Method and system for twin simulation of an induction hob

Abstract

The application discloses a kind of electromagnetic stove's twin simulation method and system, the present application relates to the technical field of twin simulation, input twin data source to the preset digital twin space, and in digital twin space is based on boundary condition in twin data source Trigger corresponding reduction operation, to carry out joint space-time solution to electromagnetic field, temperature field and flow field, and then generate and synchronize mapping twin simulation body of electromagnetic stove physical entity, improve the accuracy of twin simulation body.According to the autonomous optimization of state trajectory, a corresponding control instruction sequence is determined;The control instruction sequence is issued to the electromagnetic stove, and the electromagnetic stove is triggered to perform autonomous control, the actual response data after execution is obtained, and the virtual response data corresponding to the state trajectory is combined for difference comparison, to output the corresponding comparison result, and then trigger the closed-loop control of twin simulation body, the closed-loop control of twin simulation body is realized.

Description

Technical Field

[0001] This invention relates to the technical field of twin simulation, and in particular to a twin simulation method and system for an induction cooker. Background Technology

[0002] With the evolution of smart home appliances and industrial IoT technology, digital twin technology has been gradually introduced into the research and development and intelligent control of induction cookers. Digital twins aim to achieve precise monitoring and optimized control of the operating status of induction cookers by constructing a virtual model that is synchronously mapped with the physical entity.

[0003] The operation of an induction cooker is essentially a highly nonlinear dynamic coupling process involving electromagnetic fields, temperature fields, and flow fields. Existing technologies typically only collect single-modal data such as temperature or current as input, lacking in-depth alignment and feature fusion of multi-modal sensing data (electrical, temperature, vibration, etc.) in the spatiotemporal dimensions. This results in the twin data source failing to fully reflect the true boundary conditions of the physical entity, affecting the accuracy of the twin simulation and thus impacting its closed-loop control. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a twin simulation method and system for an induction cooker.

[0005] This invention provides a twin simulation method for an induction cooker, comprising: Multimodal data of the induction cooker during operation is collected, and the multimodal sensing data is spatiotemporally aligned and feature fused to construct a twin data source that reflects the real-time physical state of the induction cooker. The twin data source is input into the preset digital twin space, and the corresponding order reduction operation is triggered in the digital twin space based on the boundary conditions in the twin data source, so as to perform joint spatiotemporal calculation of electromagnetic field, temperature field and flow field, and then generate a twin simulation body that is synchronously mapped with the physical entity of the induction cooker. The corresponding timing prediction mechanism is loaded onto the twin simulation object, and the current state of the twin simulation object is used as the initial node. Combined with the preset cooking constraints, dynamic simulation is performed in the future time window to output the corresponding state trajectory. The corresponding control command sequence is determined based on the autonomous optimization of the state trajectory. The control command sequence is sent to the induction cooker, triggering the induction cooker to perform autonomous control, obtaining the actual response data after execution, and comparing the difference with the virtual response data corresponding to the state trajectory, thereby outputting the corresponding comparison result, and then triggering the closed-loop control of the twin simulation.

[0006] This invention provides a twin simulation system for an induction cooker, which is applied to the aforementioned twin simulation method for induction cookers; the twin simulation system for the induction cooker includes: The twin data source module is used to collect multimodal data of the induction cooker during operation, and to perform spatiotemporal alignment and feature fusion on the multimodal sensing data to construct a twin data source that reflects the real-time physical state of the induction cooker. The twin simulation module is used to input the twin data source into the preset digital twin space, and trigger the corresponding order reduction operation based on the boundary conditions in the twin data source within the digital twin space, thereby performing joint spatiotemporal calculations on the electromagnetic field, temperature field and flow field, and generating a twin simulation that is synchronously mapped with the physical entity of the induction cooker. The control instruction sequence module is used to load the corresponding timing prediction mechanism onto the twin simulation object, and to dynamically extrapolate the current state of the twin simulation object as the initial node, combined with the preset cooking constraints in the future time window, so as to output the corresponding state trajectory, and determine the corresponding control instruction sequence based on the autonomous optimization of the state trajectory. The closed-loop control module is used to send the control command sequence to the induction cooker and trigger the induction cooker to perform autonomous control, obtain the actual response data after execution, and compare the difference with the virtual response data corresponding to the state trajectory, thereby outputting the corresponding comparison result, and then triggering the closed-loop control of the twin simulation.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) Collect multimodal data of the induction cooker during operation, perform spatiotemporal alignment and feature fusion on the multimodal sensing data, and thus construct a twin data source that reflects the real-time physical state of the induction cooker; input the twin data source into the preset digital twin space, and trigger the corresponding order reduction operation based on the boundary conditions in the twin data source in the digital twin space, so as to perform joint spatiotemporal calculation of electromagnetic field, temperature field and flow field, and then generate a twin simulation body that is synchronously mapped with the physical entity of the induction cooker. The twin data source is introduced to further control the electromagnetic field, temperature field and flow field, and improve the accuracy of the twin simulation body.

[0008] (2) Load the corresponding timing prediction mechanism onto the twin simulation object, and take the current state of the twin simulation object as the initial node, and combine the preset cooking constraints to perform dynamic simulation in the future time window to output the corresponding state trajectory. Determine the corresponding control command sequence based on the autonomous optimization of the state trajectory. Send the control command sequence to the induction cooker and trigger the induction cooker to perform autonomous control, obtain the actual response data after execution, and compare the difference with the virtual response data corresponding to the state trajectory to output the corresponding comparison result, thereby triggering the closed-loop control of the twin simulation object, further controlling the control command sequence, fully considering the actual response data and virtual response data, and realizing the closed-loop control of the twin simulation object. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the twin simulation method for an induction cooker in an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 in the twin simulation method for an induction cooker according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating step S12 in the twin simulation method for an induction cooker according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating step S13 in the twin simulation method for an induction cooker according to an embodiment of the present invention. Figure 5 This is a flowchart illustrating step S14 of the twin simulation method for an induction cooker in an embodiment of the present invention. Figure 6 This is a schematic diagram of the structural composition of the twin simulation system for an induction cooker in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0011] Please see Figures 1 to 6 A twin simulation method for an induction cooker, applied to a twin simulation scenario; the twin simulation method for an induction cooker includes: Step S11: Collect multimodal data of the induction cooker during operation, perform spatiotemporal alignment and feature fusion on the multimodal sensing data, and thus construct a twin data source that reflects the real-time physical state of the induction cooker; Step S12: Input the twin data source into the preset digital twin space, and trigger the corresponding order reduction operation based on the boundary conditions in the twin data source in the digital twin space, so as to perform joint spatiotemporal calculation of electromagnetic field, temperature field and flow field, and then generate a twin simulation body that is synchronously mapped with the physical entity of the induction cooker. Step S13: Load the corresponding timing prediction mechanism onto the twin simulation object, and take the current state of the twin simulation object as the initial node, and combine it with the preset cooking constraints to perform dynamic simulation in the future time window to output the corresponding state trajectory, and determine the corresponding control command sequence based on the autonomous optimization of the state trajectory. Step S14: Send the control command sequence to the induction cooker and trigger the induction cooker to perform autonomous control, obtain the actual response data after execution, and compare the difference with the virtual response data corresponding to the state trajectory, thereby outputting the corresponding comparison result, and then triggering the closed-loop control of the twin simulation.

[0012] refer to Figure 2 In step S11, the specific steps are as follows: S111: Real-time monitoring of the induction cooker, and collection of corresponding multimodal sensing data during the operation of the induction cooker. The multimodal sensing data covers electrical data, temperature data and vibration data, and is combined with a multimodal spatiotemporal alignment network for timestamp registration and spatial coordinate mapping. At the same time, cross-modal feature fusion is performed on the multimodal sensing data to extract feature tensors that characterize the operating status of the induction cooker. S112: Obtain the structural parameters of the induction cooker and load them along with the feature tensor into the same graph neural network, thereby performing multiple mappings along the graph neural network. During the mapping process, combine the past usage events of the induction cooker to construct a twin data source represented by a dynamic topology graph.

[0013] In the embodiments of this application, the induction cooker is monitored in real time, and corresponding multimodal sensing data is collected during the operation of the induction cooker. The multimodal sensing data covers electrical data, temperature data and vibration data, and timestamp registration and spatial coordinate mapping are performed in combination with a multimodal spatiotemporal alignment network. At the same time, cross-modal feature fusion is performed on the multimodal sensing data to extract feature tensors that characterize the operating state of the induction cooker.

[0014] At this time, during the operation of the induction cooker, multimodal sensing data covering electrical data, temperature data, and vibration data are collected in real time through a distributed sensor array. Taking into account the high-frequency characteristics of electrical data and the low-frequency characteristics of temperature and vibration data, an anti-aliasing low-pass filter is used to truncate the original signals of each channel in the frequency domain. Based on a preset reference clock beat, the high-frequency electrical data is downsampled to extract characteristic statistics, such as mean, peak-to-peak value, and harmonic occupancy rate. The low-frequency temperature and vibration data are upsampled and reconstructed based on polynomial interpolation so that the modal data can achieve a uniform frame rate in the time dimension, forming an initial multimodal data frame sequence.

[0015] The initial multimodal data frame sequence is input into the time alignment module of the multimodal spatiotemporal alignment network, which embeds a time-series manifold alignment algorithm. Using the PWM control cycle of the induction cooker inverter bridge as the time reference anchor point, the energy integral features of electrical data within a single control cycle are extracted, and the trend and transient features of temperature and vibration data within a sliding window containing multiple control cycles are extracted. Furthermore, by calculating the cross-covariance matrix of each modal feature with respect to the time reference anchor point, the optimal time delay compensation factor is solved, and the timestamps of each modal data frame are rigidly translated and elastically stretched to eliminate the transmission and response delays caused by the multimodal sensing link and achieve strict timestamp registration.

[0016] After timestamp registration is completed, the data is transferred to the spatial alignment module. A high-dimensional spatial topological coordinate system for the physical entity of the induction cooker is pre-established, and the installation positions of each physical sensor are defined as spatial anchor points. To address the spatial sparsity of the temperature and vibration sensors, a spatial distance weight matrix is ​​constructed based on the three-dimensional structural mesh of the induction cooker. Using this weight matrix, the temperature and vibration data at the discrete anchor points are interpolated and diffused along the spatial geodesics using radial basis kernel functions. Simultaneously, combined with the prior mask of the electromagnetic heat source distribution corresponding to the electrical data, the point-like sensing data is mapped into a multi-channel feature map with uniform spatial resolution, thus completing the spatial coordinate mapping.

[0017] The shared spatiotemporal multi-channel feature map obtained after spatial coordinate mapping is input into the cross-modal feature fusion module. This module adopts a fusion architecture based on a multi-head cross-attention mechanism. The feature channel representing thermodynamic evolution is used as the query vector, and the feature channel representing electromagnetic excitation and the vibration feature channel representing mechanical response are used as the key vector and value vector, respectively. By calculating the similarity score between the query vector and the key vector, the interaction weights of different modal features in the current spatiotemporal state are dynamically assigned, so that the electromagnetic disturbance features and micro-vibration features converge towards the temperature feature space, thereby capturing the implicit correlation features under multi-field coupling. The feature dimension is compressed and nonlinearly activated by a multilayer perceptron, and a compact feature tensor representing the real-time physical operating state of the induction cooker is output.

[0018] Optionally, the spatiotemporally aligned multimodal data is organized into three feature channels: a thermal channel feature map characterizing the evolution of the temperature field, an electromagnetic channel feature map characterizing electromagnetic excitation, and a vibration channel feature map characterizing mechanical vibration. In the fusion network, the system uses the thermal channel feature map as the query vector, and the electromagnetic channel feature map and the vibration channel feature map as the key vector and value vector, respectively.

[0019] The similarity score between the query vector and the key vector is calculated. This score reflects the weight of the influence of electromagnetic disturbances and mechanical vibrations on the evolution of the temperature field at the current spatiotemporal location. This similarity score is normalized using the Softmax function and then multiplied by the value vector to achieve a weighted aggregation of vibrational and electromagnetic features.

[0020] The fusion network of this invention employs four parallel attention heads, each independently calculating the aforementioned attention weights. The outputs of the four heads are then concatenated along the feature dimension. Finally, a multilayer perceptron consisting of two fully connected layers performs feature dimension compression and nonlinear activation, outputting a compact feature tensor. The first layer of this multilayer perceptron maps the input feature dimension to the intermediate hidden layer dimension, and the second layer maps it back to the target output dimension. Both layers use the ReLU activation function to introduce nonlinearity.

[0021] Specifically, during the high-frequency heating stage of the induction cooker, the sensor array collects the coil current at a sampling rate of 100kHz, such as electrical data; it collects the temperature of the thermistors on the inner arc side and straight edge side of the induction cooker panel at a sampling rate of 10Hz, such as temperature data; and it collects the MEMS accelerometer data on the coil frame at a sampling rate of 1kHz, such as vibration data. To eliminate the sampling rate difference, the fundamental effective value within a single PWM cycle is extracted from the current data (downsampling), and the temperature data is completed to a frame rate of 1kHz based on quadratic spline interpolation (upsampling).

[0022] Due to the thermal inertia of the coil and the thermal conduction delay of the sensor, there is a physical lag of about 2 seconds between the current change and the temperature response. The time alignment module uses the timestamp of the effective value of the fundamental current wave as the reference anchor point. By calculating the extreme values ​​of the cross-correlation between current and temperature within a long sliding window, it automatically solves for a delay compensation factor of 2.15 seconds, aligning the temperature characteristics of historical moments with the electrical characteristics of the current moment on the time axis, thus achieving timestamp registration.

[0023] For asymmetric structures, a two-dimensional polar coordinate grid adapted to the contour is constructed. At this time, only two temperature anchor points are arranged on the inner arc side and the straight edge side, while the corner is a temperature measurement blind zone. The spatial alignment module uses Gaussian radial basis interpolation based on the geodesic distance of the grid to spatially diffuse the data of the two temperature anchor points to the corner blind zone grid. At the same time, combined with the Lorentz force distribution mask generated by the coil current, the electrical data is mapped to the same grid to form a spatially aligned feature map.

[0024] When a pot placed on an induction cooker experiences localized boiling, the bursting of boiling bubbles triggers micro-vibrations on the panel, such as vibration modes. Simultaneously, the absorption of latent heat causes localized temperature rise to stagnate, such as temperature modes. The cross-modal fusion module uses the temperature features on the mesh as the query and the vibration features as the key and value. The attention mechanism automatically identifies the spatially overlapping region where the high-frequency abrupt changes in vibration features and the gradual slowdown in the slope of temperature features are located, i.e., the boiling center. The vibration features are then superimposed and fused into the temperature features, ultimately outputting a feature tensor with dimensions [Time, Channel, D_Profile_Height, D_Profile_Width]. This tensor not only contains the real-time temperature field distribution of the induction cooker but also implicitly contains cross-modal coupling information between electrical excitation fluctuations and boiling micro-vibrations, accurately characterizing the current asymmetric boiling operating state of the induction cooker.

[0025] Furthermore, the structural parameters of the induction cooker are obtained and loaded onto the same graph neural network along with the feature tensor, thereby performing multiple mappings along the graph neural network. During the mapping process, the historical usage events of the induction cooker are combined to construct a twin data source represented by a dynamic topology graph, which fully considers the twin data source.

[0026] At this point, the physical structural parameters of the induction cooker are obtained. These parameters include at least geometric dimensions, material thermal conductivity, electromagnetic relative permeability, and contact thermal resistance between components. Based on the physical component division and spatial discrete grid of the induction cooker, an initial static topology graph of the graph neural network is constructed. The nodes of the graph represent independent structural units or spatial grid units of the induction cooker, and the edges of the graph represent physical adjacency or field coupling relationships between units. The structural parameters are assigned as static inherent properties to the corresponding nodes and edges. At the same time, the feature tensors extracted in step S111 are sliced ​​and distributed to the corresponding nodes according to the spatial coordinate mapping relationship as dynamic observation properties of the nodes, thereby completing the initial loading of the graph neural network.

[0027] In the initialized graph neural network, multiple information transmission and aggregation operations are performed along the topological connections of the graph. The multiple mappings include spatial heat conduction mapping, electromagnetic-heat source mapping, and thermal-vibration coupling mapping. In a single mapping iteration, for the target node, the message transmission weight coefficient is calculated based on the "inherent properties" of its adjacent edges, such as thermal conductivity or magnetic permeability. The dynamic observation features of the adjacent nodes are weighted and aggregated along the edges, and nonlinear transformation and residual splicing are performed with the features of the target node itself. Through the stacking iteration of multiple layers of graph neural networks, the electrical excitation features diffuse along the electromagnetic-heat source edges to the coil and panel nodes, the temperature gradient features are smoothed along the heat conduction edges to the adjacent grid nodes, and the vibration response features resonate with the temperature field along the thermal-vibration coupling edges, thereby realizing the implicit high-dimensional solution of multi-physics spatiotemporal coupling within the graph structure.

[0028] In the multiple mapping process, the historical usage events of the induction cooker are introduced. These historical usage events include at least the cumulative running time, the number of extreme thermal stress impacts, and the duration of strong electromagnetic load. The historical usage events are encoded into global event embedding vectors and further transformed into topology modulation factors. The topology modulation factors act on the edge features and node features of the graph neural network, applying attenuation or enhancement operations to the edge weights and offset corrections to the inherent properties of the nodes, thereby simulating the material aging, increased thermal resistance, or insulation degradation of physical entities. By interleaving event-driven attribute correction with multiple rounds of mapping iterations, the static topology of the graph neural network undergoes dynamic deformation as its life cycle evolves, generating a dynamic topology graph with time-varying edge connection strength and drifting node properties.

[0029] After the graph neural network completes multiple mappings and dynamic evolution, the updated high-dimensional hidden features of each node are extracted. The hidden features of discrete nodes are compressed and reconstructed into a graph-level representation vector with physical interpretability by using graph global average pooling or physical boundary-based partition pooling operations. This graph-level representation vector integrates the real-time multimodal coupling state, structural constraint features, and degradation information of the induction cooker throughout its entire life cycle. It is defined as a twin data source represented by a dynamic topological graph to support the subsequent reduced-order solution of the digital twin space.

[0030] Optionally, for the dynamic topology graph, based on the physical structural parameters of the induction cooker, including geometric dimensions, material thermal conductivity, relative electromagnetic permeability, and contact thermal resistance between components, the induction cooker can be discretized into multiple independent structural units. For example, the coil can be divided into straight-edge segment units and curved-edge segment units, and the microcrystalline glass panel can be divided into multiple grid units. Each unit corresponds to a node in the graph neural network, and the physical adjacency or field coupling relationship between units corresponds to the edge in the graph. For example, there is an electromagnetic heat source mapping edge between the coil unit and the upper panel unit, and there is a heat conduction edge between adjacent panel units.

[0031] During initialization, the structural parameters of the induction cooker are used as inherent attributes of the nodes, and the feature tensors extracted in the preceding steps are used as dynamic observation attributes of the nodes, thus completing the loading of the graph neural network. Regarding the introduction of historical usage events, the specific approach of this invention is as follows: historical events such as the cumulative running time of the induction cooker, the number of extreme thermal stress impacts, and the duration of strong electromagnetic load are encoded into a global event embedding vector. This vector is then transformed into multiple topology modulation factors by a transformation module consisting of two fully connected layers. These modulation factors act on different parts of the graph. For example, when historical events indicate that the straight edge of the induction cooker has experienced multiple overheat protection events, the topology modulation factor for that region will multiply the weight of its corresponding heat conduction edge by an attenuation coefficient less than one, and simultaneously add a positive offset to the equivalent heat capacity attribute of the node on that side, thereby simulating the increase in contact thermal resistance and decrease in heat storage capacity due to thermal fatigue.

[0032] In each iteration of the graph neural network's multiple mappings, these modulation factors are recalculated and the connection strength of edges and the attribute values ​​of nodes are dynamically updated, causing the originally static graph topology to dynamically deform as the induction cooker's life cycle evolves and real-time operating conditions change. The graph neural network is set to three layers, with each layer performing a neighborhood information aggregation operation. That is, each node collects features from its neighboring nodes and updates its own state. After stacking the three layers, the receptive field of each node can cover its physical third-order neighborhood.

[0033] Specifically, the structural parameters of the induction cooker are obtained, including the cross-sectional area of ​​the copper wire in the straight edge section of the coil and the wire diameter in the curved edge section, the thermal conductivity of the microcrystalline glass panel, and the relative permeability of the magnetic strip system. A polar coordinate mesh is established according to the physical contour, and the straight edge region and the curved edge region are divided into independent node sets. The structural parameters of the straight edge and the curved edge are assigned as inherent attributes to the corresponding nodes and edges. At the same time, the feature tensor output by S111, which contains asymmetric temperature rise and microseismic characteristics, is segmented according to the coordinate mapping relationship and loaded onto the corresponding nodes as dynamic observation attributes.

[0034] The graph neural network performs multiple mappings along the topology. Due to the structural asymmetry, the edge weights between straight-edge nodes and curved-edge nodes are non-uniformly distributed based on magnetic permeability and thermal conductivity. During mapping, the curved-edge nodes have a concentrated magnetic field due to their large curvature, and their high-intensity electromagnetic characteristics transfer a large number of heat source terms to the corresponding panel nodes along the "electromagnetic-heat source edge". At the same time, the panel nodes perform thermal smoothing along the "heat conduction edge" to the low-temperature nodes on the straight-edge side. The vibration nodes of the coil frame map the asymmetric expansion stress characteristics to the temperature nodes along the "thermal-vibration coupling edge". After multiple mappings, the graph structure accurately restores the biased heat flow field and stress field distribution unique to induction cookers.

[0035] Historical usage events of the induction cooker are introduced, such as: long-term placement of a large cast iron pot on the straight edge section causing multiple extreme thermal shocks to the straight edge area, such as overheat protection events, and a cumulative high-frequency operation of 3000 hours. These events are encoded as topology modulation factors, which are applied to the graph structure: the modulation factor multiplies the weight of the "heat conduction edge" between the straight edge panel and the coil by an attenuation coefficient less than 1 to simulate the increase in contact thermal resistance due to thermal fatigue. At the same time, the equivalent heat capacity property of the straight edge node is offset to simulate the decrease in heat storage capacity caused by local microcracks in the microcrystalline glass. In subsequent mapping iterations, the conduction of heat flow on the straight edge is blocked, forcing the graph network to dynamically adjust the distribution topology of isotherms, presenting a physical boundary that is completely different from the new state.

[0036] The hidden features of each node in the grid are extracted from the dynamically evolved graph neural network. Due to the different degradation levels of the straight and curved sides, the extracted feature vectors contain significant asymmetric aging bias information. The node features of the contour are partitioned, pooled, and reconstructed to output a graph-level representation vector containing real-time operating status, asymmetric structural constraints, and local thermal fatigue degradation evolution. This constructs a high-fidelity, life-cycle-evolving induction cooker twin data source represented by a dynamic topological graph.

[0037] refer to Figure 3 In step S12, the specific steps are as follows: S121: Input the twin data source into the preset digital twin space, monitor the boundary conditions in the twin data source within the digital twin space, which cover the material dimension, power dimension, and temperature dimension; when the boundary conditions meet the preset order reduction trigger threshold, adaptively match the multidimensional sequence of electromagnetic field-temperature field-flow field corresponding to the current boundary conditions; S122: Call the order reduction operator corresponding to the induction cooker, and perform orthogonal basis projection and nonlinear truncation on the multidimensional sequence to perform joint spatiotemporal solution of electromagnetic field, temperature field and flow field, thereby gradually generating a twin simulation body that is synchronously mapped with the induction cooker and has synchronous response capability during the solution process.

[0038] In the embodiments of this application, the twin data source is input into a preset digital twin space, and the boundary conditions in the twin data source are monitored in the digital twin space. The boundary conditions cover the material dimension, power dimension, and temperature dimension. When the boundary conditions meet the preset order reduction trigger threshold, the multidimensional sequence of electromagnetic field-temperature field-flow field corresponding to the current boundary conditions is adaptively matched, thus introducing a multidimensional sequence.

[0039] At this point, the twin data source represented by the dynamic topology diagram constructed in step S112 is input into the preset digital twin space, which contains a multi-physics full-order mechanism model architecture of the induction cooker. Within the digital twin space, the twin data source is deconstructed and its features are stripped to extract the multi-dimensional boundary conditions of the induction cooker at the current moment. The boundary conditions specifically cover the "material dimension, including the relative permeability, conductivity and specific heat capacity of the cookware", the "power dimension, including the real-time input power of the inverter bridge and the PWM duty cycle" and the "temperature dimension, including the initial temperature distribution of the panel and the environmental heat dissipation coefficient". The multi-dimensional boundary conditions are used as the boundary conditions and driving source terms of the multi-physics equations in the digital twin space.

[0040] A reduction trigger threshold monitoring mechanism corresponding to multi-dimensional boundary conditions is established within the digital twin space. The reduction trigger threshold is derived based on the calculation convergence characteristics and nonlinearity calibration of the full-order simulation model of the induction cooker under different operating conditions. Specifically, a permeability mutation threshold is set for the material dimension, a power change rate threshold is set for the power dimension, and a phase change latent heat judgment threshold is set for the temperature dimension. The relative change rate or absolute deviation between the current boundary condition and the boundary condition at the previous steady-state moment is calculated in real time. When the relative change rate or absolute deviation exceeds the corresponding reduction trigger threshold, it is determined that the induction cooker has entered a strong nonlinear transient process, and a reduction trigger signal is output. Conversely, when it is within the threshold range, the current steady-state reduction calculation mode is maintained.

[0041] Upon receiving the order reduction trigger signal, adaptive retrieval and matching are performed in a pre-set order reduction model library based on the specific modes and combination states of the current trigger boundary. Multi-physics orthogonal basis functions are extracted in advance from offline full-order simulation data, and multiple sets of order reduction coefficient sub-matrices are trained according to the "flow field state: laminar / turbulent," electromagnetic skin depth, and heat conduction mode under different boundary condition combinations. The current material, power, and temperature boundary condition combination is used as the retrieval key to adaptively retrieve the matching order reduction coefficient sub-matrices and operator weights, thereby generating a multi-dimensional sequence of electromagnetic field-temperature field-flow field. This multi-dimensional sequence defines the coupling order, time step allocation, and source term propagation path of the order reduction equations for each physics field in the subsequent joint solution.

[0042] Specifically, a twin data source containing the real-time operating status and aging topology of the induction cooker is loaded into the digital twin space. By parsing this data source, the boundary conditions at the current moment are extracted: in the material dimension, the cookware is identified as a cast iron pot with high magnetic permeability and low electrical conductivity, and the equivalent conductivity of the straight edge side of the coil has decreased due to historical aging; in the power dimension, the current inverter bridge is obtained to be operating at a frequency of 20kHz and a duty cycle of 80%, with an input power of 2000W; in the temperature dimension, the asymmetric initial temperature distribution of the panel is obtained, with the arc edge side temperature at 120℃ and the straight edge side temperature at 150℃.

[0043] The digital twin space monitors the reduction-order trigger thresholds for this set of boundary conditions; the threshold for the change in magnetic permeability in the material dimension is set to a relative change rate of 20%, the threshold for the power change rate in the power dimension is set to 500W / s, and the threshold for the latent heat of phase change in the temperature dimension is set to 105℃, such as the boiling point of water; when cold water is suddenly added to the pot during the operation of the induction cooker, causing the panel temperature to drop instantly to 80℃ and the local temperature gradient to change drastically, the absolute deviation and change rate of this temperature dimension both exceed the set thresholds, the system immediately determines that a strong nonlinear change in the thermodynamic state has occurred, such as a phase change and thermal shock, and outputs a reduction-order trigger signal, requiring a switch to the calculation mode.

[0044] Upon receiving the order reduction trigger signal, the system adaptively matches the order reduction model library based on the current boundary condition combination of "cast iron pot + high duty cycle power + low temperature thermal shock". Due to the asymmetry of the structure, the high-frequency skin depth on the straight side is different from that on the curved side, and the addition of cold water triggers the boiling flow field, which transforms from natural convection to nucleate boiling. The system adaptively retrieves the "asymmetric skin depth electromagnetic basis function" and the "nucleate boiling two-phase flow order reduction operator" and matches the corresponding order reduction coefficient sub-matrix. This combination generates a multi-dimensional sequence: the sequence specifies that electromagnetic field order reduction is performed along the asymmetric contour to obtain a non-uniform eddy current heat source, the heat source is transferred to the temperature field in a small time step for transient heat conduction calculation, and finally, based on the bubble detachment dynamic parameters generated by the temperature field, the flow field is driven to perform boiling two-phase flow order reduction calculation, thereby ensuring that the twin calculation of the induction cooker under complex working conditions meets the real-time requirements and accurately reproduces the physical essence of asymmetric multi-field coupling.

[0045] Furthermore, the reduction operator corresponding to the induction cooker is invoked, and orthogonal basis projection and nonlinear truncation are performed on the multidimensional sequence to jointly solve the electromagnetic field, temperature field and flow field in spatiotemporal. In the process of solving, a twin simulation object that is synchronously mapped with the induction cooker and has synchronous response capability is gradually generated, which further controls the electromagnetic field, temperature field and flow field and improves the accuracy of the twin simulation object.

[0046] At this point, based on the multidimensional sequence of adaptive matching in step S121, the corresponding order reduction operator pre-set in the digital twin space of the induction cooker is called. This order reduction operator includes the order reduction projection matrix and state evolution function of the electromagnetic field, temperature field and flow field. The high-dimensional state field data of the physical entity at the current moment is projected to the corresponding low-dimensional orthogonal basis space along the coupling path specified by the multidimensional sequence. Specifically, by extracting the orthogonal basis vector matrix of each physical field obtained in advance based on the full-order finite element offline data through singular value decomposition or eigenorthogonal decomposition, the spatial discrete node information of the high-dimensional physical field is transformed into a low-dimensional orthogonal coefficient vector, realizing the dimensionality reduction mapping of the computational domain from the high-dimensional physical space to the low-dimensional manifold space, thereby greatly reducing the scale of the algebraic equation system for the joint solution of multiple fields.

[0047] After obtaining the low-dimensional orthogonal coefficient vector, nonlinear truncation is performed on the tail components in the low-dimensional manifold space that have extremely low energy proportions and are susceptible to high-frequency noise interference. The truncation threshold is set in advance according to the eigenvalue decay rate corresponding to each order of orthogonal basis, and high-order orthogonal modes whose cumulative contribution rate of characteristic energy is lower than the "set threshold, such as 10-5". At the same time, for strong nonlinear terms caused by abrupt changes in material properties or phase transitions during multi-field coupling, instead of directly discarding them, a nonlinear hypersurface reconstruction mechanism is introduced. The pre-trained autoencoder network is used to perform nonlinear feature compensation on the truncated low-dimensional principal components to ensure that the key nonlinear dynamic features reflecting the extreme working conditions or boundary abrupt changes of the induction cooker are not lost while the dimensionality is greatly reduced.

[0048] Optionally, for the nonlinear hypersurface reconstruction mechanism, the nonlinear hypersurface reconstruction mechanism adopted in this invention is a method to compensate for the high-order nonlinear features lost due to the truncation after the low-dimensional orthogonal coefficient vector is linearly truncated. Its specific implementation is based on a pre-trained fully connected autoencoder network.

[0049] The input and output dimensions of this autoencoder network are equal, both set to fifty dimensions, which is the same as the number of low-order principal component coefficients retained. The encoder part of the network consists of three fully connected layers connected in series. The first layer maps the fifty-dimensional input to twenty-five dimensions, the second layer maps the twenty-five dimensions to twelve dimensions, and the third layer maps the twelve dimensions to six dimensions, thereby extracting the deep nonlinear compression features of the input coefficients.

[0050] Each layer of the encoder is followed by a hyperbolic tangent activation function to introduce nonlinear transformation capability. The decoder part of the network has a symmetrical structure to the encoder, consisting of three fully connected layers that sequentially map the six-dimensional compressed features back to twelve, twenty-five, and fifty dimensions. The last layer of the decoder uses a linear activation function to output the reconstructed fifty-dimensional coefficient vector.

[0051] The training process of the autoencoder network is as follows: A full-order simulation model of an induction cooker is used to simulate a series of strongly nonlinear operating conditions, including the water boiling phase transition process, abrupt changes in material properties during dry burning of the cookware, and thermal shock caused by a sudden drop in power. At each simulation time step, the first one hundred orthogonal basis coefficients are recorded as the original samples, with the first fifty coefficients serving as the network input and the last fifty coefficients as the network's desired output target. During training, the mean squared error is used as the loss function, and the network's weight parameters are continuously adjusted through the backpropagation algorithm, enabling the network to learn the nonlinear mapping relationship between low-order principal component coefficients and high-order truncation coefficients.

[0052] In actual operation, for the first fifty orthogonal basis coefficients obtained by linear projection at the current time, the system inputs them into a trained autoencoder network. The network outputs a five-dimensional coefficient vector corrected by nonlinear reconstruction. This corrected coefficient vector replaces the original linear projection coefficients and serves as the low-dimensional state vector for subsequent joint spatiotemporal solutions. Through this mechanism, the system can retain the first fifty principal components while inferring the higher-order nonlinear feature information lost due to truncation from these principal components using the autoencoder, thereby achieving an accurate representation of strongly nonlinear processes.

[0053] Based on the low-dimensional state vector after nonlinear truncation and reconstruction, a joint spatiotemporal progressive solution of the electromagnetic field, temperature field, and flow field is performed in the low-dimensional manifold space. According to the inter-field coupling topology specified by the multi-dimensional sequence, the low-dimensional eddy current loss characteristics calculated by the electromagnetic field reduction operator are mapped to the volume heat source term of the temperature field reduction equation through the electromagnetic-thermal coupling projection operator. Next, the low-dimensional temperature gradient characteristics calculated by the temperature field are transformed into buoyancy term input into the flow field reduction equation. Then, the convective heat transfer coefficient field calculated by the flow field is fed back to the temperature field boundary. An implicit time integration scheme is used to perform time-step iteration in the low-dimensional space. The above inter-field source term propagation is iterated in each time step until convergence, realizing the real-time and efficient solution of the spatiotemporal coupling of multi-physics fields.

[0054] After completing a single-step joint spatiotemporal solution in the low-dimensional manifold space, the low-dimensional multiphysics predicted state vector of the induction cooker at the next moment is obtained. The order-up reconstruction operator, which is the inverse of the orthogonal basis projection, is called to multiply the low-dimensional predicted state vector by the transpose of the corresponding orthogonal basis matrix and inversely map it back to the high-dimensional physical space. Thus, the three-dimensional electromagnetic field, temperature field and flow field with complete spatial distribution characteristics are reconstructed step by step in the digital twin space. The reconstructed high-dimensional multiphysics set constitutes a twin simulation that is isomorphic to the physical entity of the induction cooker in spatial distribution and synchronized with its temporal evolution. This simulation has the ability to respond synchronously to sudden changes in boundary conditions and control command inputs at the millisecond level, providing a high-fidelity virtual image for subsequent time-series prediction and closed-loop control.

[0055] Specifically, when the induction cooker heats the cast iron pot at 2000W, the system calls a reduction operator that matches the "asymmetric magnetic field + high power + boiling" conditions. The high-dimensional grid node data of the three-dimensional space where the coil is located is projected onto the pre-extracted orthogonal basis space. Due to the specificity of the structure, the electromagnetic field orthogonal basis vector extracted offline itself contains the asymmetric characteristics of dense magnetic field lines on straight sides and divergent magnetic field lines on curved sides. After projection, the original grid node voltage and magnetic vector potential data of hundreds of thousands of dimensions are compressed into a low-dimensional orthogonal coefficient vector containing only the first 50 principal components, achieving a reduction in the computational dimension by orders of magnitude.

[0056] The coefficient vector in the low-dimensional space of the induction cooker is nonlinearly truncated. For example, in the projection result, the energy proportion corresponding to the 51st order and later electromagnetic orthogonal modes is extremely small. However, at this time, the water in the pot boils and produces bubbles, causing the local conductivity and thermal conductivity to fluctuate violently in a time-varying nonlinear manner. When the system performs nonlinear truncation, the first 50 principal components are retained. For the nonlinear local disturbances caused by boiling, the autoencoder compensation network is activated to implicitly compensate for the local high-frequency nonlinear characteristics that the 50th order principal components fail to characterize, ensuring that the boiling dynamic characteristics of the heat flux density concentration area on the straight side of the coil are not lost.

[0057] A three-field joint spatiotemporal progressive solution is performed in a low-dimensional space. The electromagnetic field order reduction operator calculates the low-dimensional eddy current loss based on the current PWM duty cycle. Due to the characteristics of the orthogonal basis, this loss naturally exhibits a non-uniform distribution with large straight edges and small curved edges. This non-uniform loss is used as a source term to input the temperature field order reduction equation, and the asymmetric temperature gradient is calculated, which shows that the temperature rise on the straight edge side of the panel is faster than that on the curved edge side. This temperature gradient is transformed into a buoyancy force term and input to the flow field order reduction equation, driving the flow field to calculate the local strong convection state where the bubble detachment frequency on the straight edge side of the pot is higher than that on the curved edge side. The flow field then feeds back the non-uniform convection heat transfer coefficient to the temperature field. The three are iterated until convergence, completing a joint solution for one time step. The entire process is completed at the low-dimensional matrix multiplication level, with extremely short time consumption.

[0058] The state vector after convergence of the low-dimensional joint solution is inversely mapped back to the high-dimensional physical space. By multiplying it by the transposed orthogonal basis matrix specific to the induction cooker, the low-dimensional coefficients are re-expanded into three-dimensional spatial field data covering the coil and the microcrystalline panel. The resulting twin simulation not only fits the irregular structure in terms of geometric contour, but also accurately reproduces the asymmetric distribution characteristics of strong electromagnetic heat flow coupling on the straight side and weak coupling on the curved side in terms of physical mechanism. When the input power of the physical entity changes abruptly, the twin simulation can synchronously respond to the transient jump of eddy current and temperature rise in the straight side section within tens of milliseconds, forming a virtual image that is highly faithfully and synchronously mapped with the physical entity of the induction cooker.

[0059] refer to Figure 4 In step S13, the specific steps are as follows: S131: Obtain a time-series prediction mechanism based on the combination of long short-term memory and diffusion content, load it into the twin simulation, mark the current multi-field coupling state of the twin simulation, take the current multi-field coupling state of the twin simulation as the initial node, and perform multiple calculations in combination with preset cooking constraints, so as to introduce random perturbation factors in the future time window for dynamic inference, and output a set of state trajectories covering multiple potential evolution paths; the cooking constraints cover anti-dry burning constraints, energy efficiency optimization constraints, and food heating uniformity constraints; S132: Perform Pareto autonomous optimization on the set of state trajectories, and select reference state trajectories that meet the global optimum during the optimization process. Based on the identification of the reference state trajectories, determine multiple key nodes, and combine the current working sequence of the induction cooker with the corresponding multimodal data to determine the corresponding control command sequence.

[0060] In the embodiments of this application, a time-series prediction mechanism based on the combination of long short-term memory and diffused content is obtained and loaded into the twin simulation object. The current multi-field coupling state of the twin simulation object is marked. Taking the current multi-field coupling state of the twin simulation object as the initial node, multiple calculations are performed in combination with preset cooking constraints. Random perturbation factors are introduced into the future time window for dynamic inference, so as to output a set of state trajectories covering multiple potential evolution paths. The cooking constraints cover anti-dry burning constraints, optimal energy efficiency constraints, and food heating uniformity constraints, thus introducing cooking constraints.

[0061] At this point, a pre-trained temporal prediction mechanism based on the combination of long short-term memory and diffusion content is obtained. The temporal prediction mechanism consists of a long short-term memory network coupled with a stochastic differential equation compiler based on a diffusion model. The temporal prediction mechanism is loaded into the twin simulation generated in step S122. The long short-term memory network is used to extract and extend the deterministic strong evolution trend of the multi-physics field state, and the diffusion model compiler is used to superimpose random perturbation features that conform to the physical evolution distribution on the evolution trend. The hidden state parameters of the temporal prediction mechanism and the multi-field coupled state variables of the twin simulation are aligned with tensor dimensions and bound to interfaces, so that the temporal prediction mechanism has the ability to drive the twin simulation to perform forward inference.

[0062] The joint solution results of the electromagnetic field, temperature field and flow field at the current moment are extracted from the twin simulation and marked as the current multi-field coupling state. Feature dimensionality reduction and latent space encoding are performed on this state to transform it into the initial latent state vector and cell state vector of the long short-term temporal memory network. At the same time, it is used as the terminal constraint condition of the diffusion model inverse denoising process. Thus, the current multi-field coupling state of the twin simulation is anchored as the time zero point and spatial initial node of the future time window dynamic extrapolation, ensuring that all extrapolation paths start from the same high-fidelity physical snapshot.

[0063] During the dynamic simulation, pre-set cooking constraints are introduced to limit the spatial and state boundaries of the simulation trajectory. These constraints include anti-dry burning constraints, energy efficiency optimization constraints, and food heating uniformity constraints. Specifically, the anti-dry burning constraint is manifested as a hard truncation of the first-order absolute upper limit threshold of temperature and the upper limit threshold of temperature rise rate of the temperature field simulation nodes. The energy efficiency optimization constraint is manifested as a punitive feedback on the proportion of ineffective heat dissipation between the electromagnetic field and temperature field simulation nodes. The food heating uniformity constraint is manifested as a limitation on minimizing the spatial temperature variance of the temperature field simulation nodes within the food coverage area. In each step of the temporal prediction mechanism, the above constraints are transformed into Lagrange multipliers or mask matrices, and the output characteristics of the long short-term temporal memory network are nonlinearly projected and corrected to eliminate or penalize evolution directions that violate physical and cooking laws.

[0064] Starting from the initial node, dynamic extrapolation is performed within a future time window [t0, t0+ΔT]. The Long Short-Term Memory Network outputs a deterministic evolution baseline of multi-field coupling states based on historical time-series characteristics. Simultaneously, at each extrapolation step of the evolution baseline, a diffusion model introduces random noise sampled from a standard Gaussian distribution and performs conditional denoising and diffusion inverse stepping based on the current multi-field coupling state and cooking constraints to generate random perturbation factors reflecting the uncertainties of the real physical world. These random perturbation factors simulate unmodeled dynamics such as thermal convection turbulence pulsations, grid voltage fluctuations, and slight changes in the heat absorption rate of ingredients. The random perturbation factors are superimposed on the deterministic evolution baseline, and by changing the direction of the extrapolation branches, divergent multiple extrapolation sub-trajectories are generated. All sub-trajectories extrapolated within the future time window are summarized to construct and output a set of state trajectories covering multiple potential evolution paths.

[0065] Specifically, a time-series prediction mechanism based on the combination of Long Short-Term Memory (LSTM) and Diffusion model is obtained and loaded into the induction cooker twin simulation. The input dimension of LSTM is aligned with the temperature, flow velocity, and magnetic vector potential feature dimensions of the grid nodes, while the Diffusion model is attached to the output of LSTM as a parallel perturbation generator. This enables the inference mechanism to not only remember the historical thermal inertia of the induction cooker, but also to simulate unpredictable thermal turbulence fluctuations.

[0066] Extract the joint solution results at the current moment from the induction cooker twin simulation. For example, the panel temperature corresponding to the straight edge section is 180℃ and the eddy current density is high, while the panel temperature corresponding to the arc edge section is 150℃ and the eddy current density is low. The flow field inside the pot shows a convection pattern of violently rising bubbles on the straight edge side and gentle convection on the arc edge side. Encode this asymmetric multi-field coupling state into a high-dimensional hidden vector and mark it as the zero point t0t0 of the deduced time, which serves as the unique physical starting point for the subsequent generation of multiple asymmetric evolution trajectories.

[0067] Multiple computational constraints are applied in conjunction with preset cooking constraints. The current cooking scenario is "anti-dry burning and grilling". The anti-dry burning constraint sets the local temperature of the panel to not exceed 240℃ and the temperature rise rate to not exceed 5℃ / second. The food heating uniformity constraint requires the temperature difference between the straight edge and the curved edge to be controlled within 20℃. The energy efficiency optimization constraint requires that the proportion of eddy current loss converted into effective heat of the pot body is greater than 85%. When the LSTM outputs a hidden feature that the temperature of the straight edge may rise to 250℃ in each step of the extrapolation, the mask matrix of the anti-dry burning constraint forces the truncated projection of this feature value, blocking the extrapolation direction that leads to dry burning. At the same time, the heating uniformity constraint applies a penalty gradient to the extrapolation feature with excessive temperature difference, forcing the subsequent extrapolation to seek an evolution path that adjusts the power distribution or changes the flow field distribution.

[0068] Dynamic simulations were performed within a set 30-second time window. The LSTM derived deterministic baselines for the temperature and flow fields based on historical data, such as the straight-edge temperature rising at the original rate. The diffusion model introduced random perturbation factors at each step, which physically simulated the instantaneous drastic change in heat transfer coefficient caused by the rupture of the local oil film at the bottom of the pot, or the weak ripple of the coil power supply voltage. The random perturbations were superimposed on the baseline. For example, one perturbation caused a sudden increase in convective heat transfer on the straight-edge side, resulting in a slowdown in the temperature rise on the straight edge, while another perturbation caused the oil film to accumulate on the arc-edge side, resulting in local overheating of the arc-edge side. By sampling and simulating hundreds of random perturbations within 30 seconds, a set of state trajectories containing 50 potential evolution paths was output. Some of these trajectories eventually triggered a dry-burning warning, while others reached a thermally uniform and stable state under the adjustment of constraints, thus providing a rich future scenario space for the next step of autonomous optimization.

[0069] Furthermore, Pareto autonomous optimization is performed on the set of state trajectories, and reference state trajectories that satisfy global optimum are selected during the optimization process. Multiple key nodes are determined based on the identification of reference state trajectories, and the corresponding control command sequence is determined by combining the current working sequence of the induction cooker and the corresponding multimodal data, thus fully considering the control command sequence.

[0070] At this point, the set of state trajectories covering multiple potential evolution paths output in step S131 is obtained, and the preset cooking constraints are extracted as a multi-objective optimization function. The multi-objective optimization function includes at least an anti-dry burning penalty term, an energy efficiency loss term, and a heating unevenness variance term. For each evolution path in the set of state trajectories, the cumulative eigenvalues ​​of each objective function are calculated along the time dimension to construct a multi-objective evaluation vector. Based on the Pareto dominance rule, the multi-objective evaluation vector is sorted and layered in a non-dominated manner, and the subset of trajectories in the first layer, i.e., the Pareto front, is extracted. In the subset of Pareto front trajectories, a proximity aggregation strategy based on the ideal point method is adopted to calculate the Euclidean distance from each front trajectory to the absolute optimal ideal solution. The trajectory with the minimum Euclidean distance is defined as the reference state trajectory that satisfies the global optimum, thereby solving for the best compromise evolution path in the mutually constraining multi-objective constraints.

[0071] The selected reference state trajectories undergo time-dimensional state mutation and boundary criticality analysis to identify key nodes in the trajectory evolution process. Specifically, along the time axis of the reference state trajectory, the first and second time derivatives of the multiphysics state variables are calculated, and the extreme points of the margins of each constraint are monitored. When the absolute value of the first derivative of a state variable at a certain moment exceeds the transient mutation threshold, or the margin of a certain constraint drops to the safety critical threshold, the state node corresponding to that moment is marked as a key node. Key nodes represent the turning point of the system's operating state or the critical intervention point where the control strategy must intervene. The target timestamp and target multiphysics coupled state set corresponding to each key node are recorded.

[0072] The current working sequence of the induction cooker is read in real time. The working sequence includes at least the current PWM duty cycle, the current operating power level, and the fan speed. The multimodal sensing data collected in step S11 is called simultaneously. The multimodal sensing data covers real-time electrical parameters, temperature distribution, and vibration spectrum. The current working sequence and the multimodal sensing data are concatenated at the feature level in the data space to construct a dynamic anchoring feature vector that represents the current real physical operating posture of the induction cooker. This dynamic anchoring feature vector reflects the initial execution boundary and physical constraints faced by the induction cooker when it transitions from the current real state to the target state of the key node.

[0073] The dynamic anchoring feature vector is differentially compared with the extracted target multi-field coupled state set of each key node to calculate the state transition deviation from the current state to the first key node and between adjacent key nodes. The state transition deviation is input into a pre-constructed inverse kinematics control mapping model, which is based on the physical execution mechanism of an induction cooker. This model is used to inversely solve the deviation in the physical field state space into the adjustment amount in the control parameter space. Through inverse mapping, the PWM frequency adjustment command, duty cycle adjustment command, and fan speed adjustment command corresponding to each key time node are obtained. According to the timestamp order of the key nodes, the adjustment commands are sequentially arranged and incrementally encoded to generate a control command sequence that is strictly isomorphic to the reference state trajectory and has physical executability.

[0074] Specifically, Pareto optimization was performed on 50 state trajectories predicted for the induction cooker within the next 30 seconds. A comprehensive evaluation vector was calculated for each trajectory, including: anti-dry-burning penalty, energy efficiency loss, and heating unevenness variance. Due to the strong magnetic field on the straight edge of the structure, some trajectories, while having the best energy efficiency, were prone to overheating and triggering dry burning on the straight edge. Other trajectories reduced power to extremely low levels to prevent dry burning, resulting in insufficient heating and poor energy efficiency on the curved edge. Through non-dominated sorting, dominated inferior solutions were eliminated, retaining 10 non-dominated trajectories on the Pareto front. Furthermore, the ideal point method was used to calculate the Euclidean distance of each leading-edge trajectory from the absolute ideal solution of "zero dry-burning risk, 100% energy efficiency, and zero temperature difference." Finally, the trajectory with the shortest distance was selected as the reference state trajectory. This trajectory, while suppressing dry burning on the straight edge, maximized the utilization of the curved edge margin, achieving a globally optimal compromise under the asymmetric structure.

[0075] Key nodes were identified in the reference state trajectory. Analysis along the time axis revealed that at the 8th second, the first derivative of the temperature of the straight-edge side panel suddenly increased, with a temperature rise rate reaching 4.5℃ / second, approaching the critical threshold of 5℃ / second for the anti-dry-burning constraint. The system marked the 8th second as the first key node, such as the critical intervention point for anti-dry-burning. At the 18th second, the variance of the temperature difference between the straight edge and the curved edge reached its minimum. The system marked the 18th second as the second key node, serving as the heat stabilization point. The target temperature field and electromagnetic field distribution states corresponding to these two nodes were extracted as the target for subsequent control.

[0076] The current operating status of the induction cooker is obtained; the current working sequence shows that the inverter is running at full power with a frequency of 20kHz and a duty cycle of 90%, and the fan is at medium speed; the current multimodal sensing data feedback shows that the temperature of the straight edge panel is 190℃ and the vibration sensor detects slight oil film rupture and high-frequency vibration, and the temperature of the curved edge is 150℃; the above current PWM parameters, fan speed and multimodal characteristics are spliced ​​together to form a dynamic anchoring feature vector, which clarifies the initial true base point for the control system to function.

[0077] Control commands are generated based on state deviation mapping. By comparing the current state with the target state at the critical node of the 8th second, it is found that the temperature rise rate of the straight edge needs to be reduced from 4.5℃ / second to 2℃ / second, while maintaining the heat input of the arc edge. The thermodynamic deviation is input into the inverse kinematics control mapping model, and the following is calculated: For the straight edge drive circuit of the coil, the PWM duty cycle needs to be reduced from 90% to 60%, while the arc edge drive circuit remains unchanged at 85%. If it is a multi-coil independent control, the overall duty cycle can be reduced to 75% and supplemented with a microsecond-level intermittent heating strategy. If it is a single coil, the fan speed is increased to the highest level to enhance bottom heat dissipation in order to suppress local overheating caused by oil film rupture. By comparing the deviation between the 8th and 18th second nodes, the command to use low duty cycle soft wave heating after entering the heat spreader period is calculated. The above control actions of "reducing duty cycle and increasing fan speed at the 8th second" and "switching to soft wave mode at the 18th second" are arranged in sequence to generate an induction cooker-specific control command sequence that is strictly synchronized with the reference state trajectory, so as to achieve precise active intervention under asymmetric thermal field.

[0078] refer to Figure 5 In step S14, the specific steps are as follows: S141: Send the optimal control command sequence to the controller of the induction cooker and trigger the induction cooker to perform autonomous control. Monitor the autonomous control process of the induction cooker in real time and collect the actual response data after execution. At the same time, obtain the virtual response data corresponding to the state trajectory, compare the actual response data and the virtual response data to output the comparison result representing the offset shape of the twin simulation body. S142: Based on the identification of the comparison results, multiple offset contents are identified. The corresponding offset process is determined by tracing each offset content. The process is then reverse-engineered in combination with the twin simulation to output the corresponding process optimization content. This further triggers the closed-loop control of the twin simulation to present the adaptive update of the twin simulation.

[0079] In the embodiments of this application, the optimal control command sequence is sent to the controller of the induction cooker and the induction cooker is triggered to perform autonomous control. The autonomous control process of the induction cooker is monitored in real time and the actual response data after execution is collected. At the same time, the virtual response data corresponding to the state trajectory is obtained. The actual response data and the virtual response data are compared to output the comparison result representing the offset shape of the twin simulation body. The comparison result is introduced.

[0080] At this point, the optimal control command sequence determined in step S132 is sent out through a two-way real-time communication link between the digital twin system and the physical controller of the induction cooker. The communication link follows a low-latency industrial-grade protocol. After receiving the control command sequence, the controller of the induction cooker parses it into action register values ​​that can be executed by the underlying hardware, triggering the power switching transistors of the inverter bridge to operate according to the PWM duty cycle and switching frequency set by the command. At the same time, it adjusts the drive voltage of the cooling fan, thereby driving the physical entity of the induction cooker to perform autonomous control. During the execution process, the interrupt service routine of the controller monitors the timing of the command execution at the millisecond level to ensure that the physical execution action is strictly synchronized with the timing of the twin system, eliminating the execution phase difference caused by communication jitter or command parsing delay.

[0081] During the physical process of autonomous control of the induction cooker, a multi-modal sensor array deployed inside the induction cooker is used for high-frequency real-time monitoring. Based on the timestamps of key nodes in the control command sequence, the actual response data of the induction cooker is synchronously collected within the transient window and steady-state maintenance window after the execution of key control actions. The actual response data includes at least the real current transient waveform of the inverter bridge circuit, the temperature rise time sequence curve of each temperature measuring point on the microcrystalline glass panel, and the acceleration vibration spectrum of the coil support. The collected original sensor signals are denoised and resampled to align with the data frame rate of the twin simulation, thereby constructing an actual response time sequence matrix that characterizes the real dynamic evolution process of the physical entity.

[0082] Within the digital twin space, based on the same timestamp of the optimal control command sequence, the twin simulation is driven to perform forward extrapolation, extracting the state evolution data of the twin simulation within the corresponding key nodes and time windows as virtual response data. Since the virtual response data is in a three-dimensional continuous computation domain, while the actual response data is collected from discrete sensor points, a spatiotemporal alignment operation needs to be performed. In the spatial dimension, based on the installation coordinates of the physical sensors, a three-dimensional interpolation operator is used to extract the virtual physical quantities of the corresponding coordinate points from the high-dimensional field distribution of the twin simulation. In the temporal dimension, the virtual response data is resampled and aligned using the sampling clock of the physical entity as a reference. This generates a virtual response time series matrix that is strictly consistent with the actual response data in both dimension and semantics.

[0083] The actual response time series matrix and the virtual response time series matrix under the same spatiotemporal coordinates are subjected to element-wise difference operation to obtain the residual time series matrix. For the residual time series matrix, the absolute error and relative error distribution of multimodal features are calculated, and the time-varying trend features of the error are extracted. Based on the preset offset morphology classification cluster, the distribution features of the residual time series matrix are pattern matched to identify systematic steady-state offset, sudden transient offset, or spatially distributed asymmetric offset. The identified offset modes, offset magnitudes and their spatiotemporal distribution features are encapsulated, and the comparison results representing the mismatch state between the twin simulation and the physical entity are output. The comparison results not only reflect the virtual-real deviation of a single control, but also reveal the potential drift of the boundary conditions or parameter mechanisms of the twin model.

[0084] Specifically, the optimal control command sequence issued to the induction cooker includes the instruction to "reduce the PWM duty cycle from 90% to 60% at the 8th second to suppress dry burning on the straight edge." This instruction sequence is sent to the MCU of the induction cooker via a Bluetooth / Wi-Fi dual-mode link. After receiving the trigger interrupt at the 8th second, the MCU immediately modifies the value of the timer compare register to shorten the conduction time of the IGBT, thereby truly changing the high-frequency excitation of the coil and performing autonomous power reduction control. Simultaneously, the MCU increases the duty cycle of the fan drive PWM to 100% to enhance bottom heat dissipation.

[0085] After the aforementioned power reduction command was executed, the sensor array monitored the induction cooker's response in real time. The current transformer collected the effective value of the loop current, but it did not decrease smoothly as expected; instead, a brief current spike appeared. The thermistor on the straight edge of the panel collected the actual temperature rise rate, which only decreased from 4.5℃ / second to 3.0℃ / second, failing to reach the expected target of 2.0℃ / second. The accelerometer installed on the straight edge section collected the low-frequency vibration energy, which did not weaken but instead increased. The above real current, temperature, and vibration data were fused to construct the actual response time sequence matrix after execution.

[0086] Within the digital twin space, the same power reduction command is input into the twin simulation for simulation. According to the ideal reduced-order model, the virtual response data shows that the current decreases smoothly, the virtual temperature rise rate on the straight side drops to 2.0℃ / second within 2 seconds, and the virtual vibration energy decays due to the release of thermal stress. Based on the actual installation coordinates of the straight and curved thermistors, the virtual temperature of the corresponding coordinate points is extracted from the three-dimensional temperature field of the twin simulation and resampled at a frame rate of 100Hz to generate a virtual response timing matrix aligned with the physical acquisition.

[0087] The actual response matrix was compared with the virtual response matrix. The difference results showed that the relative error of the current was 8%, including transient glitches, and the absolute error of the straight edge temperature was +5℃. The actual temperature was higher than the virtual temperature, and the deviation of the straight edge vibration spectrum was concentrated in the 200Hz frequency band. Based on the offset morphology classification, the system identified this as a "local asymmetric steady-state positive offset". Its physical essence is that the thermal conductivity of the microcrystalline glass on the straight edge of the induction cooker has decreased due to historical thermal fatigue. The default thermal conductivity used in the virtual model is too high, resulting in excessively fast heat dissipation in the simulation. The comparison results accurately output the offset morphology of the twin simulation in the straight edge heat conduction mechanism and the offset magnitude of +5℃, providing solid data support for subsequent triggering of the twin closed-loop evolution and parameter correction.

[0088] Furthermore, based on the identification of the comparison results, multiple offset contents are identified, and the corresponding offset procedures are determined by tracing each offset content. In conjunction with the twin simulation, reverse deduction is performed to output the corresponding process optimization content, which further triggers the closed-loop control of the twin simulation to present the adaptive update of the twin simulation. The control command sequence is further controlled, and the actual response data and virtual response data are fully considered to realize the closed-loop control of the twin simulation.

[0089] At this point, the comparison results representing the offset morphology of the twin simulation output in step S141 are deconstructed in multiple dimensions. Based on the preset physical dimensions and field attribute classification criteria, the macroscopic residual time series matrix is ​​decomposed into specific physical and spatial domains to determine multiple independent offset contents. The offset contents include at least spatial position offset coordinates, physical field attribute labels, and offset polarity and amplitude. By introducing a mechanism causal relationship mapping dictionary, the residual performance is classified by attribution, distinguishing between transient random pseudo-offsets caused by measurement noise and steady-state mechanism true offsets caused by physical parameter drift. Pseudo-offsets are filtered out, and the set of mechanism true offset contents representing the inconsistency of intrinsic parameters between the physical entity and the twin model is retained and extracted.

[0090] For each identified offset, a reverse tracing is performed along the forward modeling chain of the digital twin system. Using the offset as the starting point of the objective function's gradient, and based on the sensitivity analysis theory of multiphysics coupling, the partial derivatives of each underlying input parameter with respect to the current offset are calculated. These underlying input parameters include material properties, boundary heat transfer coefficients, and geometric micro-distortion parameters. Sensitivity is ranked according to the absolute value of the partial derivatives, and the dominant parameters contributing most to the offset are selected. Furthermore, these dominant parameters are mapped to the virtual construction process of the digital twin to determine the specific modeling or parameter setting steps that caused the offset, thus tracing back to the corresponding offset process. Specifically, the offset process refers to the property parameter assignment step in the twin initialization stage, the boundary condition setting step in the mesh generation stage, or the source term transfer step in the multiphysics coupling solution stage.

[0091] Using currently acquired multimodal sensing data as strong boundary conditions and the dominant parameters in the traced offset process as adjustable unknown variables, a back-inference mechanism is triggered within the twin simulation. A parameter identification inverse problem is constructed with the goal of minimizing the residual between the virtual and actual responses. Nonlinear least squares iterative algorithms such as Bayesian optimization or Levenberg-Marquardt are used to perform optimization calculations within the physically permissible solution space of the parameters. When the inverse problem converges iteratively, the optimal correction value of the dominant parameter is output. The optimal correction value is combined with the original parameter settings using a differential combination to form a compensation and correction strategy for the offset process, i.e., the process optimization content. The process optimization content clearly defines the parameter identifiers to be modified, the modification magnitudes, and the effective timing, ensuring that the optimization action eliminates the current offset without compromising the global energy conservation and physical rationality of the model.

[0092] The generated process optimization content is transformed into control commands and loaded into the digital twin space, triggering the closed-loop control of the twin simulation. Based on the process optimization content, the structural parameter assignment module, boundary condition matrix, or reduction operator coefficient matrix corresponding to the offset process are dynamically overwritten and reconstructed online. After the overwriting is completed, the multi-field joint solution core of the twin simulation is restarted or hot-updated, so that the subsequent spatiotemporal solution runs based on the corrected optimization parameters. Through continuous iteration of the closed-loop process of "virtual-real comparison - offset tracing - reverse inference - parameter overwriting", the twin simulation continuously approaches the real degradation state and current boundary of the physical entity during the running cycle, thus presenting the adaptive update capability of the twin simulation to evolve synchronously with the life cycle of the physical entity.

[0093] Specifically, the comparison results output in S141 were deconstructed. The comparison results showed that during the power reduction stage, there was a steady-state positive shift of +5℃ in the temperature on the straight edge side of the panel, and the vibration energy in the 200Hz frequency band of the straight edge region was too high. Through the mechanism causal relationship mapping dictionary, the +5℃ temperature rise shift was classified as "thermal conduction suppression type shift", and the high 200Hz vibration was classified as "thermal stress constraint shift". After filtering out the transient burr pseudo shifts in the current channel, two core shift contents were confirmed: first, the heat output in the microcrystalline glass region on the straight edge side was insufficient, and second, there was abnormal thermal deformation stress in the structural components on the straight edge side.

[0094] The source of the above-mentioned deviations was traced. For the +5℃ temperature rise deviation on the straight edge, thermodynamic sensitivity analysis was used to calculate the partial derivatives of the microcrystalline glass's thermal conductivity, specific heat capacity, and contact thermal resistance with respect to temperature rise. It was found that the absolute value of the partial derivative of the contact thermal resistance was the largest, indicating that it was most sensitive to temperature rise. This led to the "thermal boundary condition setting process" in the twin modeling process. For the high 200Hz vibration, fluid-structure interaction sensitivity analysis was used to trace it back to the "coil frame constraint stiffness assignment process". It was determined that the root cause of the deviation was that the initial straight edge contact thermal resistance of the twin was too low and the frame stiffness did not take thermal decay into account.

[0095] Initiate a reverse simulation within the twin simulation body; using the measured 190℃ on the straight edge and 160℃ on the curved edge as strong constraint boundaries, and take "straight edge contact thermal resistance" and "skeleton stiffness" as variables to be inversely calculated; use the Levenberg-Marquardt algorithm to iteratively call the reduced-order thermo-coupling model, continuously adjusting the thermal resistance and stiffness inputs; after 12 iterations, the residual between the virtual temperature field and the measured temperature field converges to within 0.1℃, at which point it is found that the straight edge contact thermal resistance should increase by 0.05K / W and the skeleton stiffness coefficient should decrease by 15%; encapsulate "thermal resistance +0.05K / W" and "stiffness -15%" as process optimization content for the straight edge thermo-coupling link.

[0096] The closed-loop control of the twin simulation is triggered; the system sends the process optimization content to the model parameter library, dynamically overwrites the contact thermal resistance matrix in the original boundary condition setting process, and modifies the stiffness coefficient in the structural dynamics solution module online; after overwriting, the twin simulation does not need to be stopped and restarted, and directly adopts the updated parameters in the next solution cycle; at this time, when the twin simulation simulates the same power reduction command, its virtual straight edge temperature automatically rises to be consistent with the real physical entity, and the vibration mode characteristics are synchronously matched. This shows that the induction cooker twin simulation has successfully absorbed the physical aging information caused by long-term local overheating and realized the adaptive synchronous update of the model boundary and the degradation state of the entity.

[0097] Please see Figure 6 The twin simulation system for induction cookers is applied to the aforementioned twin simulation method for induction cookers; the twin simulation system for induction cookers includes: The twin data source module 21 is used to collect multimodal data of the induction cooker during operation, perform spatiotemporal alignment and feature fusion on the multimodal sensing data, and thus construct a twin data source that reflects the real-time physical state of the induction cooker. The twin simulation module 22 is used to input the twin data source into the preset digital twin space, and trigger the corresponding order reduction operation based on the boundary conditions in the twin data source in the digital twin space, so as to perform joint spatiotemporal calculation of electromagnetic field, temperature field and flow field, and then generate a twin simulation that is synchronously mapped with the physical entity of the induction cooker. The control instruction sequence module 23 is used to load the corresponding timing prediction mechanism onto the twin simulation object, and to dynamically extrapolate the current state of the twin simulation object as the initial node, combined with the preset cooking constraints in the future time window, so as to output the corresponding state trajectory, and determine the corresponding control instruction sequence based on the autonomous optimization of the state trajectory. The closed-loop control module 24 is used to send the control command sequence to the induction cooker and trigger the induction cooker to perform autonomous control, obtain the actual response data after execution, and compare the difference with the virtual response data corresponding to the state trajectory, thereby outputting the corresponding comparison result, and then triggering the closed-loop control of the twin simulation.

[0098] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0099] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein; this application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein; the specification and embodiments are to be considered exemplary only.

[0100] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A twin simulation method for an induction cooker, characterized in that, include: Multimodal data of the induction cooker during operation is collected, and spatiotemporal alignment and feature fusion are performed on the multimodal sensing data to construct a twin data source that reflects the real-time physical state of the induction cooker. The twin data source is input into the preset digital twin space, and the corresponding order reduction operation is triggered in the digital twin space based on the boundary conditions in the twin data source, so as to perform joint spatiotemporal calculation of electromagnetic field, temperature field and flow field, and then generate a twin simulation body that is synchronously mapped with the physical entity of the induction cooker. The corresponding timing prediction mechanism is loaded onto the twin simulation object, and the current state of the twin simulation object is used as the initial node. Combined with the preset cooking constraints, dynamic simulation is performed in the future time window to output the corresponding state trajectory. The corresponding control command sequence is determined based on the autonomous optimization of the state trajectory. The control command sequence is sent to the induction cooker, triggering the induction cooker to perform autonomous control, obtaining the actual response data after execution, and comparing the difference with the virtual response data corresponding to the state trajectory, thereby outputting the corresponding comparison result, and then triggering the closed-loop control of the twin simulation.

2. The twin simulation method for an induction cooker according to claim 1, characterized in that, The process involves collecting multimodal data from the induction cooker during its operation, performing spatiotemporal alignment and feature fusion on the multimodal sensing data, and constructing a twin data source reflecting the real-time physical state of the induction cooker, including: The system monitors the induction cooker in real time and collects corresponding multimodal sensing data during the operation phase of the induction cooker. This multimodal sensing data covers electrical data, temperature data, and vibration data. It combines a multimodal spatiotemporal alignment network to perform timestamp registration and spatial coordinate mapping. At the same time, it performs cross-modal feature fusion on the multimodal sensing data to extract feature tensors that characterize the operating status of the induction cooker.

3. The twin simulation method for an induction cooker according to claim 2, characterized in that, The process of collecting multimodal data from the induction cooker during its operation, performing spatiotemporal alignment and feature fusion on the multimodal sensing data to construct a twin data source reflecting the real-time physical state of the induction cooker, also includes: The structural parameters of the induction cooker are obtained and loaded into the same graph neural network along with the feature tensor, thereby performing multiple mappings along the graph neural network. During the mapping process, the historical usage events of the induction cooker are combined to construct a twin data source represented by a dynamic topology graph.

4. The twin simulation method for an induction cooker according to claim 1, characterized in that, The process of inputting the twin data source into a preset digital twin space, and triggering corresponding order reduction operations within the digital twin space based on the boundary conditions in the twin data source, thereby performing joint spatiotemporal calculations on the electromagnetic field, temperature field, and flow field, and generating a twin simulation object synchronously mapped with the physical entity of the induction cooker, includes: The twin data source is input into a preset digital twin space, and the boundary conditions in the twin data source are monitored in the digital twin space. These boundary conditions cover the material dimension, power dimension, and temperature dimension. When the boundary conditions meet the preset order reduction trigger threshold, the multidimensional sequence of electromagnetic field-temperature field-flow field corresponding to the current boundary conditions is adaptively matched.

5. The twin simulation method for an induction cooker according to claim 4, characterized in that, The process of inputting the twin data source into a preset digital twin space, and triggering corresponding order reduction operations within the digital twin space based on the boundary conditions in the twin data source to perform joint spatiotemporal calculations of the electromagnetic field, temperature field, and flow field, thereby generating a twin simulation object synchronously mapped to the physical entity of the induction cooker, further includes: The reduction operator corresponding to the induction cooker is invoked, and orthogonal basis projection and nonlinear truncation are performed on the multidimensional sequence to perform joint spatiotemporal solution of electromagnetic field, temperature field and flow field. In the process of solution, a twin simulation object that is synchronously mapped with the induction cooker and has synchronous response capability is gradually generated.

6. The twin simulation method for an induction cooker according to claim 1, characterized in that, The aforementioned timing prediction mechanism is applied to the twin simulation object, using the current state of the twin simulation object as the initial node and combining preset cooking constraints to dynamically extrapolate within a future time window, thereby outputting the corresponding state trajectory. Based on the autonomous optimization of the state trajectory, the corresponding control command sequence is determined, including: A time-series prediction mechanism based on the combination of long short-term and short-term memory and diffusion content is acquired and loaded into a twin simulation. The current multi-field coupling state of the twin simulation is marked. Taking the current multi-field coupling state of the twin simulation as the initial node, multiple calculations are performed in combination with preset cooking constraints. Random perturbation factors are introduced into the future time window for dynamic inference, so as to output a set of state trajectories covering multiple potential evolution paths. The cooking constraints include anti-dry burning constraints, optimal energy efficiency constraints, and food heating uniformity constraints.

7. The twin simulation method for an induction cooker according to claim 6, characterized in that, The mechanism for loading the corresponding timing prediction onto the twin simulation object, using the current state of the twin simulation object as the initial node, and dynamically extrapolating within a future time window based on preset cooking constraints, outputs the corresponding state trajectory. The corresponding control command sequence is then determined based on the autonomous optimization of the state trajectory. The mechanism further includes: Pareto autonomous optimization is performed on the set of state trajectories, and reference state trajectories that satisfy global optimum are selected during the optimization process. Multiple key nodes are determined based on the identification of reference state trajectories, and the corresponding control command sequence is determined by combining the current working sequence of the induction cooker and the corresponding multimodal data.

8. The twin simulation method for an induction cooker according to claim 1, characterized in that, The process of sending a control command sequence to the induction cooker, triggering the induction cooker to perform autonomous control, acquiring the actual response data after execution, and comparing the difference with the virtual response data corresponding to the state trajectory to output the corresponding comparison result, thereby triggering the closed-loop control of the twin simulation, includes: The optimal control command sequence is sent to the controller of the induction cooker, triggering the induction cooker to perform autonomous control. The autonomous control process of the induction cooker is monitored in real time, and the actual response data after execution is collected. At the same time, the virtual response data corresponding to the state trajectory is obtained. The difference between the actual response data and the virtual response data is compared to output the comparison result representing the offset shape of the twin simulation object.

9. The twin simulation method for an induction cooker according to claim 8, characterized in that, The process of sending a control command sequence to the induction cooker, triggering the induction cooker to perform autonomous control, acquiring the actual response data after execution, and comparing the difference with the virtual response data corresponding to the state trajectory to output the corresponding comparison result, thereby triggering the closed-loop control of the twin simulation, also includes: Based on the comparison results, multiple offset contents are identified. The corresponding offset process is determined by tracing each offset content. The process is then reverse-engineered using the twin simulation to output the corresponding process optimization content. This further triggers the closed-loop control of the twin simulation, resulting in the adaptive update of the twin simulation.

10. A twin simulation system for an induction cooker, characterized in that, The twin simulation system of the induction cooker is applied to the twin simulation method of the induction cooker as described in any one of claims 1-9; The twin simulation system for the induction cooker includes: The twin data source module is used to collect multimodal data of the induction cooker during operation, and to perform spatiotemporal alignment and feature fusion on the multimodal sensing data to construct a twin data source that reflects the real-time physical state of the induction cooker. The twin simulation module is used to input the twin data source into the preset digital twin space, and trigger the corresponding order reduction operation based on the boundary conditions in the twin data source within the digital twin space, thereby performing joint spatiotemporal calculations on the electromagnetic field, temperature field and flow field, and generating a twin simulation that is synchronously mapped with the physical entity of the induction cooker. The control instruction sequence module is used to load the corresponding timing prediction mechanism onto the twin simulation object, and to dynamically extrapolate the current state of the twin simulation object as the initial node, combined with the preset cooking constraints in the future time window, so as to output the corresponding state trajectory, and determine the corresponding control instruction sequence based on the autonomous optimization of the state trajectory. The closed-loop control module is used to send the control command sequence to the induction cooker and trigger the induction cooker to perform autonomous control, obtain the actual response data after execution, and compare the difference with the virtual response data corresponding to the state trajectory, thereby outputting the corresponding comparison result, and then triggering the closed-loop control of the twin simulation.

Images