Heating container, food material boiling method thereof, sensing data AI processing method and storage medium

Abstract

The invention relates to a heating container, a food material boiling method thereof, a sensing data AI processing method and a storage medium. The method comprises the following steps: periodically acquiring at least three capacitance values of at least two non-contact detection electrodes with different heights on the outer wall of the heating container, wherein the non-contact detection electrodes are used for detecting liquid level overflow and serve as one frame of time sequence data; each capacitance value is configured to be self-capacitance of electrodes with different heights or mutual capacitance between the electrodes with different heights, and each frame of time sequence data at least has one mutual capacitance; obtaining features of each frame of time sequence data; and taking the characteristics of the continuous frame time sequence data as the input of the neural network model, and generating output information representing a liquid level overflow identification result.

Description

Technical Field

[0001] The present invention relates to heating anti-overflow detection, and particularly to a heating container, a method for boiling ingredients thereof, an AI processing method for sensing data, and a storage medium. Background Art

[0002] As a common small kitchen appliance, the health kettle is deeply loved by consumers because it can conveniently boil various ingredients. However, during the boiling process, the liquid is likely to overflow due to boiling, which not only dirties the tabletop, causes waste, but may even pose a safety hazard. The common sensor anti-overflow method in the industry usually arranges electrodes inside the kettle wall, and judges whether it is about to overflow according to the voltage change generated when the electrode contacts the liquid surface. If the inner wall electrode part is contaminated by ingredients, the accuracy of this method is extremely vulnerable to influence, and at the same time, the contact detection also has safety hazards.

[0003] CN118687652A deploys multiple capacitance sensors from top to bottom on the outer side of the kettle wall, and relies on manual experience to specify the threshold of each capacitance signal change amount to judge whether it is about to overflow. Due to the complex and changeable boiling scenarios, it is impossible to accurately process various situations by manually writing logic, and it is easy to misjudge under the condition of ensuring no overflow.

[0004] AI is an important driving force for the new round of scientific and technological revolution and industrial transformation. With the help of AI, it is helpful to solve the generalization problem.

[0005] US11583134B2 proposes an artificial intelligence cooking device. The heater heats the cooking container placed on the plate, and a vibration sensor is arranged under the plate to detect the vibration signal of the ingredients in the cooking container. Then, based on the detected vibration signal and the learning characteristics of the vibration provided to the artificial intelligence model, it is determined whether the ingredients in the cooking container are boiling over through the artificial intelligence model that has learned the characteristics of the vibration signal. However, the difference between the vibration of overflow and non-overflow is small, and the detection accuracy of the solution is not high.

[0006] US20240210040A1 proposes a system for controlling a heating device by monitoring the state of a dish being cooked. Sensors are used to collect visual data and sound data corresponding to the heating device, and estimation is carried out through the learning program of the neural network. In the scenario of boiling ingredients, the liquid is sticky and easily adheres to the container wall, blocking the line of sight, and the steam will also affect the visual data. The visual method also has a relatively high cost. Summary of the Invention

[0007] To improve the deficiencies of the prior art, the present invention provides an AI processing method for sensing data of a heating container, including: periodically obtaining at least three capacitance values of at least two non-contact detection electrodes at different heights on the outer wall of the heating container for detecting liquid level overflow as a frame of timing data; each capacitance value is configured as the self-capacitance of electrodes at different heights or the mutual capacitance between electrodes at different heights, and there is at least one mutual capacitance in each frame of timing data; obtaining the features of each frame of timing data; using the features of consecutive frames of timing data as the input of a neural network model to generate output information representing the recognition result of liquid level overflow.

[0008] The present invention adopts a combination of a capacitance sensor and a spiking neural network on heating containers such as health kettles. With the powerful fitting ability of the neural network algorithm for non-linear data patterns and the non-contact detection ability and cost advantage of capacitance, it can accurately detect the liquid level state and respond in a timely manner, realizing the automatic anti-overflow function during the boiling process and improving the convenience and safety of using heating containers such as health kettles.

[0009] The AI processing method for sensing data of the present invention further includes the following subsidiary technical solutions:

[0010] Among them, an SNN neural network is used as the inference model and the training model; discrete pulse coding is performed on the features of the timing data, and the formed time pulse sequence is used as the model input.

[0011] Among them, the method of obtaining the features of each frame of timing data further includes: calculating the timing difference of the capacitance for each detection electrode in each frame of data; encoding the difference value into a discretized pulse signal according to the set boundary points.

[0012] Among them, at least two time windows with different time lengths are given; for each time window, the difference between the current capacitance value of the detection electrode and the maximum value and / or minimum value within the time window from the current to the past is obtained; each difference is processed according to a given set of boundary points, and multiple differences form a time pulse sequence, where the processing operation is configured such that when the number of bins generated by the boundary points is n, if the difference is in the x-th bin, the encoding result is an array of length n, where the x-th bit is 1 and the other bits are all 0, and n and x are integers.

[0013] Among them, the period is configured to be 100 ms.

[0014] Among them, the capacitance values of at least the top five non-contact detection electrodes at different heights from top to bottom on the outer wall of the heating container are periodically obtained to calculate the features.

[0015] Among them, during the data collection for training the model, the output signal of the contact liquid level sensor is used as the judgment condition for the liquid level reaching the overflow position, and the probe of the contact liquid level sensor is located at the set position for triggering anti-overflow; the output signal of the contact liquid level sensor is recorded as the label for training the model.

[0016] Among them, during the data collection for training the model, after the contact liquid level sensor comes into contact with the liquid, heating is stopped after a random duration within the first set range, and / or after the contact liquid level sensor detects that the liquid has separated, reheating is started after a random duration within the second set range.

[0017] Among them, the method for training the model further includes: dividing the timing data of the capacitance obtained by the sampling detection electrode into a training set, a validation set, and a test set; using the features of every consecutive N frames of timing data as a sample data, where N is an integer greater than 2, and the output signal of the contact liquid level sensor in the last frame of every N frames is used as the label, with the goal of training a binary classification model to determine whether the last frame is in an overflow state; the model uses the training set for learning, calculates the loss on the validation set intermittently in each round, uses the mean squared error as the loss function, after several rounds of learning, selects the version with the lowest loss on the validation set, and calculates metrics on the test set to evaluate the model performance; continuously iteratively trains and adjusts various parameters, and selects the model with the optimal metrics on the test set.

[0018] Among them, the model is a multi-layer spiking neural network, and each time it executes a loop to input consecutive N frames of features, 2 output values are obtained, respectively representing the possibilities of overflow and non-overflow.

[0019] There is also provided a method for cooking ingredients in a heating container, including the above-mentioned AI processing method for sensing data; and based on the liquid level overflow recognition result output by the neural network model, when it is determined that the liquid level reaches the overflow position, heating is stopped; when it is determined that the liquid level drops, the container heating is restarted.

[0020] Furthermore, the node for reheating is configured as: after the neural network model outputs the recognition result that the liquid level is no longer overflowing, heating is started after a delay of at most 3 seconds; or, based on the capacitance data collected by the sensor to judge the liquid level, when the height difference between the liquid level when it drops to the initial liquid level at the time of just heating is less than the threshold value, heating is started.

[0021] There is also provided a computer-readable storage medium, which stores a computer program, and when the program is executed by a processor, it implements the above-mentioned AI processing method for sensing data.

[0022] There is also provided a heating container, including a capacitance digital conversion circuit, a processor, and a non-contact detection unit; the non-contact detection unit includes at least two non-contact capacitive detection electrodes arranged at intervals in the vertical direction on the outer wall of the heating container for detecting liquid level overflow; the capacitance digital conversion circuit is respectively coupled to each detection electrode to obtain the corresponding capacitance value; the processor is coupled to the capacitance digital conversion circuit and executes the computer-readable storage medium as described above. Further, an R-SpiNNaker chip running an SNN inference model integrates the capacitance digital conversion circuit and the processor, and the inference model is written into the processor of the R-SpiNNaker chip to jointly control capacitance acquisition and inference. Description of the Drawings

[0023] Figure 1 Schematically shows the main process of the AI processing method for the sensing data of the heating container.

[0024] Figure 2 Schematically shows the exemplary specific operation process of feature acquisition.

[0025] Figure 3 The solid line part shows the local structure of the neural network.

[0026] Figure 4 Schematically shows the structural design of the neural network.

[0027] Figure 5 Schematically shows the architecture diagram of the heating container on the program system.

[0028] Figure 6 Schematically shows the structural diagram of the computer-readable storage medium. Detailed Description of the Embodiments

[0029] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0030] Figure 1 Schematically shows the main process of the AI processing method for the sensing data of the heating container. The heating container of the present invention refers to a container directly acted on by a heat source during normal operation, such as the outer shell of a health care kettle, a kettle, a multi-functional kettle, or even a combustion furnace, a gas generator, etc. During the heating process of the container, the outside is directly acted on by the heat source, and the internal substance is heated through heat transfer. Non-contact liquid detection electrodes are arranged on the outer wall of the heating container, and the design pattern of the detection electrodes can be arbitrary, such as rectangular electrodes, dot electrodes, triangular electrodes, or other shapes, which are not limited in the present invention. See Figure 1, the AI processing method for sensing data of the present invention includes the following steps S1 - S3, where:

[0031] Step S1. Periodically obtain at least three capacitance values of at least two non - contact detection electrodes at different heights on the outer wall of the heating container for detecting liquid level overflow as a frame of timing data.

[0032] In the present invention, the number of non - contact detection electrodes on the outer wall of the container is at least two, located at different heights. A vertical arrangement structure can be adopted. The lowest two electrodes construct three capacitance sampling data so that the model can identify liquid overflow. More electrode numbers form more capacitance sampling data. For example, three vertically arranged electrodes can construct at most three self - capacitances and three mutual capacitances. More capacitance parameters contribute to the accuracy of the model, but too many parameters will also bring negative effects, which will be further elaborated later.

[0033] In step S1, each capacitance value is configured as the self - capacitance of an electrode at a different height or the mutual capacitance between electrodes at different heights. Among them, each frame of timing data requires at least one mutual capacitance. In other words, as the lowest construction standard, it can include one self - capacitance and two mutual capacitances, or two self - capacitances and one mutual capacitance, or three mutual capacitances. A single self - capacitance can reflect the overall liquid level information, and multiple mutual capacitances at different heights can also reflect the liquid level height. At the same time, the mutual capacitance further reflects the situation of viscous liquid hanging on the container wall within the detection range.

[0034] Step S2. Obtain the features of each frame of timing data.

[0035] The original data emitted by the detection electrode has no concept of dimension, and the data generated by the same sensor on different health - care kettles will have a certain degree of difference (mainly introduced by the production processes of the single - chip microcomputer and the health - care kettle, which cannot be avoided). Directly using the original data as features will lead to poor generalization of the model. Therefore, extracting data features from the original data as the input of the model will be beneficial to solving the problem of model generalization.

[0036] Step S3. Use the features of consecutive frames of timing data as the input of the neural network model to generate output information representing the recognition result of liquid level overflow.

[0037] The features of consecutive frames of timing data reflect the fluctuation of the liquid level over time. Under the powerful fitting of the neural network algorithm for non - linear data laws, an efficient and accurate AI anti - overflow recognition is formed.

[0038] The present invention proposes an efficient and accurate AI anti - overflow algorithm, which can timely judge by the model every time when overflow is about to occur during the process of boiling various ingredients (such as porridge, soup, tea, etc.) in the health - care kettle, with a low misjudgment rate, that is, a high recall rate under the condition of ensuring the precision rate.

[0039] Since the model needs to be deployed on a microcontroller for operation and the available memory space is limited, the number of its parameters cannot be too large. At the same time, it is necessary to make a quick response at the moment when the liquid is about to overflow, otherwise the anti-overflow measures will lose their effect. Therefore, the operation cycle of the model should be as small as possible (empirically, if it is greater than 1 second, it is completely unacceptable). To meet the requirements of model lightweight and real-time performance, as an improved solution, in the above, the SNN neural network is used as the inference model and the training model, and the features of the time-series data are encoded into discrete pulse codes, and the formed time pulse sequence is used as the model input. By using a spiking neural network to process the real-time data collected by the capacitance sensor and taking advantage of its strong ability to process spatio-temporal information, the hidden features of the liquid state change can be mined, and the requirements of lightweight, low-power consumption, and high real-time performance of the deployed model can be achieved.

[0040] On this basis, further, the methods for obtaining the features of each frame of time-series data include: calculating the time-series difference of the capacitance for the capacitance value of each detection electrode in each frame of data; encoding the difference value into a discretized pulse signal according to the set boundary points. A spiking neural network usually takes time-series features as input and perceives the feature changes in time series. The time-series difference of capacitance introduces time-series information. Moreover, by observing the data, it is found that during the boiling process, the data of the capacitance sensor jitters locally, but the time-series difference of capacitance shows regularity. Therefore, constructing data features using the time-series difference of capacitance can be beneficial to the model accuracy while solving the generalization problem.

[0041] The method of constructing features using the capacitance time-series difference is not unique. As an optional preferred implementation Figure 2The exemplary specific operation flow of feature acquisition is illustrated, including: S21. Given at least two time windows of different time lengths, for example, given four time windows of 5 frames, 20 frames, 50 frames, and 100 frames, the number and value are selected by multiple model training to be optimal; S22. For each time window, obtain the difference between the current capacitance value of the detection electrode and the maximum and / or minimum value in the time window from the current to the past; S23. Perform processing operations on each difference according to a given set of boundary points, and multiple differences form a time pulse sequence. The processing operation is configured as when the number of buckets generated by the boundary point is n, if the difference is in the xth bucket, the encoding result is an array of length n, where the xth bit is 1, the other bits are 0, and n and x are integers. In this preferred embodiment, the local maximum and minimum values ​​of the signal sequence are tracked according to the given time window, and the difference between the current value and the maximum and minimum values ​​is calculated to characterize the degree of change, so that more differences in the amount of change can be obtained, and the amount of change signal is further discretized, so that the model ignores the small differences between the amounts of change, which not only improves the stability of the model but also satisfies the pulse input of the model. Considering the overflow time, it is necessary to respond quickly within a limited time, and the more parameters the better, but due to hardware limitations, as well as thermal inertia of overflow, etc., the sampling period can be set to 100ms, and / or the capacitance values ​​of at least the first five non-contact detection electrodes at different heights from top to bottom on the outer wall of the heating container can be periodically obtained to calculate the features. The first 5 electrodes from top to bottom are selected to calculate the features. The original data is a time series data of one frame every 100ms. Each frame of data will produce a frame of features. The amount of data will not be too large, but it is sufficient to reflect the features, further improving the model's lightweight and real-time requirements. As above, for each frame of raw data, a total of 40 features (5 sensors * 4 time windows * 2 types of differences) can be calculated.

[0042] Since different ingredients have different viscosity, surface tension, boiling characteristics, etc. at different stages of cooking, the state of "about to overflow" during cooking is very variable (the density of foam and the shape of liquid surface are different). It is difficult to maintain consistency between labels by relying on manual judgment, and it requires a high degree of concentration of the annotator (the cooking time is usually 50 to 70 minutes, and manual labeling requires constant observation). Therefore, a more consistent automated labeling method is needed. To this end, as another improvement plan, in the process of data collection for training the model, the output signal of a contact liquid level sensor such as a metal probe is used as the judgment condition for the liquid level to reach the overflow position. The probe of the contact liquid level sensor is located at the position where the overflow prevention is set, and the output signal of the contact liquid level sensor is recorded as the label of the training model, which can ensure the consistency of labeling.

[0043] Further, when collecting data for model training, after the contact liquid level sensor comes into contact with the liquid, stop heating after delaying for a random duration within the first set range. For example, the turn-off node is controlled to be a random 0 - 2 seconds after the metal probe detects liquid contact, to obtain more features of liquid overflow for training the model. The purpose of the random number is to enhance the diversity of overflow data; and / or after the contact liquid level sensor detects that the liquid has separated, start reheating after delaying for a random duration within the second set range. For example, the reheating node is controlled to be a random 3 - 5 seconds after the metal probe detects liquid separation, to obtain characteristic data of liquid fall. Similarly, the random number enhances the diversity of fall data.

[0044] In the AI algorithm of the present invention, as an exemplary model training method, the training method includes: dividing the time series data of the capacitance obtained by the sampling detection electrode into a training set, a validation set, and a test set; using the features of every consecutive N frames of time series data as a sample data, where N is an integer greater than 2, and N is preferably 10, to obtain good continuity of output pulses while ensuring real-time performance. The output signal of the contact liquid level sensor in the last frame of every N frames is used as a label, and the goal is to train a binary classification model to determine whether the last frame is in an overflow state; the model uses the training set for learning, calculates the loss on the validation set intermittently in each round, uses the mean square error as the loss function, and after several rounds of learning, selects the version with the lowest loss on the validation set, and calculates metrics on the test set to evaluate the model performance; continuously iteratively trains and adjusts various parameters, and selects the model with the best metrics on the test set. Among them, the model is a multi-layer spiking neural network, preferably a 3-layer spiking neural network. Each time the loop is executed, consecutive N frames of features are input, and 2 output values are obtained, respectively representing the possibilities of overflow and non-overflow.

[0045] By writing the model parameters and calculation logic into the heating container, within each signal period (each period is one frame), the process of

filter value acquisition & storage -> calculation of features -> execution of the model -> intervention in heating

[0046] Based on the anti-overflow judgment result of the above model, a method for cooking ingredients in a heating container is further provided, including obtaining the respective probabilities of anti-overflow and non-need for anti-overflow based on the above-mentioned AI processing method of sensing data; according to the liquid level overflow recognition result output by the neural network model, when it is judged that the liquid level reaches the overflow position, heating is stopped, and when it is judged that the liquid level drops, the container heating is restarted. The ingredients cooked using this method can obtain a better taste, realizing the AI intelligent cooking scenario of the heating container. Further, when deploying the model inference on the product, the fire-off node is configured to turn off the fire immediately when the model judges overflow; the node for restarting heating is configured as follows: when this node is judged using the AI method, after the neural network model outputs the recognition result that the liquid level no longer overflows, that is, when the probability of non-overflow is greater, heating is started with a delay of at most 3s to ensure that the liquid drops back to a suitable position under the cooking of different ingredients; or, in a non-AI manner, the liquid level is judged based on the capacitance data collected by the sensor, and heating is started when the height difference between the liquid level when it drops to the initial liquid level at the start of heating is less than the threshold value.

[0047] A computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements the above-mentioned AI processing method of sensing data.

[0048] A heating container is also provided, such as a health kettle, a kettle, a multifunctional kettle, etc., including a capacitance digital conversion circuit (CDC), a processor, and a non-contact detection unit; the non-contact detection unit includes at least two non-contact capacitive detection electrodes arranged vertically at intervals on the outer wall of the heating container for detecting liquid level overflow; the capacitance digital conversion circuit is respectively coupled to each detection electrode to obtain the corresponding capacitance value; the processor is coupled to the capacitance digital conversion circuit and executes the above-mentioned computer-readable storage medium. Further, an R-SpiNNaker chip (see CN113837354A) running an SNN inference model is used to integrate the capacitance digital conversion circuit and the processor, and the inference model is written into the processor of the R-SpiNNaker chip to jointly control capacitance acquisition and inference.

[0049] Embodiment 1

[0050] 1) Data acquisition

[0051] The single-chip microcomputer integrates multiple capacitive sensors arranged from top to bottom. The single-chip microcomputer is deployed outside the health pot, inside the pot handle, and its top is aligned with the appropriate position where you want to trigger the overflow prevention. The capacitive sensor senses changes in the external environment based on the change in capacitance between the object and the sensor. When the liquid in the health pot gradually rises and approaches the mouth of the pot, the liquid as a dielectric will change the capacitance value around the sensor. By detecting the change in this capacitance value, the liquid level information can be indirectly obtained. Due to the data labeling problem stated in the previous article, a metal probe is fixed on the pot lid and connected to the single-chip microcomputer. Its lower end is at a suitable position where you want to trigger the overflow prevention. When the liquid surface contacts the probe, due to the circuit conduction, its voltage value will change significantly. Whether it is currently in an overflow state is determined by whether its voltage exceeds the set threshold, and the voltage value is converted into a label of 0 or 1, thus achieving a highly consistent automated marking method.

[0052] A complete data collection process is as follows:

[0053] 1. Pour the specified weight of materials and water into the health pot (depending on the product usage scenario).

[0054] 2. Click the panel to start boiling. At the same time, the host computer starts recording data and periodically generates data frames, including the values ​​of each capacitive sensor and the overflow signal given by the probe.

[0055] 4. When the probe signal indicates that the liquid level is about to overflow, the program controls the health kettle to temporarily turn off the fire after an appropriate time; when the probe signal indicates that the liquid level is not overflowing, the program controls the health kettle to continue the heating process after an appropriate time.

[0056] 5. When the health pot completes the entire boiling process and automatically turns off the fire, the host computer stops recording data, and a complete data record file is generated. Then clean and cool the health pot, and repeat the above process for the next collection.

[0057] 2) Implementation methods

[0058] 2.1 Overview of Spiking Neural Networks

[0059] Spiking neural network (SNN) is a new neural network model inspired by biological neurons. It transmits and processes information with the pulse sequence emitted by neurons. Compared with traditional artificial neural networks, it is closer to the working mechanism of biological nervous systems and has the advantages of strong spatiotemporal information processing capabilities and low power consumption. In this study, spiking neural networks are used to process real-time data collected by capacitive sensors and explore the hidden characteristics of liquid state changes, which meets the requirements of lightweight, low power consumption, and high real-time performance of the deployment model.

[0060] In spiking neural networks, the modeling of neurons is one of the core contents. The most commonly used neuron model is the Leaky Integrate-and-Fire (LIF) neuron model. The LIF neuron model is based on the simplification and abstraction of the electrophysiological characteristics of biological neurons and can effectively capture the basic behaviors of the neuron membrane potential changing over time and generating spikes. LIF neurons mainly consist of several key parts and corresponding processes: the integration of the membrane potential, threshold judgment, and spike generation reset. From a mathematical perspective, the dynamic change of the membrane potential of LIF neurons can be described by the following differential equation:

[0061]

[0062] In the above formula, V(t) represents the neuron membrane potential at time t, which changes with factors such as time and external input; τ is the time constant (usually with time units, such as seconds), which determines the rate of change of the membrane potential over time; V rest represents the resting potential of the neuron, that is, the stable potential state when the neuron is not stimulated by the outside world; R is the membrane resistance (the unit is often ohms, etc.); I(t) is the input current (the unit is, for example, amperes and changes with time). This equation describes the comprehensive dynamic change of the membrane potential under the influence of the input current I(t), considering its own "leakage" decay towards the resting potential V rest When the membrane potential V(t) of the neuron accumulates to a pre-set threshold V t □, the neuron will generate a spike (usually represented by a discrete event), and then immediately reset the membrane potential to the reset potential V rest , in order to simulate physiological characteristics such as the refractory period after a biological neuron generates a spike. This spike generation and potential reset process can be presented in mathematical form by the following rules:

[0063]

[0064] A spiking neural network is composed of multiple interconnected LIF neurons. Neurons are connected through synapses, and synapses have weights to regulate the strength of signal transmission, thereby affecting the membrane potential change and spike generation of downstream neurons. Suppose neuron i is the presynaptic neuron and neuron j is the postsynaptic neuron, and the weight of their synaptic connection is w ij . When neuron i generates a spike, its spike signal can be represented by s i (t) (using the Dirac function), then the contribution of neuron i to the input current of neuron j is:

[0065] I ij (t) = w ij si (t)

[0066] Considering that the postsynaptic neuron j may receive inputs from multiple presynaptic neurons, assuming that there are a total of N presynaptic neurons connected to it, the total input current of neuron j is the sum of the input current contributions from all presynaptic neurons, and the formula is as follows:

[0067]

[0068] In the entire network, the complete formula for the dynamic change of the membrane potential of the postsynaptic neuron over time considering multiple inputs:

[0069]

[0070] And once neuron j fires a pulse, this pulse will, in turn, serve as an input and affect other downstream neurons connected to it through synaptic connections according to the rules mentioned above, and so on, to achieve the propagation of pulses in the entire network and the transmission and processing of information. Figure 3 The solid line part shows the local structure of the neural network described above.

[0071] 2.2 Feature Design

[0072] Considering that the original values of the capacitance sensor vary among different products and even at different temperatures, the aim of feature design is to capture the changes in capacitance over short and long time periods. Multiple methods for calculating the change features from the original value time series are designed, and through experimental comparison of the accuracy of the test set, the best-performing one is selected. Spiking neural networks usually take time series features as inputs and perceive the feature changes in the time series. Therefore, for each frame of the original data, one frame of features is calculated, and finally, a feature sequence is obtained as the input to the network. Feature calculation is divided into the following two steps:

[0073] 1. Calculate the change signal from the original value. By observing the data, it is found that the value of the capacitance sensor will increase sharply when the other end of the kettle wall touches the liquid, and the speed and magnitude of the increase depend on the current boiling power and the density of the foam; when the foam drops due to triggering the anti-overflow fire extinguishing, the value of the capacitance sensor will also decrease. The values often show local violent jitters, but the overall trend is monotonically increasing or decreasing. Therefore, the local maximum and minimum values of the signal sequence are tracked within a given time window, and the difference between the current value and the maximum and minimum values is calculated to describe the degree of change. Assuming a given time window Win 1 , the maximum and minimum values within the window are calculated as and Then the change feature is

[0074]

[0075] Take multiple time windows to capture the long and short time variation signals. Assuming there are N time windows, then

[0076]

[0077] Finally, for a single sensor signal sequence, a total of N*2 variation features can be obtained.

[0078] 2. Discretize the variation signal. Since the process or environment will always introduce inconsistencies, on the basis of processing the original data into variations, the variation signal is further discretized so that the model ignores the slight differences between the variations to improve the stability of the model. Given the bucket points [b1, b2, b3, b4, ...], calculate the index of the interval where the value is located. For example, if the value of feature f is between b2 and b3, the calculated index value 2 is used as the input of the subsequent model. The value of the bucket point can be obtained by statistics on the variation.

[0079] In summary, for the data with the original dimension of x (number of sensors)*t (number of time frames), the feature with the dimension of (x*2*N)*t (number of time frames) is processed.

[0080] 2.3 Network structure design

[0081] As mentioned above, the pulse neural network based on LIF neurons can be divided into input layer, hidden layer (can have multiple layers), and output layer. Figure 4 shown.

[0082] Input layer: The discretized variation features mentioned above are used as model input. In addition, an embedding layer is added to correspond each index value to a learnable floating-point variable, reducing the difficulty of model learning.

[0083] Hidden layer: Design multiple hidden layers so that the network can effectively extract and nonlinearly transform the input features and learn the inherent patterns of feature changes under different liquid states. Adjust parameters such as the number of hidden layer neurons through experiments.

[0084] Output layer: The output layer has two neurons, representing the two states of liquid "about to overflow" and "not close to overflowing". The pulse emission frequency of the output layer neurons is used to determine whether the liquid in the health pot is in a dangerous state of overflowing.

[0085] 2.4 Model Training

[0086] First, the collected samples are divided into a training set, a validation set, and a test set in units of files (one file is generated for each boiling). Then, each file is traversed, and every consecutive several frames are used as a sample, and the label of the last frame among them (the 0 / 1 signal indicating whether there is an overflow given by the probe) is used as the label of this sample. Its physical meaning is that each time the model calculates the features continuously generated in the previous period of time to determine whether the current time point is in a dangerous state of about to overflow. Since spiking neural networks usually input and output one frame at a time in a cyclic manner, their output is also several frames of spike signals.

[0087] The training is carried out using the commonly used backpropagation algorithm based on surrogate gradients in the industry. During the model training process, first, the effects of several different feature calculation methods are compared, and it is found that the effect of the aforementioned feature calculation method is significantly better than directly calculating the difference according to the time interval (experiments and adjustments are also made on different sizes and numbers of time windows). Then, the effects of different input frame numbers for a single sample are compared, and it is found that the larger the input frame number, the better the continuity of the output pulses (the 0 / 1 in the time sequence is denser), but it is not that the higher the correct rate. Considering the high real-time requirement for the actual inference of the model, the number of neural network units is adjusted on the premise of minimizing the model parameters to achieve the balance between the effect and efficiency of the model.

[0088] 2.5 Integration with the health kettle control system

[0089] The trained spiking neural network model is embedded into the control system of the health kettle, enabling it to perform feature calculation and model inference periodically according to the data transmitted by the capacitance sensor. When the model determines that the liquid is about to overflow, it controls the health kettle to turn off the fire temporarily; then, when the model determines that the liquid does not overflow, it continues to execute the predetermined heating process. Repeat this process to achieve real-time anti-overflow of the entire process controlled by the algorithm.

[0090] The anti-overflow algorithm for the health kettle proposed in this study has high feasibility and can become an effective technical solution to solve the problem of liquid overflow in the health kettle. At the same time, this paper proposes a fully automatic data acquisition scheme, which effectively combines the demand characteristics of the AI algorithm, making it possible to significantly reduce the labor cost and laying a solid foundation for the mass production of the AI intelligent anti-overflow health kettle. Finally, during the actual boiling process after deployment, the model can judge in time every time it is about to overflow, and the misjudgment rate is low, that is, the recall rate is as high as possible while ensuring an accuracy rate of 100%.

[0091] The algorithms and displays provided herein are not inherently related to any particular computer, virtual apparatus, or other device. Various general-purpose devices may also be used in conjunction with the teachings presented herein. The structure required to construct such devices will be apparent from the above description. Additionally, the present invention is not directed to any particular programming language. It should be understood that the teachings of the present invention described herein can be implemented in various programming languages, and the description of a particular language above is for the purpose of disclosing the best mode of the present invention.

[0092] In the specification provided herein, numerous specific details are set forth. However, it can be understood that embodiments of the present invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0093] Similarly, it should be understood that, in order to streamline this disclosure and assist in understanding one or more of the various inventive aspects, in the foregoing description of exemplary embodiments of the present invention, the various features of the present invention are sometimes grouped together in a single embodiment, figure, or description thereof. However, the disclosed method should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim. Rather, as the following claims reflect, the inventive aspects lie in less than all of the features of the preceding single embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of the present invention.

[0094] Those skilled in the art will appreciate that the modules in the devices in the embodiments can be adaptively changed and disposed in one or more devices different from those of the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and in addition, they can be divided into multiple sub-modules or sub-units or sub-components. Except where some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all the features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all the processes or units of any method or device so disclosed. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) can be replaced by an alternative feature providing the same, equivalent, or similar purpose.

[0095] In addition, those skilled in the art can understand that although some of the embodiments described herein include certain features included in other embodiments rather than other features, the combination of features of different embodiments is within the scope of the present invention and forms different embodiments. For example, in the following claims, any one of the claimed embodiments can be used in any combination.

[0096] Each component embodiment of the present invention can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. The present invention can also be implemented as a device or apparatus program (e.g., a computer program and a computer program product) for performing part or all of the methods described herein. Such a program for implementing the present invention can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, or provided on a carrier signal, or in any other form.

[0097] For example, Figure 5 The schematic architectural diagram of the earphone according to the present invention on a program system is shown. The earphone conventionally includes a processor 101 and a memory 102 arranged to store computer-executable instructions (program code). The memory 102 can be an electronic memory such as flash memory, EEPROM (electrically erasable programmable read-only memory), EPROM, hard disk, or ROM. The memory 102 has a storage space 103 for storing program code 104 for performing any method step in the embodiment. For example, the storage space 103 for the program code can include respective program codes 104 for implementing various steps in the above method. These program codes can be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, compact discs (CDs), memory cards, or floppy disks. Such computer program products are generally, for example Figure 6 the computer-readable storage media described. The computer-readable storage media can have storage segments, storage spaces, etc. arranged similarly to the memory 102 in the Figure 5 earphone. The program code can be compressed in an appropriate form, for example. Generally, the storage unit stores program code 111 for performing the method steps according to the present invention, that is, program code that can be read by a processor such as 101. When these program codes are run by the earphone, the earphone is caused to perform each step in the method described above.

[0098] It should be noted that the described embodiments illustrate the present invention rather than limit the present invention, and those skilled in the art can design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in the claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention can be implemented by means of hardware including several different elements and by means of a suitably programmed computer. In a unit claim listing several devices, several of these devices may be embodied by the same item of hardware. The use of the words first, second, and third, etc. does not denote any order. These words may be interpreted as names.

Claims

1. The AI ​​processing method for sensor data of a heating container is characterized in that: include: Periodically acquiring at least three capacitance values ​​of at least two non-contact detection electrodes at different heights on the outer wall of the heating container for detecting liquid level overflow as a frame of time series data; each of the capacitance values ​​is configured as a self-capacitance of electrodes at different heights or a mutual capacitance between electrodes at different heights, and at least one of the mutual capacitances exists in each frame of the time series data; Obtain the characteristics of each frame of time series data; The features of continuous frame time series data are used as the input of the neural network model to generate output information representing the liquid level overflow recognition result.

2. The sensor data AI processing method according to claim 1, characterized in that: Adopt SNN neural network as inference model and training model; The features of time series data are discretely pulse encoded, and the time pulse sequence formed by the encoding is used as the model input.

3. The sensor data AI processing method according to claim 2, characterized in that: The method of obtaining the features of each frame of time series data further includes: For each capacitance value of each detection electrode in each frame of data, the time difference of capacitance is calculated; The differential value is encoded into a discretized pulse signal according to the set boundary points.

4. The sensor data AI processing method according to claim 3, characterized in that: Given at least two time windows of different lengths; For each time window, obtaining the difference between the current capacitance value of the detection electrode and the maximum value and / or the minimum value in the time window from the current time to the past time; A processing operation is performed on each difference according to a given set of boundary points, and multiple differences form the time pulse sequence, wherein the processing operation is configured such that when the number of buckets generated by the boundary point is n, if the difference is in the x-th bucket, the encoding result is an array of length n, wherein the x-th bit is 1, and the other bits are 0, and n and x are integers.

5. The sensor data AI processing method according to claim 4, characterized in that: The period is configured to be 100 ms.

6. The sensor data AI processing method according to any one of claims 1 to 5, characterized in that: The capacitance values ​​of at least the first five non-contact detection electrodes at different heights from top to bottom of the outer wall of the heating container are periodically obtained to calculate the characteristics.

7. The sensor data AI processing method according to claim 1, characterized in that: In the process of data collection for training the model, the output signal of the contact liquid level sensor is used as the judgment condition of whether the liquid level reaches the overflow position, and the probe of the contact liquid level sensor is located at the position set to trigger the overflow prevention; The output signals of the contact level sensor are recorded as labels for the training model.

8. The sensor data AI processing method according to claim 7, characterized in that: During the data collection process of the training model, the contact liquid level sensor is controlled to stop heating after delaying a random time within a first set range after contacting the liquid, and / or the contact liquid level sensor is controlled to start reheating after delaying a random time within a second set range when detecting that the liquid is detached.

9. The sensor data AI processing method according to claim 7, characterized in that: The model training method further includes: Dividing the time series data of capacitance obtained by sampling the detection electrodes into a training set, a validation set, and a test set; The features of each N-frame continuous time series data are taken as a sample data, where N is an integer greater than 2. The output signal of the contact liquid level sensor of the last frame in each N-frame is taken as the label. The goal is to train a binary classification model to determine whether the last frame is in an overflow state. The model uses the training set for learning, and calculates the loss on the validation set after each round. The mean square error is used as the loss function. After several rounds of learning, the version with the lowest loss on the validation set is selected, and the indicators are calculated on the test set to evaluate the model performance. Continuously iterate the training to adjust various parameters and obtain the model with the best test set indicators.

10. The sensor data AI processing method according to claim 9, characterized in that: The model is a multi-layer spiking neural network. Each time it executes the loop, it inputs N consecutive frames of features and obtains two output values, which represent the possibility of overflow and no overflow respectively.

11. A method for cooking food in a heating container, characterized in that: include: The sensor data AI processing method according to any one of claims 1 to 10; and According to the liquid level overflow recognition result output by the neural network model, heating is stopped when it is determined that the liquid level reaches the overflow position; When it is determined that the liquid level has dropped, the container heating is executed again.

12. The method for cooking food according to claim 11, characterized in that: The rewarmed nodes are configured as: When the neural network model outputs the recognition result that the liquid level is no longer overflowing, the heating is started with a delay of up to 3 seconds; Alternatively, the liquid level is determined based on capacitance data collected by the sensor, and heating is started when the height difference between the liquid level dropped to the initial liquid level when the liquid level is just heated and is less than a threshold.

13. A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the sensor data AI processing method according to any one of claims 1 to 12.

14. A heating container, characterized in that: It includes a capacitance digital conversion circuit, a processor, and a contactless detection unit; The non-contact detection unit includes at least two vertically spaced non-contact capacitance detection electrodes arranged on the outer wall of the heating container for detecting liquid level overflow; The capacitance-to-digital conversion circuit is coupled to each detection electrode to obtain a corresponding capacitance value; The processor is coupled to the capacitance-to-digital conversion circuit and executes the computer-readable storage medium of claim 13 .

15. The heating container according to claim 14, characterized in that: The capacitance digital conversion circuit and processor are integrated using an R-SpiNNaker chip running an SNN inference model, and the inference model is written into the processor of the R-SpiNNaker chip to jointly control capacitance acquisition and inference.

Images