and energy utilization efficiency can be improved.

[0063] In an optional embodiment, The coupling control law associates the heating power adjustment amount with the speed adjustment amount through the coupling coefficient, wherein the heating power adjustment amount is composed of a proportional integral term of the temperature deviation and a coupling term of the estimated system state vector, and the speed adjustment amount is composed of a proportional differential term of the speed deviation and a coupling term of the estimated system state vector, and the coupling coefficient is updated in real time by designing an adaptive mechanism based on a weighted product of an estimated system state vector error and an actual system state vector error, including: Obtain an instantaneous state error vector including a temperature error, a rotation speed error, a temperature change rate error, and a rotation speed change rate error, obtain a weighted state error vector based on an exponentially weighted sliding average of the instantaneous state error vector and a historical state error vector, wherein the exponentially weighted sliding average performs attenuation processing on historical error information through a forgetting factor; based on the instantaneous state error vector, calculate a difference between state error vectors at two adjacent sampling moments to obtain a state error change amount, and divide the state error change amount by a sampling period to obtain a state error change rate; Constructing an error covariance matrix considering the cross-coupling terms of temperature and speed, the error covariance matrix including the outer product term of the weighted state error vector, and a coupling strength term consisting of the product of the temperature error and the speed change rate error, and the product of the speed error and the temperature change rate error, calculated based on the state error change rate; Designing a dynamic adaptive learning rate based on the state error change rate, wherein the dynamic adaptive learning rate decays as the state error change rate increases through an exponential function, and a minimum learning rate threshold is set; A dynamic adjustment strategy for an adaptive weight matrix is ​​designed based on the relative amplitudes of the temperature error, the speed error, the temperature change rate error and the speed change rate error, wherein the diagonal elements of the adaptive weight matrix are dynamically adjusted according to the ratio of the corresponding state error amplitude to the maximum error amplitude; A coupling coefficient update equation is designed, wherein the coupling coefficient update equation includes a product term of the error covariance matrix and the adaptive weight matrix multiplied by the dynamic adaptive learning rate, and a coupling coefficient history update increment term, wherein the coupling coefficient history update increment term is used to accelerate convergence and suppress local oscillations; upper and lower limit constraints are imposed on the coupling coefficient calculated by the coupling coefficient update equation through a projection algorithm, and the constrained coupling coefficient is substituted into the coupling control law to coordinately control the heating unit and the stirring motor unit.

[0064] Exemplarily, first, the current state of the system is obtained, including the actual temperature and the actual speed, and at the same time, the set values ​​of the system, namely the target temperature and the target speed, are obtained.

[0065] Next, calculate the temperature error and speed error. The temperature error is the difference between the target temperature and the actual temperature, and the speed error is the difference between the target speed and the actual speed. For example, if the target temperature is 100°C and the actual temperature is 90°C, then the temperature error is 10°C. If the target speed is 1000 rpm and the actual speed is 900 rpm, then the speed error is 100 rpm. In addition to the temperature and speed errors, the temperature change rate error and speed change rate error must also be calculated. The temperature change rate error is calculated by subtracting the temperature error at the previous sampling time from the temperature error at the current sampling time, and then dividing by the sampling period. The speed change rate error is calculated similarly.

[0066] Then, the historical state error information is weighted averaged. The temperature error, speed error, temperature change rate error, and speed change rate error at the current sampling moment are weighted averaged with the historical state error to obtain the weighted state error. For example, the exponential weighted sliding average method can be used, and the forgetting factor is set to 0.9. This means that the weight of the current error is 0.1, while the weight of the historical error gradually decays over time.

[0067] After that, the error covariance matrix is ​​constructed. This matrix contains the product of the weighted temperature error and the weighted speed error, the product of the weighted temperature change rate error and the weighted speed error, the product of the weighted temperature error and the weighted speed change rate error, and the product of the weighted temperature change rate error and the weighted speed change rate error. These products reflect the coupling relationship between temperature and speed. For example, if the product of the temperature error and the speed change rate error is large, it means that there is a strong coupling relationship between temperature and speed.

[0068] Next, the dynamic adaptive learning rate is calculated. The learning rate is used to adjust the update speed of the coupling coefficient. The dynamic adaptive learning rate will be adjusted according to the rate of change of the state error. For example, an exponential function can be used to achieve dynamic adjustment. When the rate of change of the state error is large, the learning rate will decrease to prevent system oscillation. At the same time, a minimum learning rate threshold is set to ensure that the learning rate is not too small and affects the system convergence speed.

[0069] At the same time, the adaptive weight matrix is ​​dynamically adjusted according to the relative amplitudes of the temperature error, speed error, temperature change rate error, and speed change rate error. For example, if the amplitude of the temperature error is large, the weight corresponding to the temperature error is increased to enhance the effect of temperature control. Assuming that the temperature error is 10 and the maximum error amplitude is 20, the weight corresponding to the temperature error can be calculated as 10 / 20=0.5.

[0070] Then, the coupling coefficient is updated. The update equation of the coupling coefficient contains the product of the error covariance matrix, the adaptive weight matrix and the dynamic adaptive learning rate, as well as the historical update increment of the coupling coefficient. The historical update increment is used to accelerate convergence and suppress local oscillations. Assuming that the error covariance matrix is ​​[[1, 0.5], [0.5, 1]], the adaptive weight matrix is ​​[[0.5, 0], [0, 0.5]], the dynamic adaptive learning rate is 0.0037, and the historical update increment of the coupling coefficient is 0.01, then the update amount of the coupling coefficient can be calculated as [[0.00185, 0.000925], [0.000925, 0.00185]]+0.01=[[0.01185, 0.010925], [0.010925, 0.01185]].

[0071] Finally, the updated coupling coefficient is substituted into the coupling control law to calculate the heating power adjustment and speed adjustment, and the heating unit and the stirring motor unit are controlled in a coordinated manner. The updated coupling coefficient is constrained with upper and lower limits to ensure the stability of the system.

[0072] The present invention can better compensate for the coupling effect inside the system by updating the coupling coefficient in real time, thereby improving the control accuracy of temperature and rotation speed; the dynamic adaptive learning rate and the adaptive weight matrix can be adjusted according to the state change of the system, thereby enhancing the system's adaptability to external interference and parameter changes; the introduction of the historical update increment of the coupling coefficient can accelerate the system convergence, thereby improving the response speed of the system.

[0073] In an optional embodiment, The controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the frequency conversion control instructions based on the switched mode until the cooking is completed, the steps comprising: Preprocessing the collected real-time operation data to construct a real-time operation parameter vector, calculating the parameter error vector between the real-time operation parameter vector and the expected operation parameter vector, and calculating the parameter change rate vector; extracting a feature index set based on the parameter error vector and the parameter change rate vector, the feature index set including temperature rise rate, temperature fluctuation amplitude, pressure change rate, pressure fluctuation amplitude, power utilization rate, power fluctuation coefficient and speed fluctuation, constructing the feature index set into an operation state feature vector and performing dimensionality reduction; A multi-layer fuzzy neural network is used for working mode recognition, wherein the input layer of the multi-layer fuzzy neural network receives the running state feature vector after dimension reduction, the first hidden layer uses the Gaussian membership function to fuzzify the running state feature vector, the second hidden layer performs reasoning operation based on fuzzy rules, and the output layer gives the degree of attribution of each mode, the cluster center of the Gaussian membership function is obtained by offline training of the K-means algorithm, and the standard deviation of the Gaussian membership function is estimated by the sample variance; a fuzzy rule base including a heating dominant mode, a stirring dominant mode and a coupling control mode is constructed, and the rule weights in the fuzzy rule base are optimized online by a back propagation algorithm, and the optimization goal is to minimize the pattern recognition error; The current optimal working mode is determined according to the degree of belonging of each mode, and a mode soft switching strategy is designed based on a negative exponential decay function. The mode soft switching strategy adopts a time decay switching function to perform weighted fusion on the control quantities before and after the switching. The switching rate parameter of the time decay switching function is adaptively adjusted according to the norm of the parameter error vector; Based on the working mode after switching, a performance index function including a tracking error term, a control energy consumption term and a control stability term is constructed, the weight coefficient of the performance index function is associated with the mode attribution degree of the multi-layer fuzzy neural network output, and is dynamically adjusted according to the current working mode; based on the performance index function, a gradient update equation of the PID parameters is constructed, and a variable frequency control instruction is generated according to the updated PID parameters.

[0074] Exemplarily, first, real-time operating data during the cooking process is collected, such as temperature, pressure, power, speed, etc. The collected data is preprocessed, such as removing noise, filtering, etc., and converted into a format that can be used for subsequent calculations. Then, the preprocessed data is constructed into a real-time operating parameter vector. At the same time, according to the preset cooking goal, the expected operating parameter vector is determined. By comparing the real-time operating parameter vector with the expected operating parameter vector, the parameter error vector is calculated, and the parameter change rate vector is calculated. For example, if the real-time temperature is 80°C and the target temperature is 100°C, the temperature error is 20°C. If the temperature at the previous moment is 75°C, the temperature change rate is 5°C / second.

[0075] Next, a characteristic index set is extracted based on the parameter error vector and the parameter change rate vector. The characteristic index set includes temperature rise rate, temperature fluctuation amplitude, pressure change rate, pressure fluctuation amplitude, power utilization, power fluctuation coefficient and speed fluctuation. These characteristic index sets are constructed as operating state characteristic vectors. In order to reduce the computational complexity, the operating state characteristic vector is subjected to dimensionality reduction processing, for example, by using principal component analysis. Assuming that the temperature rise rate is 2°C / second, the temperature fluctuation amplitude is 1°C, the pressure change rate is 0.1Pa / second, the pressure fluctuation amplitude is 0.05Pa, the power utilization is 80%, the power fluctuation coefficient is 5%, and the speed fluctuation is 2%, a seven-dimensional operating state characteristic vector can be constructed and its dimension can be reduced to, for example, three dimensions.

[0076] Then, a multi-layer fuzzy neural network is used for working mode recognition. The reduced-dimensional operating state feature vector is input into the input layer of the multi-layer fuzzy neural network. The first hidden layer uses the Gaussian membership function to fuzzify the operating state feature vector. The cluster center of the Gaussian membership function is obtained by offline training of the K-means algorithm, and the standard deviation is estimated by the sample variance. For example, assuming that the cluster center of a certain feature is 10 and the standard deviation is 2, the membership when the feature value is 12 can be calculated by the Gaussian function. The second hidden layer performs reasoning operations based on fuzzy rules. The fuzzy rule base contains multiple modes such as heating dominant mode, stirring dominant mode and coupling control mode. The rule weights in the fuzzy rule base are optimized online by the back propagation algorithm, and the optimization goal is to minimize the pattern recognition error. For example, a fuzzy rule can be: if the temperature rise rate is high and the temperature fluctuation amplitude is small, the heating dominant mode is selected. The output layer gives the degree of belonging to each mode, for example, the degree of belonging to the heating dominant mode is 0.8, the degree of belonging to the stirring dominant mode is 0.1, and the degree of belonging to the coupling control mode is 0.1.

[0077] The current optimal working mode is determined according to the degree of belonging of each mode. A mode soft switching strategy is designed based on a negative exponential decay function. The mode soft switching strategy uses a time decay switching function to perform weighted fusion of the control quantities before and after the switch. The switching rate parameter of the time decay switching function is adaptively adjusted according to the norm of the parameter error vector. For example, if the norm of the parameter error vector is large, the switching rate parameter is large, making the mode switch faster.

[0078] Finally, based on the switched working mode, a performance indicator function is constructed. The performance indicator function contains tracking error terms, control energy consumption terms, and control stability terms. The weight coefficient of the performance indicator function is associated with the mode attribution degree of the multi-layer fuzzy neural network output and is dynamically adjusted according to the current working mode. For example, in the heating-dominated mode, the weight of the tracking error term is greater. The gradient update equation of the PID (proportional-integral-differential) parameters is constructed based on the performance indicator function, and the variable frequency control command is generated according to the updated PID parameters to control the operation of the cooking equipment.

[0079] The present invention can control the cooking process more accurately and make the cooking result more in line with expectations through multi-mode adaptive switching and PID parameter optimization based on real-time operation data; by optimizing the control strategy, the cooking time and energy consumption can be reduced, the cooking efficiency can be improved, and adaptive adjustments can be made according to real-time operation data, thereby enhancing the adaptability of the cooking system to different ingredients and environments and improving the robustness of the system.

[0080] Figure 2 FIG. 1 is a schematic diagram of a multi-mode control system for an intelligent variable frequency cooking device according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to obtain food images through an image acquisition device and collect food weight data at the same time. The image acquisition device transmits the food images to a deep learning processor. The deep learning processor extracts feature information from the food images based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data, wherein the food feature data packet includes food type information, weight information and volume information; the food feature recognition model is obtained by transferring the teacher network knowledge to the student network, wherein the multi-scale feature map of the teacher network is mapped to the student network feature space through a feature alignment module, and the sample difficulty is determined based on the predicted probability distribution entropy value, feature point density value and annotation difficulty value of the sample, and a hierarchical training strategy is used to optimize the model; The second unit is used for matching the corresponding initial cooking mode parameters from a preset cooking mode database based on the food feature data packet, and collecting real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data, and the controller calculates a power-temperature dynamic response curve and a motor speed-torque response curve according to the real-time operation data; The third unit is used for the controller to generate a variable frequency control instruction according to the power-temperature dynamic response curve and the motor speed-torque response curve in combination with a preset multi-mode switching judgment rule, and coordinately adjust the heating unit and the stirring motor unit of the cooking equipment through the variable frequency control module; the controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the variable frequency control instruction based on the switched mode until cooking is completed.

[0081] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0082] A fourth aspect of the embodiments of the present invention is: 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.

[0083] 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.

[0084] 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. Intelligent variable frequency cooking equipment multi-mode control method, characterized in that: include: The image of the food is acquired through an image acquisition device, and the weight data of the food is collected at the same time. The image acquisition device transmits the food image to a deep learning processor. The deep learning processor extracts feature information from the food image based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data, wherein the food feature data packet includes food type information, weight information and volume information; the food feature recognition model is obtained by transferring the knowledge of the teacher network to the student network, wherein the multi-scale feature map of the teacher network is mapped to the feature space of the student network through a feature alignment module, and the sample difficulty is determined based on the predicted probability distribution entropy value, feature point density value and annotation difficulty value of the sample, and a hierarchical training strategy is used to optimize the model; Based on the food feature data packet, the controller matches the corresponding initial cooking mode parameters from a preset cooking mode database, and collects real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data, and the controller calculates a power-temperature dynamic response curve and a motor speed-torque response curve according to the real-time operation data; The controller generates a variable frequency control instruction based on the power-temperature dynamic response curve and the motor speed-torque response curve in combination with a preset multi-mode switching judgment rule, and coordinately adjusts the heating unit and the stirring motor unit of the cooking device through the variable frequency control module; The controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the frequency conversion control instructions based on the switched mode until cooking is completed.

2. The method according to claim 1, characterized in that The steps of acquiring food images and food weight data through an image acquisition device, wherein the image acquisition device transmits the food images to a deep learning processor, wherein the deep learning processor extracts feature information from the food images based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data include: The image of the food is acquired by an image acquisition device disposed on the inner side of the cover of the cooking device, wherein the image acquisition device comprises a CMOS sensor, a waterproof glass and a heat-insulating layer, wherein the CMOS sensor is disposed within the protection of the waterproof glass and the heat-insulating layer, and the image acquisition device uses an adaptive exposure algorithm to adjust the exposure time and gain of the acquired food image; The weight data of the food is obtained by a strain gauge weighing sensor arranged at the base of the cooking device, wherein the output end of the strain gauge weighing sensor is connected to a low-pass filter and a temperature compensation circuit, wherein the low-pass filter is used to eliminate vibration interference, and the temperature compensation circuit is used to eliminate the influence of ambient temperature on the weighing accuracy; The food image is input into a deep learning processor, and the deep learning processor processes the food image based on a pre-trained food feature recognition model, wherein the food feature recognition model adopts a ResNet50 network structure provided with a spatial pyramid pooling module, and the deep learning processor extracts features from the input food image through a feature extraction network of the food feature recognition model to obtain an image feature vector; the image feature vector is input into a feature dimension reduction module, and the feature dimension reduction module performs dimension reduction processing on the image feature vector using a principal component analysis method to obtain a reduced dimension feature vector; The reduced-dimensional feature vector and the food weight data are input into a feature fusion network, which includes a plurality of fully connected layers connected in series. The feature fusion network performs feature fusion on the reduced-dimensional feature vector and the normalized food weight data to generate a food feature data packet including food type information, weight information, volume information and appearance feature parameters, wherein the food volume information is calculated based on an image segmentation result and a pre-calibrated correspondence between pixels and actual sizes.

3. The method according to claim 1, characterized in that The steps of training the food feature recognition model include: Collecting food images and cooking process images and annotating them to obtain a training data set, wherein the annotated content includes food type information, weight information, volume information and feature point location information; using the training data set to train a feature extraction network; processing the output features of the feature extraction network through a spatial feature processing branch and a temporal feature processing branch and weighted fusion to obtain a fusion feature; The fusion feature is input into a multi-task learning network, the multi-task learning network includes a food classification module, a feature point positioning module and a weight and volume prediction module, the food classification module is trained by a cross entropy loss function, the feature point positioning module is trained by a Euclidean distance loss function, and the weight and volume prediction module is trained by a mean square error loss function to obtain a teacher network; knowledge transfer is performed on the teacher network to obtain a student network, including: mapping the multi-scale feature map of the teacher network to the feature space of the student network through a feature alignment module, constructing a spatial attention matrix and a channel attention matrix of the multi-scale feature map of the teacher network and transferring them to the student network, and transferring the prediction result of the teacher network to the student network through a soft label method, wherein the student network adopts a depth-separable convolutional layer structure, and a channel rearrangement module is set between adjacent network layers; The sample difficulty is determined according to the prediction result of the student network. The sample difficulty is obtained by weighting the entropy value of the predicted probability distribution of the sample by the student network, the density value of the feature points in the sample and the annotation difficulty value of the training data set. The training samples are stratified based on the sample difficulty. During the training process, the sampling ratio of samples of different difficulties is adjusted according to the training rounds; data enhancement processing and clarity detection, feature point detection and signal-to-noise ratio detection are performed on the sampled samples; the enhanced samples that pass the detection are input into the student network for iterative training until the loss value of the student network converges.

4. The method according to claim 1, characterized in that: Based on the food feature data packet, the controller matches the corresponding initial cooking mode parameters from a preset cooking mode database, and collects real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data. The controller calculates the power-temperature dynamic response curve and the motor speed-torque response curve according to the real-time operation data, and the steps include: Constructing a cooking mode database, the cooking mode database comprising food attribute information and corresponding cooking parameter combinations, the food attribute information comprising weight characteristic value, volume characteristic value, water content characteristic value, shape regularity, surface smoothness and hardness characteristic value; constructing a to-be-cooked food feature vector according to the food feature data packet, calculating the feature importance of the food attribute information in the to-be-cooked food feature vector according to different cooking modes, calculating a feature weight coefficient based on the feature importance, and the feature weight coefficient is obtained by mapping and normalizing the feature importance through an exponential function; The feature weight coefficient is used to weight the feature vector of the to-be-cooked food, and the weighted threshold cosine similarity algorithm is used to calculate the similarity between the weighted feature vector of the to-be-cooked food and the food attribute information in the cooking mode database, wherein the weighted threshold cosine similarity algorithm sets an adaptive threshold constraint for each feature value difference, and the adaptive threshold is obtained through data statistics in the cooking mode database; the top three matching ingredients with the highest similarity are selected, and the cooking parameter combinations corresponding to the matching ingredients are weightedly fused based on the feature weight coefficient to obtain initial cooking parameters, wherein the initial cooking parameters include a heating power parameter and a stirring speed parameter; The temperature data, pressure data, heating power data and stirring motor speed data in the cavity of the cooking equipment are collected, and the initial cooking parameters are adaptively corrected according to the deviation between the reference cooking effect in the cooking mode database and the actual cooking effect; a second-order transfer function model is established based on the temperature data and the heating power data to obtain a power-temperature dynamic response curve, and a torque model considering the moment of inertia and viscous damping is established based on the stirring motor speed data and the motor current data to obtain a motor speed-torque response curve.

5. The method according to claim 1, characterized in that The controller generates a variable frequency control instruction based on the power-temperature dynamic response curve and the motor speed-torque response curve in combination with a preset multi-mode switching judgment rule, and the step of cooperatively adjusting the heating unit and the stirring motor unit of the cooking device through the variable frequency control module includes: Construct a system state vector including temperature, rotation speed, temperature change rate and rotation speed change rate, design a Kalman state observer based on a state transfer matrix and an observation matrix to estimate the system state vector, and optimize the gain matrix of the Kalman state observer through a system noise covariance matrix; design a state coupling matrix based on the estimated system state vector, and obtain the coupling coefficient of the state coupling matrix through particle swarm optimization; Based on the estimated system state vector, the working mode is identified in combination with a preset multi-mode switching determination rule, wherein the multi-mode switching determination rule includes a temperature steady-state error threshold, a speed fluctuation range threshold, a temperature change rate threshold, and a speed change rate threshold, and the control mode to be currently adopted is determined according to the comparison result between each component of the system state vector and the corresponding threshold; A coupling control law is designed according to the state coupling matrix and the currently adopted control mode, wherein the coupling control law associates the heating power adjustment amount with the speed adjustment amount through the coupling coefficient, wherein the heating power adjustment amount is composed of a proportional integral term of the temperature deviation and a coupling term of the estimated system state vector, and the speed adjustment amount is composed of a proportional differential term of the speed deviation and a coupling term of the estimated system state vector, and the coupling coefficient is updated in real time by designing an adaptive mechanism based on a weighted product of an estimated system state vector error and an actual system state vector error; The output of the coupling control law is converted into a variable frequency control instruction, wherein the heating power variable frequency control instruction is composed of a proportional integral term of a reference frequency and a heating power adjustment amount, and the stirring motor variable frequency control instruction is composed of a proportional integral term of a reference frequency and a speed adjustment amount; the variable frequency control instruction is converted into a control signal through pulse width modulation to coordinately control the heating unit and the stirring motor unit of the cooking equipment.

6. The method according to claim 5, characterized in that The coupling control law associates the heating power adjustment amount with the speed adjustment amount through the coupling coefficient, wherein the heating power adjustment amount is composed of a proportional integral term of the temperature deviation and a coupling term of the estimated system state vector, and the speed adjustment amount is composed of a proportional differential term of the speed deviation and a coupling term of the estimated system state vector, and the coupling coefficient is updated in real time by designing an adaptive mechanism based on a weighted product of an estimated system state vector error and an actual system state vector error, including: Obtain an instantaneous state error vector including a temperature error, a rotation speed error, a temperature change rate error, and a rotation speed change rate error, obtain a weighted state error vector based on an exponentially weighted sliding average of the instantaneous state error vector and a historical state error vector, wherein the exponentially weighted sliding average performs attenuation processing on historical error information through a forgetting factor; based on the instantaneous state error vector, calculate a difference between state error vectors at two adjacent sampling moments to obtain a state error change amount, and divide the state error change amount by a sampling period to obtain a state error change rate; Constructing an error covariance matrix considering the cross-coupling terms of temperature and speed, the error covariance matrix including the outer product term of the weighted state error vector, and a coupling strength term consisting of the product of the temperature error and the speed change rate error, and the product of the speed error and the temperature change rate error, calculated based on the state error change rate; Designing a dynamic adaptive learning rate based on the state error change rate, wherein the dynamic adaptive learning rate decays as the state error change rate increases through an exponential function, and a minimum learning rate threshold is set; A dynamic adjustment strategy for an adaptive weight matrix is ​​designed based on the relative amplitudes of the temperature error, the speed error, the temperature change rate error and the speed change rate error, wherein the diagonal elements of the adaptive weight matrix are dynamically adjusted according to the ratio of the corresponding state error amplitude to the maximum error amplitude; A coupling coefficient update equation is designed, wherein the coupling coefficient update equation includes a product term of the error covariance matrix and the adaptive weight matrix multiplied by the dynamic adaptive learning rate, and a coupling coefficient history update increment term, wherein the coupling coefficient history update increment term is used to accelerate convergence and suppress local oscillations; upper and lower limit constraints are imposed on the coupling coefficient calculated by the coupling coefficient update equation through a projection algorithm, and the constrained coupling coefficient is substituted into the coupling control law to coordinately control the heating unit and the stirring motor unit.

7. The method according to claim 1, characterized in that The controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the frequency conversion control instructions based on the switched mode until the cooking is completed, the steps comprising: Preprocessing the collected real-time operation data to construct a real-time operation parameter vector, calculating the parameter error vector between the real-time operation parameter vector and the expected operation parameter vector, and calculating the parameter change rate vector; extracting a feature index set based on the parameter error vector and the parameter change rate vector, the feature index set including temperature rise rate, temperature fluctuation amplitude, pressure change rate, pressure fluctuation amplitude, power utilization rate, power fluctuation coefficient and speed fluctuation, constructing the feature index set into an operation state feature vector and performing dimensionality reduction; A multi-layer fuzzy neural network is used for working mode recognition, wherein the input layer of the multi-layer fuzzy neural network receives the running state feature vector after dimension reduction, the first hidden layer uses the Gaussian membership function to fuzzify the running state feature vector, the second hidden layer performs reasoning operation based on fuzzy rules, and the output layer gives the degree of attribution of each mode, the cluster center of the Gaussian membership function is obtained by offline training of the K-means algorithm, and the standard deviation of the Gaussian membership function is estimated by the sample variance; a fuzzy rule base including a heating dominant mode, a stirring dominant mode and a coupling control mode is constructed, and the rule weights in the fuzzy rule base are optimized online by a back propagation algorithm, and the optimization goal is to minimize the pattern recognition error; The current optimal working mode is determined according to the degree of belonging of each mode, and a mode soft switching strategy is designed based on a negative exponential decay function. The mode soft switching strategy adopts a time decay switching function to perform weighted fusion on the control quantities before and after the switching. The switching rate parameter of the time decay switching function is adaptively adjusted according to the norm of the parameter error vector; Based on the working mode after switching, a performance index function including a tracking error term, a control energy consumption term and a control stability term is constructed, the weight coefficient of the performance index function is associated with the mode attribution degree of the multi-layer fuzzy neural network output, and is dynamically adjusted according to the current working mode; based on the performance index function, a gradient update equation of the PID parameters is constructed, and a variable frequency control instruction is generated according to the updated PID parameters.

8. An intelligent variable frequency cooking equipment multi-mode control system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain food images through an image acquisition device and collect food weight data at the same time. The image acquisition device transmits the food images to a deep learning processor. The deep learning processor extracts feature information from the food images based on a pre-trained food feature recognition model, and generates a food feature data packet in combination with the food weight data, wherein the food feature data packet includes food type information, weight information and volume information; the food feature recognition model is obtained by transferring the teacher network knowledge to the student network, wherein the multi-scale feature map of the teacher network is mapped to the student network feature space through a feature alignment module, and the sample difficulty is determined based on the predicted probability distribution entropy value, feature point density value and annotation difficulty value of the sample, and a hierarchical training strategy is used to optimize the model; The second unit is used for matching the corresponding initial cooking mode parameters from a preset cooking mode database based on the food feature data packet, and collecting real-time operation data in the cooking device cavity, wherein the real-time operation data includes temperature data, pressure data, heating power data, and stirring motor speed data, and the controller calculates a power-temperature dynamic response curve and a motor speed-torque response curve according to the real-time operation data; The third unit is used for the controller to generate a variable frequency control instruction according to the power-temperature dynamic response curve and the motor speed-torque response curve in combination with a preset multi-mode switching judgment rule, and coordinately adjust the heating unit and the stirring motor unit of the cooking device through the variable frequency control module; The controller continuously collects the real-time operation data during the cooking process, performs multi-mode adaptive switching according to the real-time operation data, and optimizes the frequency conversion control instructions based on the switched mode until cooking is completed.

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