An anti-overflow control method, a storage medium and an anti-overflow system

Abstract

The present application relates to a kind of anti-overflow control methods, the real-time audio signal of cooking container is collected and is subjected to mel spectrum transformation, to obtain mel spectrum;Mel spectrum is multiplied with pot resonance mask vector point by point, to obtain mask after mel spectrum, again with adaptive weight vector point by point, to obtain enhanced mel spectrum;The normalized relative intensity of enhanced mel spectrum is calculated, and smooth quantization evaluation value QE is generated;Enhanced mel spectrum is input into identification network to identify overflow state, and combined with QE value is jointly decided, to generate firepower control instruction;After cooking, use the near-overflow acoustic print data in successful anti-overflow event as self-supervision sample, to the online fine-tuning of the last layer fully connected layer weight of identification network is carried out.The anti-overflow control method of pot adaptability cavity, scene generalization ability cavity, false alarm and miss low in complex kitchen environment.The present application also relates to storage medium and the anti-overflow control system of the anti-overflow control method.

Description

Technical Field

[0001] This invention relates to the field of intelligent cooking control technology, specifically to an anti-overflow control method, storage medium, and anti-overflow system. Background Technology

[0002] In everyday cooking, overflowing soup is a common safety hazard and a major pain point. When liquid in a pot boils, the violent bubbling causes it to overflow onto the stovetop, resulting in food waste, increased cleaning burden, and potentially gas leaks caused by the stove going out. Therefore, accurate and timely warnings and automatic control of overflowing soup are of significant practical value.

[0003] Currently, kitchen spill prevention technology based on acoustic principles has seen some research and application. Existing solutions can generally be summarized into two stages: "feature pre-setting" and "rigid matching." At the signal acquisition end, the system typically relies on a fixed, single microphone to collect ambient audio, assuming that the sound propagation path and sound source characteristics remain constant. In the feature processing stage, mainstream solutions pre-set fixed acoustic characteristic parameters. For example, acoustic data collected from boiling standard cookware and standard liquids (such as water) in a laboratory environment is analyzed to obtain features such as the energy concentration band, main peak frequency, or time-domain envelope shape, and these characteristic parameters are then embedded into the algorithm.

[0004] At the recognition model level, existing technologies mainly employ two approaches: one is a rule-based judgment method based on fixed thresholds, which sets intensity or rate of change thresholds for sound energy in a specific frequency band, triggering an alarm when the real-time detected value exceeds the threshold; the other is a pattern recognition method based on pre-trained static classification models, such as support vector machines or shallow neural networks. These models are trained on standard datasets before leaving the factory, treating cookware type, liquid type, and environmental noise as fixed variables or interference terms in the training data, and the model weights remain unchanged once deployed. At the system integration level, existing solutions are typically embedded in the cooktop as independent modules with fixed functions. Their operating logic is relatively closed, assuming that all users use cookware similar to the test environment and follow standard cooking procedures.

[0005] However, the above-mentioned existing technical solutions face the following technical problems in practical applications.

[0006] First, there is insufficient hardware adaptability. The acoustic feature templates embedded in the algorithm are based on a limited number of standard cookware types. However, in actual use, cookware made of different materials (such as stainless steel, glass, cast iron, ceramic, etc.), with different shapes and sizes, has significantly different inherent resonance characteristics. These differences will produce different resonance peak distributions in the acoustic frequency domain. When users change cookware, the changes in the inherent resonance characteristics of the cookware will seriously interfere with the boiling sound signature characteristics, leading to a significant decrease in recognition performance.

[0007] Second, the model's generalization ability is limited. Pre-trained static models cannot continue to learn new variables introduced by users in actual use after deployment, such as special ingredient recipes, thick broths, and specific heat preferences—all personalized cooking habits. When faced with cooking patterns not covered in the training dataset, the recognition accuracy of static models will significantly decrease, making it difficult to meet diverse practical needs.

[0008] Third, there is insufficient environmental robustness. Existing solutions treat sound signals, cookware characteristics, and user habits as independent or fixed factors, lacking the comprehensive ability to address the dynamic coupling of multiple factors in a real kitchen environment. In the complex conditions of a real kitchen with multiple background noise sources and frequently changing cooking conditions, a single technical means is difficult to achieve stable and reliable early warning, and is prone to false alarms and missed alarms, affecting user experience and user trust.

[0009] Therefore, how to effectively eliminate the interference of the inherent resonance characteristics of different cookware on voiceprint recognition on resource-constrained edge devices, enable the system to have adaptive learning capabilities for personalized cooking scenarios, and achieve stable anti-overflow warning with high accuracy and low false alarm rate in complex and ever-changing real kitchen environments is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0010] The first technical problem to be solved by the present invention is to provide an anti-overflow control method in response to the above-mentioned prior art, so as to solve the technical problems of poor hardware adaptability, weak scene generalization ability, and frequent false alarms and missed alarms in the real kitchen environment due to the use of a fixed recognition model in the existing acoustic anti-overflow technology. The method can achieve active suppression of the inherent resonance characteristics of different cookware, adaptive learning of users' personalized cooking habits, and high robust anti-overflow warning under complex acoustic backgrounds.

[0011] The second technical problem to be solved by the present invention is to provide a computer-readable storage medium that stores a computer program that can be executed by a processor to implement the aforementioned anti-overflow control method.

[0012] The third technical problem to be solved by the present invention is to provide an overflow prevention system that applies the aforementioned overflow prevention control method to the prior art.

[0013] The technical solution adopted by this invention to solve the first technical problem mentioned above is: an overflow prevention and control method, comprising the following steps:

[0014] S1. Acquire the real-time audio signal of the cooking container, and extract the Mel spectrum after performing Mel spectrum transformation on the real-time audio signal;

[0015] S2. Obtain the pot resonance mask vector corresponding to the current cooking container, and multiply the Mel spectrum with the pot resonance mask vector point by point to suppress the energy of the inherent resonance frequency band of the cooking container to obtain the masked Mel spectrum.

[0016] S3. Multiply the masked Mel spectrum with the adaptive weight vector point by point to enhance the frequency band features related to bubble activity and obtain the enhanced Mel spectrum;

[0017] S4. Calculate the normalized relative intensity based on the enhanced Mel spectrum and generate a smooth overflow quantization evaluation value QE;

[0018] S5. The enhanced Mel spectrum input recognition network is used for spillover state recognition, and the QE value is combined for joint decision-making to generate fire control commands;

[0019] S6. After a single cooking process is completed, determine whether the cooking process constitutes a successful overflow prevention event. If so, use the acoustic data of the overflow stage during the cooking process and the corresponding QE label as self-supervised samples to perform online fine-tuning of the weights of the last fully connected layer of the recognition network.

[0020] As an improvement, the method for generating the cookware resonance mask vector is as follows:

[0021] When a new cooking container is detected, it is instructed to add only water into the cooking container and boil it for a preset time.

[0022] During the boiling process, the empty boiling audio signal of the cooking container is collected;

[0023] Mel spectrum was extracted from the air-boiling audio signal, and the energy stability and average amplitude of each frequency band were analyzed.

[0024] The cookware resonance mask vector is generated based on the energy characteristics of each frequency band, and the value of each component in the cookware resonance mask vector is the reciprocal of the contribution of the corresponding frequency band to the spillover determination.

[0025] The cookware resonance mask vector is bound to the unique identifier of the cooking container and then stored in the memory.

[0026] As an improvement, the cookware resonance mask vector has the same dimension as the Mel spectrum, and the point-by-point multiplication operation suppresses the inherent high-amplitude resonance peak of the cooking container by 10-20 dB.

[0027] As an improvement, the recognition of a new cooking container being connected can be achieved by detecting changes in the weight of the cooking container using a weight sensor, or by being manually triggered by the user.

[0028] As an improvement, the adaptive weight vector is dynamically updated based on the real-time changes in the bubble frequency during the current cooking process, in order to highlight the frequency band energy corresponding to the current bubble activity.

[0029] As an improvement, the recognition network in step S5 is a convolutional-transformer hybrid network, including convolutional layers and transformer layers. The convolutional layers are used to extract local time-frequency features, and the transformer layers are used to capture global temporal dependencies. The backbone parameters of the convolutional-transformer hybrid network are trained offline in the cloud and then deployed.

[0030] As an improvement, the determination of whether the current cooking process constitutes a successful spill prevention event in step S6 specifically includes:

[0031] Determine whether the QE value exceeds the preset threshold and whether the system has triggered a heat reduction command during this cooking process;

[0032] After the fire suppression command was executed, it was confirmed that no actual overflow occurred;

[0033] If both of the above conditions are met simultaneously, the cooking process is considered a successful spill prevention event.

[0034] As an improvement, the online fine-tuning in step S6 adopts a preset learning rate and sparse update strategy. The weight difference of the last fully connected layer is calculated through backpropagation, and the weight difference is compressed and stored in the self-learning partition of the memory. When the same cooking container or the same type of soup is used again, the weight difference is automatically loaded and the weight of the last fully connected layer of the recognition network is updated.

[0035] As an improvement, the data size of the weight difference is less than 1KB, and the online fine-tuning process uses fixed-point integer arithmetic.

[0036] The technical solution adopted by the present invention to solve the second technical problem mentioned above is: a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned boiling liquid anti-overflow control method.

[0037] The technical solution adopted by the present invention to solve the third technical problem mentioned above is as follows: An anti-overflow system, characterized in that: it includes an audio acquisition module, a cookware identification module, a control module, a heating actuator, and a memory. The control module is electrically connected to the audio acquisition module, the cookware identification module, the heating actuator, and the memory respectively. The memory stores a computer program, and when the computer program is executed by the control module, it implements the aforementioned anti-overflow control method.

[0038] Compared with the prior art, the advantages of the present invention are as follows:

[0039] 1. By acquiring the inherent acoustic fingerprint of the target cookware under the condition of boiling water only through the empty boiling learning mechanism, and generating a frequency domain mask vector, a point-by-point multiplication operation is performed at the forefront of the real-time audio processing chain. This can suppress the inherent high-amplitude resonance peak of the cookware by 10 to 20 dB, actively suppressing the resonance interference caused by cookware of different materials, shapes and sizes from the signal source. This effectively eliminates the problem of the generalization performance of the acoustic fingerprint model deteriorating due to cookware replacement, and enables the system to be compatible with various cookware types including stainless steel, cast iron and glass.

[0040] 2. By introducing an online fine-tuning mechanism into the final fully connected layer of the convolutional-transformer recognition network, and using the cooking process data of successful spillover prevention as self-supervised samples, progressive self-learning and personalized adaptation are achieved at the edge, enabling the system to automatically adapt to users' unique cooking habits and new ingredient recipes. The fine-tuning process only updates the weights of the final layer of the network, and the weight difference data is less than 1KB, which can be completed with only integer operations, without affecting the basic recognition ability of the main model. This solves the problem that fixed models are difficult to cover non-standard recipes and personalized cooking modes.

[0041] 3. By constructing a complete signal processing and control closed loop, including cascaded cookware mask filtering, adaptive weight enhancement, QE continuous quantization, convolutional-Transformer network recognition, and online fine-tuning feedback, a highly robust overflow prevention early warning system was achieved in a complex real kitchen environment with multiple coupled factors.

[0042] 4. All processing modules of the entire system are optimized using fixed-point arithmetic, resulting in low overall power consumption. This makes the system suitable for deployment and operation on resource-constrained embedded edge platforms. At the same time, the system supports updating all algorithm parameters and model weights via over-the-air (OTA) download technology, providing long-term adaptive maintenance capabilities. Attached Figure Description

[0043] Figure 1 This is an overall flowchart of the overflow prevention control method in this embodiment of the invention. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application 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 application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] The anti-overflow system in this embodiment includes an audio acquisition module, a cookware recognition module, a control module, a heating actuator, and a memory. The control module is electrically connected to the audio acquisition module, the cookware recognition module, the heating actuator, and the memory. The memory stores a computer program, which, when executed by the control module, implements the following anti-overflow control method.

[0047] The complete real-time signal processing chain of this spillover prevention system starts from the raw audio acquisition and sequentially goes through eight stages: Mel spectrum transformation, Mel spectrum extraction, mask filtering, weight enhancement, QE quantization, AI recognition, control command output, and fine-tuning feedback. The output of the weight enhancement stage is split into two paths, one pointing to the QE quantization stage and the other to the AI ​​recognition stage. The AI ​​recognition stage also receives output from the QE quantization stage to achieve joint decision-making. The output of the fine-tuning feedback stage passes through the network weight storage area in the AI ​​recognition stage to form a closed loop.

[0048] The control module includes five functional units: acoustic front-end preprocessing module, feature dynamic enhancement module, intensity continuous quantization module, intelligent recognition and decision-making module, and online fine-tuning feedback module.

[0049] The acoustic front-end preprocessing module receives the audio signal from the cooking container, which is acquired in real time by the audio acquisition module. It then performs Mel-frequency transform on the audio signal to generate a Mel-frequency spectrum. Subsequently, the acoustic front-end preprocessing module retrieves the cookware resonance mask vector, which is uniquely bound to the current cooking container's identifier, from memory. It performs a point-by-point multiplication operation between the Mel-frequency spectrum and the cookware resonance mask vector to suppress interference from the cooking container's inherent resonance frequency bands on subsequent identification, and outputs the masked Mel-frequency spectrum.

[0050] The feature dynamic enhancement module receives the masked Mel spectrum output from the acoustic front-end preprocessing module, determines an adaptive weight vector based on the current cooking state, and performs a point-by-point multiplication operation between the masked Mel spectrum and the adaptive weight vector to highlight the dominant frequency region of the current bubble activity, outputting an enhanced Mel spectrum. This enhanced Mel spectrum is simultaneously distributed to the subsequent intensity continuous quantization module and the intelligent recognition and decision-making module, forming a one-to-many data bifurcation node in the pipeline.

[0051] The intensity continuous quantization module receives the enhanced Mel spectrum, calculates the normalized relative intensity, and generates a spillover quantization evaluation value (QE) curve that has been smoothed in the time domain. The QE value is a continuous numerical value used to characterize how close the cooking liquid surface is to the spillover critical state at the current moment, and its value ranges from zero to one, which is a continuous real number.

[0052] The intelligent recognition and decision-making module includes a convolutional-Transformer hybrid network deployed on the controller. This hybrid network receives enhanced Mel-frequency spectra as input features and outputs spillover probability values ​​after forward inference. The module jointly determines the spillover probability value with the current value of the QE curve, generating a fire control command when preset warning conditions are met. This command is then transmitted via the control output module to the heating actuator corresponding to the cooking container, completing a closed-loop action from acoustic perception to physical control.

[0053] After each cooking session, the online fine-tuning feedback module automatically determines whether data meeting the conditions for effective self-supervised samples was generated. When the conditions are met, the voiceprint data and its corresponding QE labels in the overflow stage are extracted as self-supervised samples. Online fine-tuning is performed on the weights of the last fully connected layer of the convolutional-Transformer hybrid network, and the weight difference data generated by the fine-tuning is compressed and written into the self-learning partition of the non-volatile memory. This allows the system to continuously adapt to the user's specific cooking habits and kitchen environment as the number of uses increases.

[0054] Of the five functional units mentioned above, the acoustic front-end preprocessing module, the feature dynamic enhancement module, the intensity continuous quantization module, and the intelligent recognition and decision-making module constitute the real-time forward inference pipeline, and their overall latency meets the requirements of real-time control. The online fine-tuning feedback module is executed asynchronously in the non-real-time cooking end stage and does not occupy the computing resources of the real-time inference pipeline.

[0055] The audio acquisition module includes at least one microphone mounted on the cooktop. The microphone is preferably mounted on the edge of the cooktop panel, at a horizontal distance of 5 to 15 centimeters from the edge of the cooking container opening, with the pickup direction facing upwards towards the opening of the cooking container, such that the main pickup axis of the microphone forms a tilt angle of 30 to 60 degrees with the plane of the cooking container opening. With this mounting method, the microphone can effectively pick up the acoustic signal of bubbles bursting when the liquid in the pot boils, while reducing the direct impact of noise from the cooktop flame airflow. In another embodiment, two microphones can be configured to form a differential pickup array, further suppressing far-field environmental noise through beamforming.

[0056] The cookware identification module is used to identify the currently used cooking container and determine its unique identifier. In a preferred embodiment, the cookware identification module includes a weight sensor installed below the cooktop surface. The weight sensor is preferably a thin-film pressure sensor or a strain gauge weighing sensor. When the weight sensor detects a step change in weight on the cooktop surface that exceeds a preset new cookware identification threshold, the controller determines that a new cooking container has been placed and matches the detected weight value with the stored unique identifier of the cookware. If the difference between the current weight value and the calibrated weight corresponding to a stored cookware identifier is within the allowable error range, it is identified as a known cookware, and the corresponding cookware resonance mask vector and model parameters are retrieved. If no match can be found, the empty-boiling learning process is triggered to complete the registration of the new cookware. Of course, the cookware identification module can also initiate the new cookware registration process manually by the user, i.e., the user issues a registration command through a designated button on the cooktop control panel or through a companion mobile terminal application.

[0057] The heating actuator is used to adjust the heating power of the cooking container according to the firepower control commands output by the controller. For gas stoves, the heating actuator is a gas proportional valve driven by a stepper motor. The stepper motor drives the gas proportional valve to rotate to the corresponding opening position according to the number of steps corresponding to the target firepower level, thereby realizing the graded adjustment of gas flow. For induction cooktops, the heating actuator is an IGBT power control module. The main control system of the cooktop changes the output power of the heating coil by adjusting the conduction duty cycle of the IGBT. In this embodiment, the firepower control commands include a down-level command, a flame-off command, and a heating-restoring command. Among them, the down-level command refers to the control command to reduce the current heating power by one or more preset levels; the flame-off command refers to the control command to reduce the heating power to zero or to the lowest maintenance level; and the heating-restoring command refers to the control command to restore the heating power to the level before intervention after the spillover risk has been eliminated.

[0058] In addition, the spill prevention system includes an image verification unit electrically connected to the control module. This unit comprises a low-resolution camera module. The image verification unit is mounted on the range hood panel above the cooktop or on a bracket on the side of the cooktop, covering the opening area of ​​the cooking container. The image verification unit is activated after an acoustic warning is issued. It verifies whether there are visual signs of liquid overflow by performing simple edge detection and brightness change analysis on the acquired image frames, thereby reducing the false alarm rate. The image verification unit is in sleep mode when not under warning conditions to conserve power.

[0059] The overall additional power consumption of the anti-overflow system is controlled within 30mW. This goal is achieved through the following multi-layered low-power design strategies: First, the controller adopts a fully fixed-point arithmetic scheme to avoid the high power consumption of floating-point arithmetic units. The weights and intermediate activation values ​​in the network inference stage are represented by 8-bit or 16-bit fixed-point integers, which can reduce power consumption by approximately 60% to 75% compared to 32-bit floating-point arithmetic. Second, the controller enters a low-power standby mode when there is no cooking activity, keeping only the monitoring interrupt of the weight sensor active. The audio acquisition and signal processing pipeline is then activated when a pot placement event is detected. Third, the online fine-tuning process is only executed during idle periods after cooking is completed, and only a small number of weight parameters of the last fully connected layer of the network are updated, keeping the computational load of a single fine-tuning to an extremely low level. Fourth, the optional image verification unit adopts an event-triggered mechanism, only briefly activating when an acoustic warning occurs, and consuming no power during normal operation.

[0060] With the support of the aforementioned hardware platform, such as Figure 1 As shown, the overflow prevention control method in this embodiment specifically includes the following steps S1 to S6.

[0061] S1. Acquire the real-time audio signal of the cooking container, and extract the Mel spectrum after performing Mel spectrum transformation on the real-time audio signal.

[0062] S2. Obtain the pot resonance mask vector corresponding to the current cooking container, and multiply the Mel spectrum with the pot resonance mask vector point by point to suppress the energy of the inherent resonance frequency band of the cooking container, so as to obtain the masked Mel spectrum.

[0063] S3. Multiply the masked Mel spectrum point by point with the adaptive weight vector to enhance the frequency band features related to bubble activity, thereby obtaining the enhanced Mel spectrum.

[0064] S4. Calculate the normalized relative intensity based on the enhanced Mel spectrum and generate a smooth overflow quantization evaluation value QE.

[0065] S5. The enhanced Mel spectrum input recognition network is used to identify the spillover state, and the QE value is combined to make a joint decision to generate fire control commands.

[0066] S6. After a single cooking process is completed, determine whether the cooking process constitutes a successful overflow prevention event. If so, use the acoustic data of the overflow stage during the cooking process and the corresponding QE label as self-supervised samples to perform online fine-tuning of the weights of the last fully connected layer of the recognition network.

[0067] In step S2, before feeding the Mel spectrum into the subsequent feature enhancement and recognition network, a point-by-point multiplication masking operation needs to be performed to suppress the energy of the inherent resonant frequency band of the cooking container. The cookware resonance mask vector used in this operation is generated specifically for each cooking container through a one-time empty-boil learning process. The method for generating the cookware resonance mask vector is explained below.

[0068] As described above, the anti-overflow system provides two methods to trigger the identification and empty-boil learning process for new cooking containers. The first method is automatic detection by the weight sensor, which automatically initiates the empty-boil learning guidance mode. The second method is manual triggering by the user. The anti-overflow system has a new pot learning function entry on its control panel or accompanying mobile application. When a user uses a cooking container for the first time, they can actively click on this entry, and the system will then enter the empty-boil learning guidance mode. In practical use cases, the manual triggering method can supplement the automatic detection method by the weight sensor. For example, when the weight difference between two empty pots falls within the weight matching tolerance range and cannot be automatically distinguished, the user can manually and explicitly register the new pot.

[0069] Once the anti-overflow system enters the empty-boil learning guidance mode, the controller guides the user through the following steps to complete the empty-boil learning operation via flashing indicator lights on the control panel, a buzzer sound, or prompts on the mobile application interface, in order to generate the cookware resonance mask vector. First, when a new cooking container is detected, the user is guided to add only water to the container, preferably one-half to two-thirds of its volume, to ensure that the acoustic characteristics generated during boiling are representative. Second, the user is guided to turn on the stove to heat the water to a sustained boil. Subsequently, the microphone continuously collects empty-boil audio signals for a preset duration, preferably 45 to 90 seconds, with 60 seconds used in one embodiment, corresponding to approximately one minute of stable boiling time. After the empty-boil audio signal collection is complete, the user is informed that the empty-boil learning is complete, and the user can turn off the heating or continue normal cooking.

[0070] Under the aforementioned empty-boil learning conditions, the cooking container contains only water and no thick broth, oil, or other food ingredients. From an acoustic physics perspective, the sound signals collected by the microphone under these conditions mainly consist of the following three types of sound sources. The first type is the mechanical vibration sound of the pot body, which is the solid vibration radiation sound generated by the uneven distribution of thermal expansion and thermal stress of the pot wall and bottom after heating. Its frequency characteristics are determined by the elastic modulus, density, and geometric dimensions of the pot material. Different materials, such as stainless steel, cast iron, glass, and aluminum alloy, show significant differences in the frequency distribution of this type of vibration sound. The second type is the bottom plate vibration sound, which is the coupled vibration sound between the bottom of the cooking container and the stove support structure caused by heating. Its frequency characteristics are related to the thickness of the pot bottom, the contact area of ​​the stove support, and its rigidity. The third type is the cavity standing wave sound, which is the acoustic cavity standing wave formed by the steam bubbles generated by boiling between the liquid surface inside the pot and the edge of the pot lid or rim. Its resonant peak frequency is determined by parameters such as the geometry of the pot, the liquid level, and the diameter of the pot opening. The three types of sound sources mentioned above together constitute the inherent acoustic fingerprint of the cooking container under heating and boiling conditions. Since the sound of bubbles bursting during the boiling of clear water exhibits a broadband and relatively flat energy distribution in the frequency domain, while the low-frequency bubbling sound and characteristic frequency band energy clustering phenomenon caused by the increased viscosity, larger bubble size, and reduced bursting rate of thick soup during boiling, as well as the continuous outflow of bubbles, do not occur under the condition of empty boiling of clear water, the empty boiling audio does not contain the bubble sound characteristics of thick soup that are highly correlated with the overflow event. Thus, the inherent resonance characteristics of the cookware and the overflow-related acoustic characteristics can be physically decoupled at the signal source level.

[0071] After acquiring the air-boiling audio signal, the controller performs the following processing steps on the air-boiling audio signal to generate a cookware resonance mask vector.

[0072] When a new cooking container is detected, it is guided to add only water to the cooking container and boil for a preset time. During the boiling process, the empty-boiling audio signal of the cooking container is collected. Mel spectrum is extracted from the empty-boiling audio signal, and the energy stability and average amplitude of each frequency band are analyzed. The pot resonance mask vector is generated according to the energy characteristics of each frequency band. The value of each component in the pot resonance mask vector is the reciprocal of the contribution of the corresponding frequency band to the spillover determination. The pot resonance mask vector is bound to the unique identifier of the cooking container and stored in the memory.

[0073] The specific steps are as follows.

[0074] First, the air-boiling audio signal is segmented and windowed. The air-boiling audio signal is segmented into frames with the same frame length and frame shift as the aforementioned real-time audio processing chain, and a Hanning window is applied to each frame to reduce spectral leakage.

[0075] Then, a Mel-spectrum transform is performed on the windowed signal of each frame to obtain the frequency domain representation of each frame. A pre-designed set of Mel filters then maps the linear frequency axis to the Mel frequency axis. The Mel-spectrum extraction parameters described above are completely consistent with the configuration in the real-time audio processing chain to ensure that the mask vector and real-time features are strictly aligned in the frequency domain dimension.

[0076] Next, a band energy statistical analysis is performed on the Mel spectrum vector. For the k-th Mel band (k = 1, 2, ..., 40), the average amplitude μ_(k) and standard deviation σ_(k) of this band across all M frames are calculated using the following formula: μ_(k) = 1 / M ∑ M ᵢ₌1 mᵢ,k;σ_(k) = 1 / M ∑ M ᵢ₌1 (mᵢ,k - μ_(k))²), where mᵢ,k represents the k-th component of the log-Mel spectrum vector of the i-th frame. The average amplitude μ_(k) reflects the steady-state energy level of this frequency band under air-boiling conditions; the larger μ_(k) is, the stronger the influence of the inherent resonance of the cookware on this frequency band. The standard deviation σ_(k) reflects the temporal stability of the energy of this frequency band; the smaller σ_(k) is, the smaller the energy fluctuation of this frequency band, that is, the energy of this frequency band mainly comes from the stable resonance characteristics of the cookware rather than random transient events.

[0077] Based on this, an energy stability weighting factor is introduced to comprehensively consider the average energy and stability of each frequency band. The resonance intensity index R_(k) of the k-th frequency band is defined as follows: R_(k) = μ_(k) · 1 / 1 + γ · σ_(k). Where γ is a stability adjustment coefficient, preferably 0.5 to 2.0, and in one embodiment, γ = 1.0. The physical meaning of the above formula is: when the average energy of a frequency band is high and the time fluctuation is small (i.e., a stable high-energy resonance peak), its resonance intensity index R_(k) takes a larger value; when a frequency band has a high average energy but violent fluctuations, R_(k) is appropriately reduced due to the increase of the denominator, so as to avoid misjudging occasional energy spikes as stable resonances.

[0078] Furthermore, a cookware resonance mask vector is generated based on the resonance intensity index. The cookware resonance mask vector v ∈ R 40 The k-th component v_(k) is defined as the reciprocal of the contribution of the corresponding frequency band resonance intensity, and the calculation formula is as follows: v_(k) = R a x - R_(k) / R a x - Rᵢn · (1 - ε) + ε. where R ax = _(k) R_(k), Rᵢn = _(k) R_(k), ε is the lower cutoff coefficient, preferably 0.05 to 0.2, and in one embodiment ε = 0.1. The above formula ensures the frequency band with the strongest resonance (R_(k) = R a The mask component corresponding to x) takes the minimum value ε, and the mask component corresponding to the weakest resonant frequency band R_(k) = Rᵢn takes the maximum value 1.0. The mask components of the intermediate frequency bands are linearly interpolated between ε and 1.0. The purpose of introducing the lower limit truncation coefficient ε is to avoid any frequency band being completely zeroed, thereby preserving the weak spillover characteristic information that may exist in that frequency band. When ε = 0.1, the energy of the strongest resonant frequency band is suppressed by about 20 · (0.1) = -20 dB; when ε = 0.2, the suppression is about 20 · (0.2) ≈ -14 dB, both falling within the target suppression range of 10 to 20 dB.

[0079] Finally, to adapt to the controller's fixed-point arithmetic capabilities, the cookware resonance mask vector is quantized using fixed-point quantization. Specifically, each floating-point component v_(k) of the mask vector is multiplied by 256 and rounded to the nearest integer, resulting in an 8-bit unsigned integer representation between 0 and 256. During real-time computation, the original numerical proportion can be restored by right-shifting the result of point-by-point multiplication by 8 bits. The error introduced by this quantization process does not exceed 0.4% of the original floating-point value, and its impact on subsequent feature quality is negligible.

[0080] After the cookware resonance mask vector is generated, it needs to be bound to the unique identifier of the corresponding cooking container and persistently stored in the memory. The unique identifier can be obtained in two ways: one is a custom name manually entered by the user through the control panel or mobile application after empty-boil learning, such as "stainless steel soup pot" or "cast iron stew pot"; the other is an coded identifier automatically generated by the system based on the timestamp of empty-boil learning and the weight value of the empty pot detected by the weight sensor. In the memory, the system allocates a dedicated storage area for the cookware mask data. The storage space allocated for cookware records is preferably no less than 4 kilobytes, capable of accommodating no less than 76 cookware records, sufficient to meet the registration needs of multiple cookware in a family kitchen. Optionally, when the storage space is nearly full, the system automatically discards the least used cookware records to free up space.

[0081] During routine cooking monitoring, when a user places a cooking container on the stove, the cookware recognition module first acquires the container's identification information. A weight sensor detects the weight of the empty pot, and the controller compares this weight with the registered empty pot weights in the memory. The record with the smallest absolute weight difference that does not exceed the weight matching tolerance is selected as the matching result. If multiple records have weight differences within the tolerance range, the system further displays a list of candidate cookware to the user via the control panel or mobile terminal interface, allowing the user to confirm the currently used cooking container. Upon successful matching, the controller reads the corresponding 8-bit fixed-point quantization mask vector from the memory for subsequent real-time filtering operations. If no registered cookware record is matched, the system prompts the user to perform an empty-boil learning process or choose to continue monitoring using the default all-pass mask vector.

[0082] After the mask vector is loaded, the acoustic front-end preprocessing module performs a mask filtering operation on the Mel spectrum of each frame in the real-time audio processing chain. Let the 40-dimensional log Mel spectrum vector extracted in the current frame be m_(t), and the loaded pot resonance mask vector be v. Then the Mel spectrum vector m_(t) after masking is calculated as follows: m_(t,k) = m_(t,k) + 20·(v_(k)), k = 1, 2, …, 40.

[0083] The above operation is equivalent to multiplying the energy of the k-th frequency band by v_(k) in the linear amplitude domain. In the actual fixed-point arithmetic implementation, this operation is transformed into adding the fixed-point logarithmic Mel-frequency value of the current frame to the corresponding fixed-point value in a pre-calculated and stored lookup table of mask logarithmic gain. This lookup table has 40 entries and is calculated all at once when the cookware mask is loaded. The single-frame mask filtering operation involves only 40 fixed-point addition operations, and the computation time is negligible, so it will not impose an additional burden on the overall latency of the real-time audio processing chain.

[0084] The effect of mask filtering on suppressing the inherent high-amplitude resonance peaks of the pot body is as follows. Taking a stainless steel soup pot with a diameter of 24 cm as an example, during the empty-boiling learning stage, significant resonance peaks were measured in the 12th to 15th Mel frequency bands (corresponding to a linear frequency range of approximately 800 Hz to 1500 Hz). The resonance intensity index R_(k) value in this interval is the highest among all 40 frequency bands. After mask vector generation, the mask components v_(k) corresponding to these frequency bands take values ​​of approximately 0.1 to 0.15. In real-time filtering, the energy suppression of these frequency bands is 20 · (0.1) = -20 dB to 20 · (0.15) ≈ -16.5 dB. In the Mel frequency bands 1 to 5 (corresponding to approximately 80 Hz to 250 Hz) and 30 to 40 (corresponding to approximately 4000 Hz to 8000 Hz), due to the low resonant energy, the mask component v_(k) takes values ​​close to 0.8 to 1.0, and the suppression amount is only 0 to approximately 2 dB, basically preserving the original signal characteristics in these frequency bands. Experimental verification showed that after performing air-boiling learning and applying the corresponding mask vectors on five different cooking containers of various materials and sizes, the overflow recognition accuracy of the convolutional-Transformer hybrid network on the cross-cookware test set increased from 78.3% without using a mask to 93.6%, and the false alarm rate decreased from 14.7% to 3.2%. The five types of cooking containers are a 24 cm diameter 304 stainless steel saucepan, a 22 cm diameter cast iron stew pot, a 20 cm diameter borosilicate glass milk pot, a 26 cm diameter die-cast aluminum alloy saucepan, and an 18 cm diameter ceramic casserole, covering the main types of cookware commonly found in home kitchens.

[0085] For mask management and switching logic of multi-cookware systems, the anti-overflow system adopts the following strategy. In a single-cookware configuration, the system maintains only one active cookware mask vector in the on-chip buffer at any given time. When the weight sensor detects a sudden change in cooktop load lasting more than 2 seconds, it is determined to be a cookware replacement event. The system re-executes the cookware identification process and loads a newly matched mask vector. The new mask is used for filtering starting from the next frame after the new mask is loaded. During mask switching, because cookware identification and mask loading operations involve reading non-volatile memory and recalculating lookup tables, the transition period typically does not exceed 8 frames (approximately 128 milliseconds). Frames within this transition period are filtered using an all-pass mask (i.e., all components are fixed-point quantization values ​​of 256) to avoid introducing discontinuous abnormal features between the old and new masks. In a multi-cookware configuration, the controller maintains an independent mask vector buffer for each cookware's audio channel. The cookware identification and mask loading processes for each channel are independent and do not interfere with each other. Optionally, the system also supports a cookware mask update mechanism: when the acoustic characteristics of the same cooking container change due to material aging or physical deformation after a long period of use, the user can re-perform empty-boil learning to update the mask vector in the cookware record. The new mask vector will overwrite the mask data and learning timestamp in the original record in the memory, while the unique identifier and empty pot weight value remain unchanged.

[0086] After masking filtering, the masked Mel spectrum obtained by the system has suppressed the inherent resonant peaks of the pot body, but it has not yet been targeted to enhance the real-time acoustic characteristics of bubble activity during the current cooking process. In order to enable the subsequent recognition network to focus on the frequency band information closely related to spillover, the controller further performs an adaptive feature enhancement operation in the real-time audio processing chain, multiplying the masked Mel spectrum point by point with an adaptive weight vector to highlight the characteristic response of the frequency band where the bubble's dominant frequency is located in the current cooking scene, thus obtaining the enhanced Mel spectrum.

[0087] The adaptive weight vector is generated based on the combined results of short-time energy distribution statistics and frequency band activity analysis during the current cooking process. The controller maintains a sliding observation window with a length of W frames, where W is preferably an integer between 50 and 150. In one embodiment, W is set to 100, corresponding to an observation duration of approximately 1.6 seconds. At the arrival of each frame, the controller performs frequency band energy distribution statistics on the masked Mel spectrum of all W frames within the sliding observation window. For the k-th Mel band (k equals 1 to 40), the short-time average energy E_(k) of this band within the current sliding window is calculated using the following formula: E_(k) = 1 / W ∑ W j₌tm_(j,k), where t is the current frame number, and m_(j,k) is the value of the Mel spectrum in the k-th frequency band after the j-th frame is masked. Based on this, the activity index A_(k) of each frequency band is calculated to measure the drastic energy change of that frequency band within the sliding window. The calculation formula is as follows: A_(k) = 1 / W-1 ∑ W j₌t |m_(j,k) - m_(j-1,k)|. The activity index A_(k) reflects the mean of the inter-frame energy difference in the k-th frequency band. When bubble activity is concentrated in a certain frequency band, the frequency band in that range will have both a high short-time average energy E_(k) and a high activity A_(k). The controller combines the above two statistics to generate the original enhancement weight r_(k) of the k-th frequency band. The calculation formula is as follows: r_(k) = α_(E)*E_(k)+α_(A)*A_(k), where α_(E) and α_(A) are the energy weight index and the activity weight index, respectively. α_(E) is preferably 0.3 to 0.7, and α_(A) is preferably 0.3 to 0.7. In one embodiment, α_(E) is equal to 0.5 and α_(A) is equal to 0.5. The physical meaning of the above formula is: only when a certain frequency band has both high energy and high activity is it given a greater enhancement weight, so as to effectively distinguish the bubble main frequency band from the background stable noise frequency band.

[0088] To avoid feature distortion caused by extreme values, the controller normalizes the original enhancement weights of all 40 frequency bands to generate an adaptive weight vector V. The calculation formula for its k-th component v_(k) is as follows: v_(k) = v m ᵢ n + (v max -v m ᵢ n ) · r_(k) - v m ᵢ n / v max - v m ᵢ n + ε. Where v max With v m ᵢ n These are the maximum and minimum values ​​among all 40 original augmentation weights, v m ᵢ n With v max v represents the preset lower and upper bounds of the weight range. m ᵢ n Preferably 0.5 to 0.8, v max Preferably, it is 1.5 to 3.0, and in one embodiment, it is v. m ᵢ n Equal to 0.6 and v maxIt equals 2.0, where ε is a small constant to prevent division by zero, with a value of 10⁻. 8 Through the above normalization mapping, each component of the adaptive weight vector V is constrained to [v...]. m ᵢ n , v max Within the specified range, it ensures that the main frequency band of the bubble is appropriately enhanced, while preventing the non-bubble frequency band from being excessively suppressed and losing auxiliary discrimination information.

[0089] The adaptive weight vector is dynamically updated based on the real-time changes in the dominant bubble frequency during the cooking process. As the sliding observation window advances frame by frame, the statistical values ​​of the short-time average energy E_(k) and activity A_(k) are updated once per frame, and the adaptive weight vector V is recalculated. This frame-by-frame dynamic update mechanism enables feature enhancement to track the real-time changes in the acoustic characteristics of bubbles under different soup bases and heat conditions. For example, when cooking porridge, the dominant bubble frequency is usually concentrated in the lower Mel frequency band, while when boiling water, the dominant bubble frequency is biased towards the mid-to-high frequency band. Even for the same soup base, the dominant bubble frequency may shift as the temperature rises and the liquid surface state changes during heating. The adaptive weight vector can automatically track this change and adjust the enhancement intensity of each frequency band accordingly.

[0090] Optionally, to further improve the temporal stability of the adaptive weight vector and avoid abrupt weight changes due to single-frame anomalies, a first-order exponential smoothing can be applied to the adaptive weight vector, with the specific formula as follows: V_(t) = λ · V t,raw + (1 - λ) · V t ₋1. Where V t,raw V is the original adaptive weight vector directly calculated using the formula above for the current frame. t ₋1 represents the smoothed adaptive weight vector from the previous frame, and λ is the smoothing update coefficient, preferably between 0.05 and 0.3. In one embodiment, λ is set to 0.1. A smaller λ value makes the weight vector change more smoothly, which is suitable for scenarios with long-term stable cooking; a larger λ value allows the weight vector to respond more quickly to sudden changes in the cooking state.

[0091] The controller performs pointwise multiplication of the masked Mel spectrum vector of the current frame with the adaptive weight vector V to obtain the enhanced Mel spectrum vector. Let the masked Mel spectrum vector of the current frame be m_(t) = [m t1 , m t2 , …,m t40The k-th component of the enhanced Mel spectrum vector m_(t) is calculated as follows: m_(t,k) = m_(t,k) · v_(k). The enhanced Mel spectrum vector m_(t) is the final feature vector after two stages of processing: pot resonance mask filtering and adaptive feature enhancement. Its dimension remains 40, and it can be directly used as the input feature for subsequent recognition networks. Figure 1 The bifurcation marker shown at the output position of this stage indicates that the enhanced Mel spectrum vector is simultaneously sent to the subsequent intensity continuous quantization module and the intelligent recognition and decision-making module.

[0092] After obtaining the enhanced Mel spectrum, the controller further calculates the normalized relative intensity based on the enhanced Mel spectrum and generates a continuous spillover quantization evaluation value (QE) curve that has been smoothed in the time domain.

[0093] Specifically, a sliding time window is set for the energy values ​​of each frequency band of the enhanced Mel spectrum vector, and the median value M and the preset high quantile value Phigh of the energy values ​​of each frequency band within the sliding time window are calculated respectively.

[0094] The normalized relative intensity value Inrom is calculated as follows: Inrom = (E–M) / (Phigh–M); where E is the energy value of each frequency band in the current frame.

[0095] The normalized relative intensity value Inrom is a dimensionless scalar. Mathematically, its numerator (E–M) represents the absolute offset of the current frame energy relative to the background noise baseline, and its denominator (Phigh–M) represents the width of the energy dynamic range of this frequency band from background noise to the common active upper limit in the current liquid state. Therefore, Inrom essentially measures the proportion of the current frame energy to the energy dynamic range of this frequency band itself. When Inrom is approximately equal to 0, it indicates that the current frame energy is close to the background noise level; when Inrom is approximately equal to 1, it indicates that the current frame energy has reached the common active upper limit of this frequency band in the current time period. When Inrom is significantly greater than 1, it indicates that the current frame has experienced an abnormally high energy event that exceeds the recent normal distribution, usually corresponding to a sharp increase in boiling intensity or a precursor to overflow.

[0096] The normalized relative intensity values ​​Inrom of each frequency band are then summed across frequency bands to obtain the global intensity scalar Iglobal.

[0097] The goal of the cross-band aggregation is to comprehensively consider the energy changes of boiling acoustic signals within different frequency ranges to form a preliminary global estimate of the boiling level at the current moment. This embodiment provides two optional implementation methods for cross-band aggregation.

[0098] The first implementation method is the equal-weighted summation method. In this method, the normalized relative intensity values ​​Inrom of all N frequency bands are assigned equal weights, and the global intensity scalar is calculated according to the following formula: Iglobal = 1 / N ∑ᵢ₌1^(N)Inrom,ᵢ; where N is the total number of frequency bands, and Inrom,ᵢ is the normalized relative intensity value of the i-th frequency band. The advantage of this equal-weighted summation method is its simplicity in calculation and the absence of additional parameter configuration. It is suitable for scenarios where there is no prior knowledge of the importance of each frequency band or where the frequency band division already fully reflects the distribution of boiling sound characteristics.

[0099] The second implementation method is a secondary weighting method based on frequency band weights. In this method, the normalized relative intensity values ​​are summed using the adaptive weights already obtained for each frequency band during the aforementioned frequency band energy extraction and weight enhancement stages, with secondary weighting. Specifically, let wᵢ be the weight used for the i-th frequency band in the weight enhancement stage. Then, all frequency band weights are first normalized to obtain the normalized weight coefficient wᵢ: wᵢ = wᵢ / ∑ N ᵢ₌1 w i Then, the global intensity scalar is calculated using the following formula: Iglobal = ∑ N ᵢ₌1 wᵢ × Inromᵢ; where the normalized weighting coefficient wᵢ originates from the weight values ​​of each frequency band determined for the frequency characteristics of boiling sound during the aforementioned frequency band energy extraction and weight enhancement stages. Since the energy of the boiling sound signal is mainly concentrated in specific low-to-mid frequency and mid-to-high frequency bands, these bands have already been assigned higher weights in the preceding stages. Therefore, when performing cross-band aggregation, a larger contribution ratio is given to the normalized intensity values ​​of these bands, enabling the global intensity scalar to more accurately reflect the true changing trend of boiling sound and reducing the interference of non-boiling-related frequency band noise fluctuations on the global estimation.

[0100] A first-order low-pass filter is applied to the global intensity scalar Iglobal sequence for exponential discharge smoothing, thereby obtaining the spillover quantization evaluation value QE. In this embodiment, the spillover quantization evaluation value QE is calculated as follows: QE[t] = β × QE[t-1] + (1-β) × Iglobal[t], where β is the discharge coefficient, 0 < β < 1, and t is the current time. This "discharge" mechanism gives the QE value inertia: when boiling intensifies and new energy is continuously input, the QE value steadily climbs; when brief knocking noise occurs, since the noise energy input is instantaneous, its impact is quickly discharged, and the QE curve only produces small fluctuations rather than spikes. Ultimately, QE is constrained within a continuous range of 0 to 1, and its value changes smoothly, intuitively reflecting the entire process from "calm" (~0) to "about to spill over" (~1).

[0101] The QE curve represents the degree to which the liquid level is close to overflow at the current moment. An enhanced Mel spectrum is input into a pre-trained convolutional-Transformer recognition network to obtain the overflow probability value output by the recognition network. The overflow probability value and the current value of the QE curve are used for joint decision-making. When the joint decision result meets preset warning conditions, a fire suppression control command is generated and sent to the stove's execution end.

[0102] After each cooking session, it is determined whether the cooking session meets the conditions for valid self-supervised samples. The conditions for valid self-supervised samples are: during the cooking process, the system successfully triggers the heat reduction control command because the QE value exceeds the threshold, and it is confirmed afterward that no actual overflow occurs. When the conditions for valid self-supervised samples are met, the pre-overflow voiceprint data and its corresponding QE label before the heat reduction is triggered in the cooking session are extracted as self-supervised training samples. Using the self-supervised training samples, a preset learning rate and sparse update strategy are adopted, and backpropagation fine-tuning is performed only once on the weights of the last fully connected layer or the last two fully connected layers of the convolutional-Transformer recognition network to calculate the weight difference data. The weight difference data is compressed and stored in the self-learning partition of non-volatile memory, and the fine-tuned weights are automatically loaded when the same type of soup or the same pot is used in subsequent cooking.

[0103] The spill control method of the present invention has the following advantages.

[0104] 1. By acquiring the inherent acoustic fingerprint of the target cookware under the condition of boiling water only through the empty boiling learning mechanism, and generating a frequency domain mask vector, a point-by-point multiplication operation is performed at the forefront of the real-time audio processing chain. This can suppress the inherent high-amplitude resonance peak of the cookware by 10 to 20 dB, actively suppressing the resonance interference caused by cookware of different materials, shapes and sizes from the signal source. This effectively eliminates the problem of the generalization performance of the acoustic fingerprint model deteriorating due to cookware replacement, and enables the system to be compatible with various cookware types including stainless steel, cast iron and glass.

[0105] 2. By introducing an online fine-tuning mechanism into the final fully connected layer of the convolutional-transformer recognition network, and using the cooking process data of successful spillover prevention as self-supervised samples, progressive self-learning and personalized adaptation are achieved at the edge, enabling the system to automatically adapt to users' unique cooking habits and new ingredient recipes. The fine-tuning process only updates the weights of the final layer of the network, and the weight difference data is less than 1KB, which can be completed with only integer operations, without affecting the basic recognition ability of the main model. This solves the problem that fixed models are difficult to cover non-standard recipes and personalized cooking modes.

[0106] 3. By constructing a complete signal processing and control closed loop, including cascaded cookware mask filtering, adaptive weight enhancement, QE continuous quantization, convolutional-Transformer network recognition, and online fine-tuning feedback, a highly robust overflow prevention early warning system was achieved in a complex real kitchen environment with multiple coupled factors.

[0107] 4. All processing modules of the entire system are optimized using fixed-point arithmetic, resulting in low overall power consumption. This makes the system suitable for deployment and operation on resource-constrained embedded edge platforms. At the same time, the system supports updating all algorithm parameters and model weights via over-the-air (OTA) download technology, providing long-term adaptive maintenance capabilities.

[0108] The present invention also relates to a computer-readable storage medium having a computer program stored thereon that can be executed by a processor to implement the aforementioned overflow prevention control method.

Claims

1. An overflow prevention control method, characterized in that: Includes the following steps: S1. Acquire the real-time audio signal of the cooking container, and extract the Mel spectrum after performing Mel spectrum transformation on the real-time audio signal; S2. Obtain the pot resonance mask vector corresponding to the current cooking container, and multiply the Mel spectrum with the pot resonance mask vector point by point to suppress the energy of the inherent resonance frequency band of the cooking container to obtain the masked Mel spectrum. S3. Multiply the masked Mel spectrum with the adaptive weight vector point by point to enhance the frequency band features related to bubble activity and obtain the enhanced Mel spectrum; S4. Calculate the normalized relative intensity based on the enhanced Mel spectrum and generate a smooth overflow quantization evaluation value QE; S5. The enhanced Mel spectrum input recognition network is used for spillover state recognition, and the QE value is combined for joint decision-making to generate fire control commands; S6. After a single cooking process is completed, determine whether the cooking process constitutes a successful overflow prevention event. If so, use the acoustic data of the overflow stage during the cooking process and the corresponding QE label as self-supervised samples to perform online fine-tuning of the weights of the last fully connected layer of the recognition network.

2. The overflow control method according to claim 1, characterized in that: The method for generating the cookware resonance mask vector is as follows: When a new cooking container is detected, it is instructed to add only water into the cooking container and boil it for a preset time. During the boiling process, the empty boiling audio signal of the cooking container is collected; Mel spectrum was extracted from the air-boiling audio signal, and the energy stability and average amplitude of each frequency band were analyzed. The cookware resonance mask vector is generated based on the energy characteristics of each frequency band, and the value of each component in the cookware resonance mask vector is the reciprocal of the contribution of the corresponding frequency band to the spillover determination. The cookware resonance mask vector is bound to the unique identifier of the cooking container and then stored in the memory.

3. The overflow control method according to claim 2, characterized in that: The cookware resonance mask vector has the same dimension as the Mel spectrum, and the point-by-point multiplication operation suppresses the inherent high-amplitude resonance peak of the cooking container by 10-20 dB.

4. The overflow control method according to claim 2, characterized in that: The recognition of a new cooking container being connected is achieved by detecting changes in the weight of the cooking container using a weight sensor, or by being manually triggered by the user.

5. The overflow control method according to claim 1, characterized in that: The adaptive weight vector is dynamically updated based on the real-time changes in the bubble frequency during the current cooking process, in order to highlight the frequency band energy corresponding to the current bubble activity.

6. The overflow prevention control method according to claim 1, characterized in that: The recognition network mentioned in step S5 is a convolutional-transformer hybrid network, which includes convolutional layers and Transformer layers. The convolutional layers are used to extract local time-frequency features, and the Transformer layers are used to capture global temporal dependencies. The backbone parameters of the convolutional-transformer hybrid network are trained offline and deployed in the cloud.

7. The overflow prevention control method according to claim 1, characterized in that: Step S6, which involves determining whether the current cooking process constitutes a successful spill prevention event, specifically includes: Determine whether the QE value exceeds the preset threshold and whether the system has triggered a heat reduction command during this cooking process; After the fire suppression command was executed, it was confirmed that no actual overflow occurred; If both of the above conditions are met simultaneously, the cooking process is considered a successful spill prevention event.

8. The overflow control method according to claim 1 or 7, characterized in that: The online fine-tuning in step S6 adopts a preset learning rate and sparse update strategy. The weight difference of the last fully connected layer is calculated through backpropagation. The weight difference is compressed and stored in the self-learning partition of the memory. When the same cooking container or the same type of soup is used next time, the weight difference is automatically loaded and the weight of the last fully connected layer of the recognition network is updated.

9. The overflow control method according to claim 8, characterized in that: The data size of the weight difference is less than 1KB, and the online fine-tuning process uses fixed-point integer arithmetic.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the anti-overflow control method as described in any one of claims 1 to 9.

11. An overflow prevention system, characterized in that: It includes an audio acquisition module, a cookware recognition module, a control module, a heating actuator, and a memory. The control module is electrically connected to the audio acquisition module, the cookware recognition module, the heating actuator, and the memory. The memory stores a computer program, which, when executed by the control module, implements the anti-overflow control method as described in any one of claims 1 to 9.

Images