A control method for an intelligent cooking machine and the cooking machine itself.

Abstract

This invention provides a control method for an intelligent cooking machine and the cooking machine itself. The method specifically includes: S1, the user selects the dish, and the system loads a reference trajectory y from the cloud. ref S1) Initial MPC parameters; S2) Start preheating, actuator reset to zero; S3) Add ingredients to the pot, weigh and record initial mass; S4) Collect all sensor data, timestamp synchronized; S5) Run EKF state estimation; S6) Call RLS online identification, correct MPC model parameters; S7) Fuzzy inference; S8) Call DDPG Actor network, output weight correction ΔQ; S9) Solve MPC optimization problem; S10) Send the first control quantity to the underlying PID; S11) PID quickly adjusts the actuator with a 10ms cycle; S12) If an anomaly is detected, immediately trigger safety protection; S13 Cooking complete; etc. This invention, through multi-sensor fusion and advanced state estimation algorithms, transforms the key biochemical states of ingredients that cannot be directly measured into observable and controllable variables, realizing real-time and accurate estimation of the internal state of the cooking process.

Description

Technical Field

[0001] This invention relates to the field of mixing equipment technology, specifically to a control method for an intelligent cooking machine and the cooking machine itself. Background Technology

[0002] With the deep integration of artificial intelligence and Internet of Things technologies, smart cooking machines, as the core representatives of modern kitchen automation and intelligence, are undergoing a profound transformation from "automatic execution" to "intelligent decision-making." Existing smart cooking machine technologies mainly revolve around automated operation, programmed cooking, and basic sensor feedback. Their core goal is to replace repetitive manual labor such as stir-frying and temperature control, and to solve the industry pain points of standardization and consistency of taste in Chinese cuisine.

[0003] Based on publicly available information and industry research, the technological implementation paths of intelligent cooking machines or cooking robots currently on the market can be summarized into the following categories: Pre-programmed control type: These devices have the cooking process (such as heating time, stirring speed, and ingredient addition sequence) of classic dishes pre-programmed. After the user selects a dish, the device executes the predetermined steps. For example, some devices achieve stir-frying by "simulating a chef's wok-tossing action" or "three-axis asynchronous planetary stirring technology," and control temperature fluctuations within a narrow range (such as ±1°C to ±2°C) through high-precision PID temperature control algorithms or AI temperature curve approximation algorithms. Their advantages are fixed processes and simple operation, but their disadvantages lie in the lack of perception and feedback on real-time cooking status. They cannot cope with the effects of differences in the initial state of ingredients (such as size and moisture), environmental disturbances (such as room temperature and air pressure), or equipment performance drift. Essentially, they still belong to open-loop or semi-closed-loop control, easily leading to "one-size-fits-all" or unstable quality.

[0004] Simple sensor feedback type: This type adds basic sensors (such as temperature sensors and weighing modules) for feedback on top of program control. For example, it adjusts the heating power by monitoring the pot temperature, or controls the amount of seasoning added by weighing (as mentioned in publicly available information, "the accuracy of solid seasoning is within ±0.5g"). Some high-end models have begun to integrate visual recognition systems to monitor changes in the color of food. However, this feedback is usually limited to single-point, delayed adjustments of directly measurable physical quantities, and cannot estimate in real time the key, unobservable states inside the food, such as the core temperature, internal moisture content distribution, Maillard reaction process, and doneness. Therefore, its control remains "blind adjustment," making it difficult to accurately grasp the critical points of chemical reactions that form flavor and heat.

[0005] Cloud-based recipe and remote control type: These devices download standardized recipes from the cloud via the internet, enabling updates and expansions to the cooking program (such as the "taste AI system" mentioned in the knowledge base, which includes parameters for hundreds of dishes). However, their intelligence is mainly reflected in the richness of the recipe data; the control logic of the cooking process remains fundamentally unchanged. The device still operates according to a fixed "recipe" and cannot make impromptu, adaptive adjustments based on the current device status and real-time feedback from ingredients.

[0006] In summary, the existing control methods for intelligent cooking machines generally suffer from the following core defects: (1) Model mismatch and poor adaptability: Cooking is a complex physicochemical process that is highly nonlinear, strongly coupled, and has time-varying parameters. Preset fixed models or parameters (such as heat transfer coefficients) are difficult to adapt to changes caused by different batches of ingredients, different initial conditions, and equipment aging, resulting in a decrease in control accuracy.

[0007] (2) Insufficient and lagging state perception dimensions: Existing sensing solutions focus on external and easily measurable variables (such as pot temperature and weight), but lack effective real-time estimation methods for the internal biochemical states that determine the final quality of the dish (such as core doneness, browning degree, and moisture evaporation dynamics), and lack key basis for control decisions.

[0008] (3) Limited level of decision-making intelligence: The control strategies are mostly rule-based threshold judgment or fixed PID adjustment, which cannot make multi-objective and forward-looking optimization decisions based on the comprehensive judgment of "color, aroma, taste and shape" of the cooking process, like an experienced chef.

[0009] (4) The control links are isolated and lack collaborative optimization: the actuators such as heating, stir-frying, humidifying and ventilation are often controlled independently, lacking a unified optimization framework to coordinate multiple control variables in order to achieve the best overall cooking effect (such as taste, nutrition and aroma). Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention proposes an intelligent cooking machine and its control method capable of deep perception, real-time modeling, intelligent decision-making, and collaborative optimization.

[0011] To achieve the above technical solution, the present invention provides a control method for an intelligent cooking machine, specifically including the following steps: S1. The user selects dishes, and the system loads the reference trajectory y from the cloud. ref (t) and the initial MPC parameters; S2. Start preheating, actuator returns to zero; S3. Put the ingredients into the pot, weigh and record the initial mass; S4. Collect all sensor data and synchronize timestamps; S5. Run EKF state estimation and update the true state vector. x The estimated value ; S6. Call RLS online identification to correct MPC model parameters; S7. Fuzzy Reasoning: Based on Determine the current cooking stage and adjust the reference trajectory y. ref Local targets; S8. Call the DDPG Actor network and output the weight correction ΔQ; S9. Solve the MPC optimization problem to obtain the optimal control sequence. ; S10, Set the first control quantity Send to the underlying PID; S11, PID rapidly adjusts the actuator with a 10ms cycle; S12. If an abnormality is detected, the safety protection will be triggered immediately: reduce power by 50% and turn on the exhaust fan; S13. Cooking complete. Upload the data from this run to the cloud for fine-tuning of the reinforcement learning model.

[0012] Preferably, the specific process of step S1 is as follows: (1) The device first obtains the local identifier Device ID and current status information, and then packages them together with the dish ID selected by the user into a request instruction and sends it to the cloud server; (2) After receiving the request, the cloud server first verifies the legality of the device, and then retrieves the corresponding recipe record from the recipe database based on the dish ID; (3) Package the retrieved recipe data. The packaged data includes metadata: dish name, cuisine, difficulty, estimated time, nutritional information; ingredients: ingredient list; phase_profile: cooking stage reference trajectory; mpc_params: MPC initial parameter set; hardware_adaptation: scale transformation parameters for different models of equipment; (4) The packaged data is sent to the device via MQTT or HTTPS protocol. After receiving the data, the device first performs CRC check or hash check, and then verifies the data signature to confirm that it does indeed come from the authorized cloud. (5) The device parses the data, loads the received MPC parameter set into the controller memory, sets the initial target value of the underlying PID controller to the first point of the first stage of the reference trajectory, and starts the online parameter identification module at the same time. (6) The device prompts the user via screen or voice that it is ready and waiting for the “start” command.

[0013] Preferably, in step S4, the collected sensor data includes: a microwave moisture sensor for monitoring the overall moisture content of the food; a multispectral color sensor for monitoring the surface color of the food; an infrared thermal imaging array for monitoring the temperature distribution of the pot bottom and the food surface; dual thermocouples for monitoring the pot body temperature and the surface temperature of the food; a six-axis force / torque sensor for monitoring the stirring resistance torque and axial force; a high-precision weighing module for monitoring the total mass of the food; an electronic nose for identifying burnt, raw, and aroma maturity; and an RGB-D camera for monitoring the three-dimensional volume of the food. The above sensor data is low-pass filtered, timestamped, and then sent to the data fusion module.

[0014] Preferably, in step S5, EKF state estimation is performed to update x. hat The specific process is as follows: S51. Define a 6-dimensional state vector: ; in, T core The core temperature of the ingredients; M w This represents the average moisture content. I Ma The Maillard reaction product index; S doneness To indicate doneness; ρ browning Surface browning density; V shrink This refers to the volume shrinkage rate; S52. Define a 10-dimensional observation vector: ; in, T pan This refers to the temperature at the bottom of the pot. T surf,avg The average surface temperature of the food; These are the brightness values, red-green hue, and yellow-blue hue detected by the multispectral color sensor, respectively. M w,mincrowav The overall moisture content of the ingredients; τ stir The resistance torque for stirring; W Total mass of ingredients; VOC burn This refers to the concentration of volatile organic compounds related to charring. V camera The three-dimensional volume of the food; S53. Establish a nonlinear state transition model: ; in, This is the discretized sampling time interval; C th Heat capacity of the food; R th The internal thermal resistance of the food; Q IH The power of IH electromagnetic heating directly applied to the food. This represents the rate of water evaporation. RH The relative humidity of the air inside the pot; v air The airflow velocity above the pot lid; A Ma The frequency factor of the Maillard reaction; Ea This is the activation energy of the Maillard reaction; R This is the universal gas constant; k d The growth coefficient for ripeness; For nonlinear functions, implementation can be achieved through table lookup or linear interpolation; S54. Through EKF recursion, the core temperature, doneness, moisture content, and coloring index are obtained in real time.

[0015] Preferably, in step S54, the specific process of EKF recursion includes: S541, Prediction: ,in, Let k be the state vector predicted at time k based on all information up to time k-1. This is the optimal estimate at time k-1, after incorporating the observations from time k-1. It is a nonlinear state transition function. This is the control input vector applied to the system at time k-1; S542, Predicting Covariance: ,in, Let be the covariance matrix of the prior estimation error. State transition function f Regarding the status x Jacobian matrix, Let be the covariance matrix of the posterior estimation error at the previous time step. Q The process noise covariance matrix; S543, Calculate the Jacobian: ,in, For observation function h Regarding the status x The Jacobian matrix in the predicted state The calculation is performed at the location, and its elements are... , representing the local sensitivity of the i-th observation to the j-th state variable; S544, Kalman gain: ,in, K k Here is the Kalman gain matrix. R To observe the noise covariance matrix, This is the transpose of the Jacobian matrix; S545, Update: ,in, For the final output, the posterior state estimate, z k Let k be the actual observation vector from the sensor at time k. These are the theoretical observations calculated based on the predicted state; S546, Update Covariance: ,in, Let be the covariance matrix of the posterior estimation error. I It is an identity matrix with the same dimensions as the state vector. The product of gain and observation Jacobian represents the degree to which the information obtained from the observation reduces uncertainty.

[0016] Preferably, in step S6, the specific process of calling RLS online identification and correcting the MPC model parameters is as follows: S61. Select the parameters to be identified and establish a regression model: Select the heat transfer coefficient as the parameter to be identified. k th To facilitate online identification, the following regression model is established: ; in, T core (k+1) and T core (k) The state estimates from EKF are used by RLS to correct the model parameters; S62. Initialize RLS: Set initial values ​​for each parameter to be identified at the start of cooking: : Obtain initial estimates from cloud-based recipe parameters; Set as a diagonal matrix, with the diagonal elements taking the larger value P0 = 100. Forgetting factor λ: ranging from 0.98 to 0.995; S63. Perform RLS updates in each control cycle. In the real-time control loop, execute the following sequentially: (1) Obtain the latest data: from the sensor and EKF Obtained from the weighing module W(k) ; (2) Calculate the measurement output y(k) ; (3) Constructing the regression vector ; (4) Calculate the new information: ; (5) Calculate the gain L(k) And update the covariance P(k) ; (6) Update parameters .

[0017] S64. Pass the identified parameters to MPC: In each MPC solution cycle, use the latest... Recalculate the coefficient matrix in the prediction model.

[0018] Preferably, in step S7, the reference trajectory y is adjusted. ref The specific process of local target acquisition is as follows: S71, Estimation from EKF state Extract e S , e M , e I Three key variables are used as inputs for fuzzy inference, including the degree of unfamiliarity / familiarity deviation. This represents the deviation between the current ripeness and the expected value at the corresponding moment in the reference trajectory; a positive value indicates overripeness, and a negative value indicates underripeness; moisture content deviation. This indicates the deviation between the current moisture content and the expected value; a positive value indicates that it is too wet, and a negative value indicates that it is too dry; the deviation in coloring rate. This indicates the deviation between the current coloring rate and the desired rate. A positive value indicates that the coloring is too fast, which may cause scorching; a negative value indicates that the coloring is too slow. The target value of the reference trajectory from the cloud at the current moment; S72. Based on the experience of culinary experts, 27 rules have been formulated, each in the following form: Output correction vector , in, The target temperature correction amount is expressed in °C. This is the correction amount for the water injection rate, in ml / s; The adjustment amount for the exhaust valve opening is %; S73. Using the Mamdani reasoning method, for each rule... r Calculate its activation strength: ; Then, the centroid method is used to resolve the fuzziness for each output variable: ; in, centroid(C r,T ) The output fuzzy set is the rule r. C r,T The center of gravity c water,r It is the output part of rule r. The centroid value of the specified fuzzy set, c θ,r For rule r The output fuzzy set centroid value; S74. Correct the reference trajectory using the correction amount output by fuzzy inference. y ref (t) The target value for a short period of time in the future is determined by the following method: (1) Modify the temperature setting trajectory, and change the current time and future time in the original reference trajectory. Np The temperature target value of each step is uniformly added And limit the amplitude to a safe range: ; (2) Water injection rate reference trajectory correction: Assume that in the original reference trajectory, the future... i The target value for the water injection rate of the step is The correction amount for the fuzzy inference output is The corrected target value for the water injection rate is: ; (3) Correction of exhaust valve opening reference trajectory: Assume that in the original reference trajectory, the future number of... i The target value for the exhaust opening of the step is The correction amount for the fuzzy inference output is The corrected target value for the water injection rate is: ; in, N p To predict the number of time-domain steps; S75, When fuzzy inference detects multiple consecutive times e S Approaching zero and e M and e I When the target is within a small range, it is considered that the current stage goal has been basically achieved, and the reference trajectory is actively switched to the next cooking stage.

[0019] Preferably, in step S8, the specific process of calling the DDPG Actor network and outputting the weight correction ΔQ is as follows: S81, Status Acquisition: Obtained from EKF Calculated from the reference trajectory e S , e M , e I These values ​​are normalized to form a state vector. s ; S82. Invoke the Actor network and perform forward propagation: This yields the original output, ranging from -0.5 to 0.5. S83, Update Weights: ; in, It can be obtained from cloud parameters or the value at the previous moment; S84, Pass to MPC: Send the new This will be used for subsequent MPC optimization problems.

[0020] Preferably, in step S9, the MPC optimization problem is solved to obtain the optimal control sequence. The specific process is as follows: S91. Linearized Prediction Model: Defining State Deviation Control deviation ,in x nom ,u nom Take a simple extrapolation of the previous round of MPC solution results or reference trajectory; after linearization, we get: , in, ; S92. Construct the standard form of QP and define the decision variable vector: ; S93. Solve QP using a real-time iterative strategy. At the beginning of each sampling period, calculate the linearized model matrix using the current state estimate. A i ，B i The system assembles the QP matrix H,f,L,I, and then calls an efficient QP solver to obtain the optimal solution within milliseconds. ; S94, from Recovery Then calculate: , Will It can be sent to the underlying PID controller or directly converted into a PWM signal to drive the actuator; S95. At the next sampling time k+1, shift the entire control sequence forward and apply the newly calculated... As part of the new nominal control trajectory, it is used for hot starts.

[0021] The present invention also provides an intelligent cooking machine, including the control method described above.

[0022] The invention provides a control method for an intelligent cooking machine, and the beneficial effects of the cooking machine are as follows: (1) The control method of this intelligent cooking machine uses multi-sensor fusion and advanced state estimation algorithm to turn key biochemical states such as the core cooking degree, moisture and Maillard reaction process of ingredients that cannot be directly measured into observable and controllable variables, thereby realizing real-time and accurate estimation of the internal state of the cooking process.

[0023] (2) The control method of this intelligent cooking machine enables the control model to identify and correct its own parameters online in order to adapt to differences in ingredients, environmental changes and equipment dynamics, ensure the consistency of control accuracy under different working conditions, and improve the adaptive capability of the control model.

[0024] (3) The control method of this intelligent cooking machine takes model predictive control as the core framework, coordinates and schedules multiple actuators such as heating power, stirring speed, water injection rate, and exhaust opening, and performs rolling optimization in the future time domain to achieve optimal control under multiple objectives (such as rapid cooking, energy saving, and best taste).

[0025] (4) The control method of this intelligent cooking machine enables the system to automatically compensate for changes in the characteristics of ingredients, equipment wear and tear and environmental interference through online identification via RLS, ensuring stable cooking results under different conditions and having strong robustness and adaptability.

[0026] (5) The control method of this intelligent cooking machine uses fuzzy reasoning to give the machine the ability to adjust based on rules and experience, while DDPG provides the ability to learn and optimize beyond fixed rules. The combination of the two enables the control system to not only handle routine deviations, but also to continuously optimize the control strategy by accumulating data, and the performance improves over time.

[0027] (6) The control method of this intelligent cooking machine is based on the MPC framework, which puts multiple control variables such as heating, stirring, humidification and ventilation into a unified optimization problem to solve. This avoids the conflicts that may occur in traditional single-loop control. Under the premise of satisfying various constraints, it can find the globally optimal control action sequence, taking into account efficiency, quality and energy consumption. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the control method of this intelligent cooking machine. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0030] Example 1: A control method for an intelligent cooking machine.

[0031] To address the problems of existing intelligent cooking machines having fixed heat settings, limited power levels, and an inability to flexibly fine-tune based on the real-time state of ingredients (cooking, moisture, color), this invention designs an intelligent control system and method for cooking machines based on multi-source sensor fusion and hierarchical predictive control. This system senses the physicochemical state of the ingredients in real time, uses extended Kalman filtering (EKF) for state estimation, and combines model predictive control (MPC), fuzzy adaptive PID, and parameter self-tuning based on deep reinforcement learning to achieve continuous stepless adaptive control of heating power, stir-frying speed, and humidity adjustment, achieving professional chef-level precise heat management.

[0032] To address the three major challenges of sensing "raw / cooked," "moisture content," and "coloring," this system is equipped with the following sensors (all integrated into the pot lid, pot bottom, and robotic arm):

[0033] Based on the above intelligent control system, referring to Figure 1 As shown, the present invention provides a control method for an intelligent cooking machine, which specifically includes the following steps: S1. The user selects dishes, and the system loads the reference trajectory y from the cloud. ref (t) and the initial MPC parameters, the specific process is as follows: (1) The device first obtains its local identifier (Device ID) and current status information, then packages them together with the dish ID selected by the user into a request instruction and sends it to the cloud server. In actual operation, when the user selects a dish on the device's touchscreen or mobile APP, the system does not start cooking immediately, but first initiates a recipe data request. Then the device (cooking machine) first obtains its local identifier (Device ID) and current status information (such as pot model, heating power limit, stir-fry motor specifications, etc.), and then packages them together with the dish ID selected by the user into a request instruction. Since different models of devices have differences in heating power, stir-fry speed, etc., the system needs to report these hardware parameters when making the request so that the cloud can perform parameter adaptation accordingly.

[0034] (2) After receiving a request, the cloud server first verifies the legality of the device (based on a whitelist check of the device certificate), and then retrieves the corresponding recipe record from the recipe database according to the dish ID. The recipe database in the cloud is not a simple recipe text, but rather a precise numerical instruction sequence deconstructed from the experience of professional chefs. The cloud needs to scale the standard recipe parameters according to the reported device model to ensure that devices with different hardware specifications can accurately replicate the same taste. Its core principle is that the cloud pre-stores the scale transformation information between the "standard cooking equipment" and various "target device models". When a request is received, the system looks up the conversion relationship between the target device and the standard device and performs mathematical transformations on parameters such as temperature, power, and speed. For example, if the heating power of the standard device is 2500W, while the maximum power of the target device is only 2000W, the cloud will adjust the temperature setpoint and heating rate accordingly. If the transformation information for a new device model has not yet been established, the system can adopt an adaptive method: by comparing the feedback data in actual operation with the data of the standard device, a parameter adjustment curve can be dynamically established.

[0035] (3) Pack the retrieved recipe data into a data package. The packaged data includes metadata: dish name, cuisine, difficulty, estimated time, nutritional information; ingredients: ingredient list; phase_profile: cooking stage reference trajectory; mpc_params: MPC initial parameter set; hardware_adaptation: scale transformation parameters for different models of equipment.

[0036] (4) The packaged data is sent to the device via MQTT or HTTPS protocol. After receiving the data, the device first performs CRC check or hash verification, and then verifies the data signature to confirm that it does indeed come from the authorized cloud, preventing malicious injection. Successfully parsed data packets are stored in the device's local flash memory and labeled with a version number. The purpose of this is that if the same dish is cooked again in the future and there is no version update, the device can directly use the local cache to avoid repeated downloads. The recipe generation system also supports a dual storage strategy of "recipe data being stored in the local storage module and the cloud storage module respectively" to reduce network dependence.

[0037] (5) The device parses the data and loads the received MPC parameter set into the controller memory, including the weight matrix Q / R, prediction time domain N, constraint boundaries, etc. The initial target value of the underlying PID controller is set to the first point of the first stage of the reference trajectory. At the same time, the online parameter identification module (RLS) is started, and its initial covariance matrix is ​​set according to the model parameter confidence level provided by the cloud.

[0038] (6) The device prompts the user via screen or voice that it is ready and waiting for the "start" command. At this time, the reference trajectory has been loaded, the MPC has been initialized, and the EKF state estimator is in place. Everything is ready, and after the user issues the start command, the real-time control loop will officially start running.

[0039] Through the above steps, personalized, updatable, secure, and reliable cloud deployment and loading of recipes are achieved. By including `phase_profile` (cooking phase reference trajectory) and `mpc_params` (MPC initial parameter set), the cooking process is described as a combination of dynamic trajectory and model parameters, which is more refined than a simple time-temperature sequence; the `hardware_adaptation` parameter enables the same recipe to adapt to different hardware.

[0040] S2. Start preheating, actuators return to zero; before loading a new recipe, the equipment needs to perform a hardware self-check: confirming that the heating module, stir-fry motor, water pump and other actuators are functioning normally, and that the sensor readings are within the valid range. Then, it enters the preheating stage, raising the pot temperature to the starting temperature required for the first stage of the recipe. This process is also used to calibrate the zero-point drift of the temperature sensor.

[0041] S3. Put the ingredients into the pot, weigh them and record the initial mass.

[0042] S4. Collect all sensor data and synchronize the timestamps. The collected sensor data includes: a microwave moisture sensor for monitoring the overall moisture content of the ingredients; a multispectral color sensor for monitoring the surface color of the ingredients; an infrared thermal imaging array for monitoring the temperature distribution of the pot bottom and the surface of the ingredients; dual thermocouples for monitoring the temperature of the pot body and the surface temperature of the ingredients; a six-axis force / torque sensor for monitoring the stirring resistance torque and axial force; a high-precision weighing module for monitoring the total mass of the ingredients; an electronic nose for identifying burnt, raw, and aroma maturity; and an RGB-D camera for monitoring the three-dimensional volume of the ingredients. The above sensor data is low-pass filtered, timestamped, and then sent to the data fusion module. In actual operation, the combination of the microwave moisture sensor, multispectral color sensor, infrared thermal imaging, electronic nose, RGB-D camera, and six-axis force sensor enables comprehensive and synchronous perception of the physical properties (mass, volume, temperature distribution), chemical properties (moisture, volatiles), and mechanical properties (stirring resistance) of the ingredients. This set of specific sensor combinations and their data preprocessing methods (low-pass filtering, timestamp synchronization) optimized for state estimation during the cooking process provides reliable input for subsequent advanced algorithms.

[0043] S5. Run EKF state estimation and update the true state vector. x The estimated value Since core temperature, ripeness, and internal moisture distribution cannot be directly measured, they need to be estimated from observations using EKF (Extended Kernel-Fluid Temperature). The specific process is as follows: S51. Define a 6-dimensional state vector: ; in, T core The core temperature of the food (°C); M w Average moisture content (%) I Ma The Maillard reaction product index (0~1); S doneness The degree of doneness is indicated by the following: 0 = completely raw, 1 = fully cooked. ρ browning Surface browning density ; V shrink Volume shrinkage rate (%) S52. Define a 10-dimensional observation vector: , in, T pan This refers to the temperature at the bottom of the pot, typically 25~250℃; T surf,avgThe average surface temperature of the food, typically 20~180℃; These are the brightness values, red-green hue, and yellow-blue hue detected by the multispectral color sensor, respectively. M w,mincrowav The overall moisture content of the ingredients (%); τ stir The resistance torque for stirring is typically 0~5 N·m; W The total mass of the ingredients is measured in real time with a resolution of 0.1g. VOC burn The concentration of volatile organic compounds related to charring is expressed in ppm or normalized index (0~1). V camera The volume of the food in three dimensions, in cm³ or mL; S53. Establish a nonlinear state transition model: ; in, This is the discretized sampling time interval; C th Heat capacity of the food; R th The internal thermal resistance of the food; Q IH The power of IH electromagnetic heating directly applied to the food; the rate of change of core temperature = (conducted heat power flowing in from the surface + direct heating power) / heat capacity. The conductive heat power is obtained by dividing the temperature difference between the surface and the core by the thermal resistance.

[0044] ; in, This represents the rate of water evaporation. RH The relative humidity of the air inside the pot; v air The velocity of the airflow above the pot lid is denoted by . The moisture content decreases over time, and the rate of decrease is determined by the evaporation rate. The evaporation rate increases exponentially with increasing temperature, decreases with increasing humidity, and increases with stronger exhaust ventilation.

[0045] ; in, A Ma The frequency factor of the Maillard reaction; Ea This is the activation energy of the Maillard reaction; R The constant is a universal gas constant; the Maillard reaction rate follows the Arrhenius law, and the higher the temperature, the faster the reaction; it is also affected by moisture, and the reaction is most vigorous when the moisture content is about 50%.

[0046] ; in, k d This is the ripeness growth coefficient; the ripeness increases monotonically from 0 (completely raw) to 1 (fully ripe). The growth rate is proportional to the square of the portion of the core temperature exceeding 60°C. T core Curing is not performed at temperatures below 60℃.

[0047] ; in, f τ It is a nonlinear function, implemented using table lookup or linear interpolation; EKF utilizes measured torque. τ stir,meas Compared with model predictions The differences are used to correct for ripeness and moisture estimates.

[0048] S54. Core temperature, doneness, moisture content, and color index are obtained in real time through EKF recursion (standard extended Kalman filter). The specific process of EKF recursion includes: S541, Prediction: ,in, Let k be the state vector predicted at time k based on all information up to time k-1. This is the optimal estimate at time k-1, after incorporating the observations from time k-1. It is a nonlinear state transition function. The control input vector applied to the system at time k-1 is used; the estimation from the previous time step is pushed forward one step using the system dynamic model (heat conduction, water evaporation, Maillard reaction, etc.) to obtain the "prior" state at the current time.

[0049] S542, Predicting Covariance: ,in, Let be the covariance matrix of the prior estimation error. State transition function f Regarding the status x Jacobian matrix, Let be the covariance matrix of the posterior estimation error at the previous time step. Q Let be the process noise covariance matrix; calculate the uncertainty of the predicted state and add process noise to increase the covariance in the prediction step.

[0050] S543, Calculate the Jacobian: ,in, For observation function h Regarding the status x The Jacobian matrix in the predicted state The calculation is performed at the location, and its elements are... , representing the local sensitivity of the i-th observation to the j-th state variable; linearize the observation model so that the Kalman gain can be calculated in the update step. The Jacobian matrix tells the system "how changes in the state will affect the observations".

[0051] S544, Kalman gain: ,in, Here is the Kalman gain matrix. R To observe the noise covariance matrix, Given the transpose of the Jacobian matrix, compute an optimal weight that minimizes the covariance of the posterior estimation error. The magnitude of the gain depends on the prediction uncertainty. With observation uncertainty ( R The relative magnitude of the prediction. The more uncertain the prediction and the more precise the observation, the greater the gain (more confidence in the observation).

[0052] S545, Update: ,in, For the final output, the posterior state estimate, z k Let k be the actual observation vector from the sensor at time k. The theoretical observations are calculated based on the predicted state; using Kalman gain. K k Multiply by the new information to correct the predicted state and obtain the posterior estimate.

[0053] S546, Update Covariance: ,in, Let be the covariance matrix of the posterior estimation error. I It is an identity matrix with the same dimensions as the state vector. The posterior covariance is the product of the gain and the observation Jacobian, representing the degree to which the information obtained from the observations reduces uncertainty. Due to the inclusion of observational information, the posterior covariance is typically greater than the prior covariance. Smaller (reduced uncertainty).

[0054] Through the aforementioned EKF recursion, a standard and reliable recursive algorithm is provided for optimal estimation of the complex nonlinear state defined in this embodiment under conditions of sensor noise and model uncertainty, ensuring the real-time performance and accuracy of state estimation. Furthermore, by specifically adapting the standard EKF algorithm to the specific state and observation vectors of this invention, the execution methods of steps such as prediction, updating, and covariance propagation in this specific cooking application scenario are clarified.

[0055] In this embodiment, the "internal biochemical state" of ingredients is precisely defined and mathematically modeled using EKF (Extended Kernel-Flattening Function), making indicators such as "core temperature," "doneness," and "browning density," which are previously judged by chefs based on experience, quantifiable and estimable engineering variables. The proposed 6-dimensional nonlinear state-space model, which incorporates key physicochemical processes such as heat transfer, moisture evaporation, Maillard reaction, and volume shrinkage, provides specific and innovative state equations for the application of EKF in the culinary field. Through EKF, the system can obtain key states such as core temperature, doneness, moisture content, and browning index in real time, providing reliable feedback for subsequent control.

[0056] S6. Call RLS (Recursive Least Squares) online to identify and correct MPC model parameters. RLS is a recursive linear regression algorithm that can update parameter estimates online, minimizing the weighted squared error between the model output and actual observations. In the control system of the intelligent cooking machine, Recursive Least Squares (RLS) is used online to identify model parameters (such as heat transfer coefficient, evaporation coefficient, Maillard reaction rate constant, etc.) that change with batches of ingredients and environmental variations, thereby correcting the prediction model used by MPC (Model Predictive Control) in real time and improving control accuracy. The specific process of RLS online identification is as follows: S61. Select the parameters to be identified and establish a regression model: Select the heat transfer coefficient as the parameter to be identified. k th To facilitate online identification, the following regression model is established: ; in, T core (k+1) and T core (k) The state estimates from EKF are used by RLS to correct the model parameters; S62. Initialize RLS: Set initial values ​​for each parameter to be identified at the start of cooking: : Obtain initial estimates from cloud-based recipe parameters; Set as a diagonal matrix, with the diagonal elements taking a large value P0 = 100, indicating high initial uncertainty, which the algorithm will adjust quickly; Forgetting factor λ: takes a value of 0.98 ~ 0.995; S63. Perform RLS updates in each control cycle. In the real-time control loop, execute the following sequentially: (1) Obtain the latest data: from the sensor and EKF Obtained from the weighing module W (k) ; (2) Calculate the measurement outputy(k) ; (3) Constructing the regression vector ; (4) Calculate the new information: ; (5) Calculate the gain L(k) And update the covariance P(k) ; (6) Update parameters .

[0057] S64. Pass the identified parameters to MPC: In each MPC solution cycle, use the latest... Recalculate the coefficient matrix in the prediction model. Since RLS can be updated at the same frequency as MPC, MPC always performs rolling optimization based on the most accurate model at present.

[0058] The above method solves the problem of controlling model parameters (such as heat transfer coefficient). k th To address the critical issue of drift caused by changes in ingredients, cookware, and environment, the MPC prediction model maintains high fidelity. A specific regression model and update process based on Regression-Solution Modeling (RLS) using EKF state estimates as input were designed, enabling dynamic self-correction of model parameters. Through RLS, the cooking machine can adapt to the physical characteristics of different ingredients, achieving intelligent control that becomes increasingly accurate with each use.

[0059] S7. Fuzzy Reasoning: Based on Determine the current cooking stage and adjust the reference trajectory y. ref The local objective. The main function of fuzzy inference is to estimate the current state (roughness / familiarity) based on the EKF. Moisture content Coloring index Dynamically adjust the reference trajectory y ref (t) The local target enables the system to cope with uncertainties such as differences in ingredients and environmental changes, and achieve more flexible fine-tuning of the heat.

[0060] The specific process of fuzzy reasoning is as follows: S71, Estimation from EKF state Extract e S , e M , e I Three key variables are used as inputs for fuzzy inference, including the degree of unfamiliarity / familiarity deviation. This represents the deviation between the current ripeness and the expected value at the corresponding moment in the reference trajectory; a positive value indicates overripeness, and a negative value indicates underripeness; moisture content deviation. This indicates the deviation between the current moisture content and the expected value; a positive value indicates that it is too wet, and a negative value indicates that it is too dry; the deviation in coloring rate. This indicates the deviation between the current coloring rate and the desired rate. A positive value indicates that the coloring is too fast, which may cause scorching; a negative value indicates that the coloring is too slow. The target value of the reference trajectory from the cloud at the current moment; S72. Based on the experience of culinary experts, 27 rules (3×3×3) are formulated, each rule in the following form: Output correction vector , in, The target temperature correction amount is expressed in °C. This is the correction amount for the water injection rate, in ml / s; The adjustment amount for the exhaust valve opening is %; Some of the rules are shown in the table below:

[0061] All rule weights are set to 1.0.

[0062] S73. Using the Mamdani reasoning method, for each rule... r Calculate its activation strength: , Then, the centroid method is used to resolve the fuzziness for each output variable: , in, centroid(C r,T ) The output fuzzy set is the rule r. C r,T The center of gravity c water,r It is the output part of rule r. The centroid value of the specified fuzzy set, c θ,r For rule r The output fuzzy set centroid value; S74. Correct the reference trajectory using the correction amount output by fuzzy inference. y ref (t) The target value for a short period of time in the future is determined by the following method: (1) Modify the temperature setting trajectory, and change the current time and future time in the original reference trajectory. Np The temperature target value of each step is uniformly added And limit the amplitude to a safe range: ; (2) Water injection rate reference trajectory correction: Assume that in the original reference trajectory, the future... i The target value for the water injection rate of the step is The correction amount for the fuzzy inference output is The corrected target value for the water injection rate is: ; (3) Correction of exhaust valve opening reference trajectory: Assume that in the original reference trajectory, the future number of... i The target value for the exhaust opening of the step is The correction amount for the fuzzy inference output is The corrected target value for the water injection rate is: ; in, N p To predict the number of time-domain steps; S75, When fuzzy inference detects multiple consecutive times e S Approaching zero and e M and e I Within a small range, it is considered that the current stage goal has been basically achieved, and the reference trajectory is actively switched to the next cooking stage. The core role of fuzzy reasoning in intelligent cooking machines can be summarized as follows: by using expert rules, the state deviation estimated by EKF is transformed into a local correction of the reference trajectory, enabling the system to adaptively adjust the heat, moisture and coloring rate, thereby overcoming the limitations of fixed trajectories.

[0063] In this embodiment, by encoding the adjustment experience of human chefs into machine-executable fuzzy rules, the system possesses human-like intelligence capable of fine-tuning local targets based on real-time deviations, enhancing its flexibility in responding to sudden disturbances. Furthermore, a fuzzy inference system is defined, using three core state estimation deviations—doneness deviation, moisture content deviation, and coloring rate deviation—as input. Up to 27 IF-THEN rules are formulated to output coordinated corrections for temperature, water injection, and ventilation, achieving dynamic and intelligent local optimization of the reference trajectory.

[0064] S8. Invoke the DDPG Actor network and output the weight correction ΔQ. In the intelligent cooking machine control system, the Deep Deterministic Policy Gradient (DDPG) is used to adaptively adjust the weight matrix ΔQ of the MPC online, enabling the controller to dynamically change the emphasis on targets such as temperature, moisture, and coloring according to different ingredients and different cooking stages, thereby improving control quality. In the cooking machine, the Actor network is deployed on a local embedded platform (such as Jetson Orin NX) and is invoked once every few control cycles (e.g., every 2 seconds) to output ΔQ.

[0065] The specific process of calling the DDPG Actor network and outputting the weight correction ΔQ is as follows: S81, Status Acquisition: Obtained from EKF Calculated from the reference trajectory e S , e M , e I Identify and obtain from the fuzzy stage phase The timer provides t elapsed These values ​​are normalized to form a state vector. s ; ,in, phase This represents the current cooking stage (one-hot code, 5 stages in total). t elapsed Cooking time (normalized to [0,1]); S82. Invoke the Actor network and perform forward propagation: This yields the original output, ranging from -0.5 to 0.5. S83, Update Weights: , in, It can be obtained from cloud parameters or the value at the previous moment; S84, Pass to MPC: Send the new Q =diag( q T ,q M ,q I This is used in subsequent MPC optimization problems.

[0066] In this embodiment, a self-learning mechanism is introduced to enable the weight matrix Q in the MPC controller, which weighs the importance of different state variables, to be dynamically optimized based on long-term cooking results. By combining the Actor network of deep reinforcement learning (DDPG) with the weight optimization of MPC, a method is designed to fine-tune the weights of MPC performance indicators online based on real-time states and deviations, enabling the control system to have the ability to continuously evolve.

[0067] S9. Solve the MPC optimization problem to obtain the optimal control sequence. In the control system of an intelligent cooking machine, the core of model predictive control (MPC) is to... k Based on the current state estimation and the corrected reference trajectory Solve an open-loop optimal control problem in the finite time domain to obtain the future... N Step control sequence Then only the first control variable Apply to the underlying actuators (heating, stirring, water injection, ventilation).

[0068] The specific process is as follows: S91. Linearized Prediction Model: Since the prediction model is nonlinear, directly solving the nonlinear MPC would be computationally too expensive. To meet real-time requirements (50Hz control cycle, i.e., completed within 20ms), we adopt a continuous linearization method: at each sampling time, the nonlinear model is estimated in the current state... Control of the previous moment By linearizing the neighborhood, a linear time-varying model is obtained, which approximates the original problem as a convex quadratic programming (QP) problem.

[0069] Define state deviation Control deviation ,in x nom ,u nom Take a simple extrapolation of the previous round of MPC solution results or reference trajectory; after linearization, we get: , in, ; S92. Construct the standard form of QP and define the decision variable vector: ; S93. Solve QP using a real-time iterative strategy. At the beginning of each sampling period, calculate the linearized model matrix using the current state estimate. A i ，B i The system assembles the QP matrix H,f,L,I, and then calls an efficient QP solver to obtain the optimal solution within milliseconds. ; S94, from Recovery Then calculate: , Will It can be sent to the underlying PID controller or directly converted into a PWM signal to drive the actuator; S95. At the next sampling time k+1, shift the entire control sequence forward and apply the newly calculated... As part of the new nominal control trajectory, it is used for hot starts.

[0070] In this embodiment, the complex, constrained multivariate optimization control problem is transformed into a standard form solvable in real time (within milliseconds) on an embedded system through linearization and quadratic programming (QP). A "hot start" strategy is employed to accelerate the solution, ensuring the real-time nature of the control response. Furthermore, specific linearization methods, deviation definitions, decision variable construction, and real-time iterative solution strategies are designed for the cooking process model, realizing the practical engineering application of MPC in the high-speed dynamic process of a cooking machine. Using this method, the cooking machine can calculate the optimal heating power, stir-frying speed, water injection rate, and exhaust opening in real time to minimize tracking errors and meet safety constraints, achieving dynamic and precise heat control.

[0071] S10, Set the first control quantity This is sent to the underlying PID controller. The underlying PID controller includes a heating power PID controller and a stirring speed PID controller. The heating power PID controller uses the pot temperature T measured by an infrared thermocouple as the input. pan For feedback, the target temperature T sp Provided by the upper layer. The stir-frying speed PID uses the resistance torque measured by the torque sensor as feedback to maintain sufficient tumbling of the ingredients.

[0072] S11 and PID rapidly adjust the actuator with a 10ms cycle.

[0073] S12. If an abnormality is detected, the safety protection will be triggered immediately: reduce power by 50% and turn on the exhaust fan.

[0074] S13. Cooking complete. Upload the data from this run to the cloud for fine-tuning of the reinforcement learning model.

[0075] This intelligent cooking machine's control method utilizes multi-sensor fusion and advanced state estimation algorithms to transform key biochemical states of ingredients—such as core cookedness, moisture content, and Maillard reaction processes—that are not directly measurable, into observable and controllable variables. This enables real-time and accurate estimation of the internal states during the cooking process. Furthermore, it allows the control model to identify and correct its parameters online to adapt to differences in ingredients, environmental changes, and equipment dynamics, ensuring consistent control accuracy under various operating conditions and enhancing the control model's adaptive capabilities.

[0076] This intelligent cooking machine uses model predictive control as its core framework, comprehensively scheduling multiple actuators such as heating power, stirring speed, water injection rate, and exhaust fan opening. It performs rolling optimization in the future time domain to achieve optimal control under multiple objectives (such as rapid cooking, energy saving, and best taste). Furthermore, through online RLS identification, the system can automatically compensate for changes in food characteristics, equipment wear and tear, and environmental interference, ensuring stable cooking results under different conditions and exhibiting strong robustness and adaptability.

[0077] The control method of this intelligent cooking machine employs fuzzy reasoning, which endows the machine with rule-based experience-based adjustment capabilities, while DDPG provides self-learning and optimization capabilities that go beyond fixed rules. The combination of these two approaches allows the control system to not only handle routine deviations but also continuously optimize its control strategy through data accumulation, resulting in performance improvements over time. Furthermore, based on the MPC framework, multiple control variables such as heating, stirring, humidification, and ventilation are placed in a unified optimization problem for solution, avoiding potential conflicts arising from traditional single-loop control. This allows the system to find the globally optimal sequence of control actions while satisfying various constraints, balancing efficiency, quality, and energy consumption.

[0078] Example 2: A smart cooking machine The present invention also provides an intelligent cooking machine, including the control system described in Embodiment 1. By executing the above-mentioned advanced intelligent control method, the cooking machine has excellent adaptive, self-learning, and global optimization cooking capabilities.

[0079] The above description is only a preferred embodiment of the present invention, but the present invention should not be limited to the content disclosed in the embodiments and drawings. Therefore, any equivalent or modified embodiments made without departing from the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. A control method for an intelligent cooking machine, characterized in that... Specifically, the steps include the following: S1. The user selects dishes, and the system loads the reference trajectory y from the cloud. ref (t) and the initial MPC parameters; S2. Start preheating, actuator returns to zero; S3. Put the ingredients into the pot, weigh and record the initial mass; S4. Collect all sensor data and synchronize timestamps; S5. Run EKF state estimation and update the true state vector. x The estimated value ; S6. Call RLS online identification to correct MPC model parameters; S7. Fuzzy Reasoning: Based on Determine the current cooking stage and adjust the reference trajectory y. ref Local targets; S8. Call the DDPG Actor network and output the weight correction ΔQ; S9. Solve the MPC optimization problem to obtain the optimal control sequence. ; S10, Set the first control quantity Send to the underlying PID; S11, PID rapidly adjusts the actuator with a 10ms cycle; S12. If an abnormality is detected, the safety protection will be triggered immediately: reduce power by 50% and turn on the exhaust fan; S13. Cooking complete. Upload the data from this run to the cloud for fine-tuning of the reinforcement learning model.

2. The control method for the intelligent cooking machine as described in claim 1, characterized in that, The specific process of step S1 is as follows: (1) The device first obtains the local identifier Device ID and current status information, and then packages them together with the dish ID selected by the user into a request instruction and sends it to the cloud server; (2) After receiving the request, the cloud server first verifies the legality of the device, and then retrieves the corresponding recipe record from the recipe database based on the dish ID; (3) Package the retrieved recipe data. The packaged data includes metadata: dish name, cuisine, difficulty, estimated time, nutritional information; ingredients: ingredient list; phase_profile: cooking stage reference trajectory; mpc_params: MPC initial parameter set; hardware_adaptation: scale transformation parameters for different models of equipment; (4) The packaged data is sent to the device via MQTT or HTTPS protocol. After receiving the data, the device first performs CRC check or hash check, and then verifies the data signature to confirm that it does indeed come from the authorized cloud. (5) The device parses the data, loads the received MPC parameter set into the controller memory, sets the initial target value of the underlying PID controller to the first point of the first stage of the reference trajectory, and starts the online parameter identification module at the same time. (6) The device prompts the user via screen or voice that it is ready and waiting for the "start" command.

3. The control method for the intelligent cooking machine as described in claim 1 or 2, characterized in that, In step S4, the collected sensor data includes: a microwave moisture sensor for monitoring the overall moisture content of the ingredients; a multispectral color sensor for monitoring the surface color of the ingredients; an infrared thermal imaging array for monitoring the temperature distribution of the pot bottom and the surface of the ingredients; dual thermocouples for monitoring the temperature of the pot body and the surface temperature of the ingredients; a six-axis force / torque sensor for monitoring the stirring resistance torque and axial force; a high-precision weighing module for monitoring the total mass of the ingredients; an electronic nose for identifying burnt, raw, and aroma maturity; and an RGB-D camera for monitoring the three-dimensional volume of the ingredients. After low-pass filtering and timestamp alignment, the above sensor data is sent to the data fusion module.

4. The control method for the intelligent cooking machine as described in claim 1 or 3, characterized in that, In step S5, EKF state estimation is performed to update x. hat The specific process is as follows: S51. Define a 6-dimensional state vector: ； in, T core The core temperature of the ingredients; M w This represents the average moisture content. I Ma The Maillard reaction product index; S doneness To indicate doneness; ρ browning Surface browning density; V shrink This refers to the volume shrinkage rate; S52. Define a 10-dimensional observation vector: ； in, T pan This refers to the temperature at the bottom of the pot. T surf,avg The average surface temperature of the food; These are the brightness values, red-green hue, and yellow-blue hue detected by the multispectral color sensor, respectively. M w,mincrowav The overall moisture content of the ingredients; τ stir The resistance torque for stirring; W Total mass of ingredients; VOC burn This refers to the concentration of volatile organic compounds related to charring. V camera The three-dimensional volume of the food; S53. Establish a nonlinear state transition model: ； in, This is the discretized sampling time interval; C th Heat capacity of the food; R th The internal thermal resistance of the food; Q IH The power of IH electromagnetic heating directly applied to the food. This represents the rate of water evaporation. RH The relative humidity of the air inside the pot; v air The airflow velocity above the pot lid; A Ma The frequency factor of the Maillard reaction; Ea This is the activation energy of the Maillard reaction; R This is the universal gas constant; k d The growth coefficient for ripeness; For nonlinear functions, implementation can be achieved through table lookup or linear interpolation; S54. Through EKF recursion, the core temperature, doneness, moisture content, and coloring index are obtained in real time.

5. The control method for the intelligent cooking machine as described in claim 4, characterized in that, In step S54, the specific process of EKF recursion includes: S541, Prediction: ,in, Let k be the state vector predicted at time k based on all information up to time k-1. This is the optimal estimate at time k-1, after incorporating the observations from time k-1. It is a nonlinear state transition function. This is the control input vector applied to the system at time k-1; S542, Predicting Covariance: ,in, Let be the covariance matrix of the prior estimation error. State transition function f Regarding the status x Jacobian matrix, Let be the covariance matrix of the posterior estimation error at the previous time step. Q The process noise covariance matrix; S543, Calculate the Jacobian: ,in, For observation function h Regarding the status x The Jacobian matrix in the predicted state The calculation is performed at the location, and its elements are... , representing the local sensitivity of the i-th observation to the j-th state variable; S544, Kalman gain: ,in, K k Here is the Kalman gain matrix. R To observe the noise covariance matrix, This is the transpose of the Jacobian matrix; S545, Update: ,in, For the final output, the posterior state estimate, z k Let k be the actual observation vector from the sensor at time k. These are the theoretical observations calculated based on the predicted state; S546, Update Covariance: ,in, Let be the covariance matrix of the posterior estimation error. I It is an identity matrix with the same dimensions as the state vector. The product of gain and observation Jacobian represents the degree to which the information obtained from the observation reduces uncertainty.

6. The control method for the intelligent cooking machine as described in claim 1, characterized in that, In step S6, the specific process of calling RLS online identification and correcting the MPC model parameters is as follows: S61. Select the parameters to be identified and establish a regression model: Select the heat transfer coefficient as the parameter to be identified. k th To facilitate online identification, the following regression model is established: ； in, T core (k+1) and T core (k) The state estimates from EKF are used by RLS to correct the model parameters; S62. Initialize RLS: Set initial values ​​for each parameter to be identified at the start of cooking: : Obtain initial estimates from cloud-based recipe parameters; P(0): Set as a diagonal matrix, with the diagonal elements taking the larger value P0 = 100. Forgetting factor λ: ranging from 0.98 to 0.995; S63. Perform RLS updates in each control cycle. In the real-time control loop, execute the following sequentially: (1) Obtain the latest data: from the sensor and EKF T surf (k) , RH(k) , v air (k) Obtained from the weighing module W(k) ; (2) Calculate the measurement output y(k) ; (3) Constructing the regression vector ; (4) Calculate the new information: ; (5) Calculate the gain L(k) And update the covariance P(k) ; (6) Update parameters ; S64. Pass the identified parameters to MPC: In each MPC solution cycle, use the latest... Recalculate the coefficient matrix in the prediction model.

7. The control method for the intelligent cooking machine as described in claim 1, characterized in that, In step S7, the reference trajectory y is adjusted. ref The specific process of local target acquisition is as follows: S71, Estimation from EKF state Extract e S , e M , e I Three key variables are used as inputs for fuzzy inference, including the degree of unfamiliarity / familiarity deviation. This represents the deviation between the current ripeness and the expected value at the corresponding moment in the reference trajectory; a positive value indicates overripeness, and a negative value indicates underripeness; moisture content deviation. This indicates the deviation between the current moisture content and the expected value; a positive value indicates that it is too wet, and a negative value indicates that it is too dry; the deviation in coloring rate. This indicates the deviation between the current coloring rate and the desired rate. A positive value indicates that the coloring is too fast, which may cause scorching; a negative value indicates that the coloring is too slow. The target value of the reference trajectory from the cloud at the current moment; S72. Based on the experience of culinary experts, 27 rules have been formulated, each in the following form: Output correction vector , in, The target temperature correction amount is expressed in °C. This is the correction amount for the water injection rate, in ml / s; The adjustment amount for the exhaust valve opening is %; S73. Using the Mamdani reasoning method, for each rule... r Calculate its activation strength: ； Then, the centroid method is used to resolve the fuzziness for each output variable: ； in, centroid(C r,T ) The output fuzzy set is the rule r. C r,T The center of gravity c water,r It is the output part of rule r. The centroid value of the specified fuzzy set, c θ,r For rule r The output fuzzy set centroid value; S74. Correct the reference trajectory using the correction amount output by fuzzy inference. y ref (t) The target value for a short period of time in the future is determined by the following method: (1) Modify the temperature setting trajectory, and change the current time and future time in the original reference trajectory. Np The temperature target value of each step is uniformly added And limit the amplitude to a safe range: ； (2) Water injection rate reference trajectory correction: Assume that in the original reference trajectory, the future... i The target value for the water injection rate of the step is The correction amount for the fuzzy inference output is The corrected target value for the water injection rate is: ； (3) Correction of exhaust valve opening reference trajectory: Assume that in the original reference trajectory, the future number of... i The target value for the exhaust opening of the step is The correction amount for the fuzzy inference output is The corrected target value for the water injection rate is: ； in, N p To predict the number of time-domain steps; S75, When fuzzy inference detects multiple consecutive times e S Approaching zero and e M and e I When the target is within a small range, it is considered that the current stage goal has been basically achieved, and the reference trajectory is actively switched to the next cooking stage.

8. The control method for the intelligent cooking machine as described in claim 1, characterized in that, In S8, the specific process of calling the DDPGActor network and outputting the weight correction ΔQ is as follows: S81, Status Acquisition: Obtained from EKF Calculated from the reference trajectory e S , e M , e I These values ​​are normalized to form a state vector. s ; S82. Invoke the Actor network and perform forward propagation: This yields the original output, ranging from -0.5 to 0.

5. S83, Update Weights: ； in, It can be obtained from cloud parameters or the value at the previous moment; S84, Pass to MPC: Send the new Q =diag( q T ,q M ,q I This is used in subsequent MPC optimization problems.

9. The control method for the intelligent cooking machine as described in claim 1, characterized in that, In step S9, the MPC optimization problem is solved to obtain the optimal control sequence. The specific process is as follows: S91. Linearized Prediction Model: Defining State Deviation Control deviation ,in x nom ,u nom Take a simple extrapolation of the previous round of MPC solution results or reference trajectory; after linearization, we get: ； in, ; S92. Construct the standard form of QP and define the decision variable vector: ； S93. Solve QP using a real-time iterative strategy. At the beginning of each sampling period, calculate the linearized model matrix using the current state estimate. A i ，B i The system assembles the QP matrix H,f,L,I, and then calls an efficient QP solver to obtain the optimal solution within milliseconds. ; S94, from Recovery Then calculate: , Will It can be sent to the underlying PID controller or directly converted into a PWM signal to drive the actuator; S95. At the next sampling time k+1, shift the entire control sequence forward and apply the newly calculated... As part of the new nominal control trajectory, it is used for hot starts.

10. A smart cooking machine, characterized in that: Including the control method as described in any one of claims 1-9.

Images