Intelligent electromagnetic induction heating system based on edge calculation

Abstract

The invention discloses an intelligent electromagnetic induction heating system and method based on edge calculation, and aims to realize efficient and accurate heating control through advanced technical means. According to the system, a multi-source data real-time acquisition module is integrated, and various key parameters in the heating process can be comprehensively obtained. Through the data preprocessing and feature extraction technology, the system can accurately identify the heating state, and a reliable basis is provided for subsequent control. The edge computing node serves as the core of the system, has strong real-time analysis capability, can quickly respond to changes in the heating process, and dynamically adjusts a control strategy. In addition, through an edge-cloud collaborative optimization technology, cloud resources are fully utilized, and the intelligent level of the system is further improved. The problems of feedback lag, high cloud dependence and the like in the prior art are effectively solved, the heating uniformity, the energy efficiency ratio and the abnormity recovery capability are remarkably improved, and a reliable solution is provided for the fields of industrial heating, household induction cookers and the like.

Description

Technical Field

[0001] The present invention belongs to the field of electromagnetic induction heating control, and specifically discloses an intelligent electromagnetic induction heating system based on edge computing. Background Art

[0002] In recent years, electromagnetic induction heating technology has been widely used in industrial heating, household induction cookers, and other fields due to its high efficiency and cleanliness. However, with the increasing demand for intelligent technology, existing technologies have gradually exposed limitations in real-time data processing, energy efficiency optimization, and multi-device collaborative control. For example, Chinese invention patent CN102106185A proposes an electric induction edge heating device for thick conductive plates. It achieves uniform heating through transverse magnetic flux induction coils and magnetic flux concentrators, but has the following defects:

[0003] Insufficient data feedback and dynamic adjustment capabilities: This patent relies on fixed parameters to control the heating process, does not integrate a real-time data acquisition and analysis module, and cannot adaptively adjust the power output according to the dynamic temperature changes of the heating target (such as material differences and load fluctuations), resulting in low energy efficiency and the risk of local overheating.

[0004] Lack of edge-cloud collaboration: Existing electromagnetic induction heating systems often use a centralized control architecture (e.g., CN115914307B, a cloud-based heating system). This requires uploading large amounts of raw data to the cloud for processing, resulting in high network latency and bandwidth usage, making it difficult to meet millisecond-level real-time response requirements. Furthermore, due to their high dependency on the cloud, system control may be lost if the network is interrupted, resulting in insufficient reliability.

[0005] These shortcomings are particularly prominent in high-frequency applications such as induction cookers. For example, household induction cookers must quickly respond to changes in cookware materials and food conditions. However, traditional control methods lack multi-source data fusion (such as electromagnetic field distribution and current waveforms) and localized intelligent decision-making, resulting in large fluctuations in heating efficiency and high energy consumption. Furthermore, existing technologies lack sufficient consideration for lightweight design and security and privacy protection of edge computing nodes, further limiting their applicability in complex environments. Summary of the Invention

[0006] In response to the above problems, the present invention proposes an intelligent electromagnetic induction heating method and system based on edge computing. Through multi-sensor data fusion, lightweight neural network real-time prediction and edge-cloud collaborative optimization, dynamic power regulation and energy efficiency improvement are achieved, which specifically solves the two core defects of the existing technology: feedback lag and strong cloud dependence.

[0007] The purpose of the present invention is achieved through the following technical solutions.

[0008] An intelligent electromagnetic induction heating method based on edge computing includes the following steps:

[0009] (1) Real-time multi-source data acquisition: Temperature sensors, current sensors, voltage sensors, and electromagnetic field sensors are used to synchronously collect the surface temperature of the heating target, coil current waveform, coil voltage dynamic value, and electromagnetic field distribution data during the electromagnetic induction heating process;

[0010] (2) Data preprocessing and feature extraction: De-noising and normalization of the multi-source data in step (1) are performed, and time domain features are extracted using a sliding window algorithm, combined with frequency domain analysis methods to identify electromagnetic field energy distribution characteristics;

[0011] (3) Real-time analysis of edge computing nodes: In the local edge computing node, a lightweight neural network model is used to perform fusion analysis on the pre-processed multi-source features, predict the thermal field distribution state of the heating target, and dynamically update the model weights based on the transfer learning algorithm;

[0012] (4) Feedback iterative control: Based on the prediction results of step (3), the initial power adjustment command is generated through the fuzzy control algorithm, and an iterative learning controller is constructed in combination with historical heating efficiency data to perform closed-loop feedback optimization on the power output frequency and duty cycle;

[0013] (5) Dynamic parameter adaptive adjustment: Based on the real-time load status of the edge computing node, the coil drive parameters are dynamically adjusted through a distributed reinforcement learning algorithm, including the resonant frequency matching threshold and power density distribution range;

[0014] (6) Edge-cloud collaborative optimization: The operation logs and key feature data of the edge nodes are uploaded to the cloud, and the global optimization strategy is generated through cloud training and transmitted back to the edge nodes to achieve dynamic iterative upgrades of the heating strategy.

[0015] Furthermore, in the above-mentioned intelligent electromagnetic induction heating method based on edge computing, the specific implementation method of real-time acquisition of multi-source data in step (1) includes:

[0016] a) Arrange a distributed temperature sensor array on the surface of the heating target, forming a grid temperature measurement topology with a spacing of 0.5-2 mm, and use time division multiplexing to collect temperature data at each node;

[0017] b) The electromagnetic field distribution data is collected using 12 sets of three-axis Hall sensors installed around the excitation coil. The sensors are arranged in a circular pattern at 30° intervals, and the sampling frequency of each set of sensors is no less than 20kHz.

[0018] c) Using a hardware trigger signal synchronization mechanism, a unified clock pulse is generated by the FPGA to trigger synchronous sampling of all sensors, and the data timestamp alignment accuracy reaches ±10μs;

[0019] d) When performing spatial interpolation calculations on the electromagnetic field distribution data, the inverse distance weighted interpolation algorithm is used in combination with the electromagnetic field propagation attenuation model to generate a three-dimensional electromagnetic field intensity distribution heat map.

[0020] Furthermore, in the above-mentioned intelligent electromagnetic induction heating method based on edge computing, the data preprocessing and feature extraction in step (2) further includes:

[0021] a) Current waveform denoising uses an improved wavelet threshold filtering algorithm, which performs a 5-layer decomposition using the db6 wavelet basis and implements adaptive threshold shrinkage on high-frequency coefficients;

[0022] b) The electromagnetic field frequency domain analysis uses a windowed FFT algorithm with a Blackman-Harris window as the window function and a frequency resolution of 100 Hz to extract characteristic frequencies with energy exceeding 5% in the 2-50 kHz frequency band;

[0023] c) Time domain feature extraction includes calculating the peak-to-peak value, kurtosis coefficient, and form factor of the current waveform within a 200ms sliding window, and calculating the standard deviation and rising slope of the temperature data;

[0024] d) Establish a multi-source feature correlation matrix and use the Pearson correlation coefficient to screen the parameter combinations with strong correlation between temperature changes and electromagnetic field characteristic frequencies.

[0025] Furthermore, in the above-mentioned intelligent electromagnetic induction heating method based on edge computing, the method for constructing the lightweight neural network model in step (3) includes:

[0026] a) The network structure adopts the improved MobileNetV3 architecture, replacing the standard convolutional layer with a depth-wise separable convolution and adding a bidirectional LSTM module in the last layer to process temporal features;

[0027] b) The input layer receives a three-dimensional tensor of dimension [time step × number of features × number of sensor channels], and the time step is set to 10 sampling periods;

[0028] c) The attention mechanism is specifically implemented as follows: a channel attention module is introduced in the feature fusion layer, channel weights are generated through global average pooling, and a weighting coefficient of 1.2-1.5 is applied to high-frequency electromagnetic features;

[0029] d) The model is dynamically updated using an online transfer learning strategy. When a change in the heating target material is detected, the base network layer is frozen and only the fully connected layer parameters are fine-tuned.

[0030] Furthermore, in the above-mentioned intelligent electromagnetic induction heating method based on edge computing, the feedback iterative control of step (4) specifically includes:

[0031] a) Set up a control rule base based on the fuzzy control algorithm, with the input variable being the predicted temperature deviation and its rate of change, and the output variable being the power regulation amount;

[0032] b) The iterative learning controller adopts a feedforward-feedback composite structure. After each heating cycle, the initial control parameters for the next cycle are updated based on historical error data.

[0033] c) Power output optimization is achieved through space vector pulse width modulation (SVPWM), which adjusts the dead time and switching frequency of the IGBT drive signal in real time according to the optimized duty cycle;

[0034] d) Establish a closed-loop stability monitoring mechanism. When the power fluctuation exceeds ±15% for three consecutive cycles, the control parameters are reset and the baseline parameters sent from the cloud are reloaded.

[0035] Furthermore, in the above-mentioned intelligent electromagnetic induction heating method based on edge computing, the adaptive adjustment of dynamic parameters in step (5) further includes:

[0036] a) Constructing a six-dimensional state space based on a reinforcement learning algorithm, including current load impedance, coil temperature, target heating rate, power supply voltage fluctuation, historical energy efficiency ratio, and ambient humidity;

[0037] b) The reward function is designed as: R = α × energy efficiency coefficient + β × temperature tracking accuracy - γ × power fluctuation penalty term, where α, β, and γ are adjustable weight coefficients;

[0038] c) The online learning strategy uses the double-delayed deep deterministic policy gradient TD3 algorithm, and deploys an experience replay buffer on the edge node to store the most recent 1,000 state-action pairs;

[0039] d) During the parameter adjustment execution phase, a soft update strategy is implemented for the resonance frequency matching threshold, with each adjustment amplitude not exceeding ±5% of the current value.

[0040] Furthermore, in the above-mentioned intelligent electromagnetic induction heating method based on edge computing, the specific implementation of the edge-cloud collaborative optimization of step (6) includes:

[0041] a) Data upload uses a differential privacy protection mechanism, adding Laplace noise to temperature-sensitive data. The noise intensity is inversely proportional to the spatial correlation of the data.

[0042] b) The cloud-based digital twin model construction method is: based on the physical information neural network PINN, Maxwell's equations are embedded as constraints in the network training process;

[0043] c) The global optimization strategy is generated using a federated learning framework, aggregating local model parameters from no less than 50 edge nodes and updating the global model through an adaptive weighted averaging algorithm;

[0044] d) The policy is issued in an incremental update mode, only the difference in model parameters is transmitted, and the transmission security is ensured by the elliptic curve encryption algorithm.

[0045] The present invention also discloses an intelligent electromagnetic induction heating system based on the above method, comprising:

[0046] Sensor module, edge computing module, power drive module and cloud collaboration module;

[0047] The edge computing module integrates a lightweight neural network reasoning unit and a reinforcement learning decision unit to execute the algorithm process of steps (3)-(5) in claim 1 in real time;

[0048] The cloud collaboration module is connected to the edge computing module via an encrypted communication protocol, and is used to execute the global optimization strategy generation of step (6) in claim 1.

[0049] The present invention also discloses a computer-readable storage device storing a computer program, wherein the program implements the steps of the above method when executed by a processor.

[0050] The present invention also discloses an edge computing device for an induction cooker, comprising:

[0051] the storage device, and at least one embedded processor;

[0052] The processor calls the program in the storage device to control the power output of the electromagnetic induction coil and execute the real-time calculation task of the above method.

[0053] Compared with the existing technology, the present invention has the following advantages and beneficial effects:

[0054] 1. High-precision real-time control and dynamic response

[0055] Improved temperature control accuracy: Through a distributed sensor array (0.5-2mm spacing) and multi-source data fusion, the heating target surface temperature detection accuracy reaches ±1.2°C (traditional solution ±3.5°C), reducing the risk of local overheating by 90%.

[0056] Millisecond-level response: Edge computing nodes (<15ms inference latency) combined with FPGA hardware synchronization (±10μs timestamp alignment) increase the speed of power regulation command generation to 20 times that of traditional cloud solutions.

[0057] 2. Energy efficiency optimization and significant energy saving

[0058] Dynamic load matching: Parameter adaptive adjustment based on reinforcement learning (TD3 algorithm) improves the energy efficiency ratio (η) from 0.68 to 0.83, reducing energy consumption by 22%.

[0059] Precise resonant frequency tracking: Through online transfer learning and soft update strategies, the resonant frequency matching error is reduced from ±5kHz to ±0.8kHz, and the eddy current thermal efficiency is improved by 18%.

[0060] 3. Adaptability to complex scenarios

[0061] Multi-material compatibility: A lightweight neural network (improved version of MobileNetV3) combined with frequency domain feature analysis enables automatic recognition of heating targets such as iron, stainless steel, and composite materials, shortening mode switching time from 3.2s to 0.9s.

[0062] Enhanced anti-interference capability: The closed-loop stability monitoring mechanism (power fluctuation threshold ±15%) and parameter reset function reduce the abnormal recovery time from 2.1s to 0.4s, and improve the system reliability to 98%.

[0063] 4. Edge-cloud synergy

[0064] Data traffic compression: Differential privacy protection (Laplacian noise injection) and incremental update strategies reduce the amount of cloud communication data by 82% and reduce bandwidth usage to 1 / 5 of traditional solutions.

[0065] Global optimization capability: The federated learning framework aggregates data from 50+ edge nodes and generates optimization strategies through the Physical Information Neural Network (PINN), improving the average heating efficiency of similar devices by 9.6%.

[0066] 5. Security and reliability

[0067] Privacy protection: Using elliptic curve cryptography (ECC) and federated learning, the risk of sensitive data leakage is reduced by 92%, meeting GDPR compliance requirements.

[0068] Hardware fault-tolerant design: Dynamic adjustment of IGBT drive parameters (dead time 1.2-2.0μs adaptive) and over-temperature frequency reduction protection extend the life of key components by 30%.

[0069] 6. Cost and deployment advantages

[0070] Lightweight computing: Depthwise separable convolution and INT8 quantization technology reduce the memory usage of edge computing modules by 75%, enabling deployment on embedded devices such as NVIDIA Jetson.

[0071] Modular architecture: The decoupled design of sensors, computing, and drive modules supports rapid adaptation to multiple scenarios such as industrial induction cookers and household kitchen appliances, shortening the deployment cycle by 60%. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Flowchart of the method for intelligent electromagnetic induction heating based on edge computing of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is further described in detail below. However, it should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the scope of the invention. In addition, in the following description, the description of known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention. All raw materials in the embodiments of the present invention can be obtained through commercial channels.

[0074] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present invention will be described in detail below with reference to the embodiments.

[0075] Example 1

[0076] An intelligent electromagnetic induction heating method based on edge computing, such as Figure 1 As shown, the following steps are included:

[0077] (1) Real-time multi-source data acquisition: Temperature sensors, current sensors, voltage sensors, and electromagnetic field sensors are used to synchronously collect the surface temperature of the heating target, coil current waveform, coil voltage dynamic value, and electromagnetic field distribution data during the electromagnetic induction heating process;

[0078] (2) Data preprocessing and feature extraction: De-noising and normalization of the multi-source data in step (1) are performed, and time domain features are extracted using a sliding window algorithm, combined with frequency domain analysis methods to identify electromagnetic field energy distribution characteristics;

[0079] (3) Real-time analysis of edge computing nodes: In the local edge computing node, a lightweight neural network model is used to perform fusion analysis on the pre-processed multi-source features, predict the thermal field distribution state of the heating target, and dynamically update the model weights based on the transfer learning algorithm;

[0080] (4) Feedback iterative control: Based on the prediction results of step (3), the initial power adjustment command is generated through the fuzzy control algorithm, and an iterative learning controller is constructed in combination with historical heating efficiency data to perform closed-loop feedback optimization on the power output frequency and duty cycle;

[0081] (5) Dynamic parameter adaptive adjustment: Based on the real-time load status of the edge computing node, the coil drive parameters are dynamically adjusted through a distributed reinforcement learning algorithm, including the resonant frequency matching threshold and power density distribution range;

[0082] (6) Edge-cloud collaborative optimization: The operation logs and key feature data of the edge nodes are uploaded to the cloud, and the global optimization strategy is generated through cloud training and transmitted back to the edge nodes to achieve dynamic iterative upgrades of the heating strategy.

[0083] Example 2

[0084] An intelligent electromagnetic induction heating method based on edge computing, such as Figure 1 As shown, the following steps are included:

[0085] (1) Real-time multi-source data acquisition: Temperature sensors, current sensors, voltage sensors, and electromagnetic field sensors are used to synchronously collect the surface temperature of the heating target, coil current waveform, coil voltage dynamic value, and electromagnetic field distribution data during the electromagnetic induction heating process;

[0086] The specific implementation method of real-time multi-source data acquisition in step (1) includes the following specific steps:

[0087] a) Arrange a distributed temperature sensor array on the surface of the heating target, forming a grid temperature measurement topology with a spacing of 0.5-2 mm, and use time division multiplexing to collect temperature data at each node;

[0088] Preferred: Grid topology (0.5-2mm spacing): Through dense distribution of micron-level thermocouples, local temperature changes of the heating target (such as hot spots in pots) can be captured. 0.5mm spacing can detect microscopic temperature differences of <1°C, avoiding the blind spots of traditional single-point temperature measurement. Time-division multiplexing technology: Through a multiplexer (such as the ADG1607), sensor channels are cyclically switched, polling 64 nodes at 200Hz, reducing wiring complexity by 87% (compared to parallel acquisition).

[0089] b) The electromagnetic field distribution data is collected using 12 sets of three-axis Hall sensors installed around the excitation coil. The sensors are arranged in a circular pattern at 30° intervals, and the sampling frequency of each set of sensors is no less than 20kHz.

[0090] Solution Description: 30° circular arrangement: 12 sensor groups cover 360°, ensuring complete capture of the excitation coil's radial and axial magnetic fields. Each triaxial probe (HX711) simultaneously measures the Bx, By, and Bz components with an accuracy of ±1.5mT. 20kHz sampling rate: This meets the Nyquist sampling requirement for electromagnetic field variations (highest frequency component ≤ 10kHz), preventing aliasing distortion.

[0091] c) Using a hardware trigger signal synchronization mechanism, a unified clock pulse is generated by the FPGA to trigger synchronous sampling of all sensors, and the data timestamp alignment accuracy reaches ±10μs;

[0092] Preferably, a 50MHz master clock is used to generate precise synchronization pulse signals (LVDS differential transmission), ensuring that the timestamp deviation of all sensors is ≤10μs. This improves accuracy by two orders of magnitude compared to software synchronization solutions (typically >1ms). A hardware trigger mechanism directly controls sensor sampling through the FPGA's IO ports, eliminating random delays introduced by operating system scheduling.

[0093] d) When performing spatial interpolation calculations on the electromagnetic field distribution data, the inverse distance weighted interpolation algorithm is used in combination with the electromagnetic field propagation attenuation model to generate a three-dimensional electromagnetic field intensity distribution heat map.

[0094] (2) Data preprocessing and feature extraction: De-noising and normalization of the multi-source data in step (1) are performed, and time domain features are extracted using a sliding window algorithm, combined with frequency domain analysis methods to identify electromagnetic field energy distribution characteristics;

[0095] The data preprocessing and feature extraction of step (2) further includes:

[0096] a) Current waveform denoising uses an improved wavelet threshold filtering algorithm, which performs a 5-layer decomposition using the db6 wavelet basis and implements adaptive threshold shrinkage on high-frequency coefficients;

[0097] Solution description: db6 wavelet basis: has 6th order vanishing moment, suitable for capturing transient mutations in current waveforms.

[0098] 5-layer decomposition: decomposes the signal into different frequency bands (e.g. layer 1: 25-50kHz, layer 5: 0.78-1.56kHz), retaining the effective components (e.g. layers 1-3). Adaptive threshold: dynamically adjusts the threshold according to the noise standard deviation σ:

[0099] b) The electromagnetic field frequency domain analysis uses a windowed FFT algorithm with a Blackman-Harris window as the window function and a frequency resolution of 100 Hz to extract characteristic frequencies with energy exceeding 5% in the 2-50 kHz frequency band;

[0100] Preferred: Blackman-Harris window: main lobe width 24dB, side lobe attenuation -92dB, suppressing spectrum leakage. Feature frequency extraction: within the range of 2-50kHz, filter the frequency points that meet the following conditions:

[0101] c) Time domain feature extraction includes calculating the peak-to-peak value, kurtosis coefficient, and form factor of the current waveform within a 200ms sliding window, and calculating the standard deviation and rising slope of the temperature data;

[0102] d) Establish a multi-source feature correlation matrix and use the Pearson correlation coefficient to screen the parameter combinations with strong correlation between temperature changes and electromagnetic field characteristic frequencies.

[0103] (3) Real-time analysis of edge computing nodes: In the local edge computing node, a lightweight neural network model is used to perform fusion analysis on the pre-processed multi-source features, predict the thermal field distribution state of the heating target, and dynamically update the model weights based on the transfer learning algorithm;

[0104] The method for constructing the lightweight neural network model in step (3) includes:

[0105] a) The network structure adopts the improved MobileNetV3 architecture, replacing the standard convolutional layer with a depth-wise separable convolution and adding a bidirectional LSTM module in the last layer to process temporal features;

[0106] Preferred: Depthwise separable convolution: decompose the standard convolution into depthwise convolution (3×3 kernel, channel-independent) and point-wise convolution (1×1 kernel), reducing the amount of computation to the original

[0107] Bidirectional LSTM module: It fuses forward / backward features in the time dimension, and the memory length is set to 10 cycles (2 seconds) to capture the dynamic evolution of the heating process.

[0108] b) The input layer receives a three-dimensional tensor of dimension [time step × number of features × number of sensor channels], and the time step is set to 10 sampling periods;

[0109] Preferred: For example, dimensions [10×16×12]:

[0110] Time step 10: covers the temporal evolution within 2 seconds.

[0111] Number of features: 16: includes time domain features such as Vpp, K, and S_T, and 5 frequency energy points.

[0112] Sensor channel 12: corresponds to the Bx, By, and Bz components of 12 groups of Hall sensors.

[0113] c) The attention mechanism is specifically implemented as follows: a channel attention module is introduced in the feature fusion layer, channel weights are generated through global average pooling, and a weighting coefficient of 1.2-1.5 is applied to high-frequency electromagnetic features;

[0114] d) The model is dynamically updated using an online transfer learning strategy. When a change in the heating target material is detected, the base network layer is frozen and only the fully connected layer parameters are fine-tuned.

[0115] Preferred: Freeze strategy: When a material change is detected (determined by a sudden change in frequency energy), freeze the convolutional layer parameters (accounting for 85% of the total parameters) and only fine-tune the fully connected layer. Fine-tune the learning rate: set it to 1 / 10 of the initial value (e.g., from 0.001 to 0.0001) to prevent overfitting.

[0116] (4) Feedback iterative control: Based on the prediction results of step (3), the initial power adjustment command is generated through the fuzzy control algorithm, and an iterative learning controller is constructed in combination with historical heating efficiency data to perform closed-loop feedback optimization on the power output frequency and duty cycle;

[0117] The feedback iterative control of step (4) specifically includes:

[0118] a) Set up a control rule base based on the fuzzy control algorithm, with the input variable being the predicted temperature deviation and its rate of change, and the output variable being the power regulation amount;

[0119] Preferably, the input variables are fuzzified:

[0120] Temperature deviation e: domain [-50℃, +50℃], divided into 7 levels (NB, NM, NS, ZO, PS, PM, PB)

[0121] Deviation change rate Δe: domain [-10℃ / s, +10℃ / s], division is the same as above.

[0122] b) The iterative learning controller adopts a feedforward-feedback composite structure. After each heating cycle, the initial control parameters for the next cycle are updated based on historical error data.

[0123] Optimally, feedforward path: open-loop compensation based on model prediction

[0124] Feedback channel: PID closed-loop correction (Kp=0.8, Ki=0.05, Kd=0.1).

[0125] c) Power output optimization is achieved through space vector pulse width modulation (SVPWM), which adjusts the dead time and switching frequency of the IGBT drive signal in real time according to the optimized duty cycle;

[0126] d) Establish a closed-loop stability monitoring mechanism. When the power fluctuation exceeds ±15% for three consecutive cycles, the control parameters are reset and the baseline parameters sent from the cloud are reloaded.

[0127] (5) Dynamic parameter adaptive adjustment: Based on the real-time load status of the edge computing node, the coil drive parameters are dynamically adjusted through a distributed reinforcement learning algorithm, including the resonant frequency matching threshold and power density distribution range;

[0128] The adaptive adjustment of dynamic parameters in step (5) further includes:

[0129] a) Constructing a six-dimensional state space based on a reinforcement learning algorithm, including current load impedance, coil temperature, target heating rate, power supply voltage fluctuation, historical energy efficiency ratio, and ambient humidity;

[0130] b) The reward function is designed as: R = α × energy efficiency coefficient + β × temperature tracking accuracy - γ × power fluctuation penalty term, where α, β, and γ are adjustable weight coefficients;

[0131] c) The online learning strategy uses the double-delayed deep deterministic policy gradient TD3 algorithm, and deploys an experience replay buffer on the edge node to store the most recent 1,000 state-action pairs;

[0132] d) During the parameter adjustment execution phase, a soft update strategy is implemented for the resonance frequency matching threshold, with each adjustment amplitude not exceeding ±5% of the current value.

[0133] (6) Edge-cloud collaborative optimization: Upload the operation logs and key feature data of the edge nodes to the cloud, generate a global optimization strategy through cloud training and transmit it back to the edge nodes to achieve dynamic iterative upgrades of the heating strategy;

[0134] The specific implementation of the edge-cloud collaborative optimization in step (6) includes:

[0135] a) Data upload uses a differential privacy protection mechanism, adding Laplace noise to temperature-sensitive data. The noise intensity is inversely proportional to the spatial correlation of the data.

[0136] b) The cloud-based digital twin model construction method is: based on the physical information neural network PINN, Maxwell's equations are embedded as constraints in the network training process;

[0137] c) The global optimization strategy is generated using a federated learning framework, aggregating local model parameters from no less than 50 edge nodes and updating the global model through an adaptive weighted averaging algorithm;

[0138] d) The policy is issued in an incremental update mode, only the difference in model parameters is transmitted, and the transmission security is ensured by the elliptic curve encryption algorithm.

[0139] Example 3

[0140] An electromagnetic induction heating system, comprising:

[0141] Sensor module, edge computing module, power drive module and cloud collaboration module;

[0142] The edge computing module integrates a lightweight neural network reasoning unit and a reinforcement learning decision unit to execute the algorithm process of steps (3)-(5) in claim 1 in real time;

[0143] The cloud collaboration module is connected to the edge computing module via an encrypted communication protocol to execute the global optimization strategy generation in step (6).

[0144] The method of the present invention is specifically applied through application examples below.

[0145] Application Example 1

[0146] Heating control process of commercial intelligent induction cooker.

[0147] Taking the smart induction cooker in a commercial kitchen that handles pots of different materials (iron pots, stainless steel pots) as an example, the specific implementation process is as follows.

[0148] Step (1) Real-time collection of multi-source data

[0149] Hardware configuration:

[0150] 64 micro thermocouples are arranged at the bottom of the cookware, forming an 8×8 temperature measurement grid with a spacing of 1.2 mm. Time division multiplexing technology is used to collect temperature data at a polling rate of 200 times per second.

[0151] Twelve sets of three-axis Hall sensors (model HX711) are installed around the excitation coil, with a circular arrangement of 30° and a sampling frequency of 25kHz.

[0152] The FPGA (Xilinx Zynq-7000) generates a 50 MHz master clock and sends synchronized trigger pulses to all sensors through the LVDS interface.

[0153] Working principle:

[0154] When an iron pan is placed on an induction cooker, a temperature sensor array monitors the temperature distribution at the bottom of the pan in real time, while a Hall effect sensor captures the three-dimensional electromagnetic field (Bx, By, and Bz components) generated by the coil. The FPGA's hardware synchronization mechanism ensures that all sensor data timestamps have a deviation of ≤8μs. Inverse distance weighted interpolation (IDW parameter p=2) is applied to the electromagnetic field data, combined with a coil radius attenuation model, to generate a three-dimensional electromagnetic field heat map, accurately locating areas of concentrated eddy currents.

[0155] Step (2) Data preprocessing and feature extraction

[0156] Signal processing flow:

[0157] Current denoising: The coil current signal is decomposed by db6 wavelet 5-layer, and the high-frequency coefficient threshold is set to Retains the effective components >30kHz.

[0158] In the above formula:

[0159] σ: Noise standard deviation, which measures the degree of discreteness of the noise in the current signal. The larger the standard deviation, the more severe the noise fluctuation.

[0160] N: The number of sampling points of the signal. In this example, N = 512, which means that the signal is sampled at 512 points.

[0161] ln: Natural logarithm function, the logarithm with the mathematical constant e (approximately equal to 2.71828) as the base.

[0162] Frequency domain analysis: A Blackman-Harris window (main lobe width 24 dB) was applied to the Z component of the electromagnetic field, and a 4096-point FFT was performed to identify 35.2 kHz (the main resonant frequency of the iron pan) and 42.7 kHz (the characteristic frequency of the stainless steel pan).

[0163] Feature correlation: The Pearson correlation coefficient between the calculated temperature standard deviation and the 42.7kHz frequency energy is 0.87, indicating that the current cookware is made of stainless steel.

[0164] Key parameters:

[0165] Sliding window width: 200ms (corresponding to 50 sampling periods)

[0166] Kurtosis coefficient calculation: Used to detect sudden changes in current waveform.

[0167] In the above formula:

[0168] K: Kurtosis coefficient, used to measure the degree to which the current waveform deviates from the normal distribution. It can detect sudden changes in the current waveform. When the kurtosis coefficient deviates from the normal range, it may indicate that the current waveform is abnormal.

[0169] E[]: Expected value operator, which means performing expected operations on random variables.

[0170] X: Random variable, representing the sampling value of the current signal at a certain moment or interval.

[0171] μ: The mean of the random variable X, reflecting the average level of the current signal.

[0172] σ: standard deviation of the random variable X, which measures the discreteness of the current signal.

[0173] Step (3) Real-time analysis of edge computing nodes

[0174] Neural Network Architecture:

[0175] Input tensor dimensions: [10 × 16 × 12] (time steps × number of features × sensor channels)

[0176] Depthwise separable convolution kernel: 3×3, stride 2, reduces computation by 75%.

[0177] Number of bidirectional LSTM units: 32, memory duration covers 1.5 heating cycles.

[0178] Channel attention weighting: Apply a 1.3x weight to the 35-50kHz frequency band features.

[0179] Online learning process:

[0180] When the system detects a switch in the cookware material (e.g., iron → stainless steel), it freezes the convolutional layer weights of MobileNetV3 and fine-tunes only the last two fully connected layers (with the learning rate set to 1 / 10 of the baseline value), completing model adaptation within five heating cycles.

[0181] Step (4) Feedback iterative control

[0182] Fuzzy PID control example:

[0183] Input variables:

[0184] Temperature deviation e: set value 220℃ vs measured value 205℃

[0185] Deviation change rate Δe: -3℃ / s

[0186] Fuzzy rule base activation:

[0187] IFe=NB ANDΔe=NS THENΔP=PM (If the deviation is large and the rate of change is small, the power regulation is moderate)

[0188] SVPWM modulation: Based on the optimized duty cycle (increased from 45% to 52%), the IGBT switching frequency is increased from 23kHz to 27kHz, and the dead time is shortened to 1.2μs.

[0189] Exception handling:

[0190] When a sudden change in the power supply voltage causes a power fluctuation of +18% for three consecutive cycles, the system automatically rolls back to the cloud-based benchmark parameters (switching frequency 25kHz, duty cycle 48%) and returns to stability within 0.5s.

[0191] (5) Dynamic parameter adaptive adjustment

[0192] Reinforcement learning training process:

[0193] State vector:

[0194] [Z=0.15Ω,T coil =68℃,dT / dt=3.2℃ / s,V in =215V,n=0.81,RH=45%]

[0195] in

[0196] Z: The resistance value of the coil, measured in ohms (Ω), reflects the coil's resistance to current. Changes in resistance value will affect the power output and heating of the induction cooker.

[0197] T coil: The temperature of the coil, in degrees Celsius (℃). Excessive coil temperature may affect its performance and life, so it needs to be monitored and controlled.

[0198] dT / dt: The rate of change of coil temperature, measured in degrees Celsius per second (°C / s). It indicates how quickly the coil temperature changes over time and can be used to predict the changing trend of the coil temperature so that control measures can be taken in advance.

[0199] V in : Input voltage, measured in volts (V). Fluctuations in input voltage will affect the power output and stability of the induction cooker, so it needs to be monitored and adjusted.

[0200] η: Energy efficiency ratio, which measures the efficiency of the induction cooker in converting input electrical energy into heat energy. The higher the energy efficiency ratio, the higher the energy utilization efficiency of the induction cooker and the lower the energy consumption.

[0201] RH: Relative humidity, expressed in percentage (%). Ambient humidity may affect the performance and heat dissipation of the internal components of the induction cooker, so it needs to be taken into consideration.

[0202] TD3 algorithm output action: resonant frequency shift +1.8kHz, power density increased by 12%.

[0203] Experience replay buffer sampling: Prioritize historical states with an energy efficiency ratio > 0.75.

[0204] Physical Effects:

[0205] After the adjustment, the eddy current distribution is more uniform, the temperature difference between the center and the edge of the stainless steel pot is reduced from 32°C to 17°C, and the energy efficiency ratio is improved to 0.83.

[0206] Step (6) Edge-Cloud Collaborative Optimization

[0207] Federated Learning Implementation:

[0208] Participating nodes: 58 commercial induction cooker edge devices

[0209] Parameter aggregation:

[0210]

[0211] w global : Global model parameters are the model parameters finally aggregated by federated learning and are used for deployment and use on various edge devices.

[0212] i: Index variable, ranging from 1 to 58, representing the i-th device among the 58 commercial induction cooker edge devices participating in federated learning.

[0213] ηi: Energy efficiency ratio of the i-th node, which measures the energy utilization efficiency of the edge device when processing tasks. The higher the energy efficiency ratio, the less energy the device consumes when completing the same task.

[0214] ∑η: The sum of the energy efficiency ratios of all edge devices participating in federated learning, that is, ∑i=158ηi, which is used to normalize the energy efficiency ratio of each node.

[0215] w locali : The local model parameters of the i-th node are the model parameters obtained by each edge device through training based on its own data.

[0216] Model update: Only a 3.7MB parameter difference file (6% of the original model) is transmitted each time, and the latency is <150ms after ECC encryption.

[0217] Optimization results:

[0218] The cloud-based digital twin solves Maxwell's equations through PINN and generates a new resonant frequency matching strategy, which increases the average heating efficiency of similar cookware by 9.6%.

[0219] Implementation effect verification:

[0220] In continuous operation tests in commercial kitchens, compared to traditional PID control solutions, this case achieved:

[0221] Temperature control accuracy: ±3.5℃→±1.2℃

[0222] Energy efficiency ratio: 0.68→0.83

[0223] Abnormal recovery time: 2.1s→0.4s

[0224] 82% reduction in data uploads.

[0225] Application Example 2

[0226] Household smart frying and grilling machine

[0227] Scenario description:

[0228] The grilling machine needs to quickly switch between heating modes for steaks (2-3cm thick) and pancakes (0.5cm thick). Conventional equipment uses fixed heating parameters, resulting in undercooked steaks and burnt pancakes. The present invention achieves thickness-adaptive heating.

[0229] Hardware system composition

[0230] Sensor Module:

[0231] Array thermocouple: 64 points distributed temperature measurement, spacing 1.5mm

[0232] Flexible film pressure sensor (Tekscan 3101): measures food thickness (0.1-5cm, accuracy ±0.2mm) 3D electromagnetic field sensor (TIDRV5056): integrated into the bottom of the heating plate

[0233] Edge computing module:

[0234] Processor: NVIDIA Jetson Xavier NX

[0235] Memory configuration: 8GB LPDDR4x + 16GB eMMC

[0236] Algorithm unit:

[0237] Lightweight CNN model (TensorRT acceleration)

[0238] TD3 reinforcement learning inference engine

[0239] Power drive module:

[0240] IGBT module: Infineon FF600R12ME4 (1200V / 600A)

[0241] SVPWM modulator: TI C2000 Delfino MCU

[0242] Cloud collaboration module:

[0243] AWS IoT Greengrass enables edge-to-cloud connectivity

[0244] Encryption protocol: TLS1.3+AES-256-GCM

[0245] System workflow

[0246] Data collection phase:

[0247] Pressure sensor detects food thickness (steak 2.8cm)

[0248] Thermocouple array measured initial temperature distribution (center 45°C / edge 62°C)

[0249] Electromagnetic field sensors capture eddy current distribution characteristics

[0250] Real-time decision making at the edge:

[0251] The CNN model inputs [thickness, temperature gradient, field intensity distribution] → outputs a recommended power curve. The TD3 algorithm optimizes based on historical data:

[0252] State vector: [2.8cm, 45°C, 0.6Ω, 220V, 65°C / s]

[0253] Action output: Increase the resonant frequency to 55kHz and set the duty cycle to 58%

[0254] Cloud strategy update:

[0255] Aggregate 100+ user data every week to generate new models

[0256] Incremental update package size: 2.3MB (7% of the original model)

[0257] Encrypted transmission takes less than 200ms.

[0258] In summary, compared with the traditional solution, this solution has the following advantages as shown in Table 3.

[0259] Table 3 Comparison of advantages and disadvantages of the scheme

[0260] index Traditional solution The present invention Heating uniformity (ΔT) ±25℃ ±8℃ Mode switching time 3.2s 0.9s Energy consumption (kW·h / time) 0.78 0.61 Abnormal recovery rate 72％ 98％

[0261] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, based on the innovative concept of the present invention, changes and modifications to the embodiments described herein, or equivalent structural or equivalent process transformations made using the contents of the present invention specification, directly or indirectly applying the above technical solutions to other related technical fields are all included in the scope of protection of the patent of the present invention.

Claims

1. An intelligent electromagnetic induction heating method based on edge computing, characterized in that: The following steps are involved: (1) Real-time multi-source data acquisition: Temperature sensors, current sensors, voltage sensors, and electromagnetic field sensors are used to synchronously collect the surface temperature of the heating target, coil current waveform, coil voltage dynamic value, and electromagnetic field distribution data during the electromagnetic induction heating process; (2) Data preprocessing and feature extraction: De-noising and normalization of the multi-source data in step (1) are performed, and time domain features are extracted using a sliding window algorithm, combined with frequency domain analysis methods to identify electromagnetic field energy distribution characteristics; (3) Real-time analysis of edge computing nodes: In the local edge computing node, a lightweight neural network model is used to perform fusion analysis on the pre-processed multi-source features, predict the thermal field distribution state of the heating target, and dynamically update the model weights based on the transfer learning algorithm; (4) Feedback iterative control: Based on the prediction results of step (3), the initial power adjustment command is generated through the fuzzy control algorithm, and an iterative learning controller is constructed in combination with historical heating efficiency data to perform closed-loop feedback optimization on the power output frequency and duty cycle; (5) Dynamic parameter adaptive adjustment: Based on the real-time load status of the edge computing node, the coil drive parameters are dynamically adjusted through a distributed reinforcement learning algorithm, including the resonant frequency matching threshold and power density distribution range; (6) Edge-cloud collaborative optimization: The operation logs and key feature data of the edge nodes are uploaded to the cloud, and the global optimization strategy is generated through cloud training and transmitted back to the edge nodes to achieve dynamic iterative upgrades of the heating strategy.

2. The method according to claim 1, characterized in that The specific implementation method of real-time acquisition of multi-source data in step (1) includes: a) Arrange a distributed temperature sensor array on the surface of the heating target, forming a grid temperature measurement topology with a spacing of 0.5-2 mm, and use time division multiplexing to collect temperature data at each node; b) The electromagnetic field distribution data is collected using 12 sets of three-axis Hall sensors installed around the excitation coil. The sensors are arranged in a circular pattern at 30° intervals, and the sampling frequency of each set of sensors is no less than 20kHz. c) Using a hardware trigger signal synchronization mechanism, a unified clock pulse is generated by the FPGA to trigger synchronous sampling of all sensors, and the data timestamp alignment accuracy reaches ±10μs; d) When performing spatial interpolation calculations on the electromagnetic field distribution data, the inverse distance weighted interpolation algorithm is used in combination with the electromagnetic field propagation attenuation model to generate a three-dimensional electromagnetic field intensity distribution heat map.

3. The method according to claim 1, characterized in that The data preprocessing and feature extraction of step (2) further includes: a) Current waveform denoising uses an improved wavelet threshold filtering algorithm, which performs a 5-layer decomposition using the db6 wavelet basis and implements adaptive threshold shrinkage on high-frequency coefficients; b) The electromagnetic field frequency domain analysis uses a windowed FFT algorithm with a Blackman-Harris window as the window function and a frequency resolution of 100 Hz to extract characteristic frequencies with energy exceeding 5% in the 2-50 kHz frequency band; c) Time domain feature extraction includes calculating the peak-to-peak value, kurtosis coefficient, and form factor of the current waveform within a 200ms sliding window, and calculating the standard deviation and rising slope of the temperature data; d) Establish a multi-source feature correlation matrix and use the Pearson correlation coefficient to screen the parameter combinations with strong correlation between temperature changes and electromagnetic field characteristic frequencies.

4. The method according to claim 1, wherein The method for constructing the lightweight neural network model in step (3) includes: a) The network structure adopts the improved MobileNetV3 architecture, replacing the standard convolutional layer with a depth-wise separable convolution and adding a bidirectional LSTM module in the last layer to process temporal features; b) The input layer receives a three-dimensional tensor of dimension [time step × number of features × number of sensor channels], and the time step is set to 10 sampling periods; c) The attention mechanism is specifically implemented as follows: a channel attention module is introduced in the feature fusion layer, channel weights are generated through global average pooling, and a weighting coefficient of 1.2-1.5 is applied to high-frequency electromagnetic features; d) The model is dynamically updated using an online transfer learning strategy. When a change in the heating target material is detected, the base network layer is frozen and only the fully connected layer parameters are fine-tuned.

5. The method according to claim 1, wherein The feedback iterative control of step (4) specifically includes: a) Set up a control rule base based on the fuzzy control algorithm, with the input variable being the predicted temperature deviation and its rate of change, and the output variable being the power regulation amount; b) The iterative learning controller adopts a feedforward-feedback composite structure. After each heating cycle, the initial control parameters for the next cycle are updated based on historical error data. c) Power output optimization is achieved through space vector pulse width modulation (SVPWM), which adjusts the dead time and switching frequency of the IGBT drive signal in real time according to the optimized duty cycle; d) Establish a closed-loop stability monitoring mechanism. When the power fluctuation exceeds ±15% for three consecutive cycles, the control parameters are reset and the baseline parameters sent from the cloud are reloaded.

6. The method according to claim 1, characterized in that The adaptive adjustment of dynamic parameters in step (5) further includes: a) Constructing a six-dimensional state space based on a reinforcement learning algorithm, including current load impedance, coil temperature, target heating rate, power supply voltage fluctuation, historical energy efficiency ratio, and ambient humidity; b) The reward function is designed as: R = α × energy efficiency coefficient + β × temperature tracking accuracy - γ × power fluctuation penalty term, where α, β, and γ are adjustable weight coefficients; c) The online learning strategy uses the double-delayed deep deterministic policy gradient TD3 algorithm, and deploys an experience replay buffer on the edge node to store the most recent 1,000 state-action pairs; d) During the parameter adjustment execution phase, a soft update strategy is implemented for the resonance frequency matching threshold, with each adjustment amplitude not exceeding ±5% of the current value.

7. The method according to claim 1, characterized in that The specific implementation of the edge-cloud collaborative optimization in step (6) includes: a) Data upload uses a differential privacy protection mechanism, adding Laplace noise to temperature-sensitive data. The noise intensity is inversely proportional to the spatial correlation of the data. b) The cloud-based digital twin model construction method is: based on the physical information neural network PINN, Maxwell's equations are embedded as constraints in the network training process; c) The global optimization strategy is generated using a federated learning framework, aggregating local model parameters from no less than 50 edge nodes and updating the global model through an adaptive weighted averaging algorithm; d) The policy is issued in an incremental update mode, only the differences in model parameters are transmitted, and the elliptic curve encryption algorithm is used to ensure transmission security.

8. An intelligent electromagnetic induction heating system based on the method according to any one of claims 1 to 7, characterized in that: include: Sensor module, edge computing module, power drive module and cloud collaboration module; The edge computing module integrates a lightweight neural network reasoning unit and a reinforcement learning decision unit to execute the algorithm process of steps (3)-(5) in claim 1 in real time; The cloud collaboration module is connected to the edge computing module via an encrypted communication protocol, and is used to execute the global optimization strategy generation of step (6) in claim 1.

9. A computer-readable storage device, characterized in that: A computer program is stored, and when the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. An edge computing device for an induction cooker, characterized in that: include: The memory device of claim 9, and at least one embedded processor; The processor calls a program in the storage device to control the power output of the electromagnetic induction coil and execute the real-time computing task of the method according to any one of claims 1 to 7.

Images