A multi-modal data collaborative analysis processing method and system of an electronic atomizer

Abstract

The application relates to the technical field of electronic atomizer control, and discloses a multi-modal data collaborative analysis processing method and system of an electronic atomizer, wherein the multi-modal data collaborative analysis processing method of the electronic atomizer comprises the following steps: collecting temperature response data and impedance spectrum data to form a physical modal data set after a new atomization cartridge is installed; extracting an enhanced feature vector and outputting an initial liquid species component probability distribution vector through an inference network of a band adaptation module; calculating an atomization resistance correction coefficient based on the probability distribution; collecting multi-channel data flow of a user behavior modal; performing affine transformation on a negative pressure waveform feature by using the atomization resistance correction coefficient, and then performing user identity recognition; and configuring a temperature control parameter and performing fuzzy PID closed-loop heating control. The application solves the technical problem that, in the case that multiple users share an equipment and replace a new atomization cartridge, liquid species identification and user identity recognition circularly depend on each other, and a multi-modal collaborative analysis link is broken.

Description

Technical Field

[0001] This invention relates to the field of electronic atomizer control technology, and more specifically, to a multimodal data collaborative analysis and processing method and system for electronic atomizers. Background Technology

[0002] In multi-user shared device scenarios, electronic atomizers require collaborative analysis of multimodal sensor data to achieve functions such as liquid type identification, user identification, and personalized concentration adjustment. Liquid type identification relies on matching impedance spectrum characteristics with known steady-state characteristic distributions to infer the atomized liquid composition, while user identification relies on matching the biomechanical characteristics of oral negative pressure waveforms to confirm user identity.

[0003] Existing multimodal collaborative analysis methods can operate under normal usage scenarios, but they face collaborative failure issues in complex scenarios involving multiple users sharing equipment and replacing new atomizing cartridges. When a new atomizing cartridge is used for the first time, residual air bubbles at the interface between the heating element and the atomizing liquid, as well as the oxide layer on the surface of the new heating element, cause systematic deviations in impedance spectroscopy measurements. Domain-adaptive correction of these deviations typically requires accessing the user's individual historical usage data, but this data is unavailable before user identification. Simultaneously, differences in viscosity and surface tension among different atomizing liquid formulations alter atomization resistance characteristics, causing deformation shifts in the oral negative pressure waveform for the same user under different formulation conditions. This leads to reduced biomechanical fingerprint matching similarity, and the formulation shift correction of the negative pressure waveform relies on the liquid identification results. The analysis tasks of these two modalities mutually depend on each other's outputs for correction, forming a circular dependency, resulting in a break in the multimodal collaborative analysis chain, making it impossible to correctly configure personalized concentration target parameters and temperature control parameters. Summary of the Invention

[0004] This invention provides a multimodal data collaborative analysis and processing method and system for electronic atomizers, which solves the technical problems in related technologies such as electronic atomizers being unable to automatically identify the components of the atomized liquid, being unable to distinguish different users' vaping habits, and having poor atomization effects due to fixed temperature control parameters.

[0005] This invention discloses a multi-modal data collaborative analysis and processing method for electronic atomizers, comprising the following steps: after detecting a new atomizing cartridge installation event, applying a low-power wetting pulse to the heating element and collecting temperature response data; after wetting is completed, applying a multi-frequency AC excitation signal to the heating element to obtain the first impedance spectrum data; and combining the temperature response data with the first impedance spectrum data to generate a physical property modal dataset. Feature extraction is performed on the physical property modality dataset to generate enhanced feature vectors. The enhanced feature vectors are then input into the atomized liquid composition inference network with a domain adaptation module to output the initial liquid composition probability distribution vector. Based on the initial liquid component probability distribution vector, the weighted average viscosity of the candidate formulation is calculated in a probability-weighted manner. The atomization resistance correction coefficient is calculated using the weighted average viscosity and the preset standard reference viscosity value. The atomization resistance correction coefficient is used as the formulation-related fingerprint correction parameter. When the user initiates suction, the oral negative pressure waveform signal, aerosol particle concentration signal and heating element temperature signal are collected. The original biomechanical fingerprint feature vector is generated by extracting features from the oral negative pressure waveform signal. The formula-related fingerprint correction parameters are used to perform an affine transformation on the feature components of the original biomechanical fingerprint feature vector that are affected by the formula to generate a formula-corrected biomechanical fingerprint feature vector. The formula-corrected biomechanical fingerprint feature vector is then matched with the pre-stored user prototype vectors to identify the user's identity. Based on the initial liquid component probability distribution vector and the identified user identity, temperature control parameters and concentration target parameters are configured to perform closed-loop heating control of the heating element.

[0006] Furthermore, feature extraction is performed on the aforementioned physical property modality dataset to generate enhanced feature vectors, including: The temperature response data is processed to calculate the thermal time constant and steady-state temperature rise amplitude. Based on the thermal time constant, the steady-state temperature rise amplitude, and the known heat capacity parameters and known heat dissipation area parameters of the heating element, the initial estimate of the equivalent thermal conductivity is calculated according to the lumped parameter thermal model. The thermal time constant, the steady-state temperature rise amplitude, and the initial estimate of the equivalent thermal conductivity are combined to generate auxiliary thermal property characteristics. Wherein, the thermal time constant is the time length corresponding to the temperature rise curve from the initial temperature to 63.2% of the steady-state temperature rise amplitude; the calculation method of the lumped parameter thermal model is: the initial estimate of the equivalent thermal conductivity is equal to the product of the heating element mass and the heating element specific heat capacity divided by the product of the thermal time constant and the heating element heat dissipation area; The initial impedance spectrum data was fitted using an equivalent circuit consisting of an ohmic resistor, a charge transfer resistor, and a constant-phase element, and a three-component impedance feature vector was extracted, consisting of the ohmic resistor value, the charge transfer resistor value, and the constant-phase element index. The enhanced feature vector is formed by concatenating the components of the thermophysical auxiliary feature and the three-component impedance feature vector after Z-score normalization preprocessing.

[0007] Furthermore, the atomized liquid composition inference network with the domain adaptation module includes a feature encoding unit, a domain adaptation unit, and a classification output unit connected in sequence; The feature encoding unit receives the enhanced feature vector and outputs the intermediate layer feature representation; The domain adaptation unit calculates the Mahalanobis distance between the enhanced feature vector and the pre-stored known steady-state feature distribution center, performs affine correction on the intermediate layer feature representation based on the Mahalanobis distance, and outputs the corrected intermediate layer feature representation; wherein, the Mahalanobis distance is calculated by taking the square root of the square root of the difference vector between the current enhanced feature vector and the known steady-state feature distribution center vector after performing a quadratic form operation on the inverse matrix of the known steady-state feature distribution; the affine correction is calculated as follows: The intermediate layer feature representation is multiplied by a scaling factor determined based on Mahalanobis distance, and then a translation factor determined based on Mahalanobis distance is added. The scaling factor and the translation factor are output by the mapping function obtained by pre-training in the domain adaptation unit with Mahalanobis distance as input. The classification output unit receives the corrected intermediate layer feature representation, normalizes it using the Softmax activation function, and outputs the initial liquid component probability distribution vector.

[0008] Furthermore, both the feature encoding unit and the classification output unit are multilayer perceptrons; The mapping functions for the output scaling coefficients and the output translation coefficients in the domain adaptation unit are each implemented by a single-input, single-output multilayer perceptron. The output layer activation function of the multilayer perceptron for the output scaling coefficients uses the Softplus function, while the output layer of the multilayer perceptron for the output translation coefficients is not subject to activation function constraints. The feature encoding unit, the mapping functions in the domain adaptation unit, and the classification output unit are jointly trained end-to-end, and the loss function is cross-entropy loss.

[0009] Furthermore, feature extraction is performed on the oral negative pressure waveform signal to generate an original biomechanical fingerprint feature vector, including: The negative pressure peak value, negative pressure rise slope, negative pressure duration, waveform radius of curvature, and waveform symmetry coefficient are extracted from the oral negative pressure waveform signal. The above five feature components are preprocessed by Z-score normalization and then combined to generate the original biomechanical fingerprint feature vector. Wherein, the negative pressure peak value is the maximum negative deviation of air pressure relative to ambient atmospheric pressure during a complete suction process; the negative pressure rise slope is the ratio of the air pressure change from the start of suction to the negative pressure peak value to the corresponding time length; The duration of negative pressure is the total length of the continuous time period during which the air pressure is lower than the preset negative pressure threshold; the radius of curvature of the waveform is the radius of curvature value calculated by the second derivative at the peak point after fitting the peak neighborhood with a second-order polynomial; the waveform symmetry coefficient is the ratio of the duration of the negative pressure rising segment to the duration of the negative pressure falling segment with the negative pressure peak time as the axis of symmetry. The step of performing an affine transformation on the feature components affected by the formula using the formula-related fingerprint correction parameters includes: dividing the negative pressure peak value and the negative pressure rise slope by the atomization resistance correction coefficient to obtain the corrected negative pressure peak value and the corrected negative pressure rise slope, respectively, while keeping the negative pressure duration, the waveform radius of curvature, and the waveform symmetry coefficient unchanged, which together with the corrected negative pressure peak value and the corrected negative pressure rise slope form the formula-corrected biomechanical fingerprint feature vector.

[0010] Furthermore, the biomechanical fingerprint feature vector after formula correction is matched with the pre-stored user prototype vectors to identify the user's identity, including: The cosine similarity between the biomechanical fingerprint feature vector after formula correction and each pre-stored user prototype vector is calculated to obtain the maximum similarity value and its corresponding user prototype identifier. When the maximum similarity value exceeds the preset identity verification threshold, the corresponding user ID is output; when the maximum similarity value is lower than the preset new user detection threshold, the biomechanical fingerprint feature vector after formula correction is stored as a new prototype vector and a new user ID is assigned and output; when the maximum similarity value is between the identity verification threshold and the new user detection threshold, a temporary tag is output; wherein, the identity verification threshold is higher than the new user detection threshold.

[0011] Furthermore, when the identified user identity matches a known user, the historical usage formula list and corresponding impedance spectrum feature statistics are extracted from the user's personalized parameter storage area. The prior probability weight of each candidate formula is determined using the historical usage formula list. Bayesian posterior update is performed on the initial liquid component probability distribution vector, and the liquid component probability distribution vector after user condition correction is output. Subsequent temperature control parameter configuration is based on the liquid component probability distribution vector after user condition correction. The Bayesian posterior update is calculated as follows: the probability value of the i-th candidate formulation after the update is equal to the product of the probability value of the i-th candidate formulation in the initial liquid component probability distribution vector and the prior probability weight of the i-th candidate formulation, divided by the sum of the products of the probability values ​​of all candidate formulations and their corresponding prior probability weights. The prior probability weight is determined by the proportion of the user’s historical usage of each recipe to the total usage, and a preset low probability baseline value greater than zero is taken for recipes that have not been used. When the identification result is a new user identifier or a temporary tag, the initial liquid component probability distribution vector remains unchanged.

[0012] Furthermore, configuring temperature control parameters based on the initial liquid component probability distribution vector and the identified user identity includes: When the highest probability value in the initial liquid component probability distribution vector is lower than the preset confidence level safety threshold, the lowest safe heating temperature limit and the lowest safe heating power limit are taken from the temperature control parameters corresponding to each candidate formula as the current temperature control parameters; when the highest probability value is not lower than the confidence level safety threshold, the standard temperature control parameters of the formula corresponding to the highest probability value are directly retrieved. During the closed-loop heating control process, the power and temperature dynamic response curves of the heating element are collected synchronously. When a temperature plateau range is detected, the actual boiling point temperature is extracted. The actual boiling point temperature is matched with the theoretical boiling point temperature of each candidate formula. When the absolute value of the difference between the two is lower than the preset boiling point matching threshold, the corresponding candidate formula is confirmed as the final formula. Based on the final formula, the precise temperature control parameters are retrieved to replace the current temperature control parameters. The temperature plateau interval is detected by calculating the temperature change rate of the power and temperature dynamic response curves according to time windows. When the absolute value of the temperature change rate is lower than the preset plateau judgment threshold within a consecutive preset number of time windows, it is determined to be a temperature plateau interval.

[0013] Furthermore, after the closed-loop heating control is completed, it also includes: The confirmed final formula is associated with the user ID and stored in the corresponding user's formula association file. The formula-related fingerprint correction parameters are recalculated using the accurate viscosity value of the final formula. The biomechanical fingerprint feature vector of the formula after correction obtained by the current aspiration is corrected a second time. The prototype vector of the user is updated by incremental mean. The incremental mean update is calculated as follows: the updated prototype vector is equal to the product of the number of samples accumulated by the user and the prototype vector before the update, plus the biomechanical fingerprint feature vector after secondary correction, and then divided by the number of samples accumulated by the user plus one. The concentration, airflow matching index, and suction stage data of this port are stored in the corresponding preference cache according to the user ID. After accumulating a preset number of suction records, the stage and concentration target mapping table is updated by exponential weighted moving average. The calculation method of the exponential weighted moving average update is: the updated target concentration value is equal to the product of the preset smoothing coefficient and the concentration preference statistics calculated based on the preference cache, plus one minus the product of the smoothing coefficient and the current target concentration value. The concentration preference statistics are the average aerosol particle concentration signals of suction records in the preference cache whose concentration and airflow matching index are higher than the preset matching qualification threshold within the corresponding time window; the preference cache is stored in isolation according to user ID.

[0014] This invention provides a multimodal data collaborative analysis and processing system for electronic atomizers, comprising: The physical property modal data acquisition module is used to apply a low-power wetting pulse to the heating body and acquire temperature response data after detecting a new atomized projectile installation event. After wetting is completed, a multi-frequency AC excitation signal is applied to the heating body to obtain the first impedance spectrum data. The temperature response data and the first impedance spectrum data are combined to generate a physical property modal dataset. The liquid component inference module is used to extract features from the physical property modality dataset to generate an enhanced feature vector, input the enhanced feature vector into the atomized liquid component inference network of the domain adaptation module, and output the initial liquid component probability distribution vector. The fingerprint correction parameter generation module is used to calculate the weighted average viscosity of the candidate formulation based on the probability distribution vector of the initial liquid components in a probability-weighted manner, and to calculate the atomization resistance correction coefficient as the formulation-related fingerprint correction parameter using the weighted average viscosity and the preset standard reference viscosity value. The user behavior data acquisition module is used to collect oral negative pressure waveform signals, aerosol particle concentration signals, and heating element temperature signals when the user initiates suction. The user identification module is used to extract features from the oral negative pressure waveform signal to generate an original biomechanical fingerprint feature vector, perform affine transformation on the feature components affected by the formula using the formula-related fingerprint correction parameters to generate a formula-corrected biomechanical fingerprint feature vector, and perform similarity matching between the formula-corrected biomechanical fingerprint feature vector and the pre-stored user prototype vectors to identify the user's identity. The closed-loop heating control module is used to configure temperature control parameters and concentration target parameters based on the initial liquid component probability distribution vector and the identified user identity, and to perform closed-loop heating control of the heating element.

[0015] This invention solves the technical problem of broken multimodal collaborative analysis links caused by the cyclic dependence of liquid type identification and user identity recognition in a composite cold start scenario where multiple users share equipment and replace new atomizing cartridges, by independently collecting physical property modal data and inferring liquid type composition before the user's inhalation behavior occurs. It achieves the technical effect of obtaining effective personalized concentration target parameters and temperature control parameter configurations on the first inhalation after installing a new atomizing cartridge. The domain adaptation unit compensates for the systematic deviation of the impedance spectrum when a new atomizing cartridge is used for the first time by calculating the Mahalanobis distance online and performing affine correction, achieving the technical effect of outputting a distinguishable probability distribution of liquid type composition even in the absence of individual historical data. Based on the atomization resistance correction coefficient, affine transformation is performed on the characteristic components of the negative pressure waveform affected by the formula, achieving the technical effect of eliminating the interference of waveform deformation shift caused by different formulas on user identity matching. The preference cache is stored in isolation according to the user ID and the stage and concentration target mapping table are updated independently, achieving the technical effect of avoiding oscillation of personalized concentration target parameters caused by the mixing of multi-user data. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multimodal data collaborative analysis and processing method for electronic atomizers provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the impedance spectrum real and imaginary parts as a function of frequency, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the rate of temperature rise of the heating body during the immersion process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the Z-score standardized values ​​of each component of the enhanced feature vector provided in the embodiments of the present invention; Figure 5 This is a schematic diagram comparing the probability distribution of the formulation before and after the Bayesian posterior update, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of the formula weighted viscosity calculation and the contribution distribution of each formula provided in the embodiments of the present invention; Figure 7 This is a schematic diagram showing the comparison of biomechanical fingerprint feature vectors before and after correction according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the temperature dynamic response curve of the closed-loop heating process provided in an embodiment of the present invention. Detailed Implementation

[0017] In multimodal data collaborative analysis and processing scenarios for electronic atomizers, the device typically needs to collect sensor data from multiple modalities and perform collaborative analysis to achieve functions such as liquid type identification, user identification, and personalized concentration control. Among these, the liquid type identification modality relies on matching impedance spectrum characteristics with known steady-state characteristic distributions to infer the composition of the atomized liquid, while the user identification modality relies on matching the biomechanical characteristics of the oral negative pressure waveform to confirm the user's identity.

[0018] However, in complex scenarios involving multiple users sharing equipment and replacing new atomizing cartridges, the aforementioned multimodal collaborative analysis methods face the problem of collaborative failure. Specifically, when a new atomizing cartridge is used for the first time, residual air microbubbles at the interface between the heating element and the atomizing liquid, as well as the oxide layer on the surface of the new heating element, cause systematic deviations in impedance spectroscopy measurements, leading to a decrease in the confidence level of the component inference network output. Domain-adaptive correction of this systematic deviation usually requires historical usage data of individual users as a reference, but this data cannot be accessed when the user's identity has not yet been identified. At the same time, user identification relies on the biomechanical fingerprint matching of the oral negative pressure waveform, but differences in viscosity and surface tension of different atomizing liquid formulations alter the atomization resistance characteristics, causing deformation shifts in the negative pressure waveform of the same user under different formulation conditions, resulting in a general decrease in fingerprint matching similarity. The analysis tasks of the two modalities mentioned above are mutually dependent on each other's output results as correction basis, forming a circular dependency: without knowing the liquid type, it is impossible to correct the formulation shift of the negative pressure waveform; without knowing the user, it is impossible to use individual history to assist in liquid type identification, leading to a break in the multimodal collaborative analysis link, and the inability to correctly configure personalized concentration target parameters and temperature control parameters.

[0019] Therefore, there is a need for a multimodal data collaborative analysis and processing method for electronic atomizers that can break the above-mentioned cyclic dependence and restore the integrity of the multimodal collaborative analysis link under composite cold start conditions.

[0020] It should be understood that, according to the embodiments of this implementation, the multimodal data collaborative analysis and processing method for electronic atomizers is applied to electronic atomizer devices with the following hardware conditions: the device includes a heating element and its driving circuit, a multi-frequency AC excitation signal generator, a thermistor, a microphone sensor (miniature microphone-type pressure sensor), a light scattering sensor, and an embedded processor. The embedded processor has the computational capability to execute the data processing and control logic described in this embodiment and is equipped with non-volatile memory for storing user parameter files and a recipe database. The device is also equipped with a new atomizing cartridge installation detection mechanism for triggering the control process described in this embodiment.

[0021] At least one embodiment of the present invention discloses a method for collaborative analysis and processing of multimodal data of an electronic atomizer, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect physical property modal dataset; Upon detecting a new atomized projectile installation event, a low-power wetting pulse is applied to the heating element via a drive circuit, and the temperature response data of the heating element during the wetting process is acquired using a thermistor. After wetting is completed, a multi-frequency AC excitation signal is applied to the heating element via a multi-frequency AC excitation signal generator to scan and acquire the initial impedance spectrum data. The temperature response data and the initial impedance spectrum data are combined to generate a physical property modal dataset that is not dependent on user behavior.

[0022] It should be noted that the aforementioned low-power wetting pulse refers to a preset pulse signal with power lower than the normal atomization heating power. The power level of the low-power wetting pulse is insufficient to make the atomized liquid reach the boiling point; it is only used to accelerate the penetration and wetting process of the atomized liquid onto the surface of the heating element. The determination of wetting completion is based on the characteristic that the temperature rise rate tends to stabilize in the temperature response data. That is, when the change amplitude of the temperature rise rate is continuously lower than the preset change threshold, wetting is determined to be complete.

[0023] It should be noted that the aforementioned multi-frequency AC excitation signal refers to an AC signal sequence containing multiple discrete frequency points within a preset frequency range, with the excitation amplitude at each frequency point remaining consistent, and the scanning sequence executed sequentially from low frequency to high frequency. The initial impedance spectrum data is a set of complex impedance values ​​corresponding to each frequency point.

[0024] Step 2: Extract enhanced feature vectors and infer the probability distribution of initial liquid seed components; The temperature response data is processed to calculate the thermal time constant and steady-state temperature rise amplitude. Based on the thermal time constant and steady-state temperature rise amplitude, the initial estimate of the equivalent thermal conductivity of the interface between the heating element and the atomizing liquid is calculated. The thermal time constant, steady-state temperature rise amplitude and the initial estimate of the equivalent thermal conductivity are combined to generate auxiliary thermophysical characteristics.

[0025] It should be noted that the aforementioned thermal time constant refers to the time length corresponding to the temperature rise curve in the temperature response data rising from the initial temperature to 63.2% of the steady-state temperature rise amplitude. The steady-state temperature rise amplitude refers to the temperature difference between the initial temperature and the temperature rise curve in the temperature response data after it has stabilized. The initial estimate of the equivalent thermal conductivity is calculated based on the lumped-parameter thermal model, that is, it is obtained according to the standard calculation formula of the lumped-parameter thermal model based on the known heat capacity parameters of the heating element, the known heat dissipation area parameters, and the thermal time constant extracted from the temperature response data.

[0026] Furthermore, the standard calculation formula for the above lumped parameter thermal model is: in, This is an initial estimate of the equivalent thermal conductivity. For the mass of the heating element, The specific heat capacity of the heating element. The thermal time constant, This refers to the heat dissipation area of ​​the heating element. (The above...) , , All parameters are known parameters of the heating element and are pre-stored in non-volatile memory; It was obtained from temperature response data.

[0027] Equivalent circuit fitting is performed on the initial impedance spectrum data to extract the three-component impedance eigenvector. The thermophysical auxiliary features are then concatenated with the three-component impedance eigenvector to form an enhanced eigenvector.

[0028] It should be noted that the equivalent circuit used in the above equivalent circuit fitting is a Randle circuit that includes an ohmic resistor, a charge transfer resistor, and a constant-phase element. The three-component impedance eigenvector consists of three components: the fitted ohmic resistance value, the charge transfer resistance value, and the constant-phase element exponent.

[0029] Furthermore, the complex impedance expression for the above Randles circuit is: in, For frequency The corresponding complex impedance value, For ohm resistance, For charge transfer resistor, The admittance coefficient of a constant-phase element. For constant phase element index and , The imaginary unit, Let be the angular frequency of the AC excitation signal. The sum of squared residuals between the measured complex impedance values ​​at each frequency point in the initial impedance spectrum data and the theoretical values ​​in the complex impedance expression of the Randle circuit above is minimized using the nonlinear least squares method. , and The resulting fitting result is the three components of the three-component impedance eigenvector.

[0030] Before concatenating the thermal property auxiliary features with the three-component impedance eigenvector to form the enhanced eigenvector, data preprocessing is required for each component. The thermal time constant, steady-state temperature rise amplitude, and initial estimated equivalent thermal conductivity in the thermal property auxiliary features differ significantly in dimensions and numerical range from the ohmic resistance, charge transfer resistance, and constant-phase element index components in the three-component impedance eigenvector. Directly concatenating these components and inputting them into the neural network would lead to an unbalanced impact on network training due to the different dimensions of the feature components. Therefore, each component is preprocessed using Z-score standardization to eliminate the influence of dimensional and numerical range differences on subsequent network operations before the concatenation operation is performed.

[0031] The enhanced feature vector is input into the atomized liquid composition inference network with domain adaptation module, and the initial liquid composition probability distribution vector is output.

[0032] The aforementioned atomized liquid composition inference network with a domain adaptation module comprises a feature encoding unit, a domain adaptation unit, and a classification output unit, which are connected sequentially according to the data flow direction. The feature encoding unit receives the enhanced feature vector and outputs an intermediate layer feature representation. The domain adaptation unit receives the intermediate layer feature representation, calculates the Mahalanobis distance between the enhanced feature vector and the pre-stored known steady-state feature distribution center online, performs affine correction on the intermediate layer feature representation based on the Mahalanobis distance, and outputs the corrected intermediate layer feature representation. The classification output unit receives the corrected intermediate layer feature representation and outputs the initial liquid composition probability distribution vector.

[0033] The formula for calculating the Mahalanobis distance is as follows: in, For the current enhanced feature vector, Given the center vector of the steady-state characteristic distribution, Given the covariance matrix of the steady-state characteristic distribution, Covariance matrix The inverse matrix, Indicates transpose. This is the Mahalanobis distance.

[0034] The formula for calculating the above affine correction is: in, This represents the features of the intermediate layer. This is the feature representation of the intermediate layer after correction. Based on Mahalanobis distance A defined scaling factor, Based on Mahalanobis distance Determined translation coefficients. and All of these are obtained from pre-training in the domain adaptation unit regarding... The output value of the mapping function.

[0035] Furthermore, in the aforementioned atomized liquid component inference network with a domain adaptation module, the feature encoding unit is a multilayer perceptron. The input to the feature encoding unit is the enhanced feature vector, and the output is the intermediate layer feature representation. Similarly, the classification output unit is also a multilayer perceptron. The input to the classification output unit is the corrected intermediate layer feature representation, and the output of the classification output unit, after being normalized by the Softmax activation function, yields the initial liquid component probability distribution vector. The sum of all components in the initial liquid component probability distribution vector is 1, and each component corresponds to the probability value of a candidate formulation. The mapping function in the domain adaptation unit... and Each is implemented using a single-input single-output multilayer perceptron, with The input consists of scaling and translation coefficients, which are output separately. The output layer of the multilayer perceptron for scaling coefficients is a single-neuron fully connected layer, and the Softplus activation function ensures the output value is always positive. The output layer of the multilayer perceptron for translation coefficients is also a single-neuron fully connected layer, with no activation function constraint to allow the output value to cover both positive and negative ranges. The feature encoding unit, the mapping function in the domain adaptation unit, and the classification output unit are jointly trained end-to-end. The training data consists of enhanced feature vector samples collected under steady-state conditions of various known formulations and simulated new bullet deviation conditions. The labels are the corresponding formulation categories. The loss function is cross-entropy loss, and the optimization strategy is the Adam optimization algorithm.

[0036] In this embodiment, the domain adaptation unit functions as follows: when a new atomizing projectile is used for the first time, residual air microbubbles at the interface between the heating element and the atomizing liquid, along with the oxide layer on the surface of the heating element, cause a systematic deviation in the impedance spectrum measurement, resulting in the enhanced eigenvector deviating from the known steady-state characteristic distribution. The domain adaptation unit calculates the degree of deviation (i.e., Mahalanobis distance) and performs affine correction on the intermediate layer characteristic representation accordingly, pulling the offset intermediate layer characteristic representation back to the nearest neighbor region of the steady-state distribution, thereby compensating for the systematic deviation without requiring individual historical data.

[0037] Step 3: Generate formula-related fingerprint correction parameters; Based on the initial liquid component probability distribution vector, the viscosity and surface tension values ​​corresponding to each candidate formulation are retrieved from the pre-stored component and fluid property parameter database in a probability-weighted manner, and the weighted average viscosity of the candidate formulation is calculated.

[0038] The formula for calculating the weighted average viscosity is as follows: in, The total number of candidate formulations. The first element in the initial liquid component probability distribution vector is... The probability values ​​corresponding to each candidate formula For the first The viscosity values ​​corresponding to each candidate formulation The candidate formulation number and , This is the weighted average viscosity.

[0039] Using a fluid dynamics pipe resistance model, the weighted average viscosity is... The atomization resistance correction coefficient is obtained by calculating the pre-stored geometric parameters of the atomizer airflow. .

[0040] The above calculation of the atomization resistance correction coefficient is based on a simplified form of the Hagen-Poiseuille equation: in, The preset standard reference viscosity value, Characterizes the ratio of atomization resistance under current formulation conditions to that under standard reference conditions. This is due to the weighted average viscosity. Compared with standard reference viscosity value The two have the same dimensions, and the atomization resistance correction factor is obtained by dividing them. It is a dimensionless ratio that can be directly used in subsequent affine transformation operations.

[0041] Correction factor for atomization resistance Output as formula-related fingerprint correction parameters.

[0042] It should be noted that the above standard reference viscosity values This refers to the viscosity value corresponding to the reference formula used when training the fingerprint database for user identification; the standard reference viscosity value. It is pre-stored in non-volatile memory when the device leaves the factory.

[0043] Step 4: Collect real-time data streams of user behavior modalities across multiple channels; When a user initiates their first inhalation, the system collects the oral cavity negative pressure waveform signal through the microphone sensor, the aerosol particle concentration signal through the light scattering sensor, and the heating element temperature signal through the thermistor, simultaneously generating a multi-channel real-time data stream of the user's behavior modality.

[0044] It should be noted that the aforementioned oral negative pressure waveform signal refers to the time-series data of airway pressure changes continuously collected by the microphone sensor at a preset sampling frequency during the suction process. The aerosol particle concentration signal refers to the voltage time-series data output by the light scattering sensor, which is proportional to the intensity of light scattered by aerosol particles. The heating element temperature signal refers to the temperature time-series data continuously collected by the thermistor at a preset sampling frequency during the heating process.

[0045] Step 5: Correct the oral negative pressure waveform according to the formula and identify the user; Feature extraction is performed on the oral negative pressure waveform signal to obtain the negative pressure peak value, negative pressure rise slope, negative pressure duration, waveform radius of curvature, and waveform symmetry coefficient. The above five features are combined to generate the original biomechanical fingerprint feature vector.

[0046] Before combining the five features mentioned above to generate the original biomechanical fingerprint feature vector, data preprocessing is required for each feature component. The negative pressure peak value, negative pressure rise slope, negative pressure duration, waveform radius of curvature, and waveform symmetry coefficient have different dimensions and numerical ranges. Direct combination will lead to dimensional differences that affect the rationality of subsequent cosine similarity calculations. Therefore, each component is preprocessed using Z-score standardization to eliminate dimensional differences before the combination operation is performed.

[0047] Furthermore, the aforementioned negative pressure peak value refers to the maximum negative deviation of air pressure relative to ambient atmospheric pressure during a complete aspiration process of the oral negative pressure waveform signal; the negative pressure rise slope refers to the ratio of the air pressure change to the corresponding time length between the start of aspiration and the peak value of the oral negative pressure waveform signal; the negative pressure duration refers to the total length of the continuous time period in the oral negative pressure waveform signal where the air pressure is lower than the preset negative pressure judgment threshold; the waveform curvature radius refers to the radius of curvature calculated from the second derivative of the fitted curve at the peak point after performing a second-order polynomial fitting on the peak neighborhood of the oral negative pressure waveform signal; and the waveform symmetry coefficient refers to the ratio of the duration of the negative pressure rise segment to the duration of the negative pressure fall segment with the peak value of negative pressure as the axis of symmetry.

[0048] Affine transformations were performed on the negative pressure peak and negative pressure rise slope in the original biomechanical fingerprint feature vector using formulation-related fingerprint correction parameters to generate a formulation-corrected biomechanical fingerprint feature vector.

[0049] The above affine transformation is handled as follows: in, This is the measured peak negative pressure. This is the peak negative pressure after correction. To measure the slope of the negative pressure rise, To correct the slope of the negative pressure rise. This is the atomization resistance correction factor. (Due to the atomization resistance correction factor...) Since the ratio is dimensionless, the above division operation does not introduce dimensional changes, and the corrected value is equivalent to the fingerprint value under standard resistance conditions. The negative pressure duration, waveform radius of curvature, and waveform symmetry coefficient in the original biomechanical fingerprint feature vector remain unchanged, and together with the corrected negative pressure peak value and the corrected negative pressure rise slope, they form the formula-corrected biomechanical fingerprint feature vector.

[0050] The cosine similarity is calculated between the formula-corrected biomechanical fingerprint feature vector and the pre-stored user prototype vectors to obtain the maximum similarity value and its corresponding user prototype identifier. When the maximum similarity value exceeds a preset identity verification threshold, the user identifier ID corresponding to that user prototype identifier is output. When the maximum similarity value is lower than a preset new user detection threshold, the formula-corrected biomechanical fingerprint feature vector is stored as a new prototype vector, and a new user identifier ID is assigned and output. When the maximum similarity value is between the identity verification threshold and the new user detection threshold, a temporary flag is output.

[0051] The cosine similarity is calculated using the biomechanical fingerprint feature vector after formulation correction. and the prototype vector of a user Given the input vector, output the cosine similarity value between the two vectors.

[0052] It should be noted that the aforementioned identity verification threshold is higher than the new user detection threshold. The range between the identity verification threshold and the new user detection threshold is used to identify situations where the confidence level is insufficient for confirmation but also insufficient to determine if someone is a new user. The "suspended" flag indicates that the currently extracted data will be temporarily stored, and identity verification will be performed after more extracted data has been accumulated.

[0053] In this embodiment, the reason for performing an affine transformation on the negative pressure peak and negative pressure rise slope while keeping the other three characteristic components unchanged is that the negative pressure peak and negative pressure rise slope are directly affected by changes in atomization resistance, while the negative pressure duration, waveform radius of curvature, and waveform symmetry coefficient are mainly determined by the user's own respiratory muscle mechanics and are less affected by formula changes. Therefore, only the characteristic components significantly affected by the formula are corrected to avoid introducing unnecessary transformation errors into the remaining components.

[0054] Step 6: Configure temperature control parameters and perform closed-loop heating control; Based on the initial liquid component probability distribution vector, the corresponding temperature control parameters are retrieved from the component and thermophysical property parameter database. Combined with the user ID, the stage and concentration target mapping table is loaded from the corresponding personalized parameter storage area.

[0055] The aerosol particle concentration signal and the oral cavity negative pressure waveform signal are compared to identify the current suction stage. Based on the current suction stage, the target concentration value is obtained from the stage-concentration target mapping table, and the concentration deviation between the aerosol particle concentration signal and the target concentration value is calculated. The concentration and airflow matching degree between the aerosol particle concentration signal and the oral cavity negative pressure waveform signal is calculated to obtain a matching degree index. The concentration deviation value and the matching degree index are input into a fuzzy PID controller to generate a real-time power drive signal, driving the heating element for closed-loop heating control.

[0056] It should be noted that the above-mentioned aspiration phase identification refers to dividing a complete aspiration process into an initial inhalation phase, a stable inhalation phase, and a tail decay phase based on the amplitude variation trend of the oral negative pressure waveform signal. Each phase corresponds to a different target concentration value. The concentration-airflow matching degree calculation refers to calculating the Pearson correlation coefficient between the aerosol particle concentration signal and the oral negative pressure waveform signal after normalizing them based on the range on the time axis. This Pearson correlation coefficient reflects the degree of synchronization between the aerosol generation and the user's inspiratory airflow.

[0057] It should be noted that when the user ID is a new user ID, the preset default stage and concentration target mapping table are loaded. When the output is a suspending flag, the default stage and concentration target mapping table are also loaded.

[0058] In this embodiment, to ensure heating safety when the confidence level of liquid type identification is insufficient, when configuring temperature control parameters in step 6, if the highest probability value in the initial liquid type component probability distribution vector is lower than a preset confidence level safety threshold, a conservative lower bound value is taken from the temperature control parameters corresponding to each candidate formulation as the current temperature control parameter, i.e., the lowest safe heating temperature upper limit and the lowest safe heating power upper limit among each candidate formulation are selected; if the highest probability value exceeds the confidence level safety threshold, the standard temperature control parameter of the formulation corresponding to the highest probability value is directly retrieved. This process avoids the risk of overheating and decomposition of the atomizing liquid due to excessively high temperature control parameters under uncertain liquid type identification conditions.

[0059] Furthermore, the aforementioned confidence level safety threshold refers to a probability threshold value that is pre-set and stored in non-volatile memory, and the range of the confidence level safety threshold is as follows: This is used to determine whether the current liquid type identification result has sufficient credibility to support the direct retrieval of the standard temperature control parameters for the corresponding formula. The highest probability value in the initial liquid type component probability distribution vector... When the value is below the confidence level safety threshold, it indicates that the uncertainty of the current identification result is high, and a conservative lower limit temperature control parameter should be used to avoid the risk of overheating; when When the confidence level is not lower than the safety threshold, it indicates that the identification result has sufficient confidence and the standard temperature control parameters of the corresponding formula can be used directly.

[0060] In this embodiment, to further improve the accuracy of liquid type identification under the condition that the user's identity has been identified, the following steps are also included: When the user identifier ID output in step 5 matches a known user, the historical usage recipe list and corresponding impedance spectral feature statistics are extracted from the personalized parameter storage area of ​​the known user. Using the historical usage recipe list and corresponding impedance spectral feature statistics as user prior constraints, a Bayesian posterior update is performed on the initial liquid type component probability distribution vector, outputting the liquid type component probability distribution vector after user condition correction. Subsequent temperature control parameter configuration is retrieved based on the liquid type component probability distribution vector after user condition correction.

[0061] The formula for calculating the Bayesian posterior update is as follows: in, The first element in the initial liquid component probability distribution vector is... The probability value of each candidate formula. For the first recipe determined based on the user's historical usage list Prior probability weights for each candidate formulation For the updated version The probability value of each candidate formula. The total number of candidate formulations. The candidate formulation number and , To sum the candidate recipe indices and For this user's historically frequently used recipes, the prior probability weights are... The value is higher than the baseline value; for recipes that this user has never used before, the prior probability weight... The value is lower than the benchmark value.

[0062] Furthermore, the aforementioned prior probability weights The method for determining the prior probability weight is as follows: the proportion of the historical usage count of each recipe in the user's historical recipe list to the total usage count of the user is used as the prior probability weight of the corresponding recipe. Values; prior probability weights for recipes that the user has never used before. A preset low-probability baseline value is selected. This low-probability baseline value is greater than zero to retain the possibility of identification with a small probability, and it is pre-stored in non-volatile memory.

[0063] It should be noted that when the output of step 5 is a new user identifier or a temporary tag, the above Bayesian posterior update is not performed, and the initial liquid component probability distribution vector remains unchanged.

[0064] In this embodiment, to finally verify the liquid type identification result during the closed-loop heating control process, the following steps are also included in step 6 of the closed-loop heating control execution process: Simultaneously acquiring the power and temperature dynamic response curves of the heating element, and detecting whether a temperature plateau interval exists in the power and temperature dynamic response curves. When a temperature plateau interval is detected, extracting the actual boiling point temperature corresponding to the temperature plateau interval, matching the actual boiling point temperature with the theoretical boiling point temperature of each candidate formulation, and confirming the candidate formulation as the final formulation when the absolute value of the difference between the actual boiling point temperature and the theoretical boiling point temperature of a candidate formulation is lower than a preset boiling point matching threshold. Based on the final formulation, retrieving precise temperature control parameters from the composition and thermophysical property parameter database, and replacing the currently used temperature control parameters.

[0065] It should be noted that the aforementioned temperature plateau range refers to the section in the power-temperature dynamic response curve where the temperature value remains approximately constant over a certain period of time. The phenomenon corresponding to the temperature plateau range is the physical process in which the atomizing liquid absorbs latent heat during a phase change on the surface of the heating element, and its temperature temporarily does not rise with increasing power input. The temperature value corresponding to the temperature plateau range is the actual boiling point temperature of the atomizing liquid under the current operating pressure conditions. The aforementioned boiling point matching threshold is a pre-set temperature tolerance value stored in non-volatile memory, used to accommodate sensor measurement errors and boiling point fluctuations in different batches of atomizing liquid formulations.

[0066] Furthermore, the detection method for the aforementioned temperature plateau interval is as follows: calculate the temperature change rate for the power and temperature dynamic response curves according to time windows. When the absolute value of the temperature change rate is lower than the preset plateau judgment threshold within a consecutive preset number of time windows, the current segment is determined to be a temperature plateau interval, and the average value of the temperature sampling values ​​within the temperature plateau interval is taken as the actual boiling point temperature.

[0067] In this embodiment, to maintain the personalized parameters of each user and achieve long-term adaptive optimization in a scenario where multiple users share the same device, after the closed-loop heating control in step 6 is completed, the following steps are also included: associating the confirmed final formula with the user identifier ID and storing it in the corresponding user's formula association file; recalculating the formula-related fingerprint correction parameters using the precise viscosity value of the final formula, and using the recalculated formula-related fingerprint correction parameters to perform a secondary correction on the biomechanical fingerprint feature vector of the formula obtained from the current aspiration, and performing incremental mean update on the prototype vector of the user.

[0068] The formula for calculating the incremental mean update is as follows: in, For the first time before the update The prototype vector of each user This is the biomechanical fingerprint feature vector after secondary correction. This represents the number of samples that the user has accumulated. This is the updated prototype vector.

[0069] The concentration, airflow matching index, and suction stage data of this port are stored in the corresponding preference cache according to the user's ID. After the user accumulates a preset number of suction records, an exponentially weighted moving average is performed to update the stage and concentration target mapping table.

[0070] The formula for calculating the above-mentioned index-weighted moving average update is as follows: in, For the current stage, the concentration target mapping table is number 1. The target concentration value corresponding to each aspiration stage. The first digit calculated based on the accumulated suction records in the preference cache Concentration preference statistics for each suction phase The preset smoothing coefficient and , For the updated target concentration value, This is the sequence number of the suction stage.

[0071] Furthermore, the above concentration preference statistics The calculation method is as follows: retrieve the user's preference cache at the [number]th [period]. For each suction stage, if the historical concentration and airflow matching index are higher than the preset matching qualification threshold, the average aerosol particle concentration signal in these suction records within the corresponding time window is taken as the concentration preference statistical value. If the number of historical records meeting the criteria is insufficient, then the concentration preference statistic will be used. Maintain the current target concentration value This update will not be performed.

[0072] It should be noted that the above preference cache is stored in isolation according to user ID, and the preference data of each user is independent of each other, so as to avoid the mixing of suction data of different users when multiple users share the device, which may cause oscillation of personalized concentration target parameters.

[0073] This implementation addresses the cyclic dependency failure problem in multimodal data collaborative analysis in a complex cold start scenario where multiple users share equipment and replace new atomizing cartridges. By independently completing the acquisition of physical property modal data and liquid composition inference before the user's inhalation behavior, the liquid identification process is decoupled from the user behavior modality in terms of timing. This allows the liquid identification to output valid results independently without relying on user identity information, thereby breaking the cyclic dependency between liquid identification and user identity recognition.

[0074] Specifically, step 1 acquires physical property modal data completely unaffected by user behavior differences by collecting thermal response data during the immersion stage and impedance spectrum data after immersion. In step 2, the thermophysical auxiliary features provide a supplementary physical property information dimension independent of the impedance spectrum for component inference. The domain adaptation unit quantifies the feature offset by calculating the Mahalanobis distance online and performs affine correction accordingly, compensating for the systematic deviation of the impedance spectrum caused by the interfacial air microbubbles and the oxide layer on the heating element surface when the new atomizing bullet is used for the first time. This allows the atomizing liquid component inference network with the domain adaptation module to output a discriminative initial liquid type component probability distribution vector even under zero-sample conditions lacking individual historical data. Step 3 calculates the atomization resistance correction coefficient based on the initial liquid type component probability distribution vector. Step 5 uses the atomization resistance correction coefficient to perform affine transformation to restore the feature components in the negative pressure waveform affected by the formulation viscosity, eliminating the interference of waveform deformation offset caused by different formulations on user identity matching.

[0075] Therefore, this implementation transforms the two originally mutually dependent modal analysis tasks into a unidirectional information transmission link in time sequence—after the physical property modality independently generates liquid type identification results, it transmits formula correction information to the user behavior modality; the user identification results are then updated in reverse through Bayesian posterior updates to enhance the confidence of liquid type identification—forming a sequentially mutually beneficial collaborative analysis path. The processing method of independently updating the preference cache and the stage and concentration target mapping tables, isolated by user ID, avoids interference from the mixing of multi-user data on the convergence of personalized concentration target parameters. The synergistic effect of the above links restores the integrity of the multimodal collaborative analysis link in the complex cold start scenario, enabling personalized concentration target parameters and temperature control parameters to be effectively configured on the first inhalation after the installation of a new nebulizer.

[0076] The following is an example of an application of the present invention, such as Figure 2-8 As shown, the implementation process is as follows: A smart electronic atomizer device (device number DEV, A07) supporting multi-user sharing was used alternately by User A and User B on a certain day. The device is equipped with a heating element drive circuit, a multi-frequency AC excitation signal generator, a thermistor, a microphone sensor, a light scattering sensor, and an embedded processor. The non-volatile memory pre-stores the prototype vectors and formula databases for Users A and B (including three candidate formulas: F1, F2, and F3). After User A completed one use that day, User B installed a new atomizing cartridge and prepared for their first puff. The new cartridge installation detection mechanism triggered the control process, and the device entered a combined cold start state—the liquid type was unknown and the user's identity was unconfirmed. The following is a complete data flow example of this process.

[0077] Upon detecting the installation of a new atomizing projectile, the device applies a low-power wetting pulse (18% of the normal atomizing power) to the heating element, while a thermistor continuously collects the temperature response of the heating element at a sampling frequency of 100Hz. Wetting is considered complete when the rate of temperature rise is consistently below a preset threshold of 0.05℃ / s² for five consecutive time windows, a process lasting approximately 4.2 seconds. After wetting, a multi-frequency AC excitation signal generator sequentially applies AC excitation signals of uniform amplitude at eight discrete frequency points (logarithmically uniformly distributed from 100Hz to 100kHz), collecting the complex impedance values ​​at each frequency point to form the initial impedance spectrum data. The temperature response data is combined with the initial impedance spectrum data to generate a physical property modal dataset.

[0078] Table 1. Original Data Acquired for the Physical Property Modal Dataset Extract the thermal time constant and steady-state temperature rise from the temperature response data. The steady-state temperature rise is... The time it takes for the temperature to rise to 63.2% of its steady-state amplitude (i.e., 11.57℃) is approximately 1.6 seconds; therefore, the thermal time constant is... s. The known pre-stored parameters of the heating element are: mass g, specific heat capacity J / (g·℃), heat dissipation area m², substitute into the formula of the lumped parameter thermal model to calculate the initial estimate of the equivalent thermal conductivity: This yields three components for the auxiliary thermophysical properties: thermal time constant 1.6s, steady-state temperature rise amplitude 18.3℃, and initial estimate of equivalent thermal conductivity 734.2W / (m²·℃).

[0079] Nonlinear least-squares fitting using a Randle circuit was performed on the initial impedance spectrum data to extract the three-component impedance eigenvectors. Due to the presence of residual air microbubbles at the interface of the new atomizing bomb and the oxide layer on the surface of the heating element, the fitted ohmic resistance value is higher than the steady-state reference value.

[0080] The three components of the thermophysical auxiliary feature and the six components of the three-component impedance feature vector are Z-score normalized and concatenated to form a six-dimensional enhanced feature vector, which is then input into the atomized liquid composition inference network with a domain adaptation module. The domain adaptation unit calculates the Mahalanobis distance between the current enhanced feature vector and the known steady-state feature distribution center as 2.37, and outputs a scaling factor of 0.84 and a translation factor of 0.19. After performing affine correction on the intermediate layer feature representation, the classification output unit outputs the initial liquid component probability distribution vector after Softmax normalization.

[0081] Table 2 Comparison of each component of the enhanced feature vector before and after standardization and output of the initial probability distribution. Based on the initial liquid component probability distribution vector, viscosity values ​​of three candidate formulations are retrieved from the component and fluid property parameter database, and the weighted average viscosity is calculated by probability weighting: The standard reference viscosity value pre-stored in the non-volatile memory is 40.0 mPa·s. Calculate the atomization resistance correction factor: Table 3. Formulation property parameter retrieval and weighted calculation User B initiates the first suction, and the microphone sensor collects the oral negative pressure waveform signal at a sampling frequency of 500Hz. The light scattering sensor simultaneously collects the aerosol particle concentration signal, and the thermistor collects the heating element temperature signal. The three signals synchronously constitute a multi-channel real-time data stream of the user behavior modality, with a collection time of approximately 3.8s (corresponding to the complete suction process).

[0082] Five feature components were extracted from the oral negative pressure waveform signal and combined after Z-score normalization to form the original biomechanical fingerprint feature vector. An affine transformation was performed on the negative pressure peak and negative pressure rise slope using a formulation-related fingerprint correction parameter of 1.0425. The duration of negative pressure, the radius of curvature of the waveform, and the waveform symmetry coefficient remain unchanged. Together with the corrected negative pressure peak value and the negative pressure rise slope, they constitute the biomechanical fingerprint feature vector after formula correction.

[0083] Table 4 Comparison of biomechanical fingerprint feature vectors before and after correction The cosine similarity of the biomechanical fingerprint feature vector after formula correction with the prototype vectors of User A and User B in memory is calculated respectively. The cosine similarity of User B's prototype vector is 0.923, which exceeds the identity verification threshold of 0.85, so the user identifier is output as User B.

[0084] User B has been identified as a known user. Retrieve their historical recipe usage list from their personalized parameter storage area: total usage time 47 times, with recipe F1 used 34 times, recipe F2 used 11 times, and recipe F3 used 2 times. Calculate the prior probability weights for each recipe: the prior weight for recipe F1 is 34 / 47 ≈ 0.723, the prior weight for recipe F2 is 11 / 47 ≈ 0.234, and the prior weight for recipe F3 is 2 / 47 ≈ 0.043. Perform a Bayesian posterior update, taking recipe F1 as an example: Table 5 Comparison of probability distributions before and after Bayesian posterior update The updated probability value of Formula F1 (0.845) exceeds the confidence level safety threshold of 0.75. Therefore, the standard temperature control parameters for Formula F1 are directly retrieved (maximum safe heating temperature limit 215℃, standard heating power limit 8.4W). The stage and concentration target mapping table for User B is loaded from User B's personalized parameter storage area. The target concentration value for the initial inhalation stage is 1.82 mg / m³, the target concentration value for the stable inhalation stage is 2.46 mg / m³, and the target concentration value for the tail-end decay stage is 1.35 mg / m³.

[0085] Based on the amplitude change trend of the oral negative pressure waveform signal, the current aspiration is identified as entering the stable inhalation stage, with a target concentration value of 2.46 mg / m³. The current average aerosol particle concentration signal is 2.31 mg / m³, and the concentration deviation is [missing value]. mg / m³. The concentration and airflow matching index (Pearson correlation coefficient) were calculated simultaneously and found to be 0.87. The concentration deviation value of −0.15 mg / m³ and the matching index of 0.87 were input into the fuzzy PID controller to generate a real-time power drive signal to drive the closed-loop heating of the heating element.

[0086] During the closed-loop heating process, the power and temperature dynamic response curves exhibited a temperature plateau lasting approximately 0.6 seconds around 187℃ (the absolute value of the temperature change rate was below the plateau threshold of 0.3℃ / s for eight consecutive time windows). The average temperature of this plateau range, 187.2℃, was taken as the actual boiling point temperature. The theoretical boiling point temperature of formula F1 is 186.5℃, and the absolute value of the difference is... If the temperature is below the boiling point matching threshold of 2.0℃, the final formula is confirmed as formula F1. The precise temperature control parameters of formula F1 are then used to continue the closed-loop heating control.

[0087] After suction is completed, the precise atomization resistance correction factor is recalculated using the precise viscosity value of formula F1, 38.2 mPa·s. A secondary correction is performed on the biomechanical fingerprint feature vector after formula correction to obtain the secondary corrected vector. Incremental mean update of User B's prototype vector is then performed; User B has accumulated 47 aspirations. The aspiration concentration and airflow matching index of 0.87 exceeded the acceptable matching threshold of 0.80. During the stable inhalation phase, the average aerosol particle concentration signal within the corresponding time window was 2.31 mg / m³, which was stored in the preference cache as a concentration preference statistical value. After user B accumulates the preset number of inhalations, an exponentially weighted moving average will be applied to the phase and concentration target mapping table for updating. Taking the stable inhalation phase as an example (smoothing coefficient of 0.7): The final formula F1 confirmation result is associated with User B's identifier and stored in the formula association file, completing the full control process in this composite cold start scenario.

[0088] The data flow throughout the entire implementation process demonstrates a clear unidirectional information transmission logic: the physical property modal data (temperature response and impedance spectrum) collected in step 1 are completed independently before user behavior occurs, providing step 2 with input independent of user identity; the initial liquid component probability distribution vector output from step 2 (formula F1 probability 0.58) flows into step 3, and after weighted calculation, is transformed into an atomization resistance correction coefficient of 1.0425; this coefficient is applied to the negative pressure peak (−312.4Pa→−299.7Pa) and negative pressure rise slope in step 5, eliminating the formula viscosity deviation and completing the user B's identity verification. The prior information of User B's historical formula flows backward into the Bayesian update process, increasing the probability of formula F1 from 0.58 to 0.845, exceeding the confidence level safety threshold, supporting step 6 to directly call the standard temperature control parameters; boiling point verification (187.2℃ and 186.5℃ match) finally locks in formula F1, and precise parameter replacement completes the closed-loop control; the incremental mean update and exponentially weighted moving average update after the aspiration ends (after the update, the target concentration value for the stable inhalation stage is 2.355mg / m³) will precipitate this data into User B's personalized profile, forming a long-term adaptive optimization closed loop across aspirations.

[0089] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for collaborative analysis and processing of multimodal data from an electronic atomizer, characterized in that, Includes the following steps: After detecting a new atomized projectile installation event, a low-power wetting pulse is applied to the heating element and temperature response data is collected. After wetting is completed, a multi-frequency AC excitation signal is applied to the heating element to obtain the first impedance spectrum data. The temperature response data and the first impedance spectrum data are combined to generate a physical property mode dataset. Feature extraction is performed on the physical property modality dataset to generate enhanced feature vectors. The enhanced feature vectors are then input into the atomized liquid composition inference network with a domain adaptation module to output the initial liquid composition probability distribution vector. Based on the initial liquid component probability distribution vector, the weighted average viscosity of the candidate formulation is calculated in a probability-weighted manner. The atomization resistance correction coefficient is calculated using the weighted average viscosity and the preset standard reference viscosity value. The atomization resistance correction coefficient is used as the formulation-related fingerprint correction parameter. When the user initiates suction, the oral negative pressure waveform signal, aerosol particle concentration signal and heating element temperature signal are collected. The original biomechanical fingerprint feature vector is generated by extracting features from the oral negative pressure waveform signal. The formula-related fingerprint correction parameters are used to perform an affine transformation on the feature components of the original biomechanical fingerprint feature vector that are affected by the formula to generate a formula-corrected biomechanical fingerprint feature vector. The formula-corrected biomechanical fingerprint feature vector is then matched with the pre-stored user prototype vectors to identify the user's identity. Based on the initial liquid component probability distribution vector and the identified user identity, temperature control parameters and concentration target parameters are configured to perform closed-loop heating control of the heating element.

2. The multimodal data collaborative analysis and processing method for electronic atomizers according to claim 1, characterized in that, The process of extracting features from the physical property modality dataset to generate enhanced feature vectors includes: The temperature response data is processed to calculate the thermal time constant and steady-state temperature rise amplitude. Based on the thermal time constant, the steady-state temperature rise amplitude, and the known heat capacity parameters and known heat dissipation area parameters of the heating element, the initial estimate of the equivalent thermal conductivity is calculated according to the lumped parameter thermal model. The thermal time constant, the steady-state temperature rise amplitude, and the initial estimate of the equivalent thermal conductivity are combined to generate auxiliary thermal property characteristics. Wherein, the thermal time constant is the time length corresponding to the temperature rise curve from the initial temperature to 63.2% of the steady-state temperature rise amplitude; the calculation method of the lumped parameter thermal model is: the initial estimate of the equivalent thermal conductivity is equal to the product of the heating element mass and the heating element specific heat capacity divided by the product of the thermal time constant and the heating element heat dissipation area; The initial impedance spectrum data was fitted using an equivalent circuit consisting of an ohmic resistor, a charge transfer resistor, and a constant-phase element, and a three-component impedance feature vector was extracted, consisting of the ohmic resistor value, the charge transfer resistor value, and the constant-phase element index. The enhanced feature vector is formed by concatenating the components of the thermophysical auxiliary feature and the three-component impedance feature vector after Z-score normalization preprocessing.

3. The multimodal data collaborative analysis and processing method for electronic atomizers according to claim 1, characterized in that, The atomized liquid composition inference network with the domain adaptation module includes a feature encoding unit, a domain adaptation unit, and a classification output unit connected in sequence. The feature encoding unit receives the enhanced feature vector and outputs the intermediate layer feature representation; The domain adaptation unit calculates the Mahalanobis distance between the enhanced feature vector and the pre-stored known steady-state feature distribution center, performs affine correction on the intermediate layer feature representation based on the Mahalanobis distance, and outputs the corrected intermediate layer feature representation; wherein, the Mahalanobis distance is calculated by taking the square root of the square root of the difference vector between the current enhanced feature vector and the known steady-state feature distribution center vector after performing a quadratic form operation on the inverse matrix of the known steady-state feature distribution; the affine correction is calculated as follows: The intermediate layer feature representation is multiplied by a scaling factor determined based on Mahalanobis distance, and then a translation factor determined based on Mahalanobis distance is added. The scaling factor and the translation factor are output by the mapping function obtained by pre-training in the domain adaptation unit with Mahalanobis distance as input. The classification output unit receives the corrected intermediate layer feature representation, normalizes it using the Softmax activation function, and outputs the initial liquid component probability distribution vector.

4. The multimodal data collaborative analysis and processing method for electronic atomizers according to claim 3, characterized in that, Both the feature encoding unit and the classification output unit are multilayer perceptrons. The mapping functions for the output scaling coefficients and the output translation coefficients in the domain adaptation unit are each implemented by a single-input, single-output multilayer perceptron. The output layer activation function of the multilayer perceptron for the output scaling coefficients uses the Softplus function, while the output layer of the multilayer perceptron for the output translation coefficients is not subject to activation function constraints. The feature encoding unit, the mapping functions in the domain adaptation unit, and the classification output unit are jointly trained end-to-end, and the loss function is cross-entropy loss.

5. The multimodal data collaborative analysis and processing method for electronic atomizers according to claim 1, characterized in that, The original biomechanical fingerprint feature vector is generated by extracting features from the oral negative pressure waveform signal, including: The negative pressure peak value, negative pressure rise slope, negative pressure duration, waveform radius of curvature, and waveform symmetry coefficient are extracted from the oral negative pressure waveform signal. The above five feature components are preprocessed by Z-score normalization and then combined to generate the original biomechanical fingerprint feature vector. Wherein, the negative pressure peak value is the maximum negative deviation of air pressure relative to ambient atmospheric pressure during a complete suction process; the negative pressure rise slope is the ratio of the air pressure change from the start of suction to the negative pressure peak value to the corresponding time length; The duration of negative pressure is the total length of the continuous time period during which the air pressure is lower than the preset negative pressure threshold; the radius of curvature of the waveform is the radius of curvature value calculated by the second derivative at the peak point after fitting the peak neighborhood with a second-order polynomial; the waveform symmetry coefficient is the ratio of the duration of the negative pressure rising segment to the duration of the negative pressure falling segment with the negative pressure peak time as the axis of symmetry. The step of performing an affine transformation on the feature components affected by the formula using the formula-related fingerprint correction parameters includes: dividing the negative pressure peak value and the negative pressure rise slope by the atomization resistance correction coefficient to obtain the corrected negative pressure peak value and the corrected negative pressure rise slope, respectively, while keeping the negative pressure duration, the waveform radius of curvature, and the waveform symmetry coefficient unchanged, which together with the corrected negative pressure peak value and the corrected negative pressure rise slope form the formula-corrected biomechanical fingerprint feature vector.

6. The multimodal data collaborative analysis and processing method for electronic atomizers according to claim 1, characterized in that, The biomechanical fingerprint feature vector after formula correction is matched with the pre-stored user prototype vectors to identify the user's identity, including: The cosine similarity between the biomechanical fingerprint feature vector after formula correction and each pre-stored user prototype vector is calculated to obtain the maximum similarity value and its corresponding user prototype identifier. When the maximum similarity value exceeds the preset identity verification threshold, the corresponding user ID is output; when the maximum similarity value is lower than the preset new user detection threshold, the biomechanical fingerprint feature vector after formula correction is stored as a new prototype vector and a new user ID is assigned and output; when the maximum similarity value is between the identity verification threshold and the new user detection threshold, a temporary tag is output; wherein, the identity verification threshold is higher than the new user detection threshold.

7. The multimodal data collaborative analysis and processing method for electronic atomizers according to claim 6, characterized in that, When the identified user identity matches a known user, the historical usage recipe list and corresponding impedance spectrum feature statistics are extracted from the user's personalized parameter storage area. The prior probability weight of each candidate recipe is determined by the historical usage recipe list. Bayesian posterior update is performed on the initial liquid component probability distribution vector, and the liquid component probability distribution vector after user condition correction is output. Subsequent temperature control parameter configuration is based on the liquid component probability distribution vector after user condition correction. The Bayesian posterior update is calculated as follows: the probability value of the i-th candidate formulation after the update is equal to the product of the probability value of the i-th candidate formulation in the initial liquid component probability distribution vector and the prior probability weight of the i-th candidate formulation, divided by the sum of the products of the probability values ​​of all candidate formulations and their corresponding prior probability weights. The prior probability weight is determined by the proportion of the user’s historical usage of each recipe to the total usage, and a preset low probability baseline value greater than zero is taken for recipes that have not been used. When the identification result is a new user identifier or a temporary tag, the initial liquid component probability distribution vector remains unchanged.

8. The multimodal data collaborative analysis and processing method for electronic atomizers according to claim 1, characterized in that, The configuration of temperature control parameters based on the initial liquid component probability distribution vector and the identified user identity includes: When the highest probability value in the initial liquid component probability distribution vector is lower than the preset confidence level safety threshold, the lowest safe heating temperature limit and the lowest safe heating power limit are taken from the temperature control parameters corresponding to each candidate formula as the current temperature control parameters; when the highest probability value is not lower than the confidence level safety threshold, the standard temperature control parameters of the formula corresponding to the highest probability value are directly retrieved. During the closed-loop heating control process, the power and temperature dynamic response curves of the heating element are collected synchronously. When a temperature plateau range is detected, the actual boiling point temperature is extracted. The actual boiling point temperature is matched with the theoretical boiling point temperature of each candidate formula. When the absolute value of the difference between the two is lower than the preset boiling point matching threshold, the corresponding candidate formula is confirmed as the final formula. Based on the final formula, the precise temperature control parameters are retrieved to replace the current temperature control parameters. The temperature plateau interval is detected by calculating the temperature change rate of the power and temperature dynamic response curves according to time windows. When the absolute value of the temperature change rate is lower than the preset plateau judgment threshold within a consecutive preset number of time windows, it is determined to be a temperature plateau interval.

9. The multimodal data collaborative analysis and processing method for electronic atomizers according to claim 1, characterized in that, After the closed-loop heating control is completed, it also includes: The confirmed final formula is associated with the user ID and stored in the corresponding user's formula association file. The formula-related fingerprint correction parameters are recalculated using the accurate viscosity value of the final formula. The biomechanical fingerprint feature vector of the formula after correction obtained by the current aspiration is corrected a second time. The prototype vector of the user is updated by incremental mean. The incremental mean update is calculated as follows: the updated prototype vector is equal to the product of the number of samples accumulated by the user and the prototype vector before the update, plus the biomechanical fingerprint feature vector after secondary correction, and then divided by the number of samples accumulated by the user plus one. The concentration, airflow matching index, and suction stage data of this port are stored in the corresponding preference cache according to the user ID. After accumulating a preset number of suction records, the stage and concentration target mapping table is updated by exponential weighted moving average. The calculation method of the exponential weighted moving average update is: the updated target concentration value is equal to the product of the preset smoothing coefficient and the concentration preference statistics calculated based on the preference cache, plus one minus the product of the smoothing coefficient and the current target concentration value. The concentration preference statistics are the average aerosol particle concentration signals of suction records in the preference cache whose concentration and airflow matching index are higher than the preset matching qualification threshold within the corresponding time window; the preference cache is stored in isolation according to user ID.

10. A multimodal data collaborative analysis and processing system for an electronic atomizer, used to execute the multimodal data collaborative analysis and processing method for the electronic atomizer according to any one of claims 1 to 9, characterized in that, include: The physical property modal data acquisition module is used to apply a low-power wetting pulse to the heating body and acquire temperature response data after detecting a new atomized projectile installation event. After wetting is completed, a multi-frequency AC excitation signal is applied to the heating body to obtain the first impedance spectrum data. The temperature response data and the first impedance spectrum data are combined to generate a physical property modal dataset. The liquid component inference module is used to extract features from the physical property modality dataset to generate an enhanced feature vector, input the enhanced feature vector into the atomized liquid component inference network of the domain adaptation module, and output the initial liquid component probability distribution vector. The fingerprint correction parameter generation module is used to calculate the weighted average viscosity of the candidate formulation based on the probability distribution vector of the initial liquid components in a probability-weighted manner, and to calculate the atomization resistance correction coefficient as the formulation-related fingerprint correction parameter using the weighted average viscosity and the preset standard reference viscosity value. The user behavior data acquisition module is used to collect oral negative pressure waveform signals, aerosol particle concentration signals, and heating element temperature signals when the user initiates suction. The user identification module is used to extract features from the oral negative pressure waveform signal to generate an original biomechanical fingerprint feature vector, perform affine transformation on the feature components affected by the formula using the formula-related fingerprint correction parameters to generate a formula-corrected biomechanical fingerprint feature vector, and perform similarity matching between the formula-corrected biomechanical fingerprint feature vector and the pre-stored user prototype vectors to identify the user's identity. The closed-loop heating control module is used to configure temperature control parameters and concentration target parameters based on the initial liquid component probability distribution vector and the identified user identity, and to perform closed-loop heating control of the heating element.

Images