Adaptive thermal management control system based on knowledge graph

Abstract

The application discloses a self-adaptive heat management control system based on a knowledge graph, relates to the technical field of intelligent cooking control, and comprises a graph construction module, which creates a structured knowledge graph containing global main cooking system material thermal property nodes, typical cuisine process parameter nodes and various heating stove performance specification nodes, and the nodes are connected through attribute relationship edges; a knowledge extraction module extracts associated nodes and edges in response to a user recipe selection instruction; a temperature analysis module analyzes recommended core processing temperature, maximum safety threshold and texture change temperature range; a data acquisition module acquires real-time temperature and power sequence; and a reasoning control module constructs a Bayesian probability reasoning network based on data and analysis parameters, calculates posterior probability and outputs double-loop control decisions. Through multi-dimensional knowledge association and probabilistic reasoning, the system realizes self-adaptive precise heat management in the cooking process, improves safety and achieves consistency of material texture.

Description

Technical Field

[0001] This invention belongs to the field of intelligent cooking control technology, specifically an adaptive thermal management control system based on knowledge graphs. Background Technology

[0002] Traditional cooking thermal management control systems often rely on open-loop regulation using fixed recipe preset parameters or simple sensor feedback, failing to systematically integrate the correlation information between ingredient characteristics, cooking techniques, and cooktop performance. In existing solutions, the thermal properties of ingredients, culinary process parameters, and cooktop performance specifications are scattered and independent, making it impossible to dynamically match multi-dimensional factors based on the user's selected recipe. Control strategies often employ single threshold comparisons or PID regulation, lacking dynamic evaluation of the probabilistic cooking state. This leads to problems such as temperatures exceeding safe thresholds and food texture deviating from expectations when facing different ingredient combinations, cooktop differences, or process adjustments, making it difficult to balance safety and consistent taste.

[0003] The core problem this invention aims to solve is how to construct a structured culinary knowledge graph that connects nodes representing the thermal properties of ingredients from major global culinary systems, nodes representing the process parameters of typical cuisines, and nodes representing the performance specifications of various heating appliances with attribute relationship edges, forming a multi-dimensional association system covering ingredients, processes, and appliances, and automatically extracting all associated nodes and edges when a user selects a recipe; and how to construct a Bayesian probabilistic inference network to calculate posterior probabilities based on real-time temperature measurement sequences and power output sequences, combined with the analyzed recommended core processing temperature, maximum safety threshold, and temperature range of texture changes, and output dual-loop control decisions to achieve adaptive and precise thermal management of the cooking process. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] To this end, the present invention proposes an adaptive thermal management control system based on knowledge graphs, comprising: The knowledge graph construction module creates a structured culinary knowledge graph. The structured culinary knowledge graph includes nodes representing the thermal properties of common ingredients in major global culinary systems, nodes representing the process parameters of typical cuisines, and nodes representing the performance specifications of various heating appliances. The nodes are connected by preset attribute relationship edges. The knowledge extraction module responds to the user's input recipe selection command and automatically extracts all nodes and attribute relationship edges associated with the selected recipe from the structured cooking knowledge graph; The temperature analysis module parses the recommended core processing temperature, the maximum safe temperature threshold, and the minimum and maximum temperature ranges at which the texture of the ingredients are expected to change from all the nodes associated with the selected recipe. The data acquisition module acquires the real-time temperature measurement value sequence collected by the temperature sensing component arranged on the cooking appliance, and the real-time power output value sequence collected by the current sensing component arranged in the heating coil. The inference control module constructs a Bayesian probabilistic inference network based on the real-time temperature measurement value sequence, the real-time power output value sequence, the parsed recommended core processing temperature, the maximum allowable safe temperature threshold, and the minimum and maximum temperature ranges for expected changes in food texture. It calculates the posterior probability of reaching or maintaining the target cooking state at the current moment and makes dual-loop control decisions based on the output of the Bayesian probabilistic inference network.

[0006] Furthermore, the step of automatically extracting all nodes and attribute relationship edges associated with the selected recipe from the structured culinary knowledge graph includes: All nodes associated with the selected recipe include at least the thermal properties of the ingredients, the process parameters, and the performance specifications of the cooktop. Identify the recipe name indicated in the recipe selection instruction, perform a name matching search in the process parameter node set of the structured cooking knowledge graph, and locate the target recipe process node; Starting from the process node of the target recipe, traverse along the attribute relationship edges pointing to specific ingredient nodes to extract all the thermal attribute nodes of the ingredients necessary to complete the recipe. Starting from the target recipe process node, traverse along the attribute relationship edge pointing to the specific stove type to extract the stove performance specification node applicable to the recipe; Retrieve and extract attribute information from the structured cooking knowledge graph that connects all related edges between the target recipe process node, the thermal attribute nodes of all ingredients, and the stove performance specification node. The attribute information includes, but is not limited to, cooking order dependency and temperature conduction coefficient. By integrating the target recipe process nodes, all ingredient thermal attribute nodes, stove performance specification nodes, and attribute information of all associated edges, a cooking knowledge subgraph dedicated to the current recipe is constructed.

[0007] Furthermore, the construction of the Bayesian probabilistic inference network includes: Using the process parameter nodes in the cooking knowledge subgraph dedicated to the current recipe as prior knowledge, the structure of the Bayesian probabilistic inference network is defined. The structure includes latent variables representing the target cooking state and explicit variables representing sensor observation data. The real-time temperature measurement sequence and the real-time power output sequence are input into the Bayesian probabilistic inference network as observation evidence for the current time step. The conditional probability distribution table in the Bayesian probabilistic inference network is assigned values. The parameters of the conditional probability distribution table are partly derived from the specific values ​​corresponding to the thermal attribute nodes of the ingredients in the cooking knowledge subgraph dedicated to the current recipe, and partly derived from empirical parameters obtained from the statistical analysis of historical cooking process data. The probability distribution of all latent variables in the Bayesian probabilistic inference network is updated using a variational inference algorithm. The latent variables include the internal cooking state of the food, the actual temperature field uniformity at the bottom of the pot, and the real-time estimate of the heat transfer efficiency. The expected value of the internal cooking state variable of the updated ingredient is extracted as a quantitative estimate of the current cooking degree of the ingredient.

[0008] Furthermore, the dual-loop control decision based on the output of the Bayesian probabilistic inference network includes: The posterior probability of reaching or maintaining the target cooking state at the current moment is compared with a preset cooking progress control threshold. Based on the comparison results, it is determined that the current thermal management control should be dominated by either a knowledge-driven control loop or a real-time feedback control loop. If the knowledge-driven control loop is selected as the dominant force, then a basic power control command is generated based on the preset heating trajectory of the process parameter node corresponding to the current cooking stage in the structured cooking knowledge graph. If the real-time feedback control loop is selected as the dominant control loop, then based on the deviation between the real-time temperature measurement value sequence and the target temperature, and the changing trend of the real-time power output value sequence, a compensation power control command is generated through a proportional-integral-derivative algorithm. The basic power control command and the compensation power control command are combined to generate the final power modulation command that acts on the heating coil.

[0009] Furthermore, based on the comparison results, determining whether the current thermal management control should be in a dominant mode of knowledge-driven control loop or real-time feedback control loop includes: Preset a confidence threshold for the cooking state and a deviation threshold for the cooking state; When the posterior probability of reaching or maintaining the target cooking state at the current moment is higher than the cooking state confidence threshold, it is determined that the current actual cooking process is highly consistent with the knowledge graph prediction trajectory, and the knowledge-driven control loop dominant mode is activated to make the heating process strictly follow the preset heating trajectory extracted from the structured cooking knowledge graph. When the posterior probability of reaching or maintaining the target cooking state at the current moment is lower than the cooking state confidence threshold but higher than the cooking state deviation threshold, it is determined that there is an acceptable fluctuation in the current cooking process, and the hybrid control mode of the knowledge-driven control loop and the real-time feedback control loop is activated, with the basic power control command as the main one and the compensation power control command as the auxiliary one for fine-tuning. When the posterior probability of reaching or maintaining the target cooking state at the current moment is lower than the cooking state deviation threshold, it is determined that the current cooking process has significantly deviated from the expectation, and the real-time feedback control loop dominant mode is activated, which mainly relies on real-time feedback from sensors to generate the compensation power control command to quickly correct the deviation.

[0010] Furthermore, the process of generating the final power modulation command includes identifying the sticky pan stage and invoking the adaptive pulse heating strategy: The rate of change of the real-time temperature measurement value sequence and the food adhesion state probability output by the Bayesian probabilistic inference network are continuously monitored. The food adhesion state probability is a specific variable in the set of latent variables. When the rate of change of the real-time temperature measurement value sequence is lower than the sticking risk threshold and the probability of food sticking exceeds the sticking risk probability threshold, the cooking process is determined to have entered the sticking stage. During the stage of easy-stick pan use, the structured cooking knowledge graph is used to query the anti-stick pulse heating mode template suitable for the current combination of ingredients and cookware. The anti-stick pulse heating mode template defines the basic duty cycle range and adjustment frequency. Based on the difference between the real-time temperature measurement value sequence and the target temperature value, the specific value of the pulse duty cycle for the next control cycle is dynamically calculated within the basic duty cycle range; Based on the adjustment frequency and the calculated pulse duty cycle, a pulse power control command containing periodic on / off signals is generated, which serves as the main component of the final power modulation command in the easy-stick pan stage.

[0011] Furthermore, during the stage of easy-stick cooking, a non-stick pulse heating mode template suitable for the current combination of ingredients and cookware is queried from the structured cooking knowledge graph, including: Identify all the thermal attribute nodes of the ingredients associated with the current cooking process, and extract attribute values ​​that characterize the surface characteristics of each ingredient from all the thermal attribute nodes of the ingredients. The attribute values ​​that characterize the surface characteristics of the ingredients include, but are not limited to, starch content, protein type, and surface roughness level. Identify the cooktop performance specification node associated with the current cooking process, and extract the attribute values ​​characterizing the cookware base material from the cooktop performance specification node; Using the set of attribute values ​​representing the surface characteristics of the ingredients and the attribute values ​​representing the base material of the cookware as joint query conditions, a traversal search is performed in the structured cooking knowledge graph; Locate the anti-stick pulse heating mode template node in the structured cooking knowledge graph whose matching degree with the joint query conditions exceeds a preset threshold; The basic duty cycle range and the adjustment frequency are read from the anti-sticking pulse heating mode template node. The basic duty cycle range includes a lower limit value of the pulse duty cycle and an upper limit value of the pulse duty cycle. The adjustment frequency defines the upper limit and lower limit of the frequency of the periodic on / off signal in the pulse power control command.

[0012] Further, based on the difference between the real-time temperature measurement value sequence and the target temperature value, the specific value of the pulse duty cycle for the next control cycle is dynamically calculated within the basic duty cycle range, including: Read the lower limit and upper limit of the pulse duty cycle defined in the anti-stick pulse heating mode template; Calculate the absolute value of the temperature deviation between the average value of multiple sampling points of the temperature sensor and the target temperature value for the current stage obtained from the structured cooking knowledge graph within the current control cycle; The absolute value of the temperature deviation is input into a preset fuzzy inference engine, and the output of the fuzzy inference engine is a duty cycle adjustment coefficient, which is a real number between zero and one. The difference between the lower limit of the pulse duty cycle and the upper limit of the pulse duty cycle is multiplied by the duty cycle adjustment coefficient to obtain a duty cycle increment; The pulse duty cycle increment is added to the lower limit of the pulse duty cycle to obtain the specific value of the pulse duty cycle for the next control cycle.

[0013] Furthermore, the method also includes a graph update module for adaptively updating the structured cooking knowledge graph during the cooking process, specifically including: After a complete cooking process is completed, collect the sensor time-series data, control command time-series data, and the user's final evaluation input of the cooking result throughout the entire cooking process; By performing correlation analysis on the sensor time series data and control command time series data, the key temperature inflection point, the effective power range actually used, and the total cooking time when the best cooking effect is actually achieved are identified. The identified key temperature inflection points, the actual effective power range used, and the total cooking time are compared with the pre-stored data in the cooking knowledge sub-graph dedicated to the current recipe used in this cooking, and the degree of difference is calculated. If the difference exceeds the preset update trigger threshold, the valid data identified during this cooking process will be added as a new experience record to the corresponding node of the structured cooking knowledge graph, or used to correct the parameter values ​​of the original node.

[0014] Furthermore, the user's final evaluation of the cooking outcome serves as input to guide the direction of knowledge graph updates, including: The user's evaluation input is parsed into a structured score, which includes at least a graded score for the doneness, charring degree, and texture of the dish. The structured score is correlated with the final state probability distribution output by the Bayesian probabilistic inference network at the moment cooking ends; If the structured score is lower than the expected score threshold in a specific dimension, the state probability change curve of the corresponding dimension during the cooking process is traced back to locate the time point of control decision error and related control parameters. Based on the located time point and related control parameters, combined with the structured score, parameter correction suggestions are generated for specific process parameter nodes in the structured cooking knowledge graph. The correction suggestions include adjusting the recommended temperature, modifying the heating time, or marking sensitive process stages. The proposed corrections will be applied to subsequent versions of the structured cooking knowledge graph after being confirmed manually or by pre-defined review rules in offline mode.

[0015] Compared with the prior art, the beneficial effects of the present invention are: A structured culinary knowledge graph is constructed, comprising nodes representing the thermal properties of common ingredients in major global culinary systems, nodes representing the processing parameters of typical cuisines, and nodes representing the performance specifications of various heating appliances. These nodes are connected by preset attribute relationship edges. This technology enables the system to respond to user recipe selection commands by automatically extracting all nodes and edges associated with the selected recipe. This provides a multi-dimensional data foundation covering ingredient characteristics, processing requirements, and appliance capabilities for subsequent temperature analysis, making dynamic matching of different ingredient, processing, and appliance combinations possible and avoiding the control inaccuracies caused by isolated information in traditional solutions.

[0016] Using real-time temperature and power output sequences as inputs, and combining the analyzed recommended core processing temperature, the maximum permissible safe temperature threshold, and the temperature range for changes in food texture, a Bayesian probabilistic inference network is constructed to calculate the posterior probability of reaching or maintaining the target cooking state at the current moment, and outputs a dual-loop control decision. This technology reflects the matching degree between the current state and the target state through probabilistic evaluation. The inner loop adjusts the heating power to track the temperature, while the outer loop corrects the target based on the probability deviation, achieving dynamic adaptive control of the cooking process. Compared with traditional threshold or PID control, it is better able to cope with state fluctuations in complex scenarios, improving the safety of thermal management and the consistency of food texture. Attached Figure Description

[0017] Figure 1 This is a timing diagram of the knowledge graph-based adaptive thermal management control system described in this invention. Figure 2 A flowchart for automatically extracting recipe association information from a structured culinary knowledge graph; Figure 3 A flowchart for constructing a Bayesian probabilistic inference network; Figure 4 The adaptive thermal management control diagram is for the sticky pan stage; Figure 5 This is the closed-loop response diagram for the pulse anti-sticking control stage. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 The knowledge graph construction module creates a structured culinary knowledge graph. This graph includes nodes representing the thermal properties of common ingredients in major global culinary systems, process parameters of typical cuisines, and performance specifications of various heating appliances. These nodes are connected by pre-defined attribute relationship edges. The knowledge extraction module responds to user-inputted recipe selection commands by automatically extracting all nodes and attribute relationship edges associated with the selected recipe from the structured culinary knowledge graph. The temperature analysis module parses the recommended core processing temperature, the maximum permissible safe temperature threshold, and the minimum and maximum temperature ranges for expected changes in the ingredient's texture from all nodes associated with the selected recipe. The data acquisition module acquires real-time temperature measurement sequences from temperature sensors mounted on the cooking appliance and real-time power output sequences from current sensors mounted in the heating coil. The inference control module constructs a Bayesian probabilistic inference network based on the real-time temperature measurement value sequence, the real-time power output value sequence, the parsed recommended core processing temperature, the maximum allowable safe temperature threshold, and the minimum and maximum temperature ranges for expected changes in food texture. It calculates the posterior probability of reaching or maintaining the target cooking state at the current moment and makes dual-loop control decisions based on the output of the Bayesian probabilistic inference network.

[0020] See Figure 2In one embodiment of the present invention, the process of automatically extracting all nodes and attribute relationship edges associated with a selected recipe from a structured culinary knowledge graph involves querying and traversing the structured culinary knowledge graph. The structured culinary knowledge graph includes nodes representing the thermal attributes of common ingredients in major global culinary systems, nodes representing the process parameters of typical cuisines, and nodes representing the performance specifications of various heating appliances. These nodes are connected by preset attribute relationship edges. The knowledge extraction module responds to a user-inputted recipe selection command, which provides the recipe name in the form of a text string. All nodes associated with the selected recipe include at least nodes representing the thermal attributes of ingredients, process parameters, and appliance performance specifications. The system executes the step of recognizing the recipe name indicated in the recipe selection command, and performs a name matching search in the set of process parameter nodes in the structured culinary knowledge graph. The name matching search locates the target recipe process node by comparing the similarity between the user-input string and the process parameter node name string. In specific implementations, the name matching search uses a similarity calculation formula to quantify the degree of matching. The formula is defined as:

[0021] in: This represents the string of recipe names entered by the user. This represents the name string of the process parameter node in the structured culinary knowledge graph. This represents the Jaccard similarity coefficient based on the word segmentation set. This represents the similarity value based on normalized edit distance. and They are predefined weighting coefficients and satisfy... The system will use similarity The calculation results are compared with a preset threshold. When the similarity is higher than the threshold, the corresponding process parameter node is located as the target recipe process node.

[0022] In some embodiments, starting from the target recipe process node, the system traverses along the attribute relationship edges pointing to specific ingredient nodes to extract all the ingredient thermal attribute nodes necessary to complete the recipe. The traversal process adopts a depth-first or breadth-first graph traversal algorithm. The attribute relationship edges define the "use" or "inclusion" relationship between the process node and the ingredient node. During the traversal process, the system accesses all ingredient thermal attribute nodes that are directly or indirectly connected to the target recipe process node through attribute relationship edges. Ingredient thermal attribute nodes store parameters such as specific heat capacity, thermal conductivity, and phase change temperature. In a specific implementation, starting from the target recipe process node, the system traverses along the attribute relationship edges pointing to specific stove types to extract stove performance specification nodes suitable for the recipe. Stove performance specification nodes include heating power range, thermal efficiency, and cookware material attributes. The traversal process similarly follows the edge direction in the graph. The attribute relationship edges define the "applicability" relationship between the process node and the stove node.

[0023] Optionally, the system retrieves and extracts attribute information from the structured cooking knowledge graph, which includes all associated edges connecting the target recipe process node, all ingredient thermal attribute nodes, and stove performance specification nodes. The attribute information includes cooking order dependencies and temperature conduction coefficients. The cooking order dependencies are represented by directed edges, indicating the sequential order of ingredient processing steps. The temperature conduction coefficients are attached as numerical attributes to the edges between the stove node and the ingredient node, representing the efficiency factor of heat transfer. During the extraction process, the system accesses the attribute fields of these associated edges and loads the attribute field values ​​into the memory data structure.

[0024] It is understood that the system integrates the target recipe's process nodes, all ingredient thermal attribute nodes, stove performance specification nodes, and attribute information of all associated edges to form a dedicated cooking knowledge subgraph for the current recipe. This cooking knowledge subgraph is represented in graph form, where nodes and edges carry a complete set of attributes extracted from the structured cooking knowledge graph. The cooking knowledge subgraph serves as an independent data structure for input to subsequent modules. In some embodiments, the integration process involves serializing the extracted node and edge information into a specific data format to ensure data consistency and fast access. In specific implementations, in the similarity calculation for name matching search, weighting coefficients are used... and Adjustments can be made based on the linguistic characteristics of the recipe name. For example, for Chinese recipe names, the similarity weight based on edit distance can be increased to handle synonyms or abbreviation variations. This adjustment is based on a preset language rule table and requires no user intervention.

[0025] See Figure 3In one embodiment of the present invention, the process of constructing the Bayesian probabilistic inference network uses the process parameter nodes in the cooking knowledge subgraph dedicated to the current recipe as prior knowledge. The cooking knowledge subgraph dedicated to the current recipe includes process parameters, thermal properties of ingredients, and stove performance information extracted from the structured cooking knowledge graph. The system defines the structure of the Bayesian probabilistic inference network, which is a directed acyclic graph containing latent variables representing the target cooking state and explicit variables representing sensor observation data. The latent variables include the internal cooking state of the ingredients, the actual temperature field uniformity at the bottom of the pot, and the real-time estimate of the heat transfer efficiency. The explicit variables include the temperature sensor reading sequence and the power sensor reading sequence. In practice, the real-time temperature measurement sequence and the real-time power output sequence are input into the Bayesian probabilistic inference network. The real-time temperature measurement sequence is an array of readings collected by multiple thermocouples arranged at the bottom of the pot within a fixed sampling period. The real-time power output sequence is an array of effective current values ​​flowing through the heating coil collected by the current sensing component. These two sequences serve as observation evidence for the current time step and are mapped to the observation values ​​of the corresponding manifest variables in the Bayesian probabilistic inference network.

[0026] In some embodiments, values ​​are assigned to the conditional probability distribution table in the Bayesian probabilistic inference network. The conditional probability distribution table defines the conditional probability of each node in the Bayesian probabilistic inference network under the state of its parent node. The assignment process consists of two parts: one part of the parameters comes from the specific values ​​corresponding to the thermal attribute nodes of the ingredients in the cooking knowledge subgraph specific to the current recipe. For example, the phase transition temperature range of an ingredient is converted into the conditional probability parameter of the internal cooking state variable of the ingredient. The other part of the parameters comes from empirical parameters derived from historical cooking process data statistics. For example, the historical distribution of the uniformity of the cookware temperature field under a specific power is used to initialize the conditional probability distribution table of the corresponding latent variable. In specific implementations, the fusion of the two parts of parameters is achieved through a preset weighting formula, defined as:

[0027] in: This represents the parameter vector that is ultimately assigned values ​​in the conditional probability distribution table. This represents the parameter vector obtained from parsing the cooking knowledge subgraph specific to the current recipe. This represents a vector of empirical parameters derived from statistical analysis of historical cooking process data. It is a fusion weight coefficient between 0 and 1, whose value depends on the confidence level of the current recipe in the knowledge graph, and is the parameter vector after fusion. The conditional probability distribution table is directly loaded into the Bayesian probabilistic inference network.

[0028] It can be understood that the variational inference algorithm is used to update the probability distribution of all latent variables in the Bayesian probabilistic inference network. The variational inference algorithm approximates the true posterior distribution by iteratively optimizing an approximate posterior distribution. In each control cycle, the algorithm updates the probability distribution of the latent variable set based on the current observation evidence (real-time temperature measurement sequence and real-time power output sequence). The updated latent variable probability distribution includes the probability distribution of the internal cooking state variable of the food, the probability distribution of the actual temperature field uniformity variable at the bottom of the pot, and the probability distribution of the real-time estimate of the heat transfer efficiency variable. In some embodiments, the variational inference algorithm uses a mean-field approximation, assumes that the posterior distributions of all latent variables are independent, and iteratively optimizes the variational parameters of each latent variable using the coordinate ascent method until the variational lower bound converges. Optionally, the expected value of the internal cooking state variable of the updated ingredients can be extracted as a quantitative estimate of the current cooking degree of the ingredients. The internal cooking state variable of the ingredients may be defined as a continuous value from 0 to 1, representing the degree from completely raw to fully cooked. Its probability distribution is a parameterized distribution, and the expected value is the mean of the distribution. This mean is output as a scalar, representing the point estimate of the cooking degree of the ingredients at the current moment.

[0029] In practical implementation, the empirical parameters derived from historical cooking process data come from an independent empirical database. This database stores the correspondence between sensor data and final state labels from past successful cooking cases. The system learns the parameters of the conditional probability distribution table from this database using machine learning methods. This parameter learning process is performed offline during system idle periods. Optionally, when defining the structure of the Bayesian probabilistic inference network, the network topology is pre-designed based on a general physical model of the cooking process. However, the connection strength between nodes is dynamically adjusted by a cooking knowledge subgraph specific to the current recipe. For example, the "high heat stir-fry" technique indicated in the knowledge subgraph will enhance the connection weight between the power observation variable and the latent variable of heat transfer efficiency. It can be understood that the update frequency of the variational inference algorithm is consistent with the control decision cycle. Each control cycle executes a complete variational inference iteration to provide real-time posterior probability outputs for subsequent dual-loop control decisions.

[0030] In one embodiment of the present invention, the dual-loop control decision based on the output of the Bayesian probabilistic inference network involves a series of coherent steps. The system compares the posterior probability of reaching or maintaining the target cooking state at the current moment with a preset cooking progress control threshold. The posterior probability of reaching or maintaining the target cooking state at the current moment is calculated and output by the Bayesian probabilistic inference network in each control cycle. The preset cooking progress control threshold includes a cooking state confidence threshold and a cooking state deviation threshold. These two thresholds are scalar values ​​preset by the system according to the complexity of the recipe process. In specific implementation, the system determines whether the current thermal management control should be in the dominant mode of the knowledge-driven control loop or the real-time feedback control loop based on the comparison result. The judgment logic involves branching the numerical relationship between the posterior probability and the above two thresholds. In specific implementation, when the posterior probability of reaching or maintaining the target cooking state at the current moment is higher than the cooking state confidence threshold, the system determines that the current actual cooking process is highly consistent with the knowledge graph prediction trajectory and enables the knowledge-driven control loop dominant mode, so that the heating process strictly follows the preset heating trajectory extracted from the structured cooking knowledge graph. The preset heating trajectory is stored in the process parameter node in the form of a time-temperature-power curve.

[0031] In some embodiments, when the posterior probability of reaching or maintaining the target cooking state at the current moment is lower than the cooking state confidence threshold but higher than the cooking state deviation threshold, the system determines that the current cooking process has acceptable fluctuations and enables a hybrid control mode of knowledge-driven control loop and real-time feedback control loop. In the hybrid control mode, basic power control commands are the primary method, supplemented by compensation power control commands for fine-tuning. The basic power control commands originate from the output of the knowledge-driven control loop. In some embodiments, when the posterior probability of reaching or maintaining the target cooking state at the current moment is lower than the cooking state deviation threshold, the system determines that the current cooking process has significantly deviated from expectations and enables a real-time feedback control loop-dominated mode. In the real-time feedback control loop-dominated mode, the system mainly relies on real-time feedback from sensors to generate compensation power control commands to quickly correct deviations. In specific implementations, if the system selects the knowledge-driven control loop-dominated mode, it generates basic power control commands based on the preset heating trajectory of the process parameter nodes corresponding to the current cooking stage in the structured cooking knowledge graph. The preset heating trajectory includes the target power setting value of the heating coil at a specific cooking stage.

[0032] Optionally, if the system selects a real-time feedback control loop as the dominant control, then based on the deviation between the real-time temperature measurement sequence and the target temperature, as well as the changing trend of the real-time power output sequence, a proportional-integral-derivative (PID) algorithm is used to generate a compensating power control command. The input of the PID algorithm is the difference between the target temperature and the moving average of the real-time temperature measurement sequence, and the output of the PID algorithm is a power adjustment. The changing trend of the real-time power output sequence is used to feedforward compensate the output of the PID algorithm. The process of the PID algorithm outputting the compensating power control command is described by a discrete-time formula, defined as:

[0033] in: Indicates the first The compensated power control command value is calculated in each control cycle. Indicates the first The deviation between the target temperature measured in each control cycle and the real-time average temperature , , These are the control parameters for the proportional, integral, and derivative terms, respectively. It is the system's control cycle. Indicates from the start of control to the number The sum of all deviations over a period of time, and the parameters of the proportional-integral-differential algorithm. , , Adaptive tuning is performed based on the stove performance specification node and the ingredient thermal property node in the current recipe-specific cooking knowledge subgraph.

[0034] It can be understood that the system integrates the basic power control command and the compensated power control command to generate the final power modulation command acting on the heating coil. In the knowledge-driven control loop-dominated mode, the basic power control command is the main part of the final power modulation command, while the compensated power control command is limited to a very small range or is zero. In the real-time feedback control loop-dominated mode, the final power modulation command is entirely composed of the compensated power control command, and the basic power control command is ignored. In the hybrid control mode, the final power modulation command is a weighted sum of the basic power control command and the compensated power control command, and the weighting coefficients are dynamically calculated based on the specific position where the posterior probability falls between two thresholds. In specific implementation, the fusion operation is completed in an adder module. The adder module receives the basic power control command from the knowledge-driven control loop and the compensated power control command from the real-time feedback control loop, and outputs a single power value according to the fusion strategy determined by the current control mode. This power value is converted into the actual driving signal for the heating coil by the power modulation circuit. Optionally, the specific values ​​of the cooking state confidence threshold and the cooking state deviation threshold are not fixed. The system will fine-tune them based on the "cooking tolerance" attribute marked on the process parameter node in the cooking knowledge subgraph dedicated to the current recipe. For recipes with low tolerance, the cooking state confidence threshold will be set higher, and the cooking state deviation threshold will also be increased accordingly to ensure more sensitive control mode switching. It can be understood that in the hybrid control mode, the mixing ratio of the basic power control command and the compensation power control command is determined by a linear function. The input of this linear function is the posterior probability of reaching or maintaining the target cooking state at the current moment. The closer the posterior probability is to the cooking state confidence threshold, the higher the weight of the basic power control command and the lower the weight of the compensation power control command.

[0035] In one embodiment of the present invention, the final power modulation command generation process includes the identification of the easy-sticking stage and the invocation of an adaptive pulse heating strategy. The system continuously monitors the rate of change of the real-time temperature measurement sequence and the food adhesion probability output by the Bayesian probabilistic inference network. The rate of change of the real-time temperature measurement sequence is obtained by calculating the derivative of the temperature readings within the current sampling period. The food adhesion probability is a specific variable in the set of latent variables, and its probability distribution is output by the Bayesian probabilistic inference network after each inference. The system takes the mathematical expectation of this distribution as the monitoring value. When the rate of change of the real-time temperature measurement sequence is lower than the sticking risk judgment threshold, and the food adhesion probability exceeds the sticking risk probability threshold, the system determines that the cooking process has entered the easy-sticking stage. In the easy-sticking stage, the system queries the structured cooking knowledge graph for an anti-stick pulse heating mode template suitable for the current combination of ingredients and cookware. The anti-stick pulse heating mode template defines the basic duty cycle range and adjustment frequency. The query process includes identifying all thermal attribute nodes of ingredients associated with the current cooking process, extracting attribute values ​​representing the surface characteristics of each ingredient from all thermal attribute nodes, including starch content, protein type, and surface roughness level. The system also identifies the cookware performance specification nodes associated with the current cooking process, extracting attribute values ​​representing the base material of the cookware from the cookware performance specification nodes, and using the set of attribute values ​​representing the surface characteristics of ingredients and the attribute values ​​representing the base material of the cookware as joint query conditions to traverse and search in the structured cooking knowledge graph.

[0036] In practice, the traversal search is performed by comparing the similarity between the joint query conditions and the attribute nodes linked to the anti-stick pulse heating mode template nodes in the knowledge graph. This locates the anti-stick pulse heating mode template nodes in the structured cooking knowledge graph whose matching degree with the joint query conditions exceeds a preset threshold. From these nodes, the basic duty cycle range and adjustment frequency are read. The basic duty cycle range includes a lower limit and an upper limit for the pulse duty cycle. The adjustment frequency defines the upper and lower limits of the periodic on / off signal in the pulse power control command. Table 1 shows a simplified query matching table illustrating how different joint query conditions correspond to different anti-stick pulse heating mode template parameters.

[0037] Table 1: Anti-stick Pulse Heating Mode Template Query and Matching Table

[0038] It can be understood that the system dynamically calculates the specific pulse duty cycle value for the next control cycle within a basic duty cycle range based on the difference between the real-time temperature measurement value sequence and the target temperature value. The calculation process includes reading the lower and upper limits of the pulse duty cycle defined in the anti-stick pulse heating mode template, and calculating the absolute value of the temperature deviation between the average reading of multiple sampling points of the temperature sensor within the current control cycle and the target temperature value obtained from the structured cooking knowledge graph for the current stage. In specific implementations, the absolute value of the temperature deviation is input into a preset fuzzy inference engine. The output of the fuzzy inference engine is the duty cycle adjustment coefficient, which is a real number between zero and one. The fuzzy inference engine internally defines a fuzzy set of temperature deviations and its membership function, and defines a rule-based inference mechanism to map the clear absolute value of the temperature deviation to a clear duty cycle adjustment coefficient. In some embodiments, the difference between the lower and upper limits of the pulse duty cycle is multiplied by the duty cycle adjustment coefficient to obtain a duty cycle increment. The lower limit of the pulse duty cycle is then added to the duty cycle increment to obtain the specific pulse duty cycle value for the next control cycle. This calculation relationship is defined by the following formula:

[0039] in: This indicates the calculated pulse duty cycle value for the next control cycle. This indicates the lower limit of the pulse duty cycle read from the anti-sticking pulse heating mode template. This indicates the upper limit of the pulse duty cycle read from the anti-stick pulse heating mode template. This represents the duty cycle adjustment coefficient output by the fuzzy inference engine. Optionally, the rule base of the fuzzy inference engine can be configured according to the characteristics of different cuisines. For example, for Chinese stir-frying, the rules tend to give a higher duty cycle adjustment coefficient when the temperature deviation is moderate to achieve rapid heating. The system generates a pulse power control command containing periodic on / off signals according to the adjustment frequency and the calculated pulse duty cycle value. The specific value of the adjustment frequency is within the upper and lower frequency limits defined by the anti-stick pulse heating mode template, and is dynamically selected according to the current cooking stage. Usually, a higher frequency is used in the initial stage as the main component of the final power modulation command in the easy-stick stage. In some embodiments, the pulse power control command is output to the driving circuit of the heating coil in the form of a pulse width modulation signal. Optionally, the duty cycle adjustment coefficient... The value depends not only on the current temperature deviation but may also be related to the historical trend of the rate of change of the real-time temperature measurement sequence. If the temperature continues to drop rapidly, the fuzzy inference engine will output a larger duty cycle adjustment coefficient to enhance the heating pulse. It is understandable that the sticking risk judgment threshold and the sticking risk probability threshold are not fixed. The system will read the preset risk threshold value corresponding to the current cooking step from the relevant process parameter nodes in the structured cooking knowledge graph.

[0040] See Figure 4 This is an adaptive thermal management control chart for the easy-sticking stage, showing the changes in two key indicators during cooking. The probability of food adhesion increases linearly with the control cycle, gradually rising from 0.40 to 0.85, indicating a continuously increasing risk of food adhesion to the cookware. When this probability exceeds a preset threshold (e.g., 0.8), the system determines that it has entered the easy-sticking stage. The pulse duty cycle initially rises rapidly, reaching a peak of 0.40 at 30 seconds, and then gradually decreases to 0.20. This change corresponds to the system's adaptive pulse heating strategy: when the adhesion risk is highest, the heating intensity is enhanced by increasing the duty cycle, and then dynamically adjusted according to the temperature deviation to achieve a balance between anti-sticking and temperature control. The dynamic adjustment of the pulse duty cycle demonstrates the system's ability to extract anti-sticking templates from a knowledge graph and perform adaptive control in conjunction with real-time data.

[0041] In one embodiment of the present invention, the knowledge graph update module is used to adaptively update the structured cooking knowledge graph during the cooking process. After a complete cooking process, the system collects sensor time-series data, control command time-series data, and the user's final evaluation input of the cooking result throughout the entire cooking process. The sensor time-series data includes temperature measurement value sequences and power output value sequences, and the control command time-series data records the final power modulation command issued in each control cycle. In a specific implementation, correlation analysis is performed on the sensor time-series data and the control command time-series data. The correlation analysis aligns data points of different sequences using timestamps to identify the key temperature inflection point when the optimal cooking effect is achieved, the effective power range actually used, and the total cooking time. The key temperature inflection point is located by analyzing the zero-crossing point of the first derivative of the temperature measurement value sequence and the synchronous changes of the control commands. The effective power range actually used is determined by statistically analyzing the segments in the power output value sequence where the duration exceeds a threshold and the power value is stable. The total cooking time is recorded from the cooking start signal to the cooking end signal.

[0042] In some embodiments, the identified key temperature inflection points, the actual effective power range used, and the total cooking time are compared with pre-stored data in the cooking knowledge subgraph specific to the current recipe used in this cooking. The pre-stored data in the cooking knowledge subgraph specific to the current recipe comes from the initial values ​​of the process parameter nodes and ingredient thermal property nodes in the knowledge graph. The difference is calculated, and the difference Δ is calculated using a comprehensive formula, defined as:

[0043] in: Indicates the total degree of difference. The difference component representing the time location of key temperature inflection points is calculated by comparing the average absolute error between the actual inflection point time and the preset time point in the knowledge subgraph. This represents the difference in overlap between the actual effective power range and the preset power range of the knowledge subgraph, calculated by comparing the differences in the median and width of the ranges. The difference component representing the total cooking time is the relative error between the actual cooking time and the preset cooking time. , , These are preset weighting coefficients, used to reflect the importance of different parameters in the difference assessment. If the calculated total difference... If the preset update trigger threshold is exceeded, the system will add the valid data identified during this cooking process as a new experience record to the corresponding node of the structured cooking knowledge graph, or use it to correct the parameter values ​​of the original node. When adding a new experience record, the system will add a timestamp and the environmental context label of this cooking process.

[0044] It is understandable that the user's final evaluation of the cooking result is used to guide the direction of knowledge graph updates. This includes parsing the user's evaluation input into a structured score. The user's evaluation input can be submitted through the graphical interface of the terminal device in the form of slider rating, star rating, or text feedback. The step of parsing into a structured score includes at least a graded score for the doneness, charring degree, and texture of the dish, with each dimension quantified into a numerical score. In some embodiments, the structured score is correlated with the final state probability distribution output by the Bayesian probabilistic inference network at the moment of cooking completion. The correlation operation is achieved by calculating the correlation between the structured score vector and the expected value vector of the final state probability distribution output by the Bayesian probabilistic inference network. Optionally, if the structured score is lower than the expected score threshold in a certain dimension, the system backtracks the state probability change curve of the corresponding dimension during the cooking process to locate the time point of control decision error and related control parameters. For example, if the doneness score is too low, the system backtracks the probability curve of "internal cooking state of ingredients" output by the Bayesian probabilistic inference network to locate the time point when the growth of the probability curve stagnates or abnormally decreases, and extracts the control commands and sensor readings before and after that time point. Based on the located time points and relevant control parameters, combined with structured scoring, the system generates parameter correction suggestions for specific process parameter nodes in the structured cooking knowledge graph. The correction suggestions include adjusting the recommended temperature, modifying the heating time, or marking sensitive process stages.

[0045] In practice, the correction suggestions exist in the form of structured data patches, including target node identifiers, parameters to be corrected, suggested new values, and the basis for correction. It is understood that the correction suggestions, after being confirmed manually or by preset review rules in offline mode, are applied to subsequent versions of the structured cooking knowledge graph. Preset review rules may include requiring the same correction suggestion to be verified by multiple cooking examples, or requiring the correction magnitude to not exceed safety limits. Optionally, the expected scoring threshold is not a fixed value; the system will dynamically adjust it based on the difficulty level of the current recipe and the user's historical scoring habits. For highly difficult recipes, the expected scoring threshold will be appropriately lowered. In some embodiments, when generating parameter correction suggestions, the system will refer to similar data patterns from historical successful cases, providing reference values ​​for the correction suggestions by searching for control parameters corresponding to historically high-scoring records in the knowledge graph.

[0046] See Figure 5 This is a closed-loop response diagram of the pulse anti-stick control phase, clearly showing the dynamic response process during cooking. From 0 to 20 minutes, the risk of sticking remains at an extremely low level (approximately 0.1), the pulse duty cycle is 0, and the system is in normal heating mode. From 20 to 28 minutes, the risk of sticking surges to a high-risk range of 0.7–0.9, triggering the pulse anti-stick strategy, and the pulse duty cycle enters a high-frequency adjustment state of 0.5–0.7. After 28 minutes, the risk of sticking returns to a safe level, the pulse duty cycle returns to zero, and the system exits anti-stick mode. The risk of sticking during the anti-stick phase exhibits high-frequency, small-amplitude fluctuations, reflecting the real-time changes in the adhesion state between food and cookware. The pulse duty cycle is highly synchronized with the risk of sticking, achieving precise control of cookware temperature through rapid pulse on / off switching, thereby inhibiting food adhesion. The rapid response of the pulse duty cycle effectively controls the risk of sticking within an acceptable range, preventing food from burning or sticking to the pan.

[0047] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A knowledge graph-based adaptive thermal management control system, characterized in that, include: The knowledge graph construction module creates a structured culinary knowledge graph. The structured culinary knowledge graph includes nodes representing the thermal properties of common ingredients in major global culinary systems, nodes representing the process parameters of typical cuisines, and nodes representing the performance specifications of various heating appliances. The nodes are connected by preset attribute relationship edges. The knowledge extraction module responds to the user's input recipe selection command and automatically extracts all nodes and attribute relationship edges associated with the selected recipe from the structured cooking knowledge graph; The temperature analysis module parses the recommended core processing temperature, the maximum safe temperature threshold, and the minimum and maximum temperature ranges at which the texture of the ingredients are expected to change from all the nodes associated with the selected recipe. The data acquisition module acquires the real-time temperature measurement value sequence collected by the temperature sensing component arranged on the cooking appliance, and the real-time power output value sequence collected by the current sensing component arranged in the heating coil. The inference control module constructs a Bayesian probabilistic inference network based on the real-time temperature measurement value sequence, the real-time power output value sequence, the parsed recommended core processing temperature, the maximum allowable safe temperature threshold, and the minimum and maximum temperature ranges for expected changes in food texture. It calculates the posterior probability of reaching or maintaining the target cooking state at the current moment and makes dual-loop control decisions based on the output of the Bayesian probabilistic inference network.

2. The knowledge graph-based adaptive thermal management control system according to claim 1, characterized in that, The automatic extraction of all nodes and attribute relationship edges associated with the selected recipe from the structured culinary knowledge graph includes: All nodes associated with the selected recipe include at least the thermal properties of the ingredients, the process parameters, and the performance specifications of the cooktop. Identify the recipe name indicated in the recipe selection instruction, perform a name matching search in the process parameter node set of the structured cooking knowledge graph, and locate the target recipe process node; Starting from the process node of the target recipe, traverse along the attribute relationship edges pointing to specific ingredient nodes to extract all the thermal attribute nodes of the ingredients necessary to complete the recipe. Starting from the target recipe process node, traverse along the attribute relationship edge pointing to the specific stove type to extract the stove performance specification node applicable to the recipe; Retrieve and extract attribute information from the structured cooking knowledge graph that connects all related edges between the target recipe process node, the thermal attribute nodes of all ingredients, and the stove performance specification node. The attribute information includes, but is not limited to, cooking order dependency and temperature conduction coefficient. By integrating the target recipe process nodes, all ingredient thermal attribute nodes, stove performance specification nodes, and attribute information of all associated edges, a cooking knowledge subgraph dedicated to the current recipe is constructed.

3. The knowledge graph-based adaptive thermal management control system according to claim 2, characterized in that, The construction of the Bayesian probabilistic inference network includes: Using the process parameter nodes in the cooking knowledge subgraph dedicated to the current recipe as prior knowledge, the structure of the Bayesian probabilistic inference network is defined. The structure includes latent variables representing the target cooking state and explicit variables representing sensor observation data. The real-time temperature measurement sequence and the real-time power output sequence are input into the Bayesian probabilistic inference network as observation evidence for the current time step. The conditional probability distribution table in the Bayesian probabilistic inference network is assigned values. The parameters of the conditional probability distribution table are partly derived from the specific values ​​corresponding to the thermal attribute nodes of the ingredients in the cooking knowledge subgraph dedicated to the current recipe, and partly derived from empirical parameters obtained from the statistical analysis of historical cooking process data. The probability distribution of all latent variables in the Bayesian probabilistic inference network is updated using a variational inference algorithm. The latent variables include the internal cooking state of the food, the actual temperature field uniformity at the bottom of the pot, and the real-time estimate of the heat transfer efficiency. The expected value of the internal cooking state variable of the updated ingredient is extracted as a quantitative estimate of the current cooking degree of the ingredient.

4. The knowledge graph-based adaptive thermal management control system according to claim 3, characterized in that, The dual-loop control decision based on the output of the Bayesian probabilistic inference network includes: The posterior probability of reaching or maintaining the target cooking state at the current moment is compared with a preset cooking progress control threshold. Based on the comparison results, it is determined that the current thermal management control should be dominated by either a knowledge-driven control loop or a real-time feedback control loop. If the knowledge-driven control loop is selected as the dominant force, then a basic power control command is generated based on the preset heating trajectory of the process parameter node corresponding to the current cooking stage in the structured cooking knowledge graph. If the real-time feedback control loop is selected as the dominant control loop, then based on the deviation between the real-time temperature measurement value sequence and the target temperature, and the changing trend of the real-time power output value sequence, a compensation power control command is generated through a proportional-integral-derivative algorithm. The basic power control command and the compensation power control command are combined to generate the final power modulation command that acts on the heating coil.

5. The knowledge graph-based adaptive thermal management control system according to claim 4, characterized in that, Based on the comparison results, it is determined whether the current thermal management control should be in a dominant mode of knowledge-driven control loop or real-time feedback control loop, including: Preset a confidence threshold for the cooking state and a deviation threshold for the cooking state; When the posterior probability of reaching or maintaining the target cooking state at the current moment is higher than the cooking state confidence threshold, it is determined that the current actual cooking process is highly consistent with the knowledge graph prediction trajectory, and the knowledge-driven control loop dominant mode is activated to make the heating process strictly follow the preset heating trajectory extracted from the structured cooking knowledge graph. When the posterior probability of reaching or maintaining the target cooking state at the current moment is lower than the cooking state confidence threshold but higher than the cooking state deviation threshold, it is determined that there is an acceptable fluctuation in the current cooking process, and the hybrid control mode of the knowledge-driven control loop and the real-time feedback control loop is activated, with the basic power control command as the main one and the compensation power control command as the auxiliary one for fine-tuning. When the posterior probability of reaching or maintaining the target cooking state at the current moment is lower than the cooking state deviation threshold, it is determined that the current cooking process has significantly deviated from the expectation, and the real-time feedback control loop dominant mode is activated, which mainly relies on real-time feedback from sensors to generate the compensation power control command to quickly correct the deviation.

6. The knowledge graph-based adaptive thermal management control system according to claim 5, characterized in that, The process of generating the final power modulation command includes identifying the sticky pan stage and invoking the adaptive pulse heating strategy: The rate of change of the real-time temperature measurement value sequence and the food adhesion state probability output by the Bayesian probabilistic inference network are continuously monitored. The food adhesion state probability is a specific variable in the set of latent variables. When the rate of change of the real-time temperature measurement value sequence is lower than the sticking risk threshold and the probability of food sticking exceeds the sticking risk probability threshold, the cooking process is determined to have entered the sticking stage. During the stage of easy-stick pan use, the structured cooking knowledge graph is used to query the anti-stick pulse heating mode template suitable for the current combination of ingredients and cookware. The anti-stick pulse heating mode template defines the basic duty cycle range and adjustment frequency. Based on the difference between the real-time temperature measurement value sequence and the target temperature value, the specific value of the pulse duty cycle for the next control cycle is dynamically calculated within the basic duty cycle range; Based on the adjustment frequency and the calculated pulse duty cycle, a pulse power control command containing periodic on / off signals is generated, which serves as the main component of the final power modulation command in the easy-stick pan stage.

7. The knowledge graph-based adaptive thermal management control system according to claim 6, characterized in that, During the easy-stick cooking stage, the structured cooking knowledge graph is used to query anti-stick pulse heating mode templates suitable for the current combination of ingredients and cookware, including: Identify all the thermal attribute nodes of the ingredients associated with the current cooking process, and extract attribute values ​​that characterize the surface characteristics of each ingredient from all the thermal attribute nodes of the ingredients. The attribute values ​​that characterize the surface characteristics of the ingredients include, but are not limited to, starch content, protein type, and surface roughness level. Identify the cooktop performance specification node associated with the current cooking process, and extract the attribute values ​​characterizing the cookware base material from the cooktop performance specification node; Using the set of attribute values ​​representing the surface characteristics of the ingredients and the attribute values ​​representing the base material of the cookware as joint query conditions, a traversal search is performed in the structured cooking knowledge graph; Locate the anti-stick pulse heating mode template node in the structured cooking knowledge graph whose matching degree with the joint query conditions exceeds a preset threshold; The basic duty cycle range and the adjustment frequency are read from the anti-sticking pulse heating mode template node. The basic duty cycle range includes a lower limit value of the pulse duty cycle and an upper limit value of the pulse duty cycle. The adjustment frequency defines the upper limit and lower limit of the frequency of the periodic on / off signal in the pulse power control command.

8. The knowledge graph-based adaptive thermal management control system according to claim 7, characterized in that, Based on the difference between the real-time temperature measurement sequence and the target temperature value, the specific value of the pulse duty cycle for the next control cycle is dynamically calculated within the basic duty cycle range, including: Read the lower limit and upper limit of the pulse duty cycle defined in the anti-stick pulse heating mode template; Calculate the absolute value of the temperature deviation between the average value of multiple sampling points of the temperature sensor and the target temperature value for the current stage obtained from the structured cooking knowledge graph within the current control cycle; The absolute value of the temperature deviation is input into a preset fuzzy inference engine, and the output of the fuzzy inference engine is a duty cycle adjustment coefficient, which is a real number between zero and one. The difference between the lower limit of the pulse duty cycle and the upper limit of the pulse duty cycle is multiplied by the duty cycle adjustment coefficient to obtain a duty cycle increment; The pulse duty cycle increment is added to the lower limit of the pulse duty cycle to obtain the specific value of the pulse duty cycle for the next control cycle.

9. The knowledge graph-based adaptive thermal management control system according to claim 8, characterized in that, The method further includes a graph update module for adaptively updating the structured cooking knowledge graph during the cooking process, specifically including: After a complete cooking process is completed, collect the sensor time-series data, control command time-series data, and the user's final evaluation input of the cooking result throughout the entire cooking process; By performing correlation analysis on the sensor time series data and control command time series data, the key temperature inflection point, the effective power range actually used, and the total cooking time when the best cooking effect is actually achieved are identified. The identified key temperature inflection points, the actual effective power range used, and the total cooking time are compared with the pre-stored data in the cooking knowledge sub-graph dedicated to the current recipe used in this cooking, and the degree of difference is calculated. If the difference exceeds the preset update trigger threshold, the valid data identified during this cooking process will be added as a new experience record to the corresponding node of the structured cooking knowledge graph, or used to correct the parameter values ​​of the original node.

10. The knowledge graph-based adaptive thermal management control system according to claim 9, characterized in that, The user's final evaluation of the cooking result serves as a guide for the direction of knowledge graph updates, including: The user's evaluation input is parsed into a structured score, which includes at least a graded score for the doneness, charring degree, and texture of the dish. The structured score is correlated with the final state probability distribution output by the Bayesian probabilistic inference network at the moment cooking ends; If the structured score is lower than the expected score threshold in a specific dimension, the state probability change curve of the corresponding dimension during the cooking process is traced back to locate the time point of control decision error and related control parameters. Based on the located time point and related control parameters, combined with the structured score, parameter correction suggestions are generated for specific process parameter nodes in the structured cooking knowledge graph. The correction suggestions include adjusting the recommended temperature, modifying the heating time, or marking sensitive process stages. The proposed corrections will be applied to subsequent versions of the structured cooking knowledge graph after being confirmed manually or by pre-defined review rules in offline mode.

Images