Intelligent control system and method for cooking abnormity

Abstract

The invention discloses a cooking abnormity intelligent control system and method, belongs to the technical field of cooking intelligent control, and constructs a double-layer judgment mechanism fusing a global behavior model and a user local model; according to the system, multi-dimensional real-time data in the operation process of cooking equipment is collected, a global model is used for recognizing a general cooking rule, a local model is built based on user personalized data, and dynamic adaptation is achieved through online learning; when an abnormal state is recognized, judgment results of the two models are fused to generate an abnormal credibility score, matched control response is executed, and local model parameters are continuously optimized in combination with user feedback, so that the abnormal recognition accuracy and response flexibility are effectively improved, the false alarm rate is remarkably reduced, and the user experience is improved. And the self-adaptive capability of the system to user habits, environment changes and cookware differences is enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for cooking, specifically to an intelligent control system and method for cooking anomalies. Background Technology

[0002] Intelligent control of cooking anomalies refers to the system's ability to monitor the equipment's operating status and cooking parameters (such as temperature, time, smoke, and odor) in real time during the cooking process using sensors and intelligent algorithms. Once an anomaly is detected (such as dry burning, overflowing, or abnormal heat), the system can automatically identify and take corresponding control measures, such as alarms, adjusting heat, or automatic power-off, to ensure cooking safety and improve cooking efficiency and user experience.

[0003] The existing technology has the following shortcomings:

[0004] In existing technologies, when using historical cooking data and behavioral models to identify anomalies, user cooking habits, environmental conditions, or cookware types often change, causing deviations between current normal operations and historical models. Since the models cannot adapt in real time, they are prone to misjudging normal behavior as abnormal. For example, if a user changes the cooking pot or the temperature changes, causing a shift in the heating curve, the system might incorrectly trigger a power outage, affecting the cooking process. This type of misjudgment arises from the model's weak generalization ability and lack of online adaptive capability. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent control system and method for cooking abnormalities, so as to overcome the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent control of cooking anomalies, comprising:

[0007] Collect multi-dimensional real-time data generated during the operation of cooking equipment;

[0008] A global model is constructed, which is formed through offline training based on historical cooking data from multiple users;

[0009] Based on current user cooking behavior data, a local user model is constructed and continuously updated. The local model is incrementally trained using online learning methods.

[0010] By comparing real-time collected data with historical predicted behavior curves, and combining the global model and local model to output corresponding global anomaly confidence scores and local anomaly confidence scores, the final anomaly confidence score is generated after fusion processing to identify abnormal states.

[0011] When an abnormal state is detected, corresponding control measures are implemented according to the type of abnormality, and the local model is continuously optimized based on user feedback behavior data.

[0012] Preferably, the multidimensional real-time data includes, but is not limited to, temperature, humidity, smoke concentration, heating power, and user operation records.

[0013] Preferably, the real-time data collected during the current cooking process is segmented and processed using a fixed sliding time window to extract local abnormal behavior features;

[0014] Localized anomalous behavior characteristics include the frequency spectral density of fire fluctuations and the temperature response delay deviation index; the method for obtaining the frequency spectral density of fire fluctuations is as follows:

[0015] Within a fixed-length sliding time window T, a continuous heating power data sequence P = {p1, p2, ..., p...} is collected. N The sampling frequency is f. s ;

[0016] Performing a Fast Fourier Transform on the power sequence yields its frequency domain amplitude spectrum A = {|F1|,|F2|,...,|F...} M |}, where M = N / 2; normalize the amplitude spectrum to form a probability distribution: In the formula, p i The spectral entropy represents the energy percentage of the i-th frequency component; the spectral entropy is calculated as the power fluctuation frequency spectral density, expressed as: In the formula, HPSE represents the power fluctuation frequency spectral density.

[0017] The preferred method for obtaining the temperature response delay deviation index is as follows:

[0018] Record the event timestamp tu each time a user action event occurs;

[0019] In the data sequence of the bottom temperature sensor, find the time point tr at which the temperature curve first shows a significant change after the operation, and define the response point:

[0020] The response point determination condition is that the rate of temperature change dT exceeds the set threshold δ; calculate the response delay time: Δt = tr - tu; and calculate the temperature response delay deviation index, the expression of which is: In the formula, TRLD represents the temperature response delay deviation exponent, μ Δt σ Δt The historical average response time and standard deviation for users using the same cookware / mode.

[0021] Preferably, the local abnormal behavior features are input into the global model and the local model respectively to obtain the global anomaly credibility score and the local anomaly credibility score; the final anomaly credibility score is obtained by using a weighted average summation.

[0022] Preferably, anomaly judgment is made based on the anomaly confidence score of the final output:

[0023] If the score is lower than the first preset threshold, it is considered normal behavior;

[0024] If the score is higher than the second preset threshold, it is judged as abnormal behavior and a control response is triggered;

[0025] If the score falls between the first and second thresholds, prompt the user to confirm the current status.

[0026] Preferably, driven by user feedback, the feature decision weights in the local model are dynamically adjusted through a reinforcement learning mechanism, including:

[0027] The alarm behavior is treated as an action in reinforcement learning, and user feedback is used as a reward signal.

[0028] The long-term reward value is estimated using the time difference method. If the current alarm strategy corresponds to frequent negative rewards, the strategy weight of the feature is reduced.

[0029] When three or more false alarms are detected consecutively and all are clearly reported as normal by the user, the system will automatically enter the fault tolerance learning phase.

[0030] Preferred fault-tolerant learning mechanisms include:

[0031] Aggregate and analyze the characteristic time window data involved in multiple consecutive false alarms, and re-cluster based on similarity to construct a high-temperature stir-fry behavior feature cluster;

[0032] The feature clusters are removed from the original abnormal categories and relabeled as normal user preferences. They are then written into the local model as training samples to replace the classification criteria that previously caused false alarms.

[0033] Based on users' historical false alarm data, the tolerance threshold for this scenario will be increased, and the allowed duration will be extended.

[0034] The present invention also provides an intelligent control system for cooking anomalies, including a data acquisition module, a global behavior modeling module, a user local model update module, an anomaly scoring fusion judgment module, and a feedback optimization module;

[0035] Data acquisition module: Collects multi-dimensional real-time data generated by the cooking equipment during operation;

[0036] Global Behavior Modeling Module: Constructs a global model, which is formed through offline training based on historical cooking data from multiple users;

[0037] User Local Model Update Module: Based on the current user's cooking behavior data, a user local model is constructed and continuously updated. The local model is incrementally trained through online learning methods.

[0038] Anomaly scoring fusion judgment module: It compares real-time collected data with historical predicted behavior curves, combines the global model and local model to output corresponding anomaly credibility scores, and then fused them to generate the final anomaly credibility score in order to identify abnormal states.

[0039] Feedback optimization module: When an abnormal state is detected, corresponding control measures are implemented according to the type of abnormality, and the local model is continuously optimized based on user feedback behavior data.

[0040] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0041] 1. This invention collects multidimensional real-time data of cooking equipment during operation and extracts local behavioral features (such as the frequency spectrum density of fire fluctuation and the temperature response delay deviation index) using a sliding time window. This invention not only achieves refined anomaly identification, but also dynamically combines the scoring results of global and personalized models to output continuous credibility judgments, thereby improving the accuracy and adaptability of anomaly detection.

[0042] 2. This invention introduces a user feedback participation mechanism after detecting anomalies. Combining reinforcement learning and fault-tolerant learning strategies, the local model is continuously self-optimized, thereby supporting the model's learning and acceptance of users' personalized behaviors such as high-temperature cooking, and avoiding repeated misjudgments. This feedback-driven closed-loop structure endows the system with high intelligence, personalization, and adaptability, significantly improving the robustness, safety, and user satisfaction of intelligent kitchen equipment in complex real-world usage environments. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0044] Figure 1 This is a mind map of the method of the present invention.

[0045] Figure 2 This is a mind map of the system modules of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1, please refer to Figure 1 As shown in this embodiment, a cooking anomaly intelligent control method includes:

[0048] Collect multi-dimensional real-time data generated during the operation of cooking equipment;

[0049] A global model is constructed, which is formed through offline training based on historical cooking data from multiple users;

[0050] Based on current user cooking behavior data, a local user model is constructed and continuously updated. The local model is incrementally trained using online learning methods.

[0051] By comparing real-time collected data with historical predicted behavior curves, and combining the global and local models to output corresponding anomaly confidence scores, the final anomaly confidence score is generated after fusion processing to identify abnormal states.

[0052] When an abnormal state is detected, corresponding control measures are implemented according to the type of abnormality, and the local model is continuously optimized based on user feedback behavior data.

[0053] Cooking equipment uses various built-in sensor modules and control units to collect key operating parameters in real time during operation, and transmits the data to a central control system or cloud platform via a local processor or wireless communication module (such as Wi-Fi, Bluetooth, Zigbee). The main data collection methods include:

[0054] Various sensors are embedded in cookware, heating panels, control boards, and other locations to collect physical parameters in real time.

[0055] The control chip obtains data from the internal control bus (such as I) via a sampling program. 2 (C, SPI, UART) Read device status and user operation data.

[0056] User actions via touchscreen, knob, voice commands, or remote control via app are recorded in real time.

[0057] Preprocessing and caching of some data at the local edge reduces communication load and enables real-time response.

[0058] The multidimensional data collected by this method includes, but is not limited to, the following categories:

[0059] Thermal parameters include: pot bottom temperature, heating plate temperature, ambient temperature; temperature change rate (derived characteristics); oil temperature or liquid temperature (via infrared or contact temperature sensor).

[0060] Humidity and gas dimensions: air humidity (reflects whether there is overflow or steam evaporation); smoke concentration (identifies phenomena such as burnt food and oil fumes); combustible gas concentration (used for gas leak monitoring).

[0061] Electrical parameters: heating power, current, voltage; power fluctuation mode (analysis of induction cooker heating mode and load changes);

[0062] Time dimension: cooking duration; start / end timestamps for each stage; user departure or pause time records.

[0063] User behavior dimensions: operation frequency and sequence (e.g., adding water, stir-frying, opening the lid); selected cooking mode (e.g., boiling, stir-frying, deep-frying); operation preferences (e.g., whether to frequently adjust the heat).

[0064] Environmental and equipment status dimensions: current cookware type (determined by RFID identification or temperature rise characteristics); ambient light / sound information (which can help determine whether the user is present); current equipment operating status (standby, heating, heat preservation, etc.).

[0065] The data acquisition system consists of a data acquisition layer, an edge processing layer, and an upload interaction layer: The data acquisition layer consists of sensor units and operating interfaces distributed inside the device that capture data in real time; The edge processing layer consists of a local processor that performs preliminary anomaly detection, data cleaning, and caching; The upload interaction layer consists of a communication module that transmits preprocessed data to the cloud-based model judgment module in real time.

[0066] Constructing a global model is one of the core technologies for realizing the intelligent control of this invention. The global model is used to characterize the "normal cooking behavior patterns" that are prevalent among a wide range of users, and to provide a basic reference for subsequent personalized modeling.

[0067] Historical data is collected from smart cooking devices of multiple real users, covering different dishes, cookware, operating habits and environmental conditions;

[0068] The data format is standardized, and the fields include temperature change curves, humidity / smoke concentration, power curves, user operation behavior, cooking modes, and anomaly markers.

[0069] Data cleaning and normalization: handling null values, outliers, and duplicate values; standardizing units and formats for data from different devices; aligning time series data (e.g., unifying sampling intervals through interpolation).

[0070] Time series feature extraction: Extract one-dimensional or multi-dimensional time series statistical features such as "rate of temperature rise", "frequency of power change", and "duration of sustained high temperature"; use sliding window to extract local dynamic features, such as maximum / minimum values ​​and coefficient of variation within the window.

[0071] Behavioral feature modeling: Construct a combination of multi-dimensional feature vectors to represent the behavioral curve of the complete cooking cycle; generate typical cooking behavior templates (e.g.: cooking rice ≈ slow heating → maintaining boiling point → cooling and heat preservation).

[0072] Use expert annotations or device-defined rules (such as alarm records) to add “abnormal” or “normal” behavior tags to historical data; the tags can be further refined into multiple categories such as “dry burning”, “overflowing”, “excessively high oil temperature”, and “normal cooking”.

[0073] Depending on the specific data scale and feature dimensions, one or a combination of the following models can be selected:

[0074] Decision trees / random forests: used for structured rule determination, easy to interpret; Support vector machines (SVM): suitable for binary classification anomaly detection;

[0075] Neural network models: MLP (Multilayer Perceptron) is used for static feature classification; LSTM (Long Short-Term Memory) network is used to capture sequential patterns such as temperature / humidity that change over time; Transformer is used for more complex long-term behavior modeling.

[0076] Clustering models (such as K-Means and DBSCAN): discover high-frequency "normal behavior groups" under unsupervised learning.

[0077] The processed data is input into the selected model for training; cross-validation and grid search are used to optimize parameters and improve accuracy and generalization ability.

[0078] Use evaluation metrics such as accuracy, recall, and F1 score; pay special attention to false positive rate to prevent normal behavior from being misidentified as abnormal.

[0079] The trained model is abstracted to output "standard behavior pattern templates" corresponding to dish types, cooking stages, etc.; each template includes: temperature range, rise curve shape, power adjustment range, etc.

[0080] Lightweight processing such as quantization and pruning is performed on neural network models to facilitate deployment to edge devices; the models can be deployed on local devices for real-time initial screening or deployed in the cloud for depth judgment.

[0081] Building and continuously updating a local user model is a key technological approach to achieving personalized cooking anomaly detection. This local user model is incrementally trained using online learning methods, constantly adapting to changes in user operating habits, cookware characteristics, and cooking environment, thereby effectively reducing misjudgments and improving user experience.

[0082] Behavioral data collection begins the moment a user first uses the device, including temperature / humidity change curves, operation frequency, cookware type, and selected cooking mode. If the user agrees to authorize, historical cooking records related to their account can also be accessed in the cloud.

[0083] If there is insufficient data to train the model, global model transfer learning is used: global model parameters are loaded as initial weights for the local model; the global model is fine-tuned with the current user data to form a personalized initial model.

[0084] Each cooking operation generates a complete time series data, including: early cooking characteristics (heating rate, cookware recognition); mid-stage behavior (heat adjustment, operation frequency); late-stage temperature changes and control behavior (keeping warm, ending); and user feedback (such as whether false alarms occurred or cooking was interrupted).

[0085] Instead of retraining the local model entirely, the system achieves rapid adaptation through the following online learning methods:

[0086] It employs algorithmic structures that support online updates, such as: Online XGBoost; Stochastic Gradient Descent (SGD) fine-tuning of neural networks; Hoeffding Tree (for incremental learning of decision trees); after each new data input, only the relevant model parameters are adjusted, maintaining high training efficiency and low memory usage.

[0087] The local model uses only the most recent N (e.g., 10 to 50) cooking behavior data as the sample pool; old data is periodically discarded to prevent the model from overfitting to historical abnormal behaviors.

[0088] Extract user feature vectors from continuously recorded behaviors, such as: preferred heat level (medium, high); average heating time; frequency of stirring and percentage of time spent with the lid off; whether they prefer preheating / fast heating mode.

[0089] By using user feature vectors as part of the model input and concatenating them with real-time behavioral features, the model's sensitivity to "user habits" is enhanced, achieving true personalization.

[0090] When the system misjudges a user's action as abnormal: it prompts the user to confirm whether it is a false alarm; after the user confirms "continue cooking" or "false alarm," the system labels this sample and feeds it back into the model for incremental training. For repeated "false alarm behaviors," the system will gradually lower its abnormal confidence level; the model will gradually learn that "this is a normal user operation" and avoid repeated interruptions.

[0091] By comparing real-time collected data with historical predicted behavior curves, and combining global and local models, an anomaly confidence score is output and processed, including:

[0092] The real-time data collected during the current cooking process is segmented and processed by using a fixed sliding time window to extract local abnormal behavior features;

[0093] Localized anomalous behavior characteristics include the frequency spectral density of fire fluctuations and the temperature response delay deviation index; the method for obtaining the frequency spectral density of fire fluctuations is as follows:

[0094] Within a fixed-length sliding time window T, a continuous heating power data sequence P = {p1, p2, ..., p...} is collected. N The sampling frequency is f. s .

[0095] Performing a Fast Fourier Transform (FFT) on the power sequence yields its frequency domain amplitude spectrum A = {|F1|,|F2|,...,|F...}. M |}, where M = N / 2; normalize the amplitude spectrum to form a probability distribution: In the formula, p i The spectral entropy represents the energy percentage of the i-th frequency component; the spectral entropy is calculated as the power fluctuation frequency spectral density, expressed as: In the formula, HPSE represents the power fluctuation frequency spectral density.

[0096] The method for obtaining the temperature response delay deviation index is as follows:

[0097] Record the event timestamp tu each time a user action event (such as "heating start", "pause", "firepower adjustment") occurs.

[0098] In the data sequence of the bottom temperature sensor, find the time point tr at which the temperature curve first shows a significant change after this operation, and define the response point:

[0099] The response point determination condition is that the rate of temperature change dT exceeds the set threshold δ; calculate the response delay time: Δt = tr - tu; and calculate the temperature response delay deviation index, the expression of which is: In the formula, TRLD represents the temperature response delay deviation exponent, μ Δt σ ΔtThe historical average response time and standard deviation for users using the same cookware / mode.

[0100] The local abnormal behavior features are input into the global model and the local model respectively to obtain the global anomaly credibility score and the local anomaly credibility score; the final anomaly credibility score is obtained by using a weighted average summation.

[0101] Based on the set weighted fusion strategy, the two scores are merged according to their weights to generate the final anomaly credibility score. The weighted fusion strategy includes: calculating the similarity between the current local abnormal behavior features and the historical average of the user's local abnormal behavior; if the similarity is higher than a preset threshold, the weight of the local model score is increased and the weight of the global model score is decreased; if the similarity is lower than the threshold, the weight of the global model score is increased.

[0102] Before generating the anomaly credibility score, the following steps are further included:

[0103] The Dynamic Time Warping (DTW) algorithm is used to align the real-time behavior curve with the historical predicted behavior curve.

[0104] The aligned behavior deviation value is calculated and used as an adjustment factor for the anomaly credibility score to improve the system's fault tolerance to nonlinear behavior.

[0105] The final output anomaly confidence score is represented by a continuous real number between 0 and 1, and the judgment is made based on the score as follows:

[0106] a) If the score is lower than the first preset threshold, it is judged as normal behavior;

[0107] b) If the score is higher than the second preset threshold, it is judged as abnormal behavior and a control response is triggered;

[0108] c) If the score is between the first and second thresholds, prompt the user to confirm the current status.

[0109] By combining "anomaly credibility score" and "anomaly type label", the corresponding control response strategy is selected from the predefined anomaly type-response mapping library, and the local model is continuously optimized based on user feedback behavior, thus realizing intelligent and personalized security control.

[0110] Taking a smart induction cooker as an example, during the "frying" operation, the system collects real-time data on the temperature changes at the bottom of the pot through temperature sensors and a power monitoring module. This data is then input into the global behavior model and the user's local model for joint analysis. The system determines the current state as "abnormally high oil temperature" and calculates an anomaly confidence score of 0.82, exceeding the system's preset high-risk threshold (0.75). The corresponding anomaly type is labeled as "overheating risk".

[0111] At this point, the control strategy selection module will call the exception type-response mapping library. Based on the current exception type "overheating risk" and its corresponding response strategy level, the system will prioritize matching strategy e: "immediate power off". However, due to the large number of previous user feedback records of "high-temperature stir-frying is normal operation", the system will lower the response level and prioritize calling strategy e: "delay processing and prompt user confirmation".

[0112] The system pops up a prompt: "Abnormal temperature rise detected, continue cooking?" The user clicks "Continue," indicating that the action was a false alarm. This action is treated as user feedback and is fed back into the user's local model by the system, triggering the following two steps of processing:

[0113] First, the system dynamically fine-tunes the confidence boundary value of the "abnormal rise in oil temperature" type, raising the trigger threshold for the current user from the original 0.75 to 0.83 to adapt to their high-temperature preference characteristics and reduce false alarms in similar situations in the future.

[0114] Secondly, the system marks the feature vectors extracted within this sliding time window, such as the rate of temperature rise and power stability, as "non-anomaly" in the local model and reduces their contribution weight to the model's anomaly classification results. This way, when similar features reappear in the future, the system is more likely to classify them as normal behavior unique to the user.

[0115] This feedback processing is based on the dynamic fusion of historical behavior and current data, ensuring that the system has both general anomaly response capabilities and can be continuously optimized for individual users, thus achieving the goal of highly reliable and highly personalized intelligent cooking control.

[0116] The user-local model not only receives feedback to correct parameter thresholds, but also introduces a reinforcement learning mechanism to dynamically adjust model weights. In the event of continuous misjudgments, it automatically enters the fault-tolerant learning stage to ensure that the control strategy continuously adapts to user habits.

[0117] Taking the example of multiple false alarms during the "stir-frying chili peppers" process, the system judged "overheating + abnormal oil smoke" based on the characteristics of a sharp rise in the temperature of the bottom of the pot and an increase in the concentration of smoke in each operation. The credibility score of the abnormality often exceeded 0.80. The original strategy of the system was to "reduce the heat and issue a prompt".

[0118] However, the user consistently provided clear feedback indicating "no abnormality" and clicked "continue cooking," demonstrating their preferred "high-temperature stir-fry" behavior. At this point, the system initiates a reinforcement learning mechanism to optimize the local model weight configuration:

[0119] The system treats the action of "whether to issue an alarm" as an agent action in reinforcement learning, while the user's feedback to the system's judgment (continue / interrupt / ignore) serves as a reward signal. For example: user selects "continue" → negative reward (the current alarm is a false alarm); user selects "interrupt" → positive reward (the alarm is a genuine anomaly); no feedback → 0 reward or light handling.

[0120] The system uses the Temporal Difference (TD) algorithm to update the policy value function. That is, after several consecutive user feedbacks, it predicts whether the long-term reward of the system executing the "alarm" policy under the current feature input will decrease. If the decrease is significant, the behavior policy function will be adjusted.

[0121] After the strategy function is updated, the system makes corresponding fine adjustments to the weights of various features (such as high temperature rate, smoke slope, etc.) involved in alarm judgment in the model, reducing their weight in judging "high temperature stir-fry" type behavior, and enhancing the model's ability to recognize and adapt to the user's operating style.

[0122] Furthermore, when the system detects three or more false alarms that are explicitly reported as "non-abnormal" by the user, it will automatically enter the fault-tolerance learning phase and initiate the following operations:

[0123] Aggregate and analyze the characteristic time window data involved in multiple consecutive false alarms, and re-cluster based on similarity to construct a high-temperature stir-fry behavior feature cluster;

[0124] The feature cluster is removed from the original abnormal category and relabeled as normal user preferences. It is then written into the local model as a training sample to replace the classification criteria that previously caused false alarms.

[0125] Based on the user's historical false alarm data, the tolerance threshold for this type of scenario is dynamically increased. For example, the abnormal temperature warning line is raised from 190℃ to 205℃, and the allowable duration is extended, so that false alarms will no longer occur in similar scenarios in the future.

[0126] Through the aforementioned mechanism, the system not only enhances the model's sensitivity to real anomalies but also proactively adapts to users' unique cooking behaviors, improving alarm accuracy and user experience. This feedback loop + strategy evolution control system is significantly different from traditional static judgment logic, exhibiting high intelligence and adaptability.

[0127] Example 2, please refer to Figure 2As shown in the figure, the intelligent control system for cooking anomalies described in this embodiment includes a data acquisition module, a global behavior modeling module, a user local model update module, an anomaly scoring fusion judgment module, and a feedback optimization module.

[0128] Data acquisition module: Collects multi-dimensional real-time data generated by the cooking equipment during operation;

[0129] Global Behavior Modeling Module: Constructs a global model, which is formed through offline training based on historical cooking data from multiple users;

[0130] User Local Model Update Module: Based on the current user's cooking behavior data, a user local model is constructed and continuously updated. The local model is incrementally trained through online learning methods.

[0131] Anomaly scoring fusion judgment module: It compares real-time collected data with historical predicted behavior curves, combines the global model and local model to output corresponding anomaly credibility scores, and then fused them to generate the final anomaly credibility score in order to identify abnormal states.

[0132] Feedback optimization module: When an abnormal state is detected, corresponding control measures are implemented according to the type of abnormality, and the local model is continuously optimized based on user feedback behavior data.

[0133] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0134] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0135] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0136] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent control of cooking anomalies, characterized in that: include: Collect multi-dimensional real-time data generated during the operation of cooking equipment; A global model is constructed, which is formed through offline training based on historical cooking data from multiple users; Based on current user cooking behavior data, a local user model is constructed and continuously updated. The local model is incrementally trained using online learning methods. By comparing real-time collected data with historical predicted behavior curves, and combining the global model and local model to output corresponding global anomaly confidence scores and local anomaly confidence scores, the final anomaly confidence score is generated after fusion processing to identify abnormal states. When an abnormal state is detected, corresponding control measures are implemented according to the type of abnormality, and the local model is continuously optimized based on user feedback behavior data.

2. The intelligent control method for cooking abnormalities according to claim 1, characterized in that: The multidimensional real-time data includes, but is not limited to, temperature, humidity, smoke concentration, heating power, and user operation records.

3. The intelligent control method for cooking abnormalities according to claim 1, characterized in that: The real-time data collected during the current cooking process is segmented and processed by using a fixed sliding time window to extract local abnormal behavior features; Localized anomalous behavior characteristics include the frequency spectral density of fire fluctuations and the temperature response delay deviation index; the method for obtaining the frequency spectral density of fire fluctuations is as follows: Within a fixed-length sliding time window T, a continuous heating power data sequence P = {p1, p2, ..., p...} is collected. N The sampling frequency is f. s ; Performing a Fast Fourier Transform on the power sequence yields its frequency domain amplitude spectrum A = {|F1|,|F2|,...,|F...} M |}, where M = N / 2; normalize the amplitude spectrum to form a probability distribution: In the formula, p i The spectral entropy represents the energy percentage of the i-th frequency component; the spectral entropy is calculated as the power fluctuation frequency spectral density, expressed as: In the formula, HPSE represents the power fluctuation frequency spectral density.

4. The intelligent control method for cooking abnormalities according to claim 3, characterized in that: The method for obtaining the temperature response delay deviation index is as follows: Record the event timestamp tu each time a user action event occurs; In the data sequence of the bottom temperature sensor, find the time point tr at which the temperature curve first shows a significant change after the operation, and define the response point: The response point determination condition is that the rate of temperature change dT exceeds the set threshold δ; calculate the response delay time: Δt = tr - tu; and calculate the temperature response delay deviation index, the expression of which is: In the formula, TRLD represents the temperature response delay deviation exponent, μ Δt σ Δt The historical average response time and standard deviation for users using the same cookware / mode.

5. The intelligent control method for cooking abnormalities according to claim 4, characterized in that: The local abnormal behavior features are input into the global model and the local model respectively to obtain the global anomaly credibility score and the local anomaly credibility score; the final anomaly credibility score is obtained by using a weighted average summation.

6. The intelligent control method for cooking abnormalities according to claim 5, characterized in that: Anomaly detection is performed based on the final anomaly confidence score: If the score is lower than the first preset threshold, it is considered normal behavior; If the score is higher than the second preset threshold, it is judged as abnormal behavior and a control response is triggered; If the score falls between the first and second thresholds, prompt the user to confirm the current status.

7. The intelligent control method for cooking abnormalities according to claim 6, characterized in that: Driven by user feedback, the feature decision weights in the local model are dynamically adjusted through a reinforcement learning mechanism, including: The alarm behavior is treated as an action in reinforcement learning, and user feedback is used as a reward signal. The long-term reward value is estimated using the time difference method. If the current alarm strategy corresponds to frequent negative rewards, the strategy weight of the feature is reduced. When three or more false alarms are detected consecutively and all are clearly reported as normal by the user, the system will automatically enter the fault tolerance learning phase.

8. The intelligent control method for cooking abnormalities according to claim 7, characterized in that: Fault-tolerant learning mechanisms include: Aggregate and analyze the characteristic time window data involved in multiple consecutive false alarms, and re-cluster based on similarity to construct a high-temperature stir-fry behavior feature cluster; The feature clusters are removed from the original abnormal categories and relabeled as normal user preferences. They are then written into the local model as training samples to replace the classification criteria that previously caused false alarms. Based on users' historical false alarm data, the tolerance threshold for this scenario will be increased, and the allowed duration will be extended.

9. A cooking anomaly intelligent control system, used to implement the cooking anomaly intelligent control method according to any one of claims 1-8, characterized in that: It includes a data acquisition module, a global behavior modeling module, a user local model update module, an anomaly score fusion and judgment module, and a feedback optimization module; Data acquisition module: Collects multi-dimensional real-time data generated by the cooking equipment during operation; Global Behavior Modeling Module: Constructs a global model, which is formed through offline training based on historical cooking data from multiple users; User Local Model Update Module: Based on the current user's cooking behavior data, a user local model is constructed and continuously updated. The local model is incrementally trained through online learning methods. Anomaly scoring fusion judgment module: It compares real-time collected data with historical predicted behavior curves, combines the global model and local model to output corresponding anomaly credibility scores, and then fused them to generate the final anomaly credibility score in order to identify abnormal states. Feedback optimization module: When an abnormal state is detected, corresponding control measures are implemented according to the type of abnormality, and the local model is continuously optimized based on user feedback behavior data.

Images