Electric toothbrush control method and system based on multi-sensor adaptive weighted fusion

Abstract

The application discloses a kind of electric toothbrush control method and system based on multi-sensor adaptive weighting fusion, its method includes: posture, pressure, spectrum and acoustic data are synchronously collected;After pre-processing and feature extraction, the evidence of multi-sensor is fused using D-S evidence theory, and the regional confidence distribution and behavior mode label are output;Combining reliability factor and the context factor determined by behavior mode, dynamically generate weighting coefficient, weight and fuse multi-source features, obtain comprehensive cleanliness estimate and fused spectrum characteristics;Based on spectrum characteristics, calculate plaque index, based on acoustic characteristics, calculate brushing efficiency score;With plaque index, brushing efficiency score and pressure safety threshold as constraint, establish multi-objective optimization function, solve optimal vibration frequency and amplitude in real time, realize closed-loop control, with the advantages of strong anti-interference, personalized adaptation and the like.

Description

Technical Field

[0001] This invention relates to the field of digital medical device technology, and in particular to an electric toothbrush control method and system based on multi-sensor adaptive weighted fusion. Background Technology

[0002] In recent years, smart electric toothbrushes have become an important tool for oral care. Existing technologies mostly rely on single or limited types of sensors, such as six-axis inertial measurement units (IMUs) for coarse brushing area recognition, or pressure sensors for simple overpressure warnings. However, these solutions have the following shortcomings: (1) Limited perception dimension and accuracy: Most existing solutions rely on IMU for posture recognition and region judgment. However, the internal structure of the oral cavity is complex, and a single sensor is easily affected by the user's hand tremors and the acceleration of the brush head movement, resulting in a high rate of region misjudgment and difficulty in accurately judging specific tooth surfaces. Although some studies have attempted to combine IMU and wristband data for fusion, they mostly remain at the level of simple data complementarity, lacking a dynamic weight allocation mechanism and unable to adapt to the dynamic changes in the sensor signal-to-noise ratio during brushing.

[0003] (2) Lack of cleaning effect assessment and open-loop control: Existing technologies focus on area identification and pressure control, but cannot assess key indicators such as plaque removal effect in real time. The system is in a "blind operation" state, unable to form closed-loop control based on feedback of cleaning effect, and lacks dynamic parameter adjustment based on assessment results.

[0004] (3) Poor system robustness: In complex usage scenarios (such as when the brush head is blocked or there is a lot of foam), single sensor data is very easy to fail, and the existing system lacks fault tolerance and compensation mechanisms when some sensor data is abnormal.

[0005] For example, the invention patent with publication number CN105636479A discloses a technical solution based on obtaining the reflectance spectrum using a spectral sensor and calculating the plaque index (PI) using an algorithm. However, like other existing technologies, it is still a single optical mode detection, only providing diagnostic results and having limited closed-loop control capabilities.

[0006] The present invention aims to overcome the above-mentioned defects by fusing multi-source sensor data and introducing advanced fusion algorithms and control theories to construct a closed-loop control system capable of real-time sensing, evaluation and dynamic adjustment. Summary of the Invention

[0007] A brief overview of embodiments of the invention is provided below to provide a basic understanding of certain aspects of the invention. It should be understood that this overview is not an exhaustive summary of the invention. It is not intended to identify key or essential parts of the invention, nor is it intended to limit the scope of the invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.

[0008] According to one aspect of this application, a method for controlling an electric toothbrush based on multi-sensor adaptive weighted fusion is provided, comprising: Step 1: Simultaneously sample the attitude sensor (IMU), pressure sensor, optical sensor, and acoustic sensor using the same hardware trigger signal to obtain attitude, pressure, spectral, and acoustic data; Step 2: Preprocess and extract features from the sensor data obtained in Step 1: perform low-pass filtering on the attitude data, peak detection on the pressure data, sliding window mean filtering on the spectral data, and extract Mel frequency cepstral coefficient (MFCC) features from the acoustic data. Step 3: Based on the DS (Dempster-Shafer) evidence theory, identify the brushing area and behavioral intention, assign a basic probability assignment function (BPA) to the observation results of each sensor, and fuse them through Dempster combination rules to output the area confidence distribution and behavioral patterns (such as brushing and patting), providing decision context for subsequent steps; Step 4: Based on the adaptive weighted multi-sensor fusion state estimation algorithm, the reliability factor and context factor of each sensor are calculated using the regional confidence distribution and behavior pattern output in Step 3. Weighting coefficients are dynamically generated, and the multi-source features (attitude, pressure, spectral, and acoustic features) output in Step 2 are weighted and fused to output a high-confidence comprehensive cleanliness estimate and the fused spectral features. The reliability factor is determined by the normalized similarity between the sensor's feature value in the current window and the historical normal distribution. The similarity is calculated using Bach distance or cosine similarity. The context factor is obtained from the regional confidence distribution output in Step 3. Step 5: Calculate the plaque index (PI) based on the fused spectral features to quantify the amount of plaque residue; at the same time, calculate the brushing efficiency score based on the acoustic features to quantify the efficiency of brushing actions; the two together constitute a quantitative indicator of the current cleanliness status. Step 6: Based on the plaque index, brushing efficiency score, and pressure safety threshold, the optimal control strategy is generated by solving a multi-objective optimization problem, and the vibration frequency and amplitude of the brush head are adjusted in real time; the pressure safety threshold is a preset value.

[0009] Furthermore, step 3 specifically includes: Step 31: Define the brushing area recognition framework (Refer to the 16-zone division method) and behavioral type framework ={Effective scrubbing, excessive force, patting}; Step 32: Assign a basic probability assignment function to the observations of each sensor: For attitude sensors: based on the matching degree between historical attitude data and the region model, for The probability quality of subset assignment; For pressure sensors: based on pressure modes (e.g., continuous steady-state, instantaneous spikes), for Assignment probability quality; For acoustic sensors: based on the characteristics of Mel-frequency cepstral coefficients, for Assignment probability quality; Step 33: Use the Dempster combination rule to fuse the basic probability assignment functions of each sensor output in Step 32 to obtain the synthetic basic probability assignment function. Normalize the confidence of each hypothesis, and then output the regional confidence distribution and behavioral pattern labels for use by the adaptive weighting module. Specifically, this includes: basic probability assignment functions for two sensors (such as IMU and sound). and The combination rules are as follows: ; in, It is a normalization constant; this formula can effectively handle sensor evidence conflicts and improve the reliability of recognition results.

[0010] Furthermore, in step 3, the brushing area and behavioral intention are identified based on the DS evidence theory. Specifically, the identification of the brushing area and behavioral intention is based on the temporally enhanced DS evidence theory, which includes: (1) Temporal evidence accumulation and decay model: Define sliding time window Wherein, the window size T is a system preset constant (e.g., =2s), determined based on the typical cycle of brushing motion; for each moment within the window Calculate the statute of limitations weight of the evidence. Among them, the attenuation coefficient A preset constant (e.g., δ=0.5s) is used to control the decay rate of historical evidence. This value is calibrated using experimental data to balance the system's sensitivity to recent changes with its memory of historical information. The accumulated evidence at time t... for: ; This model assigns higher weight to recent evidence, enabling the system to perceive the dynamic evolution of brushing behavior patterns. Among these, Let τ be the degree of confidence that all sensor evidence indicates the true situation belongs to proposition A. For time τ, all sensor evidence indicates the degree of trust in the true situation belonging to proposition B; the two are obtained by retrieving the original sensor data at time τ from the cache, that is, by combining the original sensor data and step 32 to obtain the result using the general classic DS rule fusion.

[0011] (2) Improved Dempster combinatorial rule: New evidence at the current time t The accumulated evidence from the previous moment combination: ; in, This represents the basic probability allocation function calculated based on the latest collected sensor data at the current time t; To determine the degree of confidence in proposition C based on the latest sensor readings at the current moment; dynamically adjust parameters. The determination is based on the degree of conflict between the current and old evidence, specifically through calculation. and The distance between Jousselme evidence (a method in DS evidence theory used to quantify the similarity between two evidence bodies (i.e., basic probability assignment functions)) And map it to the interval [0,1]: ,here, For the Sigmoid function, This is a preset scaling factor. This rule applies when there is significant conflict in the evidence ( big, When the time is short, people tend to choose a single piece of evidence with a high degree of trust to enhance the robustness of the system under transient disturbances.

[0012] Furthermore, in step 4, the adaptive weighted multi-sensor fusion state estimation algorithm (generating high-confidence states) specifically includes: Step 41: Set time The sensor set is Its observation vector is The optimal estimated state variables after fusion (such as "overall cleanliness") are: ; in, It is a sensor At any moment The adaptive weights satisfy ; Step 42: Adaptive Weights The calculation formula is as follows: ; in, The reliability factor is calculated based on the sensor signal noise variance (such as the dynamic accuracy of the accelerometer) and historical consistency. The calculation formula is as follows: ,in It should be a very small positive number to prevent division by zero. The larger the noise variance, the lower the reliability.

[0013] It is a contextual factor, determined by the current behavior pattern; , These are learnable hyperparameters that control the influence strength of reliability and context factors, respectively.

[0014] This algorithm can dynamically adjust the weights: when the signal of a certain sensor is interfered with (such as when an optical sensor is blocked by foam), its reliability decreases, its weight is automatically reduced, and the system becomes more dependent on other sensors, thereby ensuring the robustness of the overall system.

[0015] Furthermore, step 5 specifically includes: Step 51: Calculate the plaque index: Obtain the reflectance spectrum of the tooth surface using an optical sensor. Using the high-weighted optical sensor data from step 4, the Plaque Index (PI) is defined: ; in, This refers to the characteristic absorption band of dental plaque; For reference wavelength, the tooth itself reflects light strongly; The characteristic weighting function for dental plaque is obtained through laboratory calibration.

[0016] The lower the PI value, the less plaque residue remains in that area.

[0017] Step 52: Calculate brushing efficiency score: Collect sound signals through microphone. The Mel-frequency cepstral coefficients (MFCC) feature is extracted; and the brushing efficiency score is calculated by matching it with a pre-stored effective brushing acoustic model. (Between 0 and 1), the higher the score, the more effective the brushing action.

[0018] Furthermore, in step 52, the reliability of the effective brushing acoustic mode is determined by analyzing the energy distribution entropy of the Mel-frequency cepstral coefficients (MFCC) characteristics of the sound signal. Based on this, the plaque index PI in step 51 undergoes a cross-modal fusion plaque index correction step, constructing a collaborative correction mechanism between optical and acoustic sensors. This directly links the plaque index PI to the brushing efficiency score calculation, and the corrected plaque index is ultimately used for the calculation in step 6.

[0019] The high-weighted optical sensor data in step 51 is the filtered spectrum after fusion, with its weights adaptively adjusted. Specifically, it refers to the reflectance spectrum vector of the optical sensor after being filtered by the sliding window mean and assigned the current reliability weight in the multi-sensor weighted fusion process. This vector is used for both the subsequent calculation of the plaque index PI and as a contribution component of the optical channel to the overall cleanliness estimation.

[0020] Furthermore, in step 6, the optimal control strategy formula is as follows: ; Constraints: f min ≤ ≤f max A min ≤ ≤A max ; in, The control variables to be optimized are the vibration frequency and amplitude of the brush head, respectively. These are the target plaque index and the current plaque index, respectively. and These are the current pressure and the safe pressure threshold, respectively. Rate the efficiency of acoustic brushing; These are weighting coefficients used to balance cleaning effectiveness, pressure safety threshold (safety), and acoustic efficiency (user experience), respectively.

[0021] By solving this optimization problem, optimal control commands are generated in real time. It is then sent to the motor drive module.

[0022] According to a second aspect of this application, an electric toothbrush control system based on multi-sensor adaptive weighted fusion is provided, comprising: The sensor module integrates attitude sensors, pressure sensors, optical sensors, and acoustic sensors. The processing module, including an MCU, RAM, and EEPROM, is used to store and run the computer program for the electric toothbrush control method described above. The processor can be implemented using an STM32H7 MCU. Execution module: includes motor drive and brush head.

[0023] The attitude sensor is a ten-axis inertial measurement unit, comprising a three-axis accelerometer, a three-axis gyroscope, a three-axis magnetometer, and a barometer. The optical sensor can be a spectral sensor, a near-infrared photoelectric sensor, a visible light imaging sensor, or a laser speckle sensor. The pressure sensor is a piezoresistive pressure sensor, and the acoustic sensor is a MEMS (Micro-Electro-Mechanical System) microphone. The driving unit can be a motor drive, a piezoelectric ceramic drive, a linear resonant drive, or a magnetic levitation drive module.

[0024] This invention adopts the above-mentioned solution. The system employs a closed-loop intelligent control architecture of "perception-fusion-decision-execution." Its hardware foundation is an electric toothbrush integrating an attitude sensor (IMU), pressure sensor, optical sensor, and acoustic sensor. All sensor data is synchronously acquired and processed by the MCU. Simultaneously, a multi-objective optimization function incorporating cleaning effect, pressure safety threshold, and acoustic efficiency is constructed. This function is solved in real-time to generate optimal control commands (vibration frequency and amplitude), which are then sent to the actuator to achieve precise closed-loop control. Compared with existing technologies, this invention offers the following advantages: 1. High-precision perception and strong robustness: By adopting DS evidence theory and adaptive weighted fusion, the corresponding weights can be automatically reduced when some sensor signals are disturbed or fail, so as to maintain the stability of the estimation and significantly improve the reliability in complex oral environments.

[0025] 2. Closed-loop control of cleaning effect: Using the spectral plaque index and acoustic efficiency score as feedback, a real-time closed loop is formed, which can dynamically adjust the brush head vibration parameters according to the actual cleaning status, realizing the transformation from "action-driven" to "effect-driven".

[0026] 3. Personalized Adaptation: Through online collaboration of reliability factors and context factors, the system can automatically adapt to different users' grip strength, brushing speed, and gum sensitivity, providing a cleaning strategy that matches individual habits.

[0027] 4. Multi-objective synergy: Within the same optimization function, it takes into account cleaning effect, pressure safety and acoustic comfort, avoiding overpressure or noise problems caused by solely pursuing cleaning, and improving the overall user experience.

[0028] In addition, the algorithm is deployed on the MCU through model pruning and low-bit quantization, and can run offline without external computing resources, reducing hardware costs and power consumption requirements, making it widely applicable to battery-powered portable devices. Attached Figure Description

[0029] The present invention can be better understood by referring to the description given below in conjunction with the accompanying drawings, in which the same or similar reference numerals are used throughout the drawings to denote the same or similar parts. These drawings, together with the following detailed description, are incorporated in and form part of this specification, and are used to further illustrate preferred embodiments of the invention and explain the principles and advantages of the invention. In the drawings: Figure 1 This is a flowchart of the intelligent toothbrush control method of the present invention. Detailed Implementation

[0030] Embodiments of the present invention will now be described with reference to the accompanying drawings. Elements and features described in one drawing or embodiment of the invention may be combined with elements and features shown in one or more other drawings or embodiments. It should be noted that, for clarity, representations and descriptions of components and processes unrelated to the present invention and known to those skilled in the art have been omitted from the drawings and description.

[0031] The core of the system proposed in this invention adopts a closed-loop intelligent control architecture of "perception-fusion-decision-execution". Its hardware foundation is an electric toothbrush integrating a ten-axis inertial measurement unit (IMU), namely an attitude sensor (IMU) (three-axis accelerometer, three-axis gyroscope, three-axis magnetometer, and barometer), a piezoresistive pressure sensor, a miniature optical sensor, and a MEMS microphone (acoustic sensor). All sensor data is synchronously acquired and processed by a high-performance MCU (such as STM32H7). The workflow of this closed-loop intelligent system is as follows: Figure 1 As shown, it includes multi-sensor data synchronous acquisition, data preprocessing and feature extraction, cleaning effect evaluation and multi-objective optimization strategy generation.

[0032] The specific technical solution is described below: (1) Step 1: Synchronous acquisition of data from multiple sensors Hardware synchronization, using hardware trigger signals, ensures that the timestamps of the four sensor data are aligned. The timestamp difference between the sampled data from each channel is ≤1ms. (2) Step 2: Data preprocessing and feature extraction The obtained sensor data is preprocessed and features are extracted, mainly by low-pass filtering of IMU data, sliding window mean filtering of optical sensor data, and data preprocessing and feature extraction of other sensors using common sensor calibration, data alignment and noise filtering methods.

[0033] (3) Step 3: Identification of brushing area and behavioral intention based on DS evidence theory (providing decision context) This step is the starting point for intelligent sensing, and its core objective is to understand the user's "brushing intention" (i.e., which area is being brushed and how it is being brushed), and to provide crucial contextual information for subsequent steps. This solves the problem of high uncertainty in perception by a single sensor in a complex oral environment.

[0034] To address the issue of high uncertainty associated with a single sensor, this invention introduces the Dempster-Shafer (DS) evidence theory to identify brushing areas and action types (such as brushing and patting).

[0035] Recognition Framework: Defining the brushing area recognition framework (Refer to the 16-zone division method) and behavioral type framework ={Effective scrubbing, excessive force, patting}.

[0036] Basic probability assignment: Each sensor is assigned a basic probability assignment function to its observations.

[0037] Attitude sensor: Based on the matching degree between historical attitude data and the region model, for The probability quality of subset assignment. For example, representing This indicates the IMU's assessment of the current level of trust in region 1 or 2.

[0038] Pressure sensor: Based on pressure patterns (e.g., continuous steady-state, instantaneous spikes), for Assigning probability quality. For example... This indicates that the confidence level of the pressure sensor in deeming excessive force is 0.8.

[0039] Acoustic sensors: By analyzing the sound spectrum characteristics (MFCC), for Assigning probability quality. For example... This indicates that the acoustic sensor's confidence level in the effectiveness of the scrubbing is 0.6.

[0040] Evidence Combination: Multi-sensor evidence is fused using Dempster's combination rule (the core mathematical tool in DS evidence theory, which solves the problem of how to integrate information from multiple independent evidence sources into a unified and credible conclusion).

[0041] The generation of BPA for each sensor is a pattern recognition and classification process based on a pre-trained model, which is described as a specific example below: For the attitude sensor, the input is a sequence of quaternions; feature extraction: calculate the dynamic time warping (DTW) distance with 16 pre-stored "standard region attitude templates"; BPA generation method (mapping relationship) is: 1. Calculate similarity: S i =exp(-α×DTW distance)i ), S i This represents the similarity between the current pose data and the pose template of the i-th standard region. This value is used to subsequently calculate the probability quality (confidence) of the corresponding region. DTW distance i This represents the distance between the current pose sequence calculated using the Dynamic Time Warping (DTW) algorithm and the i-th pre-stored standard region pose template, used to quantify the degree of difference between the two in their pose trajectories. α is a system-preset constant (or scaling factor) used to control the strength of the influence of distance difference on similarity decay; 2. BPA Allocation: The similarity value is directly assigned to the corresponding region single proposition, and the remaining confidence is assigned to the entire set. (i.e., "unknown region", representing uncertainty). Example: Assume the current pose has a DTW distance of 1.5 with the template in region 8 and 12.0 with region 9. Calculations yield BPA: m({region 8}) = 0.7, m({region 9}) = 0.1, and the remaining confidence level m(Θ) = 1 - (0.7 + 0.1) = 0.2 (uncertainty).

[0042] For pressure sensors, the input is a real-time pressure value sequence. Feature extraction involves calculating the mean, standard deviation, kurtosis, and number of times the pressure exceeds the threshold within a time window, forming a 4-dimensional feature vector. The BPA generation method involves inputting the 4-dimensional feature vector into a pre-trained 3-class classifier (such as SVM or a lightweight neural network). The classifier's output layer is a Softmax layer, and its three output probability values ​​represent the BPA for the three basic propositions {effective scrubbing (E), excessive force (H), and patting (T)}. Example: Feature vector [0.8N, 0.15N, 0.1, 2]. Classifier output: [0.85, 0.10, 0.05]. Then the BPA is: m({E}) = 0.85, m({H}) = 0.10, m({T}) = 0.05.

[0043] For acoustic sensors, the input is an audio signal. Feature extraction involves extracting 13-dimensional MFCC coefficients and their first and second-order differences, resulting in a 39-dimensional feature vector. The BPA generation method is as follows: the 39-dimensional MFCC feature vector is input into another pre-trained 3-class classifier. This classifier also outputs three probabilities using Softmax, which serve as the BPA for {E, H, T}. Example: The MFCC feature vector, after inference by the classifier, outputs [0.70, 0.25, 0.05]. Therefore, the BPA is: m({E}) = 0.70, m({H}) = 0.25, m({T}) = 0.05.

[0044] The identification of brushing areas and behavioral intentions is specifically based on the temporally enhanced DS evidence theory, and includes: (1) Temporal evidence accumulation and decay model: Define sliding time window Wherein, the window size T is a system preset constant (e.g., =2s), determined based on the typical cycle of brushing motion; for each moment within the window Calculate the statute of limitations weight of the evidence. Among them, the attenuation coefficient A preset constant (e.g., δ=0.5s) is used to control the decay rate of historical evidence. This value is calibrated using experimental data to balance the system's sensitivity to recent changes with its memory of historical information. The accumulated evidence at time t... for: ; This model assigns higher weight to recent evidence, enabling the system to perceive the dynamic evolution of brushing behavior patterns. Among these, Let τ be the degree of confidence that all sensor evidence indicates the true situation belongs to proposition A. To determine the degree of trust in proposition B based on all sensor evidence at time τ, the method for obtaining both is to retrieve the original sensor data at time τ from the cache, i.e., to combine the original sensor data with the data obtained in step three using the general classic DS rules for fusion.

[0045] (2) BPA for two sensors (e.g., IMU and sound) and Improved Dempster combinatorial rules: New evidence at the current time t The accumulated evidence from the previous moment combination: ; When evidence conflicts (i.e., the intersection is empty), the larger function is used to allocate probability mass, thereby enhancing the system's robustness under transient disturbances.

[0046] in, This represents the basic probability allocation function calculated based on the latest collected sensor data at the current time t; Let C be the degree of trust in proposition C based on the latest sensor readings at the current moment; C is a subset of the recognition frame Θ.

[0047] Here, A, B, and C are all propositional subsets (that is, subsets of assumptions used for formula calculation and ease of understanding), which are used in the combination formula to traverse all possible combinations of subsets to calculate the normalized probability mass.

[0048] Dynamically adjust parameters The determination is based on the degree of conflict between the current and old evidence, specifically through calculation. and The distance between Jousselme evidence (a method in DS evidence theory used to quantify the similarity between two evidence bodies (i.e., basic probability assignment functions)) And map it to the interval [0,1]: ,here, For the Sigmoid function, This is the preset scaling factor. This is a normalization constant to ensure that the sum of the probability masses after combination is 1. ; To identify the framework. This rule applies when there is significant conflict of evidence ( big, When the evidence is small, the system tends to select a single piece of evidence with high confidence. When the evidence conflicts (i.e., the intersection is empty), the system uses the larger function to allocate probability mass, thereby enhancing the system's robustness under transient disturbances.

[0049] (4) Step 4: Adaptive weighted multi-sensor fusion state estimation algorithm (generating high confidence state) Step three above identified "where and what," while this step answers "what is the current actual cleaning status." This step uses the output of step three to dynamically adjust the fusion strategy to estimate a more reliable environmental state that better fits the current context.

[0050] To achieve highly robust state estimation, this invention designs an adaptive weighted fusion model to dynamically evaluate the reliability of each sensor.

[0051] Fusion formula: Set time The sensor set is Its observation vector is The optimal estimated state variables after fusion (such as "overall cleanliness") are: ; Among them, is sensor At any moment The adaptive weights satisfy ,and ≥0.

[0052] Adaptive weight calculation: ; Parameter definition: Reliability factor: Calculated based on sensor signal noise variance (e.g., accelerometer dynamic accuracy) and historical consistency. A higher noise variance indicates lower reliability. The calculation formula is as follows: ,in To prevent division by zero for extremely small positive numbers, It is the noise variance; Contextual factors are determined by the current brushing mode. For example, in "gum care mode," the contextual factor of the pressure sensor is increased to focus on monitoring brushing pressure; a "behavior mode - sensor weight coefficient" lookup table is pre-stored in the system.

[0053] The baseline value is set to 1.0 for all sensors in the "Effective Scrub" mode, indicating that there is no special context preference.

[0054] Dynamic Adjustment: When the system identifies a specific behavioral pattern, the factor of the corresponding sensor will increase or decrease to reflect changes in its importance or reliability under that pattern. Pressure Sensor: In "Excessive Force" mode, the factor increases to 1.5, indicating a need for focused monitoring of pressure safety. Acoustic Sensor: In "Slapping" mode, the factor increases to 1.3 because sound characteristics are more representative of identifying such ineffective actions. In other modes, the factor may decrease, indicating that its contribution weight can be appropriately reduced.

[0055] For example: When step three determines that the current area is the posterior molar region (high trust level) and the action is "effective brushing," the contextual factors of the optical sensor (spectral analysis) are considered because this area is prone to food residue. This will be significantly improved. At the same time, the space in this area is narrow, and the context factor of the acoustic sensor... This will also be improved because the sound reflection characteristics are more representative. Conversely, in the incisor region, the context factor of the attitude sensor may be higher. In this way, the fusion center of gravity in step four will dynamically tilt according to the recognition results in step three, making the state estimation more accurate.

[0056] , The system employs learnable hyperparameters to control the influence of reliability and context factors. Hyperparameters α and β are pre-set at the factory and trained using a supervised learning algorithm. The training data consists of multiple sets of sensor data collected in a laboratory environment. Each set includes time-series observations from attitude, pressure, optical, and acoustic sensors during brushing, along with expert-annotated "overall cleanliness" labels for the corresponding time points. By minimizing the loss function (e.g., mean squared error) between the model's output fused state estimate and the expert-annotated true state, gradient descent is used to iteratively optimize hyperparameters α and β, thereby learning the optimal influence of reliability and context factors on the fusion weights. These are then written as fixed parameters into the firmware.

[0057] This algorithm can dynamically adjust the weights: when the signal of a certain sensor is interfered with (such as when an optical sensor is blocked by foam), its reliability decreases, its weight is automatically reduced, and the system becomes more dependent on other sensors, thereby ensuring the robustness of the overall system.

[0058] (5) Step 5: Cleaning effect evaluation model based on optical and acoustic analysis Step four outputs a high-quality "comprehensive state estimate". Step five then introduces a physical model to quantitatively evaluate the "cleaning effect" based on this, and generates the optimal control command accordingly, forming a closed loop.

[0059] Optical sensors acquire the reflectance spectrum of the tooth surface. Using the high-weighted optical sensor data from step four, the Plaque Index (PI) is defined: ; Parameter definition: : The characteristic absorption band of dental plaque, for example, 540–580 nm (plaque absorption peak); Reference wavelength: Tooth body reflection is relatively strong, for example, 680–720 nm (tooth body reflection reference). The reference wavelength and absorption wavelength do not overlap. The characteristic weighting function of dental plaque is obtained by measuring and fitting the spectral data of known plaque samples in the laboratory. The specific steps are as follows: In a laboratory setting, using standard dental molds and plaque simulation paste, spectral data and reflectance spectra of tooth samples or simulated samples at different wavelengths were simultaneously acquired using standard imaging equipment (true values) to measure plaque coverage (e.g., quantified by a professional dentist or standard plaque staining agent). The acquired reflectance spectral data underwent preprocessing, such as filtering and normalization, to reduce noise. Based on the known plaque coverage (as the target value) and reflectance spectral data at different wavelengths (as features), a weighting function characterizing plaque absorption properties was fitted using regression analysis (e.g., multiple linear regression, partial least squares regression, or neural networks). This function reflects the correlation strength between the spectral reflectance intensity at different wavelengths and the amount of plaque present.

[0060] The lower the PI value, the less plaque residue remains in that area.

[0061] Acoustic assessment (brushing efficiency score): Microphone captures sound signals The Mel-frequency cepstral coefficients (MFCC) features were extracted. A brushing efficiency score was calculated by matching the results with a pre-stored effective brushing acoustic model. (Between 0 and 1), a higher score indicates a more effective brushing motion. In the lab, audio recordings of standard brushing techniques by dental experts (positive samples) and ineffective brushing (negative samples) were generated. MFCC features were extracted, and a binary classification model (such as logistic regression) was trained. The output probability of this model for the real-time audio is... .

[0062] It's worth noting that step three (DS theory) acts as the "brain," understanding the user's intent and providing decision-making context. Step four (adaptive weighting) acts as the "nervous system," dynamically allocating perceptual resources based on the context to generate high-confidence state information. Step five (effect evaluation and optimization) acts as the "command center," evaluating the effect based on the state information and issuing precise action instructions.

[0063] Furthermore, the reliability of acoustic modes This can be determined by analyzing the energy distribution entropy of the Mel-frequency cepstral coefficients (MFCC) characteristics of the sound signal: ; in, It is the entropy function. For the energy vectors of each frequency band of MFCC, This represents the preset total number of frequency bands. ln() is the natural logarithm function.

[0064] Cross-modal plaque index correction: The corrected plaque index was calculated using cross-modal fusion: ; PI stands for Plaque Index. Rate brushing efficiency. Weighting is incorporated. for , For the reliability of acoustic modes, Based on the preset proportional coefficients obtained through experimental data optimization, this model constructs a collaborative correction mechanism for optical and acoustic sensors, linking the plaque index (PI) with brushing efficiency scores. The calculation is directly linked.

[0065] The final plaque index was adopted. Now proceed with the calculations in step six.

[0066] (6) Step Six: Multi-objective Optimization Control Strategy Based on the above perception results, the control strategy is defined by a multi-objective optimization problem to simultaneously optimize cleaning effectiveness, safety, and user experience, as shown in the following formula: ; Constraints: f min ≤ ≤f max A min ≤ ≤A max ;; The parameter is defined as follows: The control variables to be optimized are the vibration frequency and amplitude of the brush head.

[0067] Target plaque index and current plaque index The high-confidence state estimate derived from step four incorporates contextual information. Target value. This can be dynamically configured. For example, if step three identifies the current area as a "high-risk area," then... It can be set to be more stringent (smaller value) and require a higher level of cleanliness.

[0068] , Current pressure and safe pressure threshold.

[0069] Acoustic brushing efficiency rating.

[0070] Weighting coefficients balance cleaning effectiveness, safety, and efficiency. Based on the product design priorities (safety > cleaning effect > user experience), and fine-tuned through offline simulation and user test data, the final parameters are solidified in the embedded program as system parameters; for example, a safety-first preset scheme is: ω1=0.5 (cleaning effect), ω2=0.4 (pressure safety), ω3=0.1 (acoustic efficiency).

[0071] By solving this optimization problem, optimal control commands are generated in real time. The data is then sent to the motor drive module. During the system development phase, calibration in the laboratory establishes the parameters based on vibration parameters (…). Based on empirical data on plaque removal effectiveness, an "optimization strategy lookup table" is pre-computed offline. This table will store the state space ( , , The control is discretized into multiple state points, and the corresponding optimal control command is calculated for each state point. During real-time operation, the system determines the status based on the currently detected state ( , , The system reads the optimal control command directly by looking up a table (such as nearest neighbor lookup) and sends it to the motor drive module.

[0072] (7) Step Seven: Embedded System Implementation and Optimization Algorithm lightweighting: The model is pruned and quantized, and converted to 8-bit integer operations to adapt to MCU resource limitations.

[0073] Real-time system development: Deploy fusion algorithms on RTOS to ensure real-time performance.

[0074] (8) Step 8: System Integration and Testing Closed-loop testing: System integration testing is performed in a simulated oral cavity environment to verify the entire closed-loop logic.

[0075] User Testing and Parameter Fine-Tuning: Conduct small-scale user testing, collect data, and fine-tune weighted parameters (such as...). (etc.) to make final fine adjustments.

[0076] In this embodiment, the attitude sensor is a ten-axis IMU, the optical sensor is a 400-800nm ​​miniature spectral sensor, the pressure sensor is a piezoresistive type, the acoustic sensor is a MEMS microphone, and the drive unit is a linear resonant motor.

[0077] Furthermore, the spectral sensor can be replaced with a near-infrared photoelectric sensor, a visible light imaging sensor, or a laser speckle sensor. If the spectral sensor fails due to foam obstruction, the system will automatically trigger an acoustic priority mode, using only the MEMS microphone to evaluate brushing motions. The linear resonant motor of the actuator can be replaced with a piezoelectric ceramic drive, a linear resonant drive, or a magnetic levitation drive module. Regardless of whether it is a rotor motor or a piezoelectric drive, the system can perform closed-loop control using universal PWM control commands.

[0078] The adaptive weighting algorithm can be replaced by a Bayesian network-based weight allocation model, a reinforcement learning (Q-Learning) control model, or an adaptive threshold model. Even in extreme dynamic noise environments, the system can still automatically learn the optimal control strategy through the reinforcement learning strategy.

[0079] The embodiments of the present invention employ the above-described solution, which has the following advantages: 1. High-precision perception and strong robustness: By adopting the strategy of "DS evidence theory + adaptive weighted fusion", the weight of each sensor is dynamically adjusted, which can effectively deal with the failure or interference of a single sensor, and the system has extremely high reliability.

[0080] 2. True closed-loop cleaning effect evaluation and control: The introduction of spectral and acoustic sensors directly evaluates the plaque removal effect and brushing action efficiency, realizing a leap from "blind operation" to "precise control based on effect".

[0081] 3. Embedded-friendly and low-power: The algorithm avoids complex calculations such as 3D reconstruction and reinforcement learning, and has been deeply optimized for the limited resources of MCUs. It supports offline operation and has low power consumption.

[0082] 4. Personalization and Adaptability: The adaptive weighting algorithm enables the system to automatically adapt to fluctuations in sensor performance and different brushing habits of users, achieving truly personalized care. The method of this invention is not limited to being executed in the chronological order described in the specification, but can also be executed in other chronological orders, in parallel, or independently. Therefore, the execution order of the method described in this specification does not constitute a limitation on the technical scope of this invention.

[0083] Although the invention has been disclosed above through the description of specific embodiments, it should be understood that all the embodiments and examples described above are exemplary and not restrictive. Those skilled in the art can design various modifications, improvements, or equivalents to the invention within the spirit and scope of the appended claims. These modifications, improvements, or equivalents should also be considered to be included within the protection scope of the invention.

Claims

1. A control method for an electric toothbrush based on multi-sensor adaptive weighted fusion, characterized in that: include: Step 1: Simultaneously sample the attitude sensor, pressure sensor, optical sensor, and acoustic sensor using the same hardware trigger signal to obtain attitude, pressure, spectral, and acoustic data; Step 2: Preprocess and extract features from the sensor data obtained in Step 1: perform low-pass filtering on the attitude data, peak detection on the pressure data, sliding window mean filtering on the spectral data, and extract Mel frequency cepstral coefficient features from the acoustic data. Step 3: Identify brushing areas and behavioral intentions based on Dempster evidence theory, assign a basic probability assignment function to the observation results of each sensor, and fuse them through Dempster combination rules to output the area confidence distribution and behavioral pattern; Step 4: Based on the adaptive weighted multi-sensor fusion state estimation algorithm, the reliability factor and context factor of each sensor are calculated using the regional confidence distribution and behavior pattern output in Step 3. Weighting coefficients are dynamically generated, and the attitude, pressure, spectral and acoustic features output in Step 2 are weighted and fused to output the comprehensive cleanliness estimate and the fused spectral features. Step 5: Calculate the plaque index based on the fused spectral features, and calculate the brushing efficiency score based on the acoustic features; Step 6: Based on the plaque index, brushing efficiency score, and pressure safety threshold, the optimal control strategy is generated by solving a multi-objective optimization problem, and the vibration frequency and amplitude of the brush head are adjusted in real time; the pressure safety threshold is a preset value.

2. The electric toothbrush control method according to claim 1, characterized in that: Step 3 specifically includes: Step 31: Define the brushing area recognition framework and behavioral type framework ={Effective scrubbing, excessive force, patting}; Step 32: Assign a basic probability assignment function to the observations of each sensor: For attitude sensors: Based on the matching degree between historical attitude data and region models, probability quality is assigned to a subset of the brushing region recognition framework; For pressure sensors: assign probabilistic mass to subsets of the behavior type framework based on pressure patterns; For acoustic sensors: Based on the characteristics of Mel frequency cepstral coefficients, assign probabilistic quality to subsets of the behavior type framework; Step 33: Use Dempster's combination rule to fuse the basic probability assignment functions of each sensor output in Step 32 to obtain the synthetic basic probability assignment function. Normalize the confidence of each hypothesis and then output the regional confidence distribution and behavior pattern label.

3. The electric toothbrush control method according to claim 2, characterized in that: Step 3 involves identifying brushing areas and behavioral intentions based on the DS evidence theory. Specifically, this identification is based on the temporally enhanced DS evidence theory, and includes: (1) Temporal evidence accumulation and decay model: Define sliding time window Wherein, the window size T is a system preset constant, determined based on the typical cycle of brushing action; for each moment within the window Calculate the statute of limitations weight of the evidence. Among them, the attenuation coefficient A preset constant is set for the system to control the rate of decay of historical evidence; its value is calibrated using experimental data. Then, the accumulated evidence at time t... for: ； in, Let τ be the degree of confidence that all sensor evidence indicates the true situation belongs to proposition A. Let τ be the degree of confidence that all sensor evidence indicates the true situation belongs to proposition B. For identification framework; (2) Dempster's combination rule: New evidence at the current time t The accumulated evidence from the previous moment combination: ； in, This represents the basic probability allocation function calculated based on the latest collected sensor data at the current time t; To determine the degree of confidence in proposition C based on the latest sensor readings at the current moment; dynamically adjust parameters. The determination is based on the degree of conflict between the current and old evidence, specifically through calculation. and The distance between Jousselme evidence And map it to the interval [0,1]: ,in, For the Sigmoid function, This is the preset scaling factor.

4. The electric toothbrush control method according to claim 1, characterized in that: Step 4 of the adaptive weighted multi-sensor fusion state estimation algorithm specifically includes: Step 41: Set time The sensor set is Its observation vector is The optimal estimated state variables after fusion are: ; in, It is a sensor At any moment The adaptive weights satisfy ; Step 42: Adaptive Weights The calculation formula is as follows: ； in, The reliability factor is calculated based on the sensor signal noise variance and historical consistency. The calculation formula is as follows: ,in To ensure that the number is a very small positive number, and to prevent division by zero, It is the noise variance; It is a contextual factor, determined by the current behavior pattern; , These are hyperparameters that control the influence strength of reliability and context factors, respectively.

5. The electric toothbrush control method according to claim 1, characterized in that: Step 5 specifically includes: Step 51: Calculate the plaque index: Obtain the reflectance spectrum of the tooth surface using an optical sensor. Using the spectral features fused in step 4, the plaque index PI is defined: ； in, This refers to the characteristic absorption band of dental plaque. For reference wavelength, the tooth itself reflects light strongly; The characteristic weighting function for dental plaque was obtained through laboratory calibration. Step 52: Calculate brushing efficiency score: Collect sound signals through microphone. The mellitus frequency cepstral coefficients are extracted; the brushing efficiency score is calculated by matching them with a pre-stored effective brushing acoustic model. .

6. The electric toothbrush control method according to claim 1, characterized in that: In step 6, the formula for the optimal control strategy is as follows: ； Constraints: f min ≤ ≤f max A min ≤ ≤A max ; in, The control variables to be optimized are the vibration frequency and amplitude of the brush head, respectively. These are the target plaque index and the current plaque index, respectively. and These are the current pressure and the safe pressure threshold, respectively. Rate the efficiency of acoustic brushing; These are weighting coefficients used to balance cleaning effectiveness, pressure safety threshold, and acoustic efficiency, respectively.

7. An electric toothbrush control system based on multi-sensor adaptive weighted fusion, characterized in that: include: The sensor module integrates attitude sensors, pressure sensors, optical sensors, and acoustic sensors. Execution module: includes drive unit and brush head; A processing module is used to execute the electric toothbrush control method as described in any one of claims 1-6.

8. The electric toothbrush control system according to claim 7, characterized in that: The attitude sensor is a ten-axis inertial measurement unit, which includes a three-axis accelerometer, a three-axis gyroscope, a three-axis magnetometer, and a barometer.

9. The electric toothbrush control system according to claim 7, characterized in that: The optical sensor is a spectral sensor, a near-infrared photoelectric sensor, a visible light imaging sensor, or a laser speckle sensor.

10. The electric toothbrush control system according to claim 7, characterized in that: The pressure sensor is a piezoresistive pressure sensor, and the acoustic sensor is a MEMS microphone.

Images