Method for controlling an induction cooker based on voice recognition
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHONGSHAN YOULONG KITCHEN APPLIANCES CO LTD
- Filing Date
- 2025-08-27
- Publication Date
- 2026-06-02
Smart Images
Figure CN120954405B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of voice recognition and smart home appliance control technology, and in particular to a voice recognition-based control method for induction cookers. Background Technology
[0002] The voice recognition induction cooker control system is based on acoustic signal processing technology, converting user voice input into executable commands. When a user issues a command, such as adjusting the temperature or starting heating, the integrated microphone captures the sound waveform. Noise is filtered out through preprocessing algorithms, and then pattern recognition methods are applied to extract features and map them to a preset semantic model to identify the intent and parameters. The system then generates digital control signals and transmits them to the induction cooker's microcontroller unit. The microcontroller analyzes the signals and drives the power conversion circuit, dynamically adjusting the induction coil current to achieve precise temperature control. This intelligent interaction mechanism reduces the burden of manual operation and improves cooking efficiency and safety.
[0003] The technical challenge of voice recognition induction cookers in terms of control and adjustment lies in the fact that the accuracy of voice command recognition is affected by environmental noise, which may lead to erroneous execution of power control commands, thereby causing safety risks or operational errors. Specifically, pattern recognition algorithms struggle to effectively distinguish useful voice features from background noise under strong interference. Incorrectly mapping the semantic model results in the transmission of biased control signals to the microcontroller. When the power conversion circuit adjusts the induction coil current accordingly, it may produce unexpected changes in thermal output. For example, in a noisy kitchen environment, the user's voice command "reduce the heat" may be misinterpreted as "increase the heat," causing the microcontroller's drive circuit to increase the current above the set threshold, ultimately leading to localized overheating and increasing the possibility of burning food or damaging the equipment. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a voice recognition-based induction cooker control method, which solves the problems of erroneous triggering of control commands and inaccurate power adjustment in voice recognition induction cookers under complex noise environments.
[0005] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:
[0006] The present invention provides a voice recognition-based induction cooker control method, comprising:
[0007] Simultaneously acquire user voice signals and lip movement signals, and generate multimodal perception data by aligning with timestamps;
[0008] Based on the multimodal perception data, cross-modal noise reduction processing is performed, and a fused feature tensor containing denoised speech features and lip movement features is output.
[0009] The fused feature tensor is input into the semantic understanding process to perform instruction parsing and hazard pattern recognition, generating risk filtering control instructions containing temperature and firepower parameters.
[0010] In response to the parameters of the risk filtering control command, the thermodynamic dynamic model is invoked to calculate the power control parameters;
[0011] The power control parameters are converted into a space vector modulation waveform to drive the induction cooker load circuit, while simultaneously collecting the actual temperature data of the pot bottom.
[0012] The actual temperature data is compared with the target temperature. When the deviation exceeds a threshold, the noise reduction process and the parameter update of the thermodynamic model are triggered.
[0013] Furthermore, in the speech recognition-based induction cooker control method of the present invention, the acquisition of user speech signals and lip movement signals includes:
[0014] Directional beamforming technology is used to process the sound waves received by the microphone array to generate a time-domain speech signal stream; millimeter-wave radar is used to analyze the micro-Doppler effect of the lips and output the coordinate data of the lip movement trajectory.
[0015] A timestamp interpolation alignment operation is performed on the time-domain speech signal stream and lip movement displacement trajectory coordinate data to generate the multimodal sensing data bound with a synchronization timestamp.
[0016] Furthermore, in the speech recognition-based induction cooker control method of the present invention, the cross-modal noise reduction processing includes:
[0017] Speech noise separation is performed on the acoustic signals in the multimodal sensing data to extract noise reduction features;
[0018] Perform continuous frame lip movement analysis on the visual signals in the multimodal perception data to extract lip movement visual feature vectors;
[0019] The noise reduction features and the lip movement visual feature vector are fused through a domain adaptive adversarial training process, and the fused feature tensor is generated by applying a cross-modal weight allocation strategy.
[0020] Furthermore, the voice recognition-based induction cooker control method of the present invention generates risk filtering control commands including:
[0021] Perform end-to-end semantic mapping on the fused feature tensor to generate initial instruction text;
[0022] Search the pre-built cooking knowledge graph to obtain the relationship between the current cookware status and the firepower threshold constraint;
[0023] The system detects whether the initial instruction text violates the firepower threshold constraint relationship through associative reasoning. If it does, the instruction parameters are corrected and the risk filtering control instruction containing temperature / firepower parameters is generated.
[0024] Furthermore, in the speech recognition-based induction cooker control method of the present invention, the search of the pre-constructed cooking knowledge graph includes:
[0025] Retrieve data records from the cookware material thermal conductivity database;
[0026] Perform attention calculations on graph nodes to identify the relationship between cookware status and heat control commands;
[0027] When a combination of an empty pot state and a maximum heat command is detected, the heat parameters in the risk filtering control command are forcibly reset.
[0028] Furthermore, in the voice recognition-based induction cooker control method of the present invention, the calculation of power control parameters by calling the thermodynamic dynamic model includes:
[0029] Based on the temperature parameters in the risk filtering control command and the three-dimensional thermal conductivity of the cookware, a partial differential equation for heat conduction is established.
[0030] The equation is iteratively solved within a preset time window using a rolling optimization algorithm to generate a PWM duty cycle sequence.
[0031] The thermal inertia compensation factor is calculated based on the historical temperature change rate, the PWM duty cycle sequence is adjusted, and the space vector modulation waveform parameters are output.
[0032] Furthermore, in the speech recognition-based induction cooker control method of the present invention, the triggering noise reduction process and the parameter update of the thermodynamic model include:
[0033] Based on the synchronously collected actual temperature data of the bottom of the pot, a temperature field distribution cloud map of the bottom of the pot is constructed;
[0034] A reinforcement learning strategy is used to calculate the control deviation compensation value between the actual temperature and the target temperature.
[0035] When the control deviation exceeds the threshold, the parameters of the thermodynamic model are recalibrated using a meta-learning strategy.
[0036] Furthermore, in the speech recognition-based induction cooker control method of the present invention, the step of calculating the control deviation compensation value using a reinforcement learning strategy includes:
[0037] The control deviation is analyzed by evaluating the network;
[0038] Compensation parameters are generated by outputting the network.
[0039] Calculate the loss function that simultaneously optimizes temperature tracking accuracy and power fluctuation rate, and update the parameters of the evaluation network and the generation network.
[0040] Furthermore, the voice recognition-based induction cooker control method of the present invention also includes a federated evolution mechanism:
[0041] Collect local noise features and encrypt and upload the noise pattern feature vector;
[0042] The cloud performs multi-device data aggregation to generate a global noise reduction model;
[0043] The model parameters for cross-modal noise reduction processing are updated based on the global noise reduction model, and the updated model parameters are periodically sent to the induction cooker terminal.
[0044] Furthermore, in the voice recognition-based induction cooker control method of the present invention, the cloud-based multi-device data aggregation includes:
[0045] Perform differential privacy encryption on the uploaded noise pattern feature vector;
[0046] The parameter transmission amount of the global noise reduction model is compressed by applying characteristic distillation technology;
[0047] When updating the noise reduction model parameters, set the semantic parsing process parameters to a frozen state.
[0048] Beneficial effects of this invention;
[0049] This invention achieves significant benefits through a triple collaborative technology mechanism: a multimodal noise immunity verification mechanism, through millimeter-wave radar lip movement tracking and acoustic feature cross-verification, effectively reduces the false trigger rate of commands caused by noise interference, improving the robustness of speech recognition; knowledge graph-constrained thermodynamic modeling actively intercepts dangerous commands at the semantic parsing layer and dynamically compensates for heat conduction hysteresis at the power control layer, improving power regulation accuracy; federated evolutionary closed-loop optimization recalibrates local model parameters based on real-time temperature deviations and synchronously aggregates noise features from multiple devices to update the global noise immunity model, continuously enhancing the system's ability to adapt to complex environments. This solution achieves technological breakthroughs in three dimensions: command recognition accuracy, power control stability, and system evolution capability, reducing the security risks of voice control. Attached Figure Description
[0050] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.
[0051] Figure 1 A flowchart of an induction cooker control method based on voice recognition provided in an embodiment of the present invention. Detailed Implementation
[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings. To better understand the objectives of this invention, it will be described in further detail below.
[0053] Please see Figure 1 The present invention provides a voice recognition-based induction cooker control method, comprising:
[0054] Step 1: Simultaneously acquire user voice signals and lip movement signals, and generate multimodal perception data by aligning with timestamps;
[0055] Step 2: Based on the multimodal perception data, perform cross-modal noise reduction processing and output a fused feature tensor containing denoised speech features and lip movement features;
[0056] Step 3: Input the fused feature tensor into the semantic understanding process, execute instruction parsing and hazard pattern recognition, and generate risk filtering control instructions containing temperature parameters and firepower parameters;
[0057] Step 4: In response to the parameters of the risk filtering control command, call the thermodynamic dynamic model to calculate the power control parameters;
[0058] Step 5: Convert the power control parameters into a space vector modulation waveform to drive the induction cooker load circuit, while simultaneously collecting the actual temperature data of the pot bottom.
[0059] Step 6: Compare the actual temperature data with the target temperature. When the deviation exceeds the threshold, trigger the noise reduction process and update the parameters of the thermodynamic model.
[0060] This invention utilizes a linearly arranged microphone array to directionally receive acoustic signals and combines this with a generalized sidelobe cancellation algorithm to enhance the sound source in the target direction, generating a time-domain speech signal stream. Simultaneously, a millimeter-wave radar transmits frequency-modulated continuous waves to detect lip movements, and the lip movement trajectory coordinate data is analyzed based on the micro-Doppler effect. A timing alignment module performs timestamp interpolation alignment on both types of signals, binding the speech signal stream to the lip movement trajectory coordinates with synchronized timestamps, thus constructing multimodal sensing data. This multi-source physical signal collaborative acquisition mechanism provides a spatiotemporally unified input foundation for subsequent noise reduction processing.
[0061] Speech noise separation is performed on the acoustic component of multimodal perception data, and a deep neural network is used to extract the denoised Mel-spectral coefficient features. Continuous frame lip movement analysis is performed on the visual component, and a 3D convolutional network is used to extract lip opening and movement velocity feature vectors. A domain-adaptive adversarial training process is used to fuse the acoustic and visual feature vectors, and a cross-modal weight allocation strategy is applied to dynamically balance the contributions of the two types of features, ultimately outputting a fused feature tensor. This cross-modal processing significantly improves the feature robustness in noisy environments.
[0062] The process involves inputting fused feature tensors into an end-to-end semantic mapping flow to generate initial instruction text. A pre-built cooking knowledge graph is queried to obtain the relationship between the thermal conductivity of the cookware material and the firepower threshold constraint, and entity associations are identified through graph node attention calculations. When association inference detects that the initial instruction text triggers a dangerous mode, the temperature and firepower parameters are corrected to generate risk filtering control instructions. This dual-path decision-making mechanism achieves synergy between semantic understanding and safety verification.
[0063] Based on the temperature parameters of the risk-filtering control command and the three-dimensional thermal conductivity of the cookware, a partial differential equation for heat conduction is established. A rolling optimization algorithm is used to iteratively solve the equation within a preset time window, generating an initial duty cycle sequence. The thermal inertia compensation factor is calculated based on the historical temperature change rate, and the space vector modulation waveform parameters of the duty cycle sequence output are adjusted. This predictive power control strategy effectively solves the response hysteresis problem in the thermodynamic system.
[0064] The space vector modulation waveform parameters are converted into drive signals to control the power conversion circuit and adjust the current of the induction coil. Simultaneously, an infrared array scans the bottom area of the pot to construct a temperature field distribution cloud map and collect actual temperature data. This synchronous execution mechanism achieves closed-loop linkage between power output and temperature monitoring.
[0065] The system compares the deviation between the actual temperature data and the target temperature. When the deviation exceeds a preset threshold, a parameter update process is triggered. A reinforcement learning strategy is used to calculate the control deviation compensation value, combined with a meta-learning strategy to recalibrate the thermodynamic model parameters. Simultaneously, a federated evolution mechanism is used to update the cross-modal noise immunity model parameters, forming a dynamic optimization closed loop. This dual update strategy continuously improves the system's noise immunity and temperature control accuracy.
[0066] A four-microphone linear array is used to receive acoustic signals. A generalized sidelobe cancellation algorithm is employed to enhance the sound source in the target direction, outputting a 16kHz sampling rate time-domain speech signal stream. A millimeter-wave radar transmits frequency-modulated continuous waves to detect lip movements, and the lip movement trajectory coordinates are analyzed based on the micro-Doppler effect. A temporal alignment operation performs timestamp interpolation on the speech signal stream and lip movement coordinates, binding synchronization timestamps to generate multimodal sensing data. This collaborative acquisition mechanism provides a spatiotemporally aligned input basis for cross-modal noise immunity.
[0067] Acoustic signals are processed for speech noise separation, and the denoised Mel-spectral coefficients and their first-order difference features are extracted. Continuous-frame lip movement analysis is performed on visual signals, and lip opening and movement velocity feature vectors are extracted using a 3D convolutional network. A domain-adaptive adversarial training process is employed to construct an adversarial learning mechanism between the feature extractor and the domain classifier, minimizing the feature distribution differences between the laboratory environment and a real kitchen. A cross-modal weight allocation strategy is applied to output a fused feature tensor. This processing significantly improves the feature robustness under complex noise environments.
[0068] The end-to-end semantic mapping process transforms the fused feature tensor into an initial instruction text sequence. A pre-built cooking knowledge graph is queried to obtain the thermal conductivity of the cookware material and the current temperature, and a firepower threshold constraint is loaded. An associative reasoning process detects whether the initial instruction triggers a dangerous mode. If a combination of an empty pot state and a maximum firepower instruction is detected, the firepower parameters are corrected to generate a risk-filtering control instruction. This dual-path verification mechanism blocks the execution path of unsafe instructions.
[0069] The system retrieves material data records for cast iron, stainless steel, and other cookware materials from a database of thermal conductivity. It then performs graph node attention calculations to identify the relationship between cookware status and power settings, calculating the attention weight distribution of each node. When a combination of an empty cookware state and a maximum power setting is detected, the power settings are forcibly reset to a safe threshold. This mechanism eliminates the risk of dry burning through knowledge graph constraints.
[0070] Based on the temperature parameters of the risk-filtered control command and the three-dimensional thermal conductivity of the cookware, a partial differential equation for heat conduction incorporating spatial coordinates is established. The equation is iteratively solved within a 200ms time window using a rolling optimization algorithm to generate an initial PWM duty cycle sequence. The thermal inertia compensation factor is calculated based on the near-end historical temperature change rate, and the output space vector modulation waveform parameters of the duty cycle sequence are dynamically adjusted. This predictive control strategy addresses the response hysteresis problem in the thermodynamic system.
[0071] Based on the synchronously acquired actual temperature data of the pot bottom, a temperature field distribution cloud map is constructed through infrared array scanning. A dual-delay deep deterministic gradient algorithm is used to calculate the control deviation compensation value between the actual temperature and the target temperature. When the control deviation exceeds a threshold, a model-independent element learning strategy is applied to recalibrate the parameters of the thermodynamic model. This closed-loop optimization mechanism continuously improves the temperature control accuracy.
[0072] The numerical distribution characteristics of the control bias in the network analysis are evaluated to generate the adjustment amount of the network output compensation parameters. The loss function is simultaneously optimized to improve temperature tracking accuracy and power fluctuation rate, and the weight parameters of the evaluation network and the generation network are updated through gradient backpropagation. This collaborative optimization process balances temperature stability and system energy consumption.
[0073] After cross-modal noise reduction processing, local noise features are collected, and the noise pattern feature vectors are encrypted and uploaded to the cloud. The cloud aggregates data uploaded from multiple devices and generates a global noise reduction model using a federated averaging algorithm. The model parameters for cross-modal noise reduction processing are updated based on the global noise reduction model, and the updated parameters are periodically distributed to the terminal devices. This mechanism enables continuous evolution of noise suppression capabilities.
[0074] Differential privacy encryption is achieved by adding Laplacian noise to the uploaded noise pattern feature vectors. Feature distillation is applied to compress the parameter transmission volume of the global denoising model while retaining key feature dimensions. During the update of the noise-reducing model parameters, the model weights in the semantic parsing process are frozen to prevent parameter updates from interfering with instruction parsing. This design ensures the security of the system upgrade process.
[0075] This invention solves the problems of command mis-triggering and power inaccuracy in noisy environments through a triple technical mechanism, and its technical implementation logic is as follows:
[0076] This design employs a beamforming microphone array to capture speech signals while simultaneously analyzing lip movement trajectories using millimeter-wave radar. The dual-source signals are then time-stamped to generate multimodal perception data. Noise separation based on a deep neural network is performed on the acoustic signals to extract denoised speech features; concurrently, continuous-frame lip movement analysis is conducted on the visual signals to extract visual lip movement features. Domain-adaptive adversarial training is used to fuse the dual-modal features, and a cross-modal weight allocation strategy generates an interference-resistant fused feature tensor. Through acoustic-visual signal cross-validation, this design significantly improves the robustness of command recognition in noisy environments, preventing false command triggering caused by speech feature distortion at its source.
[0077] The initial instruction text is generated through an end-to-end semantic mapping process using fused feature tensors as input, while simultaneously querying the relationship between cookware status and firepower threshold constraints in the cooking knowledge graph. Entity associations are identified through graph node attention calculations. When a dangerous instruction pattern such as "empty pot - maximum firepower" is detected, the firepower parameters are forcibly reset. Responding to the filtered safety instruction parameters, a three-dimensional heat conduction equation is established using a thermodynamic dynamic model. A thermal inertia compensation factor is calculated based on historical temperature change rates, dynamically adjusting the power output curve. This dual-verification mechanism intercepts dangerous instructions at the semantic parsing layer and compensates for heat conduction hysteresis at the power execution layer, eliminating the risk of inaccurate power regulation.
[0078] Based on the infrared temperature field distribution cloud map of the pot bottom, the deviation between the actual temperature and the target temperature is calculated. When the deviation exceeds a threshold, a meta-learning strategy is triggered to recalibrate the thermodynamic model parameters, and a reinforcement learning strategy is simultaneously adopted to optimize the control deviation compensation value. Local noise features are collected, encrypted, and uploaded to the cloud. Data from multiple devices is aggregated through federated learning to generate a global noise reduction model, and the terminal noise reduction processing parameters are periodically updated. This closed-loop system continuously improves noise suppression capabilities and temperature control accuracy through local real-time optimization and cloud-based co-evolution, forming a self-optimizing technical closed loop of noise reduction, control, and evolution.
Claims
1. A method for controlling an induction cooker based on voice recognition, characterized in that, include: Simultaneously acquire user voice signals and lip movement signals, and generate multimodal perception data by aligning with timestamps; Based on the multimodal perception data, cross-modal noise reduction processing is performed, and a fused feature tensor containing denoised speech features and lip movement features is output. The fused feature tensor is input into the semantic understanding process to perform instruction parsing and hazard pattern recognition, generating risk filtering control instructions containing temperature and firepower parameters. The generation of risk filtering control instructions includes: Perform end-to-end semantic mapping on the fused feature tensor to generate initial instruction text; Search the pre-built cooking knowledge graph to obtain the relationship between the current cookware status and the firepower threshold constraint; The initial instruction text is detected by association reasoning to determine whether it violates the firepower threshold constraint relationship. If it does, the instruction parameters are corrected and the risk filtering control instruction containing temperature / firepower parameters is generated. In response to the parameters of the risk filtering control command, the thermodynamic dynamic model is invoked to calculate the power control parameters; The power control parameters are converted into a space vector modulation waveform to drive the induction cooker load circuit, while simultaneously collecting the actual temperature data of the pot bottom. The actual temperature data is compared with the target temperature. When the deviation exceeds a threshold, the noise reduction process and the parameter update of the thermodynamic model are triggered. The search pre-built cooking knowledge graph includes: Retrieve data records from the cookware material thermal conductivity database; Perform attention calculations on graph nodes to identify the relationship between cookware status and heat control commands; When a combination of an empty pot state and a maximum heat command is detected, the heat parameters in the risk filtering control command are forcibly reset. The calculation of power control parameters by calling the thermodynamic dynamic model includes: Based on the temperature parameters in the risk filtering control command and the three-dimensional thermal conductivity of the cookware, a partial differential equation for heat conduction is established. The equation is iteratively solved within a preset time window using a rolling optimization algorithm to generate a PWM duty cycle sequence. The thermal inertia compensation factor is calculated based on the historical temperature change rate, the PWM duty cycle sequence is adjusted, and the space vector modulation waveform parameters are output.
2. The induction cooker control method based on voice recognition according to claim 1, characterized in that, The acquisition of user voice signals and lip movement signals includes: Directional beamforming technology is used to process the sound waves received by the microphone array to generate a time-domain speech signal stream; The micro-Doppler effect of the lip is analyzed by millimeter-wave radar, and the coordinate data of the lip movement trajectory are output. A timestamp interpolation alignment operation is performed on the time-domain speech signal stream and lip movement displacement trajectory coordinate data to generate the multimodal sensing data bound with a synchronization timestamp.
3. The induction cooker control method based on voice recognition according to claim 1, characterized in that, The cross-modal noise reduction processing includes: Speech noise separation is performed on the acoustic signals in the multimodal sensing data to extract noise reduction features; Perform continuous frame lip movement analysis on the visual signals in the multimodal perception data to extract lip movement visual feature vectors; The noise reduction features and the lip movement visual feature vector are fused through a domain adaptive adversarial training process, and the fused feature tensor is generated by applying a cross-modal weight allocation strategy.
4. The induction cooker control method based on voice recognition according to claim 1, characterized in that, The triggered noise reduction process and the parameter update of the thermodynamic model include: Based on the synchronously collected actual temperature data of the bottom of the pot, a temperature field distribution cloud map of the bottom of the pot is constructed; A reinforcement learning strategy is used to calculate the control deviation compensation value between the actual temperature and the target temperature. When the control deviation exceeds the threshold, the parameters of the thermodynamic model are recalibrated using a meta-learning strategy.
5. The induction cooker control method based on voice recognition according to claim 4, characterized in that, The calculation of the control bias compensation value using a reinforcement learning strategy includes: The control deviation is analyzed by evaluating the network; Compensation parameters are generated by outputting the network. Calculate the loss function that simultaneously optimizes temperature tracking accuracy and power fluctuation rate, and update the parameters of the evaluation network and the generation network.
6. The induction cooker control method based on voice recognition according to claim 1, characterized in that, It also includes the federated evolution mechanism: Collect local noise features and encrypt and upload the noise pattern feature vector; The cloud performs multi-device data aggregation to generate a global noise reduction model; The model parameters for cross-modal noise reduction processing are updated based on the global noise reduction model, and the updated model parameters are periodically sent to the induction cooker terminal.
7. The induction cooker control method based on voice recognition according to claim 6, characterized in that, The cloud-based multi-device data aggregation includes: Perform differential privacy encryption on the uploaded noise pattern feature vector; The parameter transmission amount of the global noise reduction model is compressed by applying characteristic distillation technology; When updating the noise reduction model parameters, set the semantic parsing process parameters to a frozen state.