Electronic cigarette intelligent voice interaction and control system based on natural language understanding

Abstract

The application relates to the technical field of electronic cigarette control and voiceprint recognition, and particularly discloses an electronic cigarette intelligent voice interaction and control system based on natural language understanding, which comprises a voice collection and front-end processing module, a multi-modal feature fusion analysis module, a dynamic semantic understanding and intention decision module, an active health intervention strategy generation module and a control execution and feedback module. By extracting acoustic features from voice, fusing and analyzing multi-dimensional features such as voiceprints, physiological states and use behaviors, and combining semantic understanding to make context awareness decisions, the system can intelligently correct instructions or generate personalized health intervention strategies, and realizes active guidance and health risk intervention on user use behaviors.

Description

Technical Field

[0001] This invention belongs to the field of electronic cigarette control and voiceprint recognition technology, specifically relating to an intelligent voice interaction and control system for electronic cigarettes based on natural language understanding. Background Technology

[0002] In the consumer electronics and smart hardware sector, the intelligence and convenience of human-computer interaction are the core driving forces for improving user experience. As a new type of atomization product, e-cigarettes are evolving from traditional physical button operation to more intelligent control and interaction methods. Voice interaction technology, due to its ability to eliminate the need for manual operation, is considered an important technological direction for improving ease of use and safety.

[0003] A natural language understanding-based intelligent voice interaction system aims to achieve precise control and information feedback of e-cigarette devices by recognizing and understanding users' voice commands. This system typically includes modules for voice acquisition, command recognition, semantic parsing, and control execution, with the goal of building a natural and efficient human-computer dialogue interface.

[0004] The voice interaction system for e-cigarettes primarily focuses on executing basic control commands, such as power on / off, adjusting power, or switching modes. It lacks the ability to deeply perceive and proactively intervene in user behavior and health status. The system cannot effectively identify potential physiological changes implied in the user's voice, such as abnormal voiceprint characteristics due to overuse, thus missing the opportunity for early health risk warnings. Furthermore, existing solutions are relatively passive in their interaction logic, only responding to explicit user commands and failing to proactively provide personalized usage suggestions or risk warnings based on analysis of user habits and status. This results in significant deficiencies in fulfilling health guidance and social responsibility. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent voice interaction and control system for e-cigarettes based on natural language understanding, so as to solve the problems of existing e-cigarette voice interaction systems having limited functions, lacking in-depth perception and proactive intervention capabilities of users' physiological states and usage behaviors, and being unable to achieve personalized health guidance.

[0006] This invention provides an intelligent voice interaction and control system for e-cigarettes based on natural language understanding. The system includes a voice acquisition and front-end processing module, a multimodal feature fusion and analysis module, a dynamic semantic understanding and intent decision-making module, an active health intervention strategy generation module, and a control execution and feedback module. By constructing a closed-loop processing flow from the acquisition of raw voice signals to the final generation of active intervention strategies, the system achieves deep perception of the user's state and intelligent guidance.

[0007] The voice acquisition and front-end processing module is used to acquire user voice command signals in real time and preprocess the acquired raw voice signals. The preprocessing process includes noise reduction, pre-emphasis, framing, and windowing. Furthermore, this module is also used to extract basic acoustic feature vectors from the preprocessed voice signals. The basic acoustic feature vectors include Mel-frequency cepstral coefficients, linear prediction coefficients, and short-time energy.

[0008] The multimodal feature fusion analysis module receives the basic acoustic feature vectors output by the speech acquisition and front-end processing module and performs in-depth analysis to generate multidimensional state features. This module includes a voiceprint feature analysis submodule, a physiological state inference submodule, and a usage behavior modeling submodule. The voiceprint feature analysis submodule performs temporal modeling on the basic acoustic feature vectors, extracts static voiceprint features representing user identity and long-term acoustic habits, and calculates the dynamic deviation between the current speech frame and the user's historical voiceprint feature templates; this dynamic deviation serves as a voiceprint anomaly indicator.

[0009] The physiological state inference submodule infers the user's real-time physiological state parameters based on specific parameters in the fundamental acoustic feature vector. Specifically, this includes calculating laryngeal tension based on fundamental frequency trajectory stability analysis and calculating oral and nasal cavity moisture based on formant bandwidth variation analysis. The behavior modeling submodule accesses device usage log data, including single usage duration, daily usage frequency, and usage time distribution. Based on this data, it constructs a time-series model of user behavior to predict the intensity of the user's current usage tendency.

[0010] The Dynamic Semantic Understanding and Intent Decision-Making module performs semantic parsing of user voice commands and combines this with multi-dimensional state features output by the multi-modal feature fusion analysis module to make a comprehensive intent decision. This module includes a Natural Language Understanding submodule and a Context-Aware Decision-Making submodule. The Natural Language Understanding submodule uses a deep neural network model based on an attention mechanism to convert speech recognition text into structured semantic slot filling results. Semantic slots include the operation command type, target parameter value, and modifiers. The Context-Aware Decision-Making submodule fuses the semantic slot filling results with multi-dimensional state features, including voiceprint anomaly indicators, laryngeal tension indicators, oral and nasal cavity moisture indicators, and user usage tendency intensity. This submodule has a built-in decision rule engine that selects the final execution intent from a preset decision set based on the fused feature vector. The decision set includes directly executing the user command, executing the modified command, ignoring the current command, and triggering a health reminder.

[0011] The proactive health intervention strategy generation module generates specific proactive intervention strategies in response to the decision results output by the dynamic semantic understanding and intent decision-making module, which require triggering health reminders. This module includes a strategy knowledge base and a strategy matching engine. The strategy knowledge base stores multiple predefined intervention strategy templates, each associated with a set of threshold conditions for multi-dimensional state features. The strategy matching engine compares the current multi-dimensional state features with the threshold conditions in the strategy knowledge base. When the triggering conditions of a specific template are met, the template is activated and instantiated into a specific combination of interactive dialogue and control commands. The instantiation process includes filling the variable positions of the dialogue template with the user's state parameter values ​​and generating the corresponding device control command sequence.

[0012] The control execution and feedback module is used to execute the final intention decision result or proactive intervention strategy. This module includes a device control submodule and a voice synthesis feedback submodule. The device control submodule drives the actuators of the electronic cigarette device, including a power adjustment unit, an atomizer switch, and indicator lights, based on the received control command sequence. The voice synthesis feedback submodule converts text information that needs to be fed back to the user, including confirmation messages, inquiries, or health reminders, into natural and fluent speech signals and plays them through a speaker.

[0013] As one embodiment of the present invention, the specific process of temporal modeling in the voiceprint feature analysis submodule is as follows: First, a deep convolutional neural network is used to extract high-level acoustic features from the basic acoustic feature vector; then, the high-level acoustic features are input into a bidirectional long short-term memory network for temporal dependency modeling; finally, a fixed-dimensional vector is extracted from the last hidden layer state of the bidirectional long short-term memory network as the user's static voiceprint feature. The dynamic deviation is calculated using a cosine similarity metric, calculating the similarity between the high-level acoustic feature vector of the current speech frame and the pre-registered user static voiceprint feature vector. The result of subtracting this similarity value from 1 is defined as the voiceprint anomaly index.

[0014] As one embodiment of the present invention, the calculation process of the laryngeal tension index in the physiological state inference submodule is as follows: First-order difference calculation is performed on the fundamental frequency trajectory to obtain the fundamental frequency change rate sequence; the standard deviation of this sequence within the sliding time window is calculated; this standard deviation is compared with the baseline standard deviation obtained through statistical analysis of a large number of normal samples, and the ratio is mapped using the Sigmoid function to output a laryngeal tension index value ranging from 0 to 1. The calculation of the oral and nasal cavity moisture index involves analyzing the bandwidth values ​​of the first three resonance peaks; the broadening of the resonance peak bandwidth is positively correlated with the degree of mucosal dryness; the normalized difference between the current resonance peak bandwidth and the user's baseline bandwidth is calculated, and the average value is taken, then mapped using the Sigmoid function to obtain the moisture index.

[0015] In one embodiment of the present invention, the decision rule engine in the context-aware decision submodule is based on a multilayer perceptron-based classifier. The input layer of this classifier is a fused feature vector, which is composed of semantic slot vectors and multidimensional state feature vectors. The output layer is the probability distribution of each decision category in the decision set. The decision rule engine selects the category with the highest probability value as the final decision. Furthermore, during training, the classifier uses training sample labels that include not only whether the instruction is executed correctly but also whether health intervention should be performed, enabling the model to learn the mapping relationship from composite features to complex decisions.

[0016] In one embodiment of the present invention, the strategy knowledge base in the proactive health intervention strategy generation module is represented using production rules. Each rule takes the form: if the voiceprint abnormality index is greater than the threshold of 0.3, and the user's usage tendency intensity is greater than the threshold of 0.7, then an overuse risk warning strategy template is triggered. The strategy template includes a voice prompt template and device control instructions. The voice prompt template indicates that your vocal cords may be fatigued and that you have been using the device frequently recently, suggesting a 30-minute rest. The device control instruction limits the device power to 50% of its current value for 1800 seconds.

[0017] In one embodiment of the present invention, the system further includes a continuous learning and adaptive update module. This module is used to dynamically adjust the parameters in the multimodal feature fusion analysis module and the proactive health intervention strategy generation module based on the user's feedback behavior to the proactive intervention strategy. The specific process is as follows: After a health reminder is executed, the system records whether the user follows the advice to suspend use or ignores the advice and continues high-intensity use; these feedback behaviors are used as reinforcement learning signals to update the weights of the classifier in the decision rule engine and adjust the trigger thresholds of relevant rules in the strategy knowledge base, thereby making the system's intervention strategy more personalized and effective.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention breaks through the limitation of existing electronic cigarette voice interaction systems that can only respond to surface-level commands. By constructing a multimodal feature fusion analysis module, it deeply mines voiceprint features, physiological state indicators and usage behavior tendencies from the user's voice signal, realizing non-invasive, real-time perception of the user's physical and mental state and usage habits. This deep perception capability provides a solid data foundation for subsequent intelligent decision-making and proactive intervention, enabling the system to understand the user's potential state rather than just explicit commands.

[0019] 2. This invention upgrades traditional semantic parsing to a comprehensive decision-making process that combines real-time context through a dynamic semantic understanding and intent decision-making module. The system can intelligently modify, delay, or override user control commands based on the user's current health risk status and proactively trigger health reminders. This design transforms human-computer interaction from a passive command-response mode to a proactive, health-guided collaborative mode, significantly improving the product's social responsibility fulfillment capabilities and user experience safety.

[0020] 3. This invention achieves precision and personalization of intervention strategies through a proactive health intervention strategy generation module and a continuous learning mechanism. The system can not only generate targeted reminders and control commands based on a defined rule base, but also adaptively optimize decision thresholds and strategy content through continuous learning of user feedback. This allows the system's health guidance behavior to increasingly align with users' personal habits and acceptance levels over time, forming an effective and positive behavioral intervention loop, thereby positively impacting users' healthy usage habits in the long term. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the multimodal feature fusion analysis module in this invention; Figure 3 This is a logical flowchart of the dynamic semantic understanding and intent decision-making module in this invention; Figure 4 This is a schematic diagram illustrating the interaction between the proactive health intervention strategy generation module and the execution feedback in this invention; Figure 5 This is a schematic diagram illustrating the closed-loop optimization principle of the continuous learning and adaptive update module in this invention. Detailed Implementation

[0022] Example 1: This invention provides an intelligent voice interaction and control system for electronic cigarettes based on natural language understanding. Its overall technical architecture is shown in the attached figure. Figure 1 As shown in the figure. This system achieves deep fusion perception and intelligent response to user physiological states, usage behaviors, and voice commands by constructing a complete closed-loop processing flow from raw voice signal acquisition, multimodal feature fusion analysis, dynamic semantic understanding and intent decision-making, proactive health intervention strategy generation to control execution and feedback. The following will combine the attached... Figure 1 To be continued Figure 5 This section provides a detailed implementation description of each functional module of the system, expanding upon it layer by layer.

[0023] First, the voice acquisition and front-end processing module, serving as the input to the entire system, is responsible for capturing user-issued voice commands in real time and performing a series of standardized signal preprocessing operations to ensure the accuracy and robustness of subsequent feature extraction. This module incorporates a high-sensitivity microphone array for directional voice pickup in complex acoustic environments, suppressing environmental noise interference. The acquired raw voice signal is a continuous-time analog signal, which is converted into a digital signal by an analog-to-digital converter at a sampling rate of 16 kHz before entering the preprocessing pipeline. The preprocessing process strictly executes four stages sequentially: noise reduction, pre-emphasis, framing, and windowing.

[0024] The noise reduction process employs an adaptive noise suppression algorithm based on spectral subtraction, which estimates the background noise power spectrum in real time and subtracts it from the speech spectrum. The pre-emphasis processing enhances high-frequency components using a first-order high-pass filter to compensate for energy attenuation in the high-frequency band of the speech signal; its transfer function is... The frame-segmentation process divides the continuous speech signal into an overlapping frame sequence with a length of 25 milliseconds and a frame shift of 10 milliseconds to satisfy the short-time stationarity assumption. The windowing process applies a Hamming window function to each frame to reduce spectral leakage. After the above preprocessing, this module further extracts a basic acoustic feature vector from each speech frame. This vector consists of three core parameters: Mel-frequency cepstral coefficients (first 13 dimensions), linear prediction coefficients (first 12 orders), and short-time energy (in decibels). All feature vectors are stored in floating-point form and organized into a feature matrix in chronological order, serving as the input data stream for subsequent modules.

[0025] Please refer to the attached document. Figure 2 The multimodal feature fusion analysis module receives the basic acoustic feature vectors output from the speech acquisition and front-end processing module, and performs multi-level, multi-dimensional state feature mining and modeling based on these vectors. This module consists of three tightly coupled sub-modules: a voiceprint feature analysis sub-module, a physiological state inference sub-module, and a usage behavior modeling sub-module. The core task of the voiceprint feature analysis sub-module is to establish an acoustic representation of the user's identity and monitor the degree of deviation between the current speech and historical voiceprint patterns. Its implementation process involves two steps: The first step is to use a five-layer deep convolutional neural network to extract spatial features from the input basic acoustic feature matrix. This network contains alternating convolutional layers and max pooling layers, and finally outputs a high-level acoustic feature tensor. The second step is to flatten the tensor along the time dimension and input it into a bidirectional long short-term memory network (BiLSTM) with 128 hidden units. The forward and backward states are concatenated at the end of the time step to form a 256-dimensional context-aware feature sequence.

[0026] Static voiceprint features are extracted from the hidden state of the last time step of the sequence, normalized using L2, and then stored in the user's voiceprint template library. Dynamic deviation is calculated based on the high-level acoustic feature vector of the current speech segment. (Taking the output of the intermediate layer of BiLSTM) and the registered static voiceprint feature vector Cosine similarity is used for measurement: Voiceprint anomaly index is defined as Its value ranges from 0 to 1. The larger the value, the more significant the difference between the current voice and the user's normal voiceprint, which may reflect vocal cord fatigue, emotional fluctuations, or non-personal operation.

[0027] The physiological state inference submodule focuses on retrieving the user's real-time physiological parameters from the speech signal. The calculation of the laryngeal tension index is based on the stability analysis of the fundamental frequency (F0) trajectory. First, the fundamental frequency value is accurately extracted from each speech frame using the autocorrelation function method to form a fundamental frequency time series; then, a first-order difference operation is performed on this series to obtain the fundamental frequency change rate series. Next, calculations are performed within a sliding time window of 2 seconds with a step size of 0.5 seconds. Standard deviation Finally, this standard deviation is compared with the user-individualized baseline standard deviation. (The ratio is calculated using the average of the past 7 days of normal usage) and then input into the Sigmoid function for non-linear mapping: in, The scaling factor is set to 4 to ensure the output value is between 0 and 1; a higher value indicates greater laryngeal muscle tension. The moisture levels of the oral and nasal cavities are obtained through formant bandwidth analysis. The system uses a linear predictive coding model to fit the speech spectrum envelope and extracts the center frequencies and 3dB bandwidths of the first three formants (F1, F2, F3). Studies show that mucosal dryness leads to formant energy diffusion, manifested as increased bandwidth. Therefore, the current bandwidth of each formant is calculated. Compared with user baseline bandwidth Normalized differences After taking the average value of i=1,2,3, the moisture index is obtained by mapping it through the Sigmoid function. The lower the value, the drier the oral cavity and nasal cavity are.

[0028] The behavior modeling submodule does not rely on voice signals but instead accesses the e-cigarette device's internal log database to obtain structured usage behavior data. This data includes single usage duration (in seconds), total daily usage, and hourly usage frequency distribution over the past 24 hours. The module uses a Long Short-Term Memory (LSTM) network to construct a time-series prediction model. The input is the daily usage feature vector from the past 7 days, and the output is the user's current usage tendency strength. This strength value is processed using Min-Max normalization and is limited to a range of 0 to 1; a higher value indicates a stronger likelihood of the user continuing to use the device. The outputs of all three submodules—voiceprint anomaly index, laryngeal tension index, moisture index, and usage tendency strength—together constitute a multi-dimensional state feature vector, which is then passed to the next module.

[0029] Please refer to the attached document. Figure 3 The dynamic semantic understanding and intent decision-making module is responsible for converting user voice commands into executable system actions. However, its decision-making process is not isolated but deeply integrated with multi-dimensional state features. This module includes a natural language understanding submodule and a context-aware decision-making submodule. The natural language understanding submodule first calls the embedded speech recognition engine to convert the preprocessed speech signal into a text string. Subsequently, the text is fed into a deep neural network model based on an attention mechanism (specifically, a Transformer encoder structure with 6 layers and 8 heads). This model has been fine-tuned on a large corpus of e-cigarette commands and can accurately identify three types of key semantic slots: operation command type (e.g., increase power, turn off device, check battery level), target parameter value (e.g., 50%, level 3), and modifiers (e.g., slightly, immediately, gently). The slot filling result is encoded into a fixed-length semantic slot vector. For example, one-hot encoding is used to represent the command type, floating-point numbers represent parameter values, and boolean values ​​represent the presence or absence of modifiers.

[0030] The context-aware decision-making submodule receives semantic slot vectors and multidimensional state feature vectors, concatenating them along the feature dimensions to form a fused feature vector. This vector is fed as input into a three-layer fully connected multilayer perceptron classifier with 128 and 64 hidden layer neurons, respectively, using ReLU activation function. The number of output layer neurons equals the number of categories in the decision set. The decision set contains three mutually exclusive options: directly execute the user's instruction, execute the modified instruction, and ignore the current instruction and trigger a health reminder. The classifier outputs a probability distribution for each category, and the system selects the category with the highest probability as the final decision. During the training phase, the classifier uses labeled historical interaction data. The labels not only indicate whether the instruction should be executed but also specify the physiological and behavioral states under which health intervention should be triggered. For example, if a user says "take another puff," but at this time the laryngeal tension index is 0.85, the usage tendency intensity is 0.9, and the user has used the device 5 times in the past hour, the label is "trigger a health reminder." Through this supervised learning mechanism, the model learns the decision logic of prioritizing the user's health in high-risk states.

[0031] Once the context-aware decision-making submodule outputs a decision to ignore the current instruction and trigger a health reminder, the system activates the proactive health intervention strategy generation module. Please refer to the appendix. Figure 4 This module consists of a strategy knowledge base and a strategy matching engine. The strategy knowledge base stores multiple intervention strategy templates in the form of production rules. Each rule defines a set of threshold conditions for multi-dimensional state features and a corresponding response action. For example, rule R1 is defined as: if the voiceprint abnormality index > 0.3 and the use tendency intensity > 0.7, then trigger the overuse risk warning strategy template; rule R2 is defined as: if the laryngeal tension index > 0.8 and the moisture index < 0.3, then trigger the vocal cord protection warning strategy template. Each time intervention is needed, the strategy matching engine iterates through the condition parts of all rules, substituting the current multi-dimensional state features for Boolean judgment. When all conditions of a rule are met, the rule is activated. After activation, the engine reads the speech script template and device control instructions from the corresponding strategy template. The script template is a string with placeholders, such as "We have detected that your vocal cords may be fatigued and have been used frequently recently. We suggest you rest for {duration} minutes." Here, {duration} is a variable that is dynamically calculated by the system based on the intensity of the usage tendency (e.g., if the intensity is >0.8, it is set to 30; otherwise, it is set to 15).

[0032] The control execution and feedback module is responsible for translating the final decision or intervention strategy into physical actions and user-perceptible feedback. This module includes a device control submodule and a speech synthesis feedback submodule. The device control submodule parses the received control command sequence and sends electrical signals to the actuators through the device driver interface. The actuators include a power adjustment unit (controlling the heating wire power via pulse width modulation signals), an atomizer switch (controlling the atomization start / stop triggered by the airflow sensor), and multi-color LED indicators (used to display device status or health alert levels). For example, when a power limit command is received, this submodule modifies the power limit register in the device firmware and starts a timer, automatically restoring the original setting after 1800 seconds. The speech synthesis feedback submodule feeds the text to be played (whether it is a confirmation command such as "power adjusted to 50%" or a health reminder) into an end-to-end neural speech synthesis model (such as the Tacotron2 architecture) to generate a high-quality, natural, and fluent speech waveform. This waveform is then converted from digital to analog and played through a built-in miniature speaker, with the volume automatically adjusted according to the ambient noise level to ensure clear hearing for the user.

[0033] In addition, please refer to the appendix. Figure 5 This system also integrates a continuous learning and adaptive update module to achieve long-term optimization of intervention strategies. This module continuously monitors the user's subsequent behavior in response to health reminders in the background: if the user does not trigger any usage within the next 30 minutes after receiving the reminder to rest for 30 minutes, it is recorded as positive feedback; if the user restarts the device and uses it intensively within 5 minutes, it is recorded as negative feedback. These feedback behaviors are quantified as reinforcement learning signals for parameter updates at two levels: First, by backpropagation, the weights of the multilayer perceptron classifier in the context-aware decision-making submodule are fine-tuned to enable it to more accurately predict users’ true intentions and health needs in similar future scenarios. Second, the trigger thresholds of rules in the strategy knowledge base are dynamically adjusted. For example, if a user is effectively intervened multiple times even when the abnormal voiceprint index is 0.25, the system automatically lowers the threshold of the rule corresponding to that user from 0.3 to 0.22 to achieve personalized adaptation.

[0034] All update operations are performed during device idle periods to avoid impacting real-time interaction performance, and user data is anonymized using differential privacy technology to ensure privacy and security.

[0035] In summary, this embodiment, through the collaborative work of the aforementioned modules, constructs an intelligent voice interaction and control system for e-cigarettes with capabilities of deep perception, intelligent decision-making, proactive intervention, and continuous evolution. The system can not only accurately understand explicit user commands but also proactively guide users to develop healthy usage habits based on physiological and behavioral cues contained in the voice, fundamentally enhancing the intelligence level and social responsibility attributes of e-cigarette products.

[0036] Example 2: Building upon Example 1, this example expands the functionality of the usage behavior modeling submodule within the multimodal feature fusion analysis module by introducing the perception and fusion of external environmental factors to further enhance the comprehensiveness of the state assessment and the accuracy of intervention strategies. Specifically, the e-cigarette device in this example additionally integrates an environmental sensor array, including a temperature and humidity sensor, an atmospheric pressure sensor, and an ambient light sensor. These sensors continuously collect environmental data at a frequency of 1 Hz and synchronize the data stream to the usage behavior modeling submodule.

[0037] The behavior modeling submodule adds an environmental context embedding layer to the existing time series prediction model. This embedding layer first standardizes the raw environmental data: in, For temperature, Relative humidity, Atmospheric pressure, The ambient light intensity.

[0038] The processed four-dimensional environment vector is fed into a two-layer fully connected network, outputting a 32-dimensional environment context embedding vector. This embedding vector is aligned temporally with the feature vectors from the past 7 days of usage logs and serves as input to an extended Long Short-Term Memory (LSTM) network. The network structure is adjusted accordingly to have dual-channel input, processing behavioral sequences and environment sequences separately, with feature fusion performed at a higher level. This design allows the model to learn the influence of environmental factors on user behavior. For example, in hot, dry environments, users might mistakenly identify sucking as a usage command due to thirst, leading to false triggers; or in low-light conditions at night, user usage frequency decreases significantly. After learning these patterns, the model can more accurately predict the true intensity of usage tendency under current environmental conditions.

[0039] Furthermore, the strategy knowledge base of the proactive health intervention strategy generation module has been expanded to include environmentally aware rules. For example, a new rule R3 has been added: If the ambient temperature is >35℃, the relative humidity is <40%, and the humidity index is <0.4, then the 'High Temperature and Dry Environment Hydration Reminder' strategy template is triggered. The message is: The current environment is high temperature and dry; your oral mucosa has shown signs of dryness. It is recommended to replenish fluids and pause use for 15 minutes. The control command is to temporarily disable the nebulizer function for 600 seconds. When judging rules, the strategy matching engine not only checks multi-dimensional state characteristics but also queries the current environmental sensor readings to ensure the timeliness and scenario adaptability of the intervention.

[0040] This extension scheme significantly enhances the system's robustness and user-friendliness in complex real-world scenarios from a single user-centric perspective.

Claims

1. An electronic cigarette intelligent voice interaction and control system based on natural language understanding, characterized in that, include: The voice acquisition and front-end processing module is used to acquire the user's voice command signal in real time and preprocess the acquired raw voice signal. The preprocessing process includes noise reduction, pre-emphasis, frame segmentation and windowing, as well as extracting basic acoustic feature vectors from the preprocessed voice signal. The basic acoustic feature vectors include Mel frequency cepstral coefficients, linear prediction coefficients and short-time energy. The multimodal feature fusion analysis module is used to receive the basic acoustic feature vector output by the speech acquisition and front-end processing module, and perform in-depth analysis on it to generate multidimensional state features; The dynamic semantic understanding and intent decision-making module is used to perform semantic parsing of the user's voice commands and combine them with the multi-dimensional state features output by the multi-modal feature fusion analysis module to make a comprehensive intent decision. The proactive health intervention strategy generation module is used to generate specific proactive intervention strategies in response to the decision results that need to trigger health reminders output by the dynamic semantic understanding and intent decision-making module. The control execution and feedback module is used to execute the final intentional decision result or proactive intervention strategy; The multimodal feature fusion analysis module includes a voiceprint feature analysis submodule, a physiological state inference submodule, and a usage behavior modeling submodule. The voiceprint feature analysis submodule is used to perform temporal modeling on the basic acoustic feature vector, extract static voiceprint features that represent user identity and long-term acoustic habits, and calculate the dynamic deviation between the current speech frame and the user's historical voiceprint feature template. The dynamic deviation is used as a voiceprint anomaly indicator. The physiological state inference submodule is used to infer the user's real-time physiological state parameters based on specific parameters in the basic acoustic feature vector. Specifically, it includes calculating the laryngeal tension index based on the stability analysis of the fundamental frequency trajectory and calculating the moistness index of the oral cavity and nasal cavity based on the change analysis of the formant bandwidth. The usage behavior modeling submodule is used to access the device's usage log data, which includes single usage duration, daily usage frequency, and usage time distribution. Based on this data, a time series model of user usage behavior is constructed to predict the intensity of user usage tendency at the current moment.

2. The natural language understanding based electronic cigarette intelligent voice interaction and control system according to claim 1, characterized in that, The dynamic semantic understanding and intent decision-making module includes a natural language understanding submodule and a context-aware decision-making submodule; The natural language understanding submodule uses a deep neural network model based on an attention mechanism to convert speech recognition text into structured semantic slot filling results. The semantic slots include operation instruction types, target parameter values, and modifiers. The context-aware decision submodule is used to fuse the semantic slot filling result with the multi-dimensional state features. The multi-dimensional state features include voiceprint abnormality index, laryngeal tension index, oral and nasal cavity moisture index, and user usage tendency intensity. The context-aware decision submodule has a built-in decision rule engine. The decision rule engine selects the final execution intent from a preset decision set based on the fused feature vector. The decision set includes directly executing the user command, executing the modified command, ignoring the current command and triggering a health reminder. 3.The natural language understanding based electronic cigarette intelligent voice interaction and control system according to claim 2, characterized in that, The proactive health intervention strategy generation module includes a strategy knowledge base and a strategy matching engine; the strategy knowledge base stores multiple predefined intervention strategy templates, each template being associated with a set of threshold conditions for multidimensional state features; The strategy matching engine is used to compare the current multidimensional state features with the threshold conditions in the strategy knowledge base. When the triggering conditions of a specific template are met, the template is activated and instantiated into a specific combination of interactive dialogue and control instructions. The instantiation process includes filling the user's state parameter values ​​into the variable positions of the dialogue template and generating the corresponding device control instruction sequence.

4. The natural language understanding based electronic cigarette intelligent voice interaction and control system according to claim 3, characterized in that, The control execution and feedback module includes a device control submodule and a voice synthesis feedback submodule; the device control submodule drives the actuator of the electronic cigarette device according to the received control command sequence, and the actuator includes a power adjustment unit, an atomizer switch and an indicator light; The speech synthesis feedback submodule is used to convert the text information that needs to be fed back to the user, including confirmation information, inquiry information or health reminders, into natural and fluent speech signals and play them through a speaker.

5. The natural language understanding based electronic cigarette intelligent voice interaction and control system according to claim 4, characterized in that, The specific process of temporal modeling in the voiceprint feature analysis submodule is as follows: high-level acoustic features are extracted from the basic acoustic feature vector using a deep convolutional neural network; The high-level acoustic features are input into a bidirectional long short-term memory network for temporal dependency modeling; a fixed-dimensional vector is extracted from the last hidden layer state of the bidirectional long short-term memory network as the user's static voiceprint features.

6. The natural language understanding based electronic cigarette intelligent voice interaction and control system according to claim 5, characterized in that, The calculation process of the laryngeal tension index in the physiological state inference submodule is as follows: The fundamental frequency trajectory is calculated using the first-order difference to obtain the fundamental frequency rate of change sequence; the standard deviation of this fundamental frequency rate of change sequence within the sliding time window is then calculated. The standard deviation is compared with the baseline standard deviation obtained through statistics from a large number of normal samples. The ratio is then mapped by the Sigmoid function to output a laryngeal tension index value ranging from 0 to 1.

7. The natural language understanding based electronic cigarette intelligent voice interaction and control system according to claim 6, characterized in that, The decision rule engine in the context-aware decision submodule is a classifier based on a multilayer perceptron. The input layer of the classifier is a fused feature vector, which is formed by concatenating a semantic slot vector and a multidimensional state feature vector. The output layer of the classifier is the probability distribution of each decision category in the decision set. The decision rule engine selects the category with the highest probability value as the final decision. 8.The natural language understanding based electronic cigarette intelligent voice interaction and control system according to claim 7, characterized in that, During training, the classifier uses training sample labels that include not only whether instructions were executed correctly, but also whether health interventions should be performed. 9.The natural language understanding based electronic cigarette intelligent voice interaction and control system according to claim 8, characterized in that, The strategy knowledge base in the proactive health intervention strategy generation module is represented by production rules; each rule is in the form of: if the abnormal voiceprint index is greater than the threshold of 0.3 and the user's usage tendency intensity is greater than the threshold of 0.7, then the overuse risk reminder strategy template is triggered; the strategy template includes voice script templates and device control instructions.

Images