A method for continuous monitoring of attention based on smart glasses double inertial sensor

Abstract

The present application relates to a kind of attention continuous monitoring method based on smart glasses double inertia sensor, belong to wearable smart computing and digital medical cross technical field.The present application sets up double node inertia sensor at the nose support and the leg of smart glasses, synchronously acquires head and neck three-axis acceleration and angular velocity signal;Macroscopic posture and the weak tremor signal of attention related are separated by parallel signal channel realization multi-stage;Double pipeline defense mechanism based on rigid body space geometric constraint is constructed, and invalid data generated by glasses distortion, slide is removed in real time using angular velocity correlation and acceleration spatial divergence;After dynamic baseline is removed, feature is adaptively weighted and is inferred by light weight timing model, extract three core quantitative indexes, such as the proportion of continuous concentration duration, microtremor burst frequency, posture instability entropy.The present application realizes low cost, non-invasive and privacy-friendly attention continuous monitoring, and is suitable for the monitoring of human attention state and hyperkinesia tremor in natural life scene.

Description

Technical Field

[0001] This invention relates to a human cognition and attention state monitoring technology, specifically to a method for continuous attention monitoring that utilizes dual-node inertial sensors in smart glasses to achieve separation of micro-movements of the human head and neck and prevents slip interference. It belongs to the interdisciplinary fields of wearable smart computing and digital healthcare. Background Technology

[0002] Attention is a fundamental cognitive ability essential for human learning and work. For individuals with attention deficit hyperactivity disorder (ADHD), persistent attention deficit severely restricts their social adaptability.

[0003] Currently, clinical diagnosis of attention relies heavily on questionnaires and computerized testing in laboratory settings. However, questionnaire assessments are susceptible to observer bias and memory lapses; computerized testing is a static assessment, and a subject's performance in front of a screen in a clinic often fails to represent their state in real-world scenarios. This lack of universality in real-world situations leads to delays in treatment plans. Therefore, developing an objective and continuous method for monitoring attention that can be used in daily life is crucial.

[0004] In recent years, with the rise of wearable sensing technology, technicians have attempted to use EEG or eye-tracking devices to identify attentional states. However, these methods have significant limitations. EEG devices are bulky and easily affected by subtle human movements; eye-tracking technology not only faces problems of light sensitivity and high power consumption, but also has blind spots for identifying wandering attention, and is not convenient for wearing around the clock.

[0005] In contrast, inertial sensors (including accelerometers and gyroscopes) have become an ideal solution for everyday behavior sensing due to their low cost, low power consumption, and high degree of privacy protection. When users experience inattention or cognitive overload, their head and neck often exhibit specific posture changes or subtle tremors.

[0006] However, existing single-point inertial monitoring solutions face two major challenges: First, large-scale autonomous head rotations can easily mask subtle tremors, leading to the loss of key features; second, during prolonged wear, the device inevitably slips due to sweating or movement. The signal abrupt changes caused by this device slippage are highly similar to the impulsive head-turning movements of ADHD patients, resulting in a high false alarm rate.

[0007] In summary, existing attention monitoring technologies all have various shortcomings, and new methods are urgently needed to overcome their limitations. Summary of the Invention

[0008] The purpose of this invention is to overcome the technical defects of existing technologies in human attention monitoring, such as limited evaluation environment, large subjective bias, wearing discomfort, serious signal interference, and data pollution caused by easy device slippage. The invention creatively proposes a continuous attention monitoring method based on dual inertial sensors of smart glasses.

[0009] This method utilizes the unique asymmetric physical structure of smart glasses to perform collaborative sensing at the nose pad and temples. By separating large-scale posture movements from extremely weak tremor signals and establishing a rigid body spatial geometric constraint mechanism, it enables digital assessment and monitoring of human attention state, hyperactivity tremor, and macroscopic posture in natural life scenarios.

[0010] The innovations of this invention include: constructing a collaborative perception architecture that uses dual inertial sensors (accelerometer and gyroscope) in smart glasses to establish a nose pad stability benchmark and a temple lever amplification. Through algorithms, large-amplitude postural movements (such as head tilting and turning) generated by users in daily activities are separated with high fidelity from the subtle tremors characteristic of attention deficit, effectively solving the problem of large movements covering weak features. Simultaneously, utilizing the spatial geometric relationships generated by the rigid structure of the glasses frame, invalid interference data generated by non-human movements such as pushing glasses or device slippage is identified and blocked in real time. Based on this, a lightweight temporal modeling and meta-learning framework that can quickly adapt to individual differences is introduced. Personalized adaptation for new users can be achieved with only short-term data calibration, ultimately transforming complex physical signals into a quantitative indicator profile with clinical interpretability.

[0011] The objective of this invention is achieved through the following technical solutions.

[0012] A method for continuous attention monitoring based on dual inertial sensors in smart glasses includes the following steps:

[0013] Step 1: Use dual-node inertial sensors installed on the nose pads and temples of the smart glasses to synchronously collect motion signals of the head and neck.

[0014] The collected signals include triaxial linear acceleration and triaxial angular velocity data.

[0015] Step 2: Perform multi-level separation on the acquired raw mixed signal to construct two independent signal processing channels to extract macroscopic attitude signal and weak tremor signal respectively.

[0016] The multi-level signal separation is achieved using the following method:

[0017] First, in the posture processing channel, the original signal is processed by a low-pass filter to extract smooth human body large-amplitude daily posture signals, such as head turning or body swaying.

[0018] Then, in the tremor processing channel, mode decomposition and wavelet denoising algorithms are used to suppress interference caused by large-amplitude movements and extract characteristic micro-tremor signals representing attention deviance.

[0019] Step 3: Perform slip defense detection and use a dual-pipeline defense mechanism to remove invalid interference data caused by the sliding of the glasses in real time.

[0020] Specifically, the dual-line defense mechanism is implemented using the following methods:

[0021] First, regarding the deformation and twisting phenomenon of the equipment caused by force, the correlation coefficient of the angular velocity of the two sensor nodes is calculated. When the correlation is lower than the safety threshold, it is judged that twisting slip has occurred.

[0022] Secondly, to address the issue of the glasses slipping downwards, an interception judgment is made by calculating the spatial divergence of the acceleration at two nodal points. Once any of the above-mentioned abnormal slippages is detected, the data in the current time window is immediately marked as invalid and forcibly intercepted to prevent contamination of subsequent models.

[0023] Step 4: Extract multidimensional features of the signal and perform time series model inference.

[0024] First, an unsupervised clustering algorithm is used to track the dynamic physiological homeostasis baseline of individuals and calculate the features that eliminate individual differences.

[0025] Then, the nose pads and temples are adaptively evaluated and weighted. The enhanced features are then input into a lightweight temporal convolutional network to infer in real time whether the user is currently focused or distracted.

[0026] Step 5: Extract long-range digital phenotypes and conduct quantitative evaluation.

[0027] By combining the underlying physical signals with the results output by the inference model, three core evaluation indicators were calculated: the first is the percentage of sustained focus time, which represents macroscopic focus ability; the second is the frequency of micro-tremor bursts, which reflects high-frequency fluctuations in attention; and the third is the posture instability entropy, which quantifies the degree of head and neck posture disorder.

[0028] Based on three indicators, digital assessment and monitoring of human attention status, hyperactivity, tremors, and macroscopic posture can be achieved in natural life scenarios.

[0029] Beneficial effects

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] 1. This invention achieves low-cost, non-invasive, and privacy-friendly continuous attention monitoring using only two inertial sensors mounted on glasses. The invention analyzes head and neck movement signals during task performance, demonstrating that microtremor energy, postural stability, and orientation behavior characteristics can be used to identify specific behavioral patterns of attention deficit hyperactivity disorder (ADHD). The invention achieves an average end-to-end latency of only 47.1 milliseconds during monitoring, enabling relatively rapid personalized state inference.

[0032] 2. This invention achieves the extraction of weak tremor signals through specific signal separation and denoising algorithms, effectively distinguishing between specific micro-movements and autonomous large movements; and constructs a dual-pipeline defense mechanism based on rigid body spatial geometric constraints, utilizing the angular velocity correlation and acceleration spatial divergence between the two sensing nodes to intercept and eliminate invalid interference data generated by wearing slippage in real time, thereby improving the robustness and reliability of this invention.

[0033] 3. This invention explores a new digital phenotypic system based on cognitive state fluctuations and head and neck dynamics characteristics. It combines the proportion of sustained focus time, the frequency of microtremor outbreaks, and the entropy of postural instability, and introduces a meta-learning framework to analyze the heterogeneous behavioral patterns of different users, thus solving the adaptation problem under the "cold start" of new users.

[0034] 4. The attention monitoring of the present invention is effective and generalizable, achieving an average accuracy of 98.12% in a real-world experiment involving 15 subjects (including 5 confirmed patients and 10 healthy controls). Attached Figure Description

[0035] Figure 1 This is a system architecture diagram of an embodiment of the present invention;

[0036] Figure 2 This is a measured waveform diagram of multi-level signal separation and micro-flutter extraction in a complex motion scene according to an embodiment of the present invention;

[0037] Figure 3 This is a model diagram of a dual inertial sensor according to an embodiment of the present invention;

[0038] Figure 4 This is a measured diagram of the dual-pipeline wearing sliding dynamic detection and hard-blocking defense mechanism according to an embodiment of the present invention;

[0039] Figure 5 This is a comparison chart of the ablation experimental performance of the core modeling mechanism in this invention embodiment;

[0040] Figure 6 This is a graph showing the performance convergence and recovery curves of an embodiment of the present invention under fine-tuning with a very small sample size;

[0041] Figure 7This is a radar chart for multidimensional evaluation of individual digital phenotypes in an embodiment of the present invention. Detailed Implementation

[0042] The principles and features of the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0043] Figure 1 The system architecture of an embodiment of the present invention is illustrated. When a person performs daily activities in a natural setting, macroscopic head movements and subtle neck tremors caused by attention deficits are transmitted through bones and muscles. This invention utilizes the physical structure of smart glasses to continuously monitor attention by sensing the differences in these movement characteristics.

[0044] Example

[0045] A method for continuous attention monitoring based on dual inertial sensors in smart glasses includes the following steps:

[0046] Step 1: Use a dual-node inertial sensor to synchronously acquire head and neck motion signals.

[0047] The main sensor node is installed at the nose pad of the smart glasses, and the slave sensor nodes are installed at the temples behind the ears. The system synchronously collects triaxial acceleration and triaxial angular velocity data at a high sampling rate of 100Hz.

[0048] Step 2: Perform multi-level state separation on the acquired raw mixed signal to extract the macroscopic attitude signal and the weak tremor signal.

[0049] Since normal large movements can easily mask minute tremors associated with inattention, this embodiment constructs parallel posture and tremor channels for signal separation.

[0050] Step 2.1: In the attitude channel, a 5th-order Butterworth low-pass filter (cutoff frequency set to 2Hz) is used to extract the smooth macroscopic attitude signal.

[0051] Specifically, to eliminate the phase delay of the signal, a bidirectional filtering process is employed to eliminate the time delay error. The squared magnitude of the transfer function of this filter... As shown in Equation 1:

[0052] (1)

[0053] in, This is the filter order (taken as 5 in this embodiment). The corresponding cutoff angular frequency, The input signal frequency.

[0054] Step 2.2: In the flutter channel, the low-frequency attitude baseline is initially stripped using the variational mode decomposition algorithm. A bandwidth-limiting penalty factor is applied to forcibly suppress energy leakage caused by large movements, automatically separating and extracting specific flutter signals with frequencies between 2 and 5 Hz. Subsequently, the extracted signals are subjected to multi-scale decomposition and soft-threshold denoising using the db4 wavelet basis function.

[0055] Specifically, the soft threshold function is shown in Equation 2:

[0056] (2)

[0057] in, These are the wavelet coefficients after soft thresholding. These are the original wavelet coefficients. For symbolic functions, An adaptive threshold is set based on the noise variance (the threshold varies with the noise variance). Through this step, the system can accurately extract weak flutter signals with energy concentrated in the 2-5Hz frequency band.

[0058] like Figure 2 The diagram illustrates the waveform evolution process of multi-level signal separation and micro-flutter extraction in a complex motion scenario, as shown in this embodiment. Figure 2 (a) is the raw mixed angular velocity signal acquired by the sensor, containing autonomous large movements and slight jitter; after the raw signal is filtered by the attitude channel in this embodiment, as shown... Figure 2 As shown in (b), high-frequency features were removed, and a smooth macroscopic attitude signal was successfully extracted. Simultaneously, in the parallel tremor channel, by combining variational mode decomposition and wavelet lossless extraction algorithms, the interference of large-amplitude movements on the weak tremor signal was effectively blocked during periods of intense movement. Furthermore, highly concentrated waveform features were extracted during the bursts of paroxysmal microtremors, thus separating the microtremor signal, as shown in [example]. Figure 2 As shown in (c).

[0059] Step 3: Implement dual-pipeline slip defense based on spatial geometry constraints to eliminate invalid interference data caused by the sliding of glasses in real time.

[0060] like Figure 3 As shown, under normal wear, the nose pads and temples of the glasses form a relatively fixed rigid system.

[0061] Step 3.1: When the equipment is subjected to external force pushing or pulling and undergoes non-rigid body torsion, the system first locks the dominant axis with the largest absolute value of the current kinetic energy, and calculates the Pearson correlation coefficient between the two nodes on that dominant axis. :

[0062] (3)

[0063] in, Represents the mathematical expectation. , These are the angular velocity sequences for the nose pad node and the temple node, respectively. and These represent their respective means and standard deviations. This represents the standard deviation of the angular velocity sequence at the temple nodes. When the correlation coefficient... When the value is below the safety threshold of 0.75, twisting slip is determined to have occurred.

[0064] Step 3.2: When the glasses slide down the bridge of the nose due to sweating or other reasons, calculate the extreme value of the fluctuation of the difference in linear acceleration between the two nodes along the corresponding axes (i.e., the spatial divergence). (To intercept dirty data)

[0065] Specifically, the system calculates the standard deviation of the linear acceleration difference between two nodes along each of the X, Y, and Z axes in three-dimensional space, and extracts the maximum value of these three standard deviations as the spatial divergence parameter.

[0066] (4)

[0067] in, Indicates spatial divergence, and Represents the linear acceleration components of two nodes in three-dimensional space. These represent the X, Y, and Z axes in three-dimensional space, respectively.

[0068] when Exceeding the preset experience threshold Time (of which) In this embodiment, gravitational acceleration is used. Values It should be noted that the present invention is not limited to the specific values ​​mentioned above. (Other reasonable values ​​do not deviate from the protection scope of this invention), and a translational slippage event is determined to have occurred. Once a slippage anomaly is detected in either of the two pipelines, a hard block is immediately triggered, marking the data in the current time window as invalid to prevent contamination of subsequent models.

[0069] Figure 4 This document presents the experimental results of the dual-pipeline wearing slippage dynamic detection and hard-blocking defense mechanism based on rigid body spatial geometric constraints in this embodiment. It uses a real-world example of a glasses slippage event. Figure 4 (a) shows the angular velocity change curves synchronously recorded by two sensor nodes, the main node of the nose pad and the secondary node of the temple, during translational sliding; Figure 4 (b) In the corresponding pipeline, the spatial correlation of the dominant axes of the two nodes is not lower than the set safety threshold. This causes detection methods that rely solely on angular velocity to fail; however, in Figure 4 (c) In the corresponding pipeline two, the spatial divergence of the linear acceleration calculated by the system ( The system successfully broke through the 0.04g g slippage threshold, and pipeline two successfully intercepted the data and immediately triggered a hard block to remove the dirty data, thus effectively preventing slippage interference from contaminating the subsequent cognitive reasoning model.

[0070] Step 4: Extract multidimensional features and perform dynamic baseline debiasing and personalized temporal modeling inference.

[0071] Because there are significant individual differences in the neuromotor characteristics of different users, this invention introduces an adaptive adjustment mechanism.

[0072] Step 4.1: Use the mean-drift unsupervised adaptive algorithm for dynamic baseline tracking.

[0073] In practice, the system maintains a historical sliding window of a preset time length and extracts a multidimensional feature vector set within this window. By using a Gaussian kernel function to calculate the local density of feature points in space, the system adaptively tracks changes in the region where the feature data distribution is most concentrated, using this as a baseline for an individual's daily habits. .

[0074] Mean drift vector As shown in Equation 5:

[0075] (5)

[0076] in, For the extracted multidimensional feature vector set, To calculate the Gaussian kernel function for local density, This represents the total number of multidimensional feature vectors within the sliding window. After obtaining the baseline, the relative fluctuation characteristics of the feature data relative to that baseline are calculated. ,in This represents the original multidimensional feature data extracted at the current moment. This represents the dynamic baseline of daily habits tracked at the current moment.

[0077] Step 4.2: Simultaneously merge the relative features calculated from the nose pad and temples, and input them into the feature weight allocation module.

[0078] The feature weight allocation module can evaluate the reliability of two sensor data in real time based on the specific circumstances of the current action and automatically assign them corresponding importance ratios. The weight generation is shown in Equation 6:

[0079] (6)

[0080] in, This represents the attention weight matrix that the system assigns to the features from the two sensors. and These are the feature matrices for the nose pads and temples, respectively. and To learn the weights and biases.

[0081] Step 4.3: Input the weighted feature data into the time series classification model.

[0082] Time series classification models can combine the consistent behavioral characteristics of a person over a period of time for comprehensive analysis, and calculate and output the probability value of whether the person is currently in a "focused" or "distracted / impulsive" state in real time.

[0083] Step 5: Quantitative assessment and long-term digital phenotypic extraction.

[0084] After completing the real-time inference of attention state, in order to provide long-term assessment data for medical reference, the basic sensor signals are combined with the real-time calculation results of the model to extract three core quantitative evaluation indicators. The specific calculation method is as follows:

[0085] Step 5.1: Calculate the sustained attention ratio (SAR).

[0086] SAR is used to quantify a subject's ability to maintain cognitive resources and focus over a macroscopic timescale. Its calculation formula is shown in Equation 7:

[0087] (7)

[0088] in, To maintain focus on time allocation, This represents the total number of actual effective monitoring windows after removing slip interference. For the first The posterior probability of distraction / impulse smoothed by exponential moving average within a valid time window. The threshold for the mapping of the focused state. This is an indicator function.

[0089] Step 5.2: Calculate the Micro-tremor Burst Frequency (MBF).

[0090] MBF is used to quantify the intermittent bursts of high-frequency energy per unit time to identify potential subtle hyperactivity. Its calculation formula is shown in Equation 8:

[0091] (8)

[0092] in, This is an index representing the frequency of microtremor outbreaks. To effectively monitor the total number of microtremor outbreaks captured within a given time period, This represents the total net duration of actual effective monitoring.

[0093] Step 5.3: Calculate the Postural Instability Entropy (PIE).

[0094] Based on the information entropy theory, PIE quantifies the degree of head and neck posture disorder in subjects by using the probability distribution of triaxial resultant acceleration, reflecting the brain's inhibitory control over body movement. Its calculation formula is shown in Equation 9:

[0095] (9)

[0096] in, The entropy value is the attitude instability value. The total number of discretized infinitesimal elements that divide the range of the triaxial resultant acceleration amplitude distribution. The dynamic combined acceleration amplitude of the subject falls within the first The probability density of each infinitesimal interval.

[0097] This achieves objective and noise-resistant continuous monitoring of the human body's attention state.

[0098] Instance verification

[0099] To verify the performance of this invention, a prototype dual-node smart glasses was developed. The edge sensing end uses an Arduino Uno microcontroller equipped with two MPU6050 six-axis inertial sensors, with a sampling rate set to 100Hz. A total of 15 adult subjects were recruited to participate in the core experiment, including 5 clinically diagnosed patients with attention deficit hyperactivity disorder (ADHD) and 10 healthy control volunteers of similar age. Subjects wore the prototype device and completed a 15-minute standardized continuous performance test and a 20-minute natural scene reading task to obtain macroscopic and microscopic signals containing real-world motion interference, generating a total of 67,074 frames of high-dimensional temporal feature data.

[0100] Accuracy, precision, recall, and F1 score are used to evaluate the performance of an instantaneous binary classification system. Accuracy is defined as the ratio of correctly classified samples to the total number of samples. Precision is defined as the ratio of correctly predicted positive samples to the total number of samples predicted as positive. Recall is defined as the ratio of correctly predicted positive samples to the total number of positive samples. The F1 score is defined as the harmonic mean of precision and recall. A higher F1 score indicates a better balance between precision and recall, resulting in better overall classification performance.

[0101] First, the overall performance of the attention state inference model of this invention was tested. Under 15-fold cross-validation, relying on the spatial complementarity of the dual-node sensors and the algorithm's accurate extraction of core features, the average accuracy of this invention remained stable at 98.12%. More importantly, the F1 score, which is extremely sensitive to abnormal events, jumped significantly to 94.21%. Figure 5 The ablation experiment results show that, compared with the 81.17% F1 score when using only a single-node sensor basic model, the dual-node sensor architecture of the present invention improves the sensitivity and reliability of the system in capturing distracted abnormal states under complex action backgrounds.

[0102] Next, the generalization and rapid fine-tuning performance of this invention in the "cold start" scenario for new users were tested. If a generic, unpersonalized model is directly used to infer for new users, the F1 score will drop to 66.6%. To address this bottleneck, such as... Figure 6 As shown, this system introduces a meta-learning algorithm framework that supports rapid adaptation with limited data to address the issue of rapid fine-tuning and adaptation for new users. Experiments show that when a new user first wears the device, only about 3 minutes of extremely short interaction data is required for calibration, and the algorithm can quickly restore the model accuracy to 91.2% within a few gradient updates; if 5 minutes of calibration data is provided, the accuracy can further climb to 92.8%. This confirms that the present invention can generate a high-precision personalized monitoring model within a short initial calibration time.

[0103] Finally, the effectiveness of the digital phenotypic indicators extracted in this invention in distinguishing the attentional characteristics of different population groups was tested. For example... Figure 7As shown, there were highly significant differences between the two groups of subjects in the core indicators of long-term monitoring. The percentage of sustained focus time, representing macroscopic cognitive stability, remained high in the healthy group (mean 0.86), while it was significantly lower in the attention deficit hyperactivity disorder (ADHD) group (mean 0.68). Regarding the frequency of microtremor bursts, which reflects high-frequency fluctuations in attention independent of gross motor skills, the ADHD group had a higher frequency of 16.19 bursts / hour, far exceeding the 3.69 bursts / hour in the healthy group. Simultaneously, the postural instability entropy index (2.50 vs 1.06), characterizing the degree of global motor instability, also showed significant differentiation. Furthermore, the total latency of a single decision at the edge computing end was only 47.1 milliseconds. These experimental results demonstrate that this invention not only meets the real-time interactive requirements of natural and imperceptible monitoring but also quantifies and screens the specific manifestations of attention deficit disorder from easily imperceptible physical signals.

Claims

1. A method for continuous attention monitoring based on dual inertial sensors in smart glasses, characterized in that, Includes the following steps: Step 1: Use dual-node inertial sensors installed on the nose pads and temples of the smart glasses to synchronously collect motion signals of the head and neck; The collected signals include triaxial linear acceleration and triaxial angular velocity data; Step 2: Perform multi-level separation on the acquired raw mixed signal to construct two independent signal processing channels to extract macroscopic attitude signals and weak jitter signals respectively; The multi-level signal separation is achieved using the following method: In the posture processing channel, the original signal is processed by a low-pass filter to extract smooth human body large-amplitude daily posture signals. In the tremor processing channel, mode decomposition and wavelet denoising algorithms are used to suppress interference caused by large-amplitude movements and extract characteristic micro tremor signals representing attention deviance. Step 3: Perform slip defense detection and use a dual-pipeline defense mechanism to remove invalid interference data caused by the sliding of the glasses in real time; The dual-line defense mechanism is implemented using the following methods: To address the deformation and torsion phenomenon of equipment caused by force, the correlation coefficient of the angular velocities of two sensor nodes is calculated. When the correlation is lower than the safety threshold, it is determined that torsion slip has occurred. To address the issue of glasses slipping downwards, the system calculates the spatial divergence of acceleration at two nodal points to determine if the glasses are slipping. If any of these abnormal slips are detected, the data in the current time window is immediately marked as invalid and forcibly blocked. Step 4: Extract multidimensional features of the signal and perform time-series model inference; Unsupervised clustering algorithms are used to track the dynamic physiological homeostasis baseline of individuals and calculate the features that eliminate individual differences. Adaptively evaluate and assign weights to nose pads and temples, then input the enhanced features into a lightweight temporal convolutional network to infer in real time whether the user is currently focused or distracted. Step 5: Extract long-range digital phenotypic patterns and perform quantitative evaluation; By aggregating the underlying physical signals with the results output by the inference model, three core evaluation indicators are calculated: first, the percentage of sustained focus time, which represents macroscopic focus ability; second, the frequency of micro-tremor bursts, which reflects high-frequency fluctuations in attention; and third, the posture instability entropy, which quantifies the degree of head and neck posture disorder. Based on three indicators, digital assessment and monitoring of human attention status, hyperactivity, tremors, and macroscopic posture can be achieved in natural life scenarios.

2. The method for continuous attention monitoring based on dual inertial sensors in smart glasses as described in claim 1, characterized in that, In step 1, the system synchronously acquires triaxial acceleration and triaxial angular velocity data at a high-frequency sampling rate of 100Hz.

3. The method for continuous attention monitoring based on dual inertial sensors in smart glasses as described in claim 1, characterized in that, In step 2, a bidirectional filter is used to eliminate the time delay error of the signal; the squared magnitude of the transfer function of this filter... As shown in Equation 1: （1） in, Let the filter order be . The corresponding cutoff angular frequency, The input signal frequency; In the flutter channel, the low-frequency attitude baseline is initially stripped using the variational mode decomposition algorithm, and the energy leakage caused by large movements is forcibly suppressed by applying a bandwidth limiting penalty factor. Specific flutter signals with frequencies between 2 and 5 Hz are automatically separated and extracted. The extracted signals are then subjected to multi-scale decomposition and soft thresholding denoising using the db4 wavelet basis function. The soft threshold function is shown in Equation 2: （2） in, These are the wavelet coefficients after soft thresholding. These are the original wavelet coefficients. For symbolic functions, An adaptive threshold set based on noise variance.

4. The method for continuous attention monitoring based on dual inertial sensors in smart glasses as described in claim 1, characterized in that, In step 3, when the device is subjected to external force pushing or pulling and undergoes non-rigid body torsion, the system first locks the dominant axis with the largest absolute value of the current kinetic energy and calculates the Pearson correlation coefficient between the two nodes on that dominant axis. : （3） in, Represents the mathematical expectation. , These are the angular velocity sequences for the nose pad node and the temple node, respectively. and These represent their respective means and standard deviations. The standard deviation of the temple node angular velocity sequence; when the correlation coefficient When the value falls below the safety threshold, it is determined that a twisting slip has occurred; When glasses slide down the bridge of the nose due to sweating or other reasons, dirty data is intercepted by calculating the extreme value of the fluctuation of the difference in linear acceleration between the two nodes in the corresponding axis. The system calculates the standard deviation of the linear acceleration difference between two nodes along each of the X, Y, and Z axes in three-dimensional space, and extracts the maximum value of these three standard deviations as the parameter of spatial divergence. （4） in, Indicates spatial divergence, and Represents the linear acceleration components of two nodes in three-dimensional space. These represent the X, Y, and Z axes in three-dimensional space, respectively. when Exceeding the preset experience threshold When a translational slippage event occurs, it is determined that a slippage anomaly has occurred. Once a slippage anomaly is detected in either of the two pipelines, a hard block is immediately triggered, and the data in the current time window is marked as invalid to prevent contamination of subsequent models.

5. The method for continuous attention monitoring based on dual inertial sensors in smart glasses as described in claim 4, characterized in that, In step 3, the safety threshold is 0.

75.

6. The method for continuous attention monitoring based on dual inertial sensors in smart glasses as described in claim 4, characterized in that, gravitational acceleration Values .

7. The method for continuous attention monitoring based on dual inertial sensors in smart glasses as described in claim 1, characterized in that, Step 4 includes: Step 4.1: Use the mean-shift unsupervised adaptive algorithm for dynamic baseline tracking; The system maintains a historical sliding window of a preset time length and extracts a multidimensional feature vector set within the window. By using a Gaussian kernel function to calculate the local density of feature points in space, the system adaptively tracks changes in the region where the feature data is most concentrated, using this as a baseline for an individual's daily habits. ; Mean drift vector As shown in Equation 5: （5） in, For the extracted multidimensional feature vector set, To calculate the Gaussian kernel function for local density, This represents the total number of multidimensional feature vectors within the sliding window; after obtaining the baseline, the relative fluctuation characteristics of the feature data relative to that baseline are calculated. ,in This represents the original multidimensional feature data extracted at the current moment. This represents the dynamic baseline of daily habits tracked at the current moment; Step 4.2: Simultaneously merge the relative features calculated from the nose pads and temples, and input them into the feature weight allocation module; The feature weight allocation module can evaluate the reliability of the two sensor data in real time according to the specific circumstances of the current action, and automatically assign them corresponding importance ratios; the weight generation is shown in Equation 6: （6） in, This represents the attention weight matrix that the system assigns to the features from the two sensors. and These are the feature matrices for the nose pads and temples, respectively. and To learn the weights and biases; Step 4.3: Input the weighted feature data into the time series classification model; Time series classification models can combine the consistent behavioral characteristics of a person over a period of time for comprehensive analysis, and calculate and output the probability value of whether the person is currently in a "focused" or "distracted / impulsive" state in real time.

8. The method for continuous attention monitoring based on dual inertial sensors in smart glasses as described in claim 1, characterized in that, In step 5, the calculation methods for the three core quantitative evaluation indicators are as follows: SAR (Support Specific Time) is used to quantify a subject's ability to maintain cognitive resources and focus on a macro-timescale, as shown in Equation 7. （7） in, To maintain focus on time allocation, This represents the total number of actual effective monitoring windows after removing slip interference. For the first The posterior probability of distraction / impulse smoothed by exponential moving average within a valid time window. The threshold for the mapping of the focused state. For indicator functions; Microtremor burst frequency index (MBF): Used to quantify the paroxysmal burst characteristics of high-frequency energy per unit time, in order to identify potential micro-hysteresis features, as shown in Equation 8: （8） in, This is an index representing the frequency of microtremor outbreaks. To effectively monitor the total number of microtremor outbreaks captured within a given time period, This refers to the total net duration of actual effective monitoring. Postural instability entropy (PIE): Based on information entropy theory, it quantifies the degree of head and neck posture disorder in subjects by using the probability distribution of triaxial resultant acceleration, reflecting the brain's inhibitory control ability over body movement, as shown in Equation 9: （9） in, The entropy value is the attitude instability value. The total number of discretized infinitesimal elements that divide the range of the triaxial resultant acceleration amplitude distribution. The dynamic combined acceleration amplitude of the subject falls within the first The probability density of each infinitesimal interval.

Images