Bluetooth earphone supporting AI voice intelligence

Abstract

The invention discloses a Bluetooth earphone supporting AI voice intelligence. The Bluetooth earphone comprises a sound source separation and noise reduction module, an edge model compression module, a multi-mode emotion interaction module, a heat dissipation control module and a predictive maintenance module. Belongs to the technical field of Bluetooth earphones, and particularly relates to a Bluetooth earphone supporting AI voice intelligence. According to the Bluetooth earphone supporting AI voice intelligence, a bone conduction signal and an air conduction signal are fused through a three-dimensional sound field mapping and dynamic mask generation technology, and pure voice can be extracted from a complex sound field; the translation model is compressed based on progressive neural architecture search and mixed precision quantification, and low-delay edge calculation of dialects and foreign languages is achieved; a space-time diagram convolutional network and near-end strategy optimization are combined, translation data, tactile operation frequency data and eyeball trajectory data are coordinated, and a heat dissipation strategy is dynamically regulated and controlled, so that equipment stably runs in the environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of Bluetooth earphones, and particularly relates to a Bluetooth earphone supporting AI voice intelligence. BACKGROUND

[0002] Traditional Bluetooth earphones are long limited by single function and static technical architecture, and are difficult to meet the intelligent interaction demand in complex scenes. In the aspect of acoustic processing, the noise reduction algorithm thereof depends on fixed frequency band filtering or single microphone beam forming, and cannot effectively separate mixed sound sources such as multi-person conversation, wind noise and mechanical noise, so that the voice command recognition rate sharply decreases in a noisy environment (such as subway scene mis-triggering rate > 30%). At the same time, the traditional device lacks multi-modal signal fusion capability, and only interacts through voice commands, so it is difficult to capture touch operation habits (such as double-click pause) or eye gaze intention (such as staring at the screen to trigger translation), and the interaction mode is rigid and has low fault tolerance.

[0003] In the aspect of computing power and system design, the traditional scheme faces the contradiction between high power consumption and low real-time performance of edge AI tasks. The cloud-dependent translation model leads to a response delay of more than 500 ms, and cannot guarantee privacy and security; the heat dissipation design mostly adopts passive heat conduction, and when running under high load, the chip temperature rises to more than 60 DEG C, causing performance downshifting and even hardware damage. In addition, the lack of battery health monitoring and environmentally friendly material life prediction causes users to frequently encounter a cliff-like decrease in battery life or accidental degradation and cracking of the shell, and the product reliability is questioned.

[0004] Therefore, it is imperative to develop a Bluetooth earphone supporting AI voice intelligence. SUMMARY

[0005] In view of the above, in order to overcome the defects of the prior art, the application provides a Bluetooth earphone supporting AI voice intelligence, which fuses bone conduction vibration signals and air sound wave characteristics through three-dimensional sound field mapping and dynamic mask generation technology, can extract pure voice in a complex sound field, relies on progressive neural architecture search and mixed precision quantization to compress the translation model, realizes low-delay edge computing of dialects and foreign languages, combines spatio-temporal graph convolution network and near-end strategy optimization to cooperatively analyze voice, tactile and eye movement data, and dynamically controls the heat dissipation strategy, so that the device stably operates in the environment.

[0006] The Bluetooth earphone supporting AI voice intelligence provided by the application comprises a sound source separation and noise reduction module, an edge model compression module, a multi-modal emotional interaction module, a heat dissipation control module and a predictive maintenance module.

[0007] The sound source separation and noise reduction module is used for sound source separation and noise reduction, fuses bone conduction signals and air conduction signals through a multi-head self-attention Transformer network, constructs a three-dimensional sound field mapping, adopts a dynamic mask generation algorithm to separate noise, human voice and environmental sound in real time, outputs pure voice stream data, and sends the pure voice stream data to the edge model compression module;

[0008] The edge model compression module is used for edge model compression and real-time translation, automatically distills a lightweight model from a cloud model through progressive neural architecture search; applies mixed precision quantization to compress the model volume, matches the computing power of the chip, supports real-time translation of dialects or foreign languages, and sends the translated data to the multi-modal emotional interaction module;

[0009] The multi-modal emotional interaction module is used for multi-modal emotional interaction, fuses translation data, haptic operation frequency data and eye tracking data through a spatio-temporal graph convolution network, generates a user emotional polarity and intent confidence matrix, and triggers a differentiated interaction strategy;

[0010] The heat dissipation control module is used for heat dissipation control, dynamically adjusts the active heat dissipation period of the phase change material by fusing the heat dissipation data of the sound source separation and noise reduction module, the edge model compression module and the multi-modal emotional interaction module, and the near-end strategy optimization; integrates IMU sensor data to construct a heat conduction digital twin model, so that the Bluetooth earphone can effectively control heat dissipation and stabilize temperature;

[0011] The predictive maintenance module is used for predictive maintenance, analyzes the multi-dimensional features of the battery charge and discharge curve through a deep time convolution network; deploys a physical information neural network to simulate the molecular chain breakage rate of biodegradable materials in a humidity range of 30%-90%, so that the Bluetooth earphone can push battery charging warnings and optimize the use environment in advance.

[0012] Further, the dynamic mask generation algorithm specifically includes the following:

[0013] Signal feature extraction: the bone conduction signal acquires energy value through a jaw bone vibration sensor ;

[0014] The air conduction signal calculates the environmental noise power through a 5-microphone array ;

[0015] Weight dynamic allocation:

[0016] ;

[0017] In the formula, represents a calibration coefficient, represents the energy integral of the bone conduction vibration signal, The sum of frequency domain energy representing environmental noise;

[0018] The directivity mask enhances the target speech separation precision in combination with the sound source azimuth angle constraint. The beamforming and the azimuth matrix generation are performed as follows: the sound arrival direction is calculated through the 5-microphone array, and the azimuth matrix is constructed.

[0019] The transformer attention layer is optimized:

[0020] ;

[0021] In the formula, Query and Key matrices are represented, and the dimensions are The feature vector length is represented as The sound source direction matrix is represented, and the main direction signal is enhanced by adding the attention score; The azimuth enhancement coefficient is represented, and the dynamic range is 0.5-2.0, which is adaptively adjusted according to the noise type.

[0022] Further, the light-weight model is automatically distilled from the cloud model, and the specific steps are as follows:

[0023] The progressive neural architecture search constructs a multi-objective optimization function, synchronously compresses the model parameter quantity and the inference delay, adopts the evolutionary algorithm for iterative search, and generates an 8-layer micro-BERT architecture:

[0024] ;

[0025] In the formula, The verification set accuracy is represented, which is measured in the dialect translation task through the fine-tuned BERT model;

[0026] The single inference delay is represented, which is based on the measured results of the target chip;

[0027] The parameter quantity is represented, which directly affects the memory occupation and power consumption;

[0028] The weight coefficient: 0.7 / 0.2 / 0.1 balances the performance-efficiency demand and adapts to the resource limitation of Bluetooth earphones.

[0029] Further, the application of mixed precision quantization compresses the model volume, specifically, the Query / Key matrix in the attention mechanism is kept at FP16 precision, and the Value matrix is compressed to INT8, as follows:

[0030] Step 1: Dynamic statistical quantity calculation, specific formula as follows:

[0031] ;

[0032] In the formula, represents the original floating-point value of the first element in the Value matrix;

[0033] represents the Value matrix mean calculated layer by layer, used for center alignment quantization;

[0034] represents the Value matrix standard deviation calculated layer by layer, used for dynamic adjustment of scaling factors;

[0035] Step 2: Value matrix quantization compression, specific formula as follows:

[0036] ;

[0037] In the formula, represents the original Value matrix;

[0038] represents the quantized INT8 matrix, ranging from -128 to 127;

[0039] 127 represents the maximum scaling value of INT8 symmetric quantization, preserving the accuracy of the dynamic range;

[0040] Step 3: Mixed precision fusion, Query / Key retains FP16 to ensure high accuracy of attention weight calculation and avoid distortion of probability distribution caused by quantization error; Value is compressed to INT8 to take advantage of low-bit operation when participating in matrix multiplication, reducing memory bandwidth pressure; Specific formula as follows:

[0041] ;

[0042] In the formula, , represents the half-precision Query / Key matrix;

[0043] Q (Query) represents the query matrix, used to capture target semantic information; For example, the intent keyword in "navigate to the airport";

[0044] K (Key) represents the key matrix, storing context features; Such as time, place, and related information;

[0045] FP16 represents 16-bit floating-point precision, preserving the high resolution of attention weight calculation;

[0046] represents the feature length of the Key vector, used to scale the dot product result;

[0047] represents the 8-bit integer Value matrix;

[0048] represents a normalization function.

[0049] Further, the computing power of the matching chip is specifically optimized by dynamically adjusting the core frequency and the hardware instruction set to realize low-delay processing and energy efficiency balance for complex tasks, thereby significantly improving the endurance.

[0050] Further, the tactile operation frequency data is specifically based on touch behavior modeling and multi-modal data collaborative analysis to accurately identify high-frequency interaction intentions and trigger emergency operations, thereby greatly reducing the mis-touch rate.

[0051] Further, the eye trajectory data is specifically combined with a high-frame-rate sensor and deep learning technology to analyze user gaze behavior characteristics and actively adapt to scene requirements to improve interaction efficiency.

[0052] Further, the generated user sentiment polarity and intention confidence matrix triggers differentiated interaction strategies, specifically including the following:

[0053] Multi-modal graph structure construction:

[0054] The multi-modal graph structure node includes a pure speech stream data node, a tactile operation frequency data node, and an eye trajectory data node.

[0055] The pure speech stream data node extracts high-dimensional semantic vectors through a pre-trained language model to represent user instruction content and emotional tendencies.

[0056] The tactile operation frequency data node encodes the timing characteristics (such as double-click interval and pressure peak) of touch operations based on an LSTM network to model interaction behavior patterns.

[0057] The eye trajectory data node uses a lightweight Vision Transformer to analyze the attention weights of the gaze point heat map to generate a spatial attention vector.

[0058] Dynamic edge weight calculation, specifically as follows:

[0059] ;

[0060] wherein, represents the feature vectors of nodes and ;

[0061] represents the time synchronization deviation of multi-modal events;

[0062] represents a trainable weight matrix, is a Sigmoid activation function;

[0063] Multi-head cross-attention joint representation, the specific formula is as follows:

[0064] ;

[0065] In the formula, The speech node feature is taken as the query vector, focusing on the semantic core; such as the action instruction in "turn up the volume";

[0066] The eye movement node feature is taken as the key vector, associated with the visual focus; such as the user's gaze at the position of the "volume bar";

[0067] The haptic node feature is taken as the value vector, injecting physical interaction preferences; such as the force habit of sliding adjustment;

[0068] The feature vector dimension scaling factor prevents the gradient from being unstable due to the large value of the dot product;

[0069] Intention decision and strategy triggering, the specific formula is as follows:

[0070] Sentiment polarity classification: ;

[0071] In the formula, The sentiment classification weight matrix outputs the positive, neutral, and negative probability distribution.

[0072] Confidence threshold control:

[0073] High confidence (>0.8): directly execute the instruction; such as detecting "pause playing" + staring at the play icon;

[0074] Medium confidence (0.5-0.8): haptic vibration feedback confirmation; such as the user's vague instruction "turn down the volume a little;

[0075] Low confidence (<0.5): start voice interrogation; such as "Do you mean to lower the media volume, or the call volume?".

[0076] Further, the near-end strategy optimization specifically includes the following: dynamically adjusting the active heat dissipation period of the phase change material: based on the chip load and environmental temperature changes, the heat dissipation mode is dynamically adjusted through the reinforcement learning framework, balancing the device temperature control stability and energy efficiency;

[0077] Multi-objective collaborative optimization: taking user sensory comfort, chip performance, and endurance as joint optimization goals, generating adaptive heat dissipation action strategies to ensure that the device performance does not degrade in extreme scenarios;

[0078] Building a heat conduction digital twin model: integrating sensor data to build a thermodynamic simulation model, and updating the heat dissipation system to adapt to user behavior and environmental state in real time through policy gradient.

[0079] Among them, the IMU data fusion detects user behavior patterns through accelerometers and gyroscopes:

[0080] Stationary office, through the use of low-frequency fans;

[0081] Running, through the activation of phase change material heat storage; in the 35℃ outdoor running scene, the chip temperature is stable at 41±1℃, without the feeling of hot ears.

[0082] High temperature environment, through the superposition of graphene heat conduction.

[0083] Further, the deployment of physical information neural network simulates the molecular chain breakage rate of biodegradable materials in the humidity range of 30%-90%, specifically including the following:

[0084] Physical information time series convolution network, embedding battery chemical degradation equation as regular constraint:

[0085] ;

[0086] In the formula, represents the molecular weight of the polymer (unit: g / mol), which characterizes the change of material integrity over time;

[0087] represents the humidity and temperature dependent degradation rate constant, defined as ;

[0088] represents the pre-factor, which is related to the material structure;

[0089] represents the degradation activation energy (kJ / mol), which reflects the energy required for molecular chain breakage;

[0090] represents the environmental humidity, ranging from 30% to 90%, driving the hydrolysis reaction;

[0091] represents the humidity influence index, calibrated by experiment;

[0092] represents the reaction order, describing the degree of dependence of the degradation reaction on molecular weight;

[0093] represents the regularization weight, balancing the strength of experimental data fitting and physical equation constraint;

[0094] Predicting molecular chain breakage rates, quantifying material lifetime, and constructing a biodegradable PINN model:

[0095] ;

[0096] In the formula, This represents the rate of change of molecular weight, or the polymer molecular weight per unit time. The reduction in molecular weight represents the real-time rate of material degradation; the negative sign indicates that the molecular weight decreases over time. It represents the molecular weight of the polymer and characterizes the length and integrity of the material's molecular chains; The environmentally dependent degradation rate constant, which combines the accelerating effects of temperature T and humidity H on degradation, is composed of the Arrhenius equation and a humidity correction term: .

[0097] The beneficial effects achieved by adopting the above solution are as follows:

[0098] (1) To address the issues of high distortion rate and weak anti-interference capability in speech separation of traditional Bluetooth headsets in complex sound fields, the sound source separation and noise reduction module achieves a breakthrough through three-dimensional sound field mapping and dynamic mask generation technology. Based on the multi-head self-attention Transformer network, bone conduction vibration signal and air conduction signal are fused, and combined with the sound source orientation matrix constraint, a joint mask in the frequency domain, time domain and spatial domain is dynamically generated to accurately separate noise, human voice and ambient sound. In extreme scenarios with wind speed >8m / s, the speech signal-to-noise ratio is improved to more than 25dB and the distortion rate is less than 5%, solving the problems of speech truncation and low-frequency information loss caused by traditional fixed filtering, and providing broadcast-grade speech streams for scenarios such as meetings and sports.

[0099] (2) Edge model compression and multimodal interaction module: To address the pain points of traditional solutions such as high latency due to cloud dependence, significant privacy risks, and high single-modal false touch rate, the module is optimized through a dual-engine approach of progressive neural architecture search and spatiotemporal graph convolutional network. Automated distillation technology compresses the dialect translation model to 50MB, supports 20ms-level low-latency edge computing, and reduces power consumption by 60%. Multimodal dynamic graph fusion of translation data, tactile operation frequency data, and eye trajectory data generates an intent confidence matrix using multi-head cross-attention, triggering a graded strategy (high-confidence execution / low-confidence follow-up questioning), reducing the false touch rate from 30% to 1.3%, significantly improving the operational safety in driving and sports scenarios.

[0100] (3) Heat dissipation and predictive maintenance module, aiming at the problems of high load overheating and uncontrollable battery / material life, innovation is realized through near-end strategy optimization and physical information neural network. The heat conduction digital twin model dynamically adjusts the heat dissipation period of the phase change material, and the chip temperature fluctuation is controlled within ±1℃; the deep time convolution network analyzes the battery charging and discharging characteristics, and the capacity attenuation is warned 7 days in advance; the humidity-driven molecular chain rupture equation predicts the service life of the biodegradable shell, and the user can optimize the use environment (such as avoiding storage in >80% humidity), and prolong the service life of the device to more than 3 years. BRIEF DESCRIPTION OF DRAWINGS

[0101] Figure 1 A schematic diagram of a Bluetooth earphone supporting AI voice intelligence is provided.

[0102] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0103] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0104] Embodiment one, refer to Figure 1 The present application provides a Bluetooth earphone supporting AI voice intelligence, which comprises a sound source separation and noise reduction module, an edge model compression module, a multi-modal emotional interaction module, a heat dissipation control module and a predictive maintenance module.

[0105] The sound source separation and noise reduction module is used for sound source separation and noise reduction. Through a multi-head self-attention Transformer network, bone conduction signals and air conduction signals are fused to construct a three-dimensional sound field mapping. A dynamic mask generation algorithm is adopted to separate noise, human voice and environmental sound in real time, output pure voice stream data, and send the pure voice stream data to the edge model compression module.

[0106] The edge model compression module is used for edge model compression and real-time translation. Through progressive neural architecture search, a lightweight model is automatically distilled from a cloud model. The model volume is compressed by applying mixed precision quantization to match the computing power of the chip, support real-time translation of dialects or foreign languages, and send the translated data to the multi-modal emotional interaction module.

[0107] The multi-modal emotional interaction module is configured to perform multi-modal emotional interaction, generate a user emotional polarity and an intention confidence matrix by fusing translation data, haptic operation frequency data and eye trajectory data through a spatio-temporal graph convolution network, and trigger a differentiated interaction strategy.

[0108] The heat dissipation control module is configured to perform heat dissipation control, dynamically adjust an active heat dissipation period of the phase change material by fusing heat dissipation data of the sound source separation and noise reduction module, the edge model compression module and the multi-modal emotional interaction module and a near-end strategy optimization, integrate IMU sensor data, and construct a heat conduction digital twin model to enable effective heat dissipation control of the Bluetooth earphone and temperature stability.

[0109] The predictive maintenance module is configured to perform predictive maintenance, analyze multi-dimensional features of a battery charge-discharge curve through a deep time convolution network, and deploy a physical information neural network to simulate a molecular chain breakage rate of biodegradable materials in a humidity range of 30%-90% to enable the Bluetooth earphone to push a battery charging warning and optimize a use environment in advance.

[0110] In Embodiment Two, the dynamic mask generation algorithm specifically includes the following:

[0111] Signal feature extraction: the bone conduction signal acquires an energy value through a jaw bone vibration sensor ;

[0112] Air conduction signal: the environmental noise power is calculated through a 5-microphone array ;

[0113] Weight dynamic allocation:

[0114] ;

[0115] In the formula, denotes a calibration coefficient, denotes an energy integral of the bone conduction vibration signal, denotes a total frequency energy of the environmental noise;

[0116] Directional mask reinforcement: combined with a sound source azimuth angle constraint, the target speech separation precision is improved, and beamforming and azimuth matrix generation are performed: the sound arrival direction is calculated through a 5-microphone array, and an azimuth matrix is constructed.

[0117] Transformer attention layer optimization:

[0118] ;

[0119] In the formula, denotes a Query and Key matrix, and the dimension is a feature vector length; denotes the sound source direction matrix, and is added to the attention score to strengthen the main direction signal; denotes the azimuth enhancement coefficient, the dynamic range is 0.5-2.0, and is adaptively adjusted according to the noise type.

[0120] Embodiment three, based on the above embodiment, the lightweight model is automatically distilled from the cloud model, and specifically includes the following steps:

[0121] Progressive neural architecture search, construct multi-objective optimization function, synchronize compression model parameter quantity and inference delay, adopt evolutionary algorithm iterative search, generate 8-layer micro BERT architecture:

[0122] ;

[0123] In the formula, denotes the verification set accuracy, which is measured in the dialect translation task by the BERT model after fine-tuning;

[0124] denotes the single inference delay, which is based on the measured results of the target chip;

[0125] denotes the parameter quantity, which directly affects the memory occupation and power consumption;

[0126] Weight coefficient: 0.7 / 0.2 / 0.1 balance performance-efficiency demand, adapt to Bluetooth earphone resource limit.

[0127] Embodiment four, based on the above embodiment, the application of mixed precision quantization compresses the model volume, specifically for the Query / Key matrix in the attention mechanism to retain FP16 precision, and the Value matrix is compressed to INT8, as follows:

[0128] Step 1: Dynamic statistical calculation, the specific formula is as follows:

[0129] ;

[0130] In the formula, denotes the original floating-point value of the i-th element in the Value matrix;

[0131] denotes the Value matrix mean calculated layer by layer, used for center alignment of quantization;

[0132] denotes the Value matrix standard deviation calculated layer by layer, used for dynamic adjustment of the scaling factor;

[0133] Step 2: Value matrix quantization compression, the specific formula is as follows:

[0134] ​ ;

[0135] In the formula, represents the original Value matrix;

[0136] represents the quantized INT8 matrix, ranging from -128 to 127;

[0137] 127 represents the maximum scaling value of INT8 symmetric quantization, preserving the dynamic range accuracy;

[0138] Step 3: Mixed precision fusion, Query / Key retains FP16, ensuring high accuracy of attention weight calculation, avoiding distortion of probability distribution due to quantization error of Softmax; Value is compressed to INT8, taking advantage of low-bit operation when participating in matrix multiplication, reducing memory bandwidth pressure; The specific formula is as follows:

[0139] ;

[0140] In the formula, , represents the half-precision Query / Key matrix;

[0141] Q (Query) represents the query matrix, used to capture the target semantic information (such as the intent keyword in "navigate to the airport");

[0142] K (Key) represents the key matrix, storing context features (such as time, location, and related information);

[0143] FP16 represents 16-bit floating-point precision, preserving the high resolution of attention weight calculation;

[0144] represents the feature length of the Key vector, used to scale the dot product result;

[0145] represents the 8-bit integer Value matrix;

[0146] represents the normalization function.

[0147] Example five, based on the above-mentioned embodiment, the matching chip's computing power, specifically through dynamic adjustment of core frequency and hardware instruction set optimization, realizes low-delay processing and energy efficiency balance of complex tasks, significantly improving the endurance.

[0148] Example six, based on the above-mentioned embodiment, the haptic operation frequency data, specifically based on touch behavior modeling and multi-modal data collaborative analysis, accurately identifies high-frequency interaction intent and triggers emergency operations, significantly reducing the mis-touch rate.

[0149] Embodiment seven, based on the above embodiment, the eye tracking data, specifically, combined with high frame rate sensor and deep learning technology, analyzes the user's gaze behavior characteristics, actively adapts to the scene demand to improve the interaction efficiency.

[0150] Embodiment eight, based on the above embodiment, the generation of user sentiment polarity and intention confidence matrix triggers differentiated interaction strategy, specifically including the following:

[0151] Multi-modal graph structure construction:

[0152] Constructing multi-modal graph structure nodes, including pure speech stream data nodes, touch operation frequency data nodes, and eye tracking data nodes;

[0153] Pure speech stream data node: extract high-dimensional semantic vector through pre-trained language model, represent user instruction content and emotional tendency;

[0154] Touch operation frequency data node: based on LSTM network to encode the timing characteristics of touch operation (such as double-click interval, pressure peak), model interaction behavior pattern;

[0155] Eye tracking data node: use lightweight Vision Transformer to analyze the attention weight of fixation point heat map, generate spatial attention vector;

[0156] Dynamic edge weight calculation, the specific formula is as follows:

[0157] ;

[0158] In the formula, represent the feature vectors of nodes and ;

[0159] represents the time synchronization deviation of multi-modal events;

[0160] represents the trainable weight matrix, Sigmoid activation function;

[0161] Multi-head cross attention joint representation, the specific formula is as follows:

[0162] ;

[0163] In the formula, represent the speech node features as query vectors, focusing on the semantic core (such as action instruction in "turn up the volume");

[0164] Eye movement node features as key vectors, associated with visual focus (e.g. user gaze at the position of the "volume bar");

[0165] Haptic node features as value vectors, injecting physical interaction preferences (e.g. force habit for sliding adjustment);

[0166] Feature vector dimension scaling factor to prevent gradient instability caused by excessively large dot product values;

[0167] Intention decision and policy triggering, specifically formulated as follows:

[0168] Sentiment polarity classification: ;

[0169] In the formula, Sentiment classification weight matrix, output positive, neutral, negative probability distribution.

[0170] Confidence threshold control:

[0171] High confidence (>0.8): Directly execute the instruction (e.g. detect "pause playing" + gaze at the play icon);

[0172] Medium confidence (0.5-0.8): Haptic vibration feedback confirmation (e.g. user vague instruction "lower the volume a little bit");

[0173] Low confidence (<0.5): Start voice interrogation ("Do you mean to lower the media volume, or the call volume?")

[0174] As a specific embodiment of the present scheme:

[0175] Vehicle-mounted scenario:

[0176] When the user says "navigate to the company", if the eyeball continuously gazes at the map interface, the confidence is raised to 0.95, and the route planning is automatically started;

[0177] Detecting negative user emotions, such as rapid speech + frequent touch cancel key, automatically switch to concise broadcast mode.

[0178] Barrier-free interaction:

[0179] Hearing-impaired users trigger emergency calls by haptic long press and gaze at specific icons, bypassing voice input restrictions.

[0180] Example nine, based on the above-mentioned embodiments, the near-end strategy optimization, specifically includes as follows: dynamically adjusting the active heat dissipation period of the phase change material: based on the chip load and environmental temperature changes, dynamically adjusting the heat dissipation mode through the reinforcement learning framework, balancing the device temperature control stability and energy efficiency;

[0181] Multi-objective collaborative optimization: taking user comfort, chip performance and endurance as joint optimization objectives, generating adaptive cooling action strategies to ensure that device performance is not degraded in extreme scenarios;

[0182] Building a heat conduction digital twin model: integrating sensor data to build a thermodynamic simulation model, and updating the heat dissipation system to adapt to user behavior and environmental conditions in real time through strategy gradient.

[0183] Among them, IMU data fusion detects user behavior patterns through accelerometers and gyroscopes:

[0184] Stationary office, activate low-frequency fan;

[0185] Running, activate phase change material heat storage; in a 35°C outdoor running scenario, the chip temperature is stable at 41±1°C, and there is no hot ear feeling.

[0186] High temperature environment, superimposed graphene heat conduction.

[0187] Example ten, based on the above examples, the deployment of physical information neural network, simulation of biodegradable material in the range of 30%-90% humidity molecular chain breakage rate, specifically includes the following:

[0188] Physical information timing convolution network, embedding battery chemical degradation equation as regular constraint:

[0189] ;

[0190] In the formula, M(t) represents the molecular weight of the polymer (unit: g / mol), which characterizes the change of material integrity over time;

[0191] K(H, T) represents the humidity and temperature-dependent degradation rate constant, defined as ;

[0192] K0 represents the pre-factor (related to material structure);

[0193] Ea represents the degradation activation energy (kJ / mol), reflecting the energy required for molecular chain breakage;

[0194] H represents the environmental humidity (range 30%-90%), driving the hydrolysis reaction;

[0195] H represents the humidity influence index, calibrated by experiment;

[0196] n represents the reaction order, describing the degree of dependence of the degradation reaction on molecular weight;

[0197] This represents the regularization weight, which balances the fitting of experimental data with the constraint strength of the physical equations.

[0198] Predicting molecular chain breakage rates, quantifying material lifetime, and constructing a biodegradable PINN model:

[0199] ;

[0200] In the formula, This represents the rate of change of molecular weight, or the polymer molecular weight per unit time. The reduction in molecular weight represents the real-time rate of material degradation; the negative sign indicates that the molecular weight decreases over time. It represents the molecular weight of the polymer and characterizes the length and integrity of the material's molecular chains; The environmentally dependent degradation rate constant, which combines the accelerating effects of temperature T and humidity H on degradation, is composed of the Arrhenius equation and a humidity correction term: .

[0201] This solution utilizes 3D sound field mapping and dynamic mask generation technology, fusing bone conduction and air conduction signals to extract pure speech from complex sound fields (improving the signal-to-noise ratio by more than 15dB). Leveraging progressive neural architecture search and mixed-precision quantization, the translation model is compressed to within 50MB, achieving low-latency (<20ms) edge computing for both dialects and foreign languages. Combining spatiotemporal graph convolutional networks and near-end strategy optimization, it coordinates translation data, tactile operation frequency data, and eye trajectory data, and dynamically adjusts heat dissipation strategies to ensure stable operation of the device in environments ranging from -5℃ to 45℃. This innovation not only reshapes the interaction paradigm of headphones but also provides a technological closed loop for green manufacturing and sustainable development—from molecular-level degradation prediction to user-level health management, driving consumer electronics towards a fully intelligent leap across the "perception-decision-lifecycle" chain.

[0202] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0203] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0204] The above describes the present application and its embodiments, which are not limited, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments to the technical solution, which shall belong to the protection scope of the present application.

Claims

1. A Bluetooth earphone supporting AI voice intelligence, characterized in that: The sound source separation and noise reduction module, the edge model compression module, the multi-modal emotional interaction module, the heat dissipation control module and the predictive maintenance module are comprised. The sound source separation and noise reduction module is used for sound source separation and noise reduction, fuses bone conduction signals and air conduction signals through a multi-head self-attention Transformer network, constructs a three-dimensional sound field mapping, adopts a dynamic mask generation algorithm to separate noise, human voice and environmental sound in real time, outputs pure voice stream data, and sends the pure voice stream data to the edge model compression module. The edge model compression module is used for edge model compression and real-time translation, automatically distills a lightweight model from a cloud model through progressive neural architecture search, applies mixed precision quantization to compress the model volume, matches the computing power of the chip, supports real-time translation of dialects or foreign languages, and sends the translated data to the multi-modal emotional interaction module. The multi-modal emotional interaction module is used for multi-modal emotional interaction, fuses translation data, haptic operation frequency data and eye tracking data through a spatio-temporal graph convolution network, generates a user emotional polarity and intent confidence matrix, and triggers a differentiated interaction strategy. The heat dissipation control module is used for heat dissipation control, dynamically adjusts the active heat dissipation period of the phase change material by fusing heat dissipation data and near-end strategy optimization of the sound source separation and noise reduction module, the edge model compression module and the multi-modal emotional interaction module, integrates IMU sensor data, constructs a heat conduction digital twin model, and enables effective heat dissipation control of the Bluetooth earphone with stable temperature. The predictive maintenance module is used for predictive maintenance, analyzes the multi-dimensional features of the battery charge and discharge curve through a deep time convolution network. A physical information neural network is deployed to simulate the molecular chain breakage rate of biodegradable materials in a humidity range of 30%-90%, so that the Bluetooth earphone can push battery charging warnings and optimize the use environment in advance.

2. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The dynamic mask generation algorithm specifically includes the following: Signal feature extraction, Bone conduction signal acquires energy value through jaw bone vibration sensor ; Air conduction signals are computed ambient noise power through 5 microphone array ; Dynamic weight allocation: ； wherein denotes a calibration coefficient, denotes an energy integral of the bone conduction vibration signal, denotes a sum of frequency domain energies of the ambient noise; Directional mask enhancement, combined with sound source azimuth angle constraints, improves target speech separation accuracy, beamforming and azimuth matrix generation: sound direction is calculated through a 5-microphone array to construct an azimuth matrix; Transformer attention layer optimization: ； In the formula, Query and Key matrices, dimension is the feature vector length; represents the sound source direction matrix, which is added to the attention score to enhance the main direction signal; represents the azimuth enhancement coefficient, the dynamic range is 0.5~2.0, which is adaptively adjusted by the noise type.

3. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The automatic distillation of a lightweight model from a cloud model specifically includes the following steps: Progressive neural architecture search, construct a multi-objective optimization function, synchronize model parameter compression and inference delay, use evolutionary algorithm iterative search to generate an 8-layer micro-BERT architecture: ； In the formula, denotes the validation set accuracy; denotes the single inference latency based on the measured results of the target chip; The parameter represents the amount of parameters, which directly affects the memory usage and power consumption. Weight coefficient: 0.7 / 0.2 / 0.1 balance performance-efficiency demand, adapt to the resource constraints of Bluetooth earphones.

4. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The application of mixed precision quantization to compress the model volume specifically reserves FP16 precision for the Query / Key matrix in the attention mechanism, and compresses the Value matrix to INT8, as follows: Step 1: Dynamic statistical calculation, the specific formula is as follows: ； In the formula, Represents the value matrix at position n. The original floating-point values ​​of each element; Value matrix mean computed layer by layer, used to quantify center alignment; Value matrix standard deviation representing layer-by-layer calculation for scaling factor dynamic adjustment; Step 2: Value matrix quantization compression, the specific formula is as follows: ； wherein denotes the original Value matrix; represents a quantized INT8 matrix, range -128~127; 127 represents the maximum scaling value of INT8 symmetric quantization, which preserves the dynamic range accuracy. Step 3: Mixed precision fusion, Query / Key reserved FP16 to ensure high precision in attention weight calculation, avoiding distortion of probability distribution due to quantization error; Value compressed to INT8 to take advantage of low-bit operation in matrix multiplication, reducing memory bandwidth pressure; Specific formula as follows: ； wherein , denotes a half-precision Query / Key matrix; Q represents the query matrix, which captures the target semantic information. K represents the key matrix, which stores the context features. FP16 represents 16-bit floating-point precision, which preserves the high resolution of attention weight calculation. characteristic length of the Key vector, used to scale the dot product result; represents an 8-bit integer value matrix; denotes a normalization function.

5. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The matching chip's computing power, specifically through dynamic adjustment of core frequency and hardware instruction set optimization, realizes low-latency processing and energy efficiency balance for complex tasks, significantly improving endurance.

6. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The haptic operation frequency data, specifically based on touch behavior modeling and multi-modal data collaborative analysis, accurately identifies high-frequency interaction intent and triggers emergency operations, significantly reducing the rate of accidental touch.

7. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The eye tracking data, specifically combining high-frame-rate sensors and deep learning strategies, analyzes user gaze behavior characteristics, actively adapts to scene requirements to improve interaction efficiency.

8. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The generated user sentiment polarity and intent confidence matrix triggers differentiated interaction strategies, including the following: Multi-modal graph structure construction: Construct multi-modal graph structure nodes, including pure speech stream data nodes, haptic operation frequency data nodes, and eye tracking data nodes. Pure speech stream data nodes extract high-dimensional semantic vectors through pre-trained language models to represent user instruction content and emotional tendencies. Haptic operation frequency data nodes encode the timing characteristics of touch operations based on LSTM networks to model interaction behavior patterns. Eye tracking data nodes use lightweight Vision Transformer to analyze the attention weights of gaze point heat maps to generate spatial attention vectors. Dynamic edge weight calculation, specific formula as follows: ； wherein representing a node and characteristic vector of a time synchronization bias representing a multimodal event; denotes a trainable weight matrix, is a sigmoid activation function; Multi-head cross-attention joint representation, specific formula as follows: ； In the formula, The voice node features are represented as query vectors, focusing on semantic cores. representing eye movement node features as key vectors, associated with visual focus; representing haptic node characteristics as value vectors, injecting physical interaction preferences; represents a feature vector dimension scaling factor to prevent the gradient from becoming unstable due to excessively large dot product values; Intent decision and strategy triggering, specific formula as follows: Sentiment polarity classification: ; In the formula, represents the emotion classification weight matrix, and an output positive, neutral, and negative probability distribution.

9. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The near-end strategy optimization includes the following: dynamically adjusting the active heat dissipation period of phase change materials: based on chip load and environmental temperature changes, dynamically adjusting the cooling mode through a reinforcement learning framework to balance device temperature stability and energy efficiency; Multi-objective collaborative optimization: taking user comfort, chip performance, and endurance as joint optimization objectives, generating adaptive cooling action strategies to ensure that device performance degrades without feeling in extreme scenarios; Build a heat conduction digital twin model: integrate sensor data to build a thermodynamic simulation model, and update the cooling system's real-time adaptation to user behavior and environmental conditions through policy gradient.

10. The Bluetooth earphone supporting AI voice intelligence according to claim 1, characterized in that: The deployment of physical information neural networks simulates the molecular chain breakage rate of biodegradable materials in a humidity range of 30%-90%, including the following: Physical information time series convolution network, embedding battery chemical degradation equation as a regular constraint: ； wherein Mw represents the polymer molecular weight, characterizing the material integrity over time; represents the humidity and temperature dependent degradation rate constant, defined as ; denotes a pre-factor; Ea represents the degradation activation energy, reflecting the energy required for the molecular chain to break; representing the environmental humidity, ranging from 30% to 90%, driving the hydrolysis reaction; represents a humidity influence index, calibrated by experiment; n represents the reaction order, describing the degree of dependence of the degradation reaction on the molecular weight; represents the regularization weight, balancing the strength of experimental data fitting and physical equation constraints; Predict the molecular chain breakage rate to quantify material life and build a biodegradable PINN model: ； wherein, represents the rate of change of molecular weight, the decrease in polymer molecular weight per unit time represents the rate of change of molecular weight, the decrease in polymer molecular weight per unit time represents the rate of change of molecular weight, the decrease in polymer molecular weight per unit time represents the rate of change of molecular weight, the decrease in polymer molecular weight per unit time

Images