A multi-level permission control method based on electrocardio biological characteristics and a wearable device

Abstract

The present application belongs to the field of biometric recognition and information security technology, and relates to a multi-level permission control method based on electrocardiogram biometric features and a wearable device. The technical problems solved include: the electrocardiogram signal is disturbed by physiological state fluctuations, resulting in insufficient individual recognition stability and other problems. The technical solution includes: collecting the user's electrocardiogram signal and evaluating the signal quality; using a multi-scale residual network model containing parallel multi-scale convolution branches and residual connections to process the signal and extract an identity feature vector; comparing the feature vector with the same user's multi-physiological state reference template in the pre-stored multi-level feature template library to calculate the similarity score; according to the comparison result of the similarity score and the plurality of preset threshold values, the user is mapped to the permission level divided by the continuous interval, and each level is associated with differentiated device operation and data access permissions; in the device operation, the above process is executed in a sliding window manner, realizing identity flow monitoring and dynamic adjustment of permissions.

Description

Technical Field

[0001] This invention belongs to the field of biometric identification and information security technology, and provides a multi-level access control method and wearable device based on electrocardiogram biometrics. Background Technology

[0002] With the widespread application of biometric technology in mobile payments, smart homes, and personal healthcare devices, identity security authentication has become a core element in protecting user privacy. While traditional biometric technologies such as fingerprint, facial, and iris recognition have made significant progress in convenience, they still face inherent challenges such as ease of forgery, non-liveness detection, and sensitivity to ambient light or skin conditions. In contrast, electrocardiogram (ECG) signals, as a deep biological characteristic originating from the electrophysiological activity of the heart, possess natural liveness, concealment, and difficulty in forgery, and are considered a new generation of highly secure "biological fingerprints."

[0003] Currently, some ECG monitoring devices with identification capabilities have appeared on the market, but these technologies still have significant limitations in practical applications:

[0004] First, the robustness of feature extraction is insufficient. Most existing technologies rely on simple comparisons based on morphological features of electrocardiograms (such as the amplitude and interval of P-QRS-T waves). However, human electrocardiogram signals are highly susceptible to interference from physiological states (such as tachycardia after exercise or rhythm changes caused by emotional fluctuations) and misalignment of the device. This leads to a significant drop in recognition rate when the user is not at rest, frequently resulting in the embarrassing situation of "the user cannot unlock," making it difficult to meet the need for seamless, all-day interaction.

[0005] Secondly, the access control model is too simplistic. Traditional ECG recognition systems mostly use an "all or nothing" binary judgment logic, which can only distinguish between "successful match" and "failed recognition." In home healthcare scenarios, devices are often shared by multiple family members, or there is a need to temporarily lend them to visitors for monitoring. The existing single access logic cannot achieve differentiated data access permission management for different identities (such as the main user, family members, and visitors), making it extremely easy for the main user's personal health privacy (such as past medical history and abnormal warning records) to be leaked in scenarios where multiple people share the device.

[0006] Finally, there is a lack of dynamic interaction and adaptive update mechanisms. Existing identification schemes often preset fixed judgment thresholds, which cannot be dynamically adjusted according to environmental noise or user physiological changes, and also lack continuous tracking and early warning of unknown potential risk identities.

[0007] In summary, how to utilize advanced deep learning algorithms to improve the anti-interference capability of ECG feature extraction, and on this basis, build a dynamic hierarchical permission allocation system that can automatically distinguish between primary users, temporary authorized users, and unknown identities, has become a pressing technical challenge in the field of wearable ECG monitoring.

[0008] Existing patent publication number CN113140302A discloses a method for managing access permissions for mobile ECG devices, the mobile ECG device itself, and a storage medium. This patent proposes a method for managing access permissions for mobile ECG devices. This technology primarily involves triggering a biometric data acquisition module by acquiring preset operation commands when the device is logged in. The acquired biometric information is then used for identity verification to determine whether the current user has the permission to execute the specific operation command, thus achieving secure access control based on biometrics. This solution, to some extent, addresses the problems of easy leakage and poor security associated with traditional password logins, providing a preliminary security barrier for the operation of sensitive medical data.

[0009] However, in real-world scenarios involving long-term wearable monitoring and shared use by multiple family members, the aforementioned solution still has significant limitations. First, this solution primarily relies on a passive "command-triggered" identification mode, meaning identity verification is only initiated when a user attempts to perform a specific preset operation. This mechanism cannot achieve seamless, all-weather security protection. If the device is worn or touched by someone other than the user during login, the system struggles to proactively detect identity changes and protect privacy data in real time. Second, this technical solution exhibits a "binary" characteristic in permission allocation, only capable of granting or denying single permissions, lacking tiered permission management logic for complex social scenarios (such as primary users, trusted family members, and temporary visitors). In real-world environments where multiple people share a single device, this singular identification mode cannot meet the diverse needs for privacy isolation and functional access.

[0010] More importantly, existing technologies do not adequately consider the robustness of ECG signals in dynamic environments. Due to the highly physiological sensitivity of ECG signals, a user's movement, heart rate fluctuations, or even slight shifts in the wearing position can all cause deviations in the acquired morphological features. Current solutions lack effective physiological state compensation mechanisms, leading to significant fluctuations in recognition accuracy during non-resting states. This can easily result in situations where the primary user fails verification due to signal changes during exercise or stress. In conclusion, developing an ECG fingerprint recognition system with dynamic physiological compensation capabilities, capable of proactively sensing identity changes and accurately mapping multi-level permissions, has become a critical technical challenge that urgently needs to be addressed in this field.

[0011] The technical problem to be solved by the present invention

[0012] First, this invention aims to address the problem of extremely poor robustness in ECG signal recognition under complex and dynamic environments due to fluctuations in physiological states. Existing solutions often rely on static morphological features or higher-order statistics for identity verification, which has significant drawbacks in clinical practice: when users are in states such as exercise, anxiety, fever, or a surge in heart rate variability, the waveform of the ECG signal undergoes natural nonlinear stretching or compression, causing the preset static fingerprint template to fail and frequently resulting in false rejections such as "the primary user cannot pass verification." This invention introduces a physiological state compensation model to achieve deep decoupling between rhythm and identity features during the feature extraction stage. This ensures that the system can stably extract feature vectors with cardiac anatomy uniqueness across different heart rate zones, fundamentally overcoming the interference of physiological biases on recognition accuracy.

[0013] Secondly, this invention addresses the security monitoring blind spots and privacy leakage risks inherent in traditional "passive trigger identification" solutions. Existing technologies often employ a passive verification mode based on "preset operation commands," meaning the biometric module is only invoked when the user clicks a specific function. This implies that during continuous monitoring after device login, if the wearer changes, the system cannot detect and identify the change in real time. This invention, by constructing a continuous, unobtrusive background silent monitoring mechanism, transforms "command triggering" into "semantic awareness triggering," achieving streaming monitoring of the current wearer's identity. Once real-time features deviate from the whitelist template, the system can proactively lock sensitive data and restrict operation permissions, filling the security vulnerabilities of existing technologies during non-interactive periods.

[0014] Finally, addressing the practical needs of multi-user sharing in home healthcare scenarios, this invention solves the problem of overly simplistic and crude permission management models. Existing patents, such as those only achieving binary "yes" or "no" permission judgments, cannot distinguish between trusted family members and unfamiliar visitors. This invention constructs a tiered permission redirection system through confidence scores output by deep learning, finely classifying users into four dimensions: primary users, trusted family members, temporary visitors, and unknown strangers. It dynamically switches the device interface UI and medical data visibility based on the identity level, achieving a balance between medical privacy protection and convenient device sharing. Summary of the Invention

[0015] The purpose of this invention is to solve the problems in existing wearable ECG devices, such as poor robustness of individual identification caused by interference from physiological fluctuations in ECG signals (e.g., heart rate variability and waveform distortion caused by exercise and emotional changes), blind spots in identity supervision and privacy leakage risks during non-interactive periods caused by reliance on "command-triggered" passive verification mechanisms, and the inability to achieve tiered permission allocation and dynamic privacy protection in multi-user shared scenarios (e.g., family members, temporary visitors) due to the adoption of "admission / denial" binary logic in permission management.

[0016] To achieve the above objectives, the present invention employs the following technical means:

[0017] This invention provides a multi-level access control method based on electrocardiogram biometrics, applicable to wearable devices, comprising the following steps:

[0018] Step S1: Collect the user's electrocardiogram (ECG) signal;

[0019] Step S2: Perform signal quality assessment on the electrocardiogram signal to obtain the signal quality assessment result;

[0020] Step S3: Process the ECG signal using a multi-scale residual network model to extract the identity feature vector. The multi-scale residual network model includes parallel multi-scale convolutional branches and residual connections. The parallel multi-scale convolutional branches are used to capture the morphological features of the ECG signal at different time scales. The residual connections are used to fuse the input signal and the multi-scale feature output to suppress the interference of physiological state fluctuations on identity feature extraction.

[0021] The electrocardiogram (ECG) signal is processed using a multi-scale residual network model to extract identity feature vectors. To address the issue of decoupling physiological variability from individual uniqueness, and to resolve the pain point of traditional identification schemes failing due to nonlinear stretching of heart rate during user movement or emotional fluctuations, a specific neural network structure was designed. In implementation, underlying improvements are made by introducing parallel multi-scale convolutional branches and residual connections: the parallel multi-scale convolutional branches simulate multi-focal perspectives, simultaneously capturing the morphological features of the ECG signal at different time scales using convolutional kernels of different sizes, thus solving the problem of physiological deviation interference caused by drastic heart rate fluctuations; the residual connections deeply fuse the geometric proportion information of the original input with the multi-scale feature output through identity mapping, effectively suppressing information degradation during feature extraction and ensuring that constant components reflecting the inherent properties of the heart's anatomical structure are preserved. This improvement overcomes the technical difficulties of feature shift in complex dynamic environments, achieving the technical effect of being insensitive to physiological states and stable in identity feature recognition, significantly reducing the false rejection rate in all-weather monitoring scenarios, and realizing accurate and continuous identity verification.

[0022] Step S4: Compare the identity feature vector with multiple benchmark templates in the pre-stored multi-level feature template library, and calculate the similarity score. The multiple benchmark templates correspond to the electrocardiogram features of the same user under different physiological states.

[0023] Step S5: Based on the comparison results of the similarity score and multiple preset thresholds, map the current user to the corresponding permission level. The permission level is a series of levels divided into continuous intervals, and each level is associated with differentiated device operation permissions and data access permissions.

[0024] Step S6: During continuous device operation, steps S1 to S5 are executed cyclically in a sliding window manner to achieve streaming monitoring of user identities and dynamic adjustment of permissions.

[0025] In the above scheme, the acquisition of the user's electrocardiogram signal includes the following steps:

[0026] Step 1.1: Receive the user's analog electrocardiogram signal in real time through the analog front-end circuit integrated into the wearable device;

[0027] Step 1.2: Perform analog-to-digital conversion on the ECG analog signal at a sampling rate of 250 Hz to 500 Hz to generate the original ECG signal;

[0028] The sampling rate is configured to preserve the detailed morphological features of the QRS complex in the original ECG signal to support subsequent ECG fingerprint construction.

[0029] In the above scheme, step S2 specifically includes:

[0030] To address the challenges of morphological deception caused by transient noise generated by wearable devices in dynamic environments, and the inability of traditional evaluation methods to distinguish between physiological fluctuations and physical interference due to reliance on past information, this paper employs a bidirectional gated cyclic unit combined with a high-overlap temporal window. By utilizing bidirectional hidden layer features, it simultaneously extracts the forward causality and backward evolution of the signal, achieving high-frequency scanning. This solves the problem of misjudgment caused by the difficulty in decoupling physiological signals and environmental noise in dynamic environments, and overcomes the computational waste caused by the blind operation of high-power backend recognition algorithms on low-quality data. While ensuring high confidence in the recognition results, it effectively reduces the system's ineffective computational overhead and significantly extends the battery life of the wearable terminal.

[0031] Step 2.1: Input the ECG signal into the signal quality assessment model, which is a timing assessment model based on a bidirectional gated cyclic unit;

[0032] Step 2.2: In the signal quality assessment model, the input ECG signal is sliced ​​with a window length of 2 seconds and a step size of 0.5 seconds to generate multiple ECG segment sequences. This high-overlapping temporal window slicing design, combined with the context-aware capability of the bidirectional gating unit, constitutes the key physical implementation method for achieving seamless verification and instantaneous permission isolation in this solution.

[0033] Step 2.3: Perform time-series analysis on the multiple ECG segment sequences using the bidirectional gated loop unit, and output signal quality scores. The It represents a continuous value within the interval [0,1].

[0034] Step 2.4: According to the above The numerical value executes the corresponding signal processing strategy, including:

[0035] when At that time, the electrocardiogram signal is determined to be a high-quality signal and marked as a valid signal;

[0036] when If the ECG signal is determined to be a medium-quality signal, the system will only maintain the basic heart rate monitoring function and temporarily suspend the identity recognition process.

[0037] when If the ECG signal is deemed a poor quality signal, the current data segment is discarded, and the device interface prompts the user to adjust the wearing posture.

[0038] In the above scheme, step 3 includes the following steps:

[0039] Step 3.1: Input the preprocessed ECG segment into the multi-scale residual network model (multi-scale residual block);

[0040] Step 3.2: Using the initial feature mapping layer of the multi-scale residual network model, A one-dimensional convolutional kernel performs spatial mapping on the input ECG sequence;

[0041] Step 3.3: Process the output of step 3.2 through the multi-scale residual feature extraction layer of the multi-scale residual network model. The multi-scale residual feature extraction layer consists of cascaded multi-scale residual blocks, and each multi-scale residual block performs the following operations:

[0042] Step 3.3.1: By setting up in parallel Convolutional branching and The convolutional branches process the input signal respectively, wherein... Convolutional branches are used to capture fine morphological details of electrocardiogram signals under high heart rate conditions. Convolutional branches are used to capture the global waveform envelope of ECG signals in low heart rate states;

[0043] Step 3.3.2: [The sentence is incomplete and requires more context to translate accurately.] Convolutional branches and the The outputs of the convolutional branches are concatenated along the channel dimension;

[0044] Step 3.3.3: Add the input signal of the multi-scale residual block to the splicing result of step 3.3.2 through residual connection;

[0045] Step 3.4: Perform global average pooling dimensionality reduction on the output of Step 3.3 through the global feature compression layer of the multi-scale residual network model;

[0046] Step 3.5: Through the identity vector output layer of the multi-scale residual network model, the output of Step 3.4 is mapped to a fixed-length identity feature vector through the embedding layer.

[0047] In the above scheme, step 4 includes the following steps:

[0048] Step 4.1: During the device registration phase, ECG segments of the user under low and high heart rate conditions are collected. Feature vectors are extracted using a multi-scale residual network model, and multiple feature vectors under the same physiological state are averaged to generate a baseline template for the main user. and The reference template is stored in encrypted form in the multi-level feature template library of the local security chip;

[0049] Step 4.2: During the device verification phase, obtain the identity feature vector extracted in real time by the multi-scale residual network model. ;

[0050] Step 4.3: Calculate the above Compared with the base template of the master user in the multi-level feature template library and The cosine similarity is calculated, and the maximum value is taken as the similarity score. The calculation formula is:

[0051]

[0052] in, This represents the identity feature vector extracted in real time. This represents the baseline template for the primary user in a low heart rate resting state. This represents the baseline template for the primary user in a high heart rate state after exercise. Represents the cosine similarity function. This indicates the similarity score;

[0053] Step 4.4: When the similarity score is... At that time, an incremental update is performed on the baseline template, and the update formula is:

[0054]

[0055] in, This represents the updated baseline template. This represents the baseline template before the update. This represents the identity feature vector that meets the high-quality criteria in the current verification (i.e., the one in step 4.2). ), To smoothly update the coefficients and take values ​​of .

[0056] In the above scheme, the construction steps of the multi-level feature template library include the following steps:

[0057] Addressing the industry pain point that existing technologies using a single static template fail to recognize ECG waveforms due to nonlinear compression during high heart rate states such as exercise or anxiety, this solution abandons single-sample input and adopts a dual-state collaborative sampling mode. This mode guides users to input features at rest and after exercise, extracts vectors of different dimensions, and generates benchmark templates for resting and exercising states in the feature space. This fundamentally solves the problem of ECG waveform feature shift caused by large fluctuations in heart rate, overcomes the limitation of poor adaptability of traditional solutions in non-resting states across all scenarios, significantly reduces the false rejection rate after exercise, and ensures high security and high response speed for identity verification by relying on local secure chip encrypted storage.

[0058] Step a.1: During the initial registration phase of the device, guide the user to collect ECG segments under low heart rate conditions;

[0059] Step a.2: Guide the user to collect ECG segments under high heart rate conditions after exercise;

[0060] Step a.3: Extract features from the ECG segments under the low heart rate state using a multi-scale residual network model to generate a low heart rate feature vector;

[0061] Step a.4: Extract features from the ECG segments under the high heart rate state after exercise using a multi-scale residual network model to generate a high heart rate feature vector;

[0062] Step a.5: Averaging the low heart rate feature vector and the high heart rate feature vector to generate a resting state baseline template. and motion state reference template ;

[0063] Step a.6: The resting state reference template and motion state reference template It is stored in encrypted form on a local security chip.

[0064] In the above scheme, step 5 includes the following steps:

[0065] Step 5.1: Calculate the similarity score. The user's current permission level is determined by comparing the data with multiple preset thresholds; wherein the multiple preset thresholds include a first threshold. Second threshold and the third threshold ;

[0066] Step 5.2: Perform the corresponding access control operation according to the access level:

[0067] when When the user is identified as the primary user, the highest level of privileges are granted, allowing access to complete historical medical records, export of raw electrocardiogram data, and modification of medical warning thresholds.

[0068] when When the system is deemed to be a trusted family member, it enters the family sharing mode, allowing the user to view real-time ECG waveforms and daily health summaries, but automatically hides the main user's past medical history, sensitive diagnostic conclusions, and personal privacy settings.

[0069] when If the user is identified as a temporary visitor, the system will switch to anonymous measurement mode, displaying only instantaneous heart rate and blood oxygen saturation. The measurement data will not be recorded in the main user's health record and the local cache will be automatically cleared after the measurement is completed.

[0070] when If the user is identified as an unknown user, the core interactive interface will be locked and a "Permissions Restricted" message will be displayed.

[0071] Step 5.3: Monitor the similarity score of consecutive sampling periods during continuous equipment operation. When similarity scores from two consecutive sampling periods are detected When a mutation occurs, it is determined that the wearer has changed and a permission redirection operation is triggered. The permission redirection operation includes closing the sensitive data interface and encrypting the currently running data within 1 second.

[0072] In the above scheme, step 6 includes the following steps:

[0073] Step 6.1: During continuous operation of the device, steps S1 to S4 are executed cyclically using a sliding window mechanism, wherein the sliding window continuously acquires ECG signal segments at fixed time windows;

[0074] Step 6.2: Configure the identity verification module as a low-priority background process so that it runs in the background and does not interfere with the real-time heart rate display function in the foreground of the device;

[0075] Step 6.3: In each loop, when the signal quality evaluation result meets the preset quality conditions, execute steps S3 and S4 to obtain the current similarity score. ,in This represents the score obtained by calculating the cosine similarity between the real-time identity feature vector and the pre-stored benchmark template;

[0076] Step 6.4: Monitor the similarity score of consecutive sampling periods When similarity scores from two consecutive sampling periods are detected When a mutation occurs, it is determined that the wearer has changed;

[0077] Step 6.5: Trigger permission redirection operation, which includes closing the sensitive data access interface and encrypting the currently running data within 1 second.

[0078] The present invention also provides a wearable device for implementing multi-level access control based on electrocardiogram biometrics, including:

[0079] The analog front-end circuit is configured to acquire the user's electrocardiogram signal at a sampling rate of 250 Hz to 500 Hz;

[0080] The local security chip is configured to encrypt and store a multi-level feature template library, which contains baseline templates of the same user at rest under low heart rate and high heart rate conditions after exercise. and motion state reference template ;

[0081] The processor is connected to the analog front-end circuit and the local security chip, and is configured to execute the multi-level access control method based on electrocardiogram biometrics.

[0082] The display interface is configured to dynamically present differentiated operation interfaces and data access content based on the permission levels determined by the method.

[0083] The processor is further configured to run the identity verification process as a low-priority background process, and when it is determined that the wearer's identity has changed, to trigger a permission redirection operation, close the sensitive data access interface, and encrypt the currently running data.

[0084] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0085] 1. This invention processes electrocardiogram (ECG) signals using a multi-scale residual network model (step S3). This model includes parallel multi-scale convolutional branches and residual connection structures. The multi-scale convolutional branches capture the morphological features of the ECG signal at different time scales (e.g., a 3×1 convolutional branch focuses on fine morphological details under high heart rate conditions, and a 7×1 convolutional branch extracts the global waveform envelope under low heart rate conditions). The residual connections fuse the input signal with the multi-scale feature output, achieving deep decoupling between physiological variability and individual uniqueness, effectively suppressing fluctuations in physiological state (such as those caused by exercise or emotional changes). The system addresses the interference of heart rate variability and waveform distortion on identity feature extraction. Simultaneously, it compares the same user's baseline templates stored in a multi-level feature template library (step S4) under different physiological states such as rest and exercise. This solves the technical problem of existing technologies relying on static morphological feature matching, leading to poor user identification stability and frequent false rejections of "the user cannot pass verification" in non-resting states. It achieves stable extraction of identity feature vectors reflecting the uniqueness of cardiac anatomy even when the user is in diverse physiological scenarios, significantly improving the robustness and adaptability of the identity recognition system in dynamic environments.

[0086] 2. This invention constructs a sliding window cyclic execution mechanism (step S6) to configure the identity verification process as a low-priority background process that runs continuously. Based on the signal quality assessment results (step S2), it dynamically determines whether to trigger identity recognition processing (only when the signal quality score meets preset conditions to perform subsequent verification). This solves the technical problems of blind spots in identity supervision and privacy leakage risks during non-interactive periods caused by the passive identification mode of the "command-triggered" method in the prior art. It achieves continuous and imperceptible monitoring of the wearer's identity while ensuring the smooth operation of core functions such as real-time heart rate monitoring in the foreground. When the system detects a sudden change in the similarity score of continuous sampling periods, it can promptly trigger permission redirection operations (closing the sensitive data interface and encrypting the running data), thereby constructing an active security barrier of "authentication upon wearing and protection upon removal," effectively filling the gap in identity security monitoring during continuous device operation.

[0087] 3. This invention maps user identities precisely to consecutive permission levels such as primary user, trusted family member, temporary visitor, and unknown user by performing a tiered mapping / matching of similarity scores with multiple preset thresholds (step S5). Differentiated interface presentation logic and data access strategies are configured for each level (e.g., automatically hiding the primary user's past medical history and sensitive diagnostic conclusions for trusted family members, and enabling anonymous measurement mode and automatically clearing the cache for temporary visitors). This solves the technical problem that the traditional "admission / denial" binary logic of permission management cannot adapt to the shared needs of multiple roles in family medical scenarios. It achieves refined isolation of medical privacy data and tiered opening of functional permissions based on identity credibility. While strictly protecting the core health privacy of the primary user, it also considers the convenience of health monitoring for family members and the basic measurement needs of temporary users, achieving an organic unity of privacy security and convenient sharing.

[0088] This invention introduces an incremental update mechanism for benchmark templates (step 4.4), which performs smooth iterative updates on pre-stored benchmark templates when the similarity score meets the high-quality condition. This solves the technical problem that the matching degree between preset templates and current biometrics gradually decays due to the natural evolution of user physiological characteristics over time (such as aging and cardiac physiological remodeling). This enables the multi-level feature template library to have adaptive evolution capabilities, avoids the degradation of recognition performance caused by template solidification after long-term use of the system, and extends the effectiveness of identity authentication and the continuity of user experience throughout the device's entire life cycle.

[0089] 4. This invention organically integrates the multi-scale residual feature decoupling architecture (step S3), the polymorphic benchmark template library (step S4), the signal quality adaptive triggering mechanism (steps S2 and S6), the streaming identity monitoring process (step S6), and the gradient permission mapping logic (step S5), resulting in significant synergistic effects: the physiological invariant feature vectors extracted by the multi-scale residual network provide high-confidence input for multi-template comparison; the signal quality assessment module intelligently filters effective data segments, avoiding low-quality signals from interfering with the verification process and optimizing system resource scheduling; the streaming monitoring mechanism embeds identity verification into the continuous operation loop of the device, allowing the permission status to dynamically evolve with the wearer's identity; and the gradient permission system precisely adjusts data visibility and functional openness based on real-time similarity scores. The aforementioned technical features are not simply a combination of functions, but rather form a closed-loop enhancement system encompassing "feature extraction—quality screening—template comparison—permission mapping—dynamic feedback": when a user's physiological state changes, multi-scale feature extraction and a multi-state template library work together to maintain recognition stability; when the wearer changes, streaming monitoring and permission redirection mechanisms work together to achieve a secure response; and the incremental update mechanism continuously optimizes the template library to adapt to long-term usage needs. This deep coupling enables the system to simultaneously achieve highly robust identity authentication, real-time security monitoring, and refined privacy protection in complex real-world scenarios (such as switching between motion states or multiple users taking turns using the system). It solves the challenges of dynamic adaptability, continuous security, and scenario adaptability that are difficult to achieve with a single technical approach, resulting in an overall technological advancement where "1+1>2". Attached Figure Description

[0090] Figure 1 This is a simplified flowchart of the present invention;

[0091] Figure 2 A simplified flowchart of the signal quality assessment process;

[0092] Figure 3 This is a simplified flowchart of the process for processing the electrocardiogram signal using a multi-scale residual network model. Detailed Implementation

[0093] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0094] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0095] This invention belongs to the field of biometric identification and information security technology, specifically relating to an identity recognition method and device based on electrocardiogram (ECG) signals, and more particularly to an intelligent interactive system and method that uses a deep learning model to extract individual anatomical features from ECG data to form an "ECG fingerprint," and scores the credibility of the identified identity to achieve seamless unlocking and dynamic hierarchical permission allocation on the device.

[0096] This invention belongs to the field of biometric identification and information security technology, specifically relating to an identity recognition method and device based on electrocardiogram (ECG) signals, and more particularly to an intelligent interactive system and method that uses a deep learning model to extract individual anatomical features from ECG data to form an "ECG fingerprint," and scores the credibility of the identified identity to achieve seamless unlocking and dynamic hierarchical permission allocation on the device.

[0097] The wearable device provided by this invention is the core for achieving real-time, seamless access control. Its processing flow strictly follows embedded pipeline design. As one possible implementation method, its implementation logic includes the following steps:

[0098] Step 1: The multi-dimensional ECG signal acquisition system acquires raw ECG signals at a sampling rate of 250Hz-500Hz by integrating the analog front end of the smartwatch. High sampling rates can preserve the detailed morphology of the QRS complex, which is a key anatomical feature for constructing an "ECG fingerprint".

[0099] Step 2: Signal quality assessment and adaptive guided interaction, the flowchart of which is as follows: Figure 2 As shown;

[0100] The raw signal is first input into the signal quality assessment model. If the signal quality score is lower than the set threshold, it is determined that there is serious motion artifact or poor contact. The system automatically discards the data segment and triggers the device interface to prompt the user to adjust the wearing posture.

[0101] The signal quality assessment model is a time-series evaluation model using a bidirectional gated cyclic unit. The preprocessed ECG sequence is sliced ​​in 2-second windows with a step size of 0.5 seconds. The model outputs a continuous score SQS in the range [0,1].

[0102] An SQS ≥ 0.75 indicates a high-quality signal with a clear waveform and a steep R-wave, allowing it to enter the identity verification pipeline.

[0103] A value of 0.4 ≤ SQS < 0.75 indicates a medium-quality signal with some interference. The system will only maintain baseline heart rate monitoring and temporarily suspend identity recognition.

[0104] If SQS < 0.4, it is considered a poor signal and is determined to be invalid interference data. The system will execute the "data discard" policy and trigger user interaction guidance.

[0105] Step 3: The MSRN-based identity feature extraction model is the core module of the invention for solving the physiological deviation problem. It achieves deep decoupling of rhythm features and identity features through a multi-scale residual network.

[0106] The model transforms preprocessed ECG segments into high-order identity vectors through cascaded convolutions and residual structures:

[0107] Initial feature mapping layer: Located at the front end of the model, it uses a 1×7 Conv1D convolutional kernel to spatially map the input ECG sequence, achieving preliminary dimensionality reduction while extracting basic morphological features.

[0108] Multi-scale residual feature extraction layer: Composed of cascaded multi-scale residual blocks, it is the core sensing unit of the model.

[0109] Multi-scale residual blocks include:

[0110] Parallel Convolution Branches: Within a block, 3×1 Conv and 7×1 Conv branches are deployed in parallel, respectively responsible for capturing fine morphological details under high heart rate and global waveform envelope under low heart rate.

[0111] Feature concatenation: The outputs of convolutions at different scales are fused along the channel dimension to achieve the integration of different physiological state morphological representations in the same feature vector.

[0112] Residual accumulation: By skipping connections, the input signal is directly accumulated to the spliced ​​features, ensuring that even if the heart rate changes, the "constant component" reflecting the physical structure of the heart can be effectively preserved.

[0113] Global Feature Compression Layer: A global average pooling layer is used to reduce the dimensionality of the high-dimensional feature map, and the robustness of the model to signal temporal shift is enhanced by calculating the channel average value.

[0114] Identity Vector Output Layer: The embedding layer maps the features into a fixed-length identity vector, which serves as the final "electrocardiogram fingerprint" and enters the verification pipeline.

[0115] The specific model structure is as follows: Figure 3 As shown, the signal achieves adaptive compensation for physiological deviations within the model through the following path:

[0116] When the original fragment enters the multi-scale residual fast stream, it is automatically split into branch streams with different receptive fields. When the heart rate increases, causing waveform compression, the signal flows to the 3×1 kernel path to preserve details; when the heart rate decreases, causing waveform stretching, it flows to the 7×1 kernel path to extract the envelope, thus automatically offsetting the effects of time axis scaling at the feature level. Through deep residual connections, the rhythmic fluctuations in the signal stream are gradually decoupled by nonlinear transformation, while the anatomically unique geometric proportions are forcibly transmitted to the ends.

[0117] The generated identity vector is compared with the template library and a similarity score S is output.

[0118] Step 4: This module receives the identity feature vector extracted by the MSRN model and calculates the cosine similarity score S by comparing it with the multi-level feature template library stored in the local security chip. Based on the range in which the similarity score S falls, the system dynamically triggers four differentiated permission modes to achieve precise isolation of medical privacy.

[0119] During the initial "registration phase" of the device, the system does not simply collect an electrocardiogram (ECG) segment, but rather guides and captures samples from the user under different physiological states.

[0120] The system captures ECG segments of users in two typical scenarios: low heart rate and high heart rate after exercise. It then uses an MSRN model to extract features from these scenarios and generates two initial baseline templates through feature averaging. These two templates are stored in encrypted form on a local secure chip.

[0121] During verification, the system calculates real-time characteristics. With two baseline templates in the library and The maximum cosine similarity is calculated using the following formula:

[0122]

[0123] in, This represents the identity feature vector extracted in real time. This represents the baseline template for the primary user in a low heart rate resting state. This represents the baseline template for the primary user in a high heart rate state after exercise. Represents the cosine similarity function. This indicates the similarity score;

[0124] To address feature drift caused by aging or cardiac remodeling, the system introduces incremental update logic. When the real-time verification score S > 0.98, the currently collected features are considered to be the user's latest, high-quality state. The system does not replace the original template but instead performs smooth iterations at a fixed ratio, as shown in the following formula:

[0125]

[0126] in The value is set to 0.01 to ensure that the template library can evolve imperceptibly with the user's physiological changes, thus avoiding the problem of decreased recognition rate after prolonged wear.

[0127] Step S5: Based on the comparison results of the similarity score and multiple preset thresholds, map the current user to the corresponding permission level. The permission level is a series of levels divided into continuous intervals, and each level is associated with differentiated device operation permissions and data access permissions.

[0128] The system has set three key thresholds: , , This accurately maps user identities to the following four dimensions:

[0129] main user The device has been identified as the legitimate owner, and the highest level of access has been granted. This includes access to complete historical medical records, exporting raw electrocardiogram data, and modifying medical alert thresholds.

[0130] Trust the family Users identified as long-term users on the whitelist, such as spouses and children, can enter "Family Sharing Mode." Viewing real-time ECG waveforms and daily health summaries is allowed, but the primary user's medical history, sensitive diagnostic conclusions, and personal privacy settings are automatically hidden.

[0131] Temporary visitors If identified as a registered temporary user, the system switches to "anonymous measurement mode." Only instantaneous heart rate and blood oxygen saturation are displayed; data is not recorded in the primary user's health profile, and local cache is automatically cleared after the measurement.

[0132] Unknown user If identified as an unauthorized operator or a completely unregistered individual, the core user interface is locked, and a "Restricted Access" pop-up is displayed.

[0133] The feature library is logically isolated according to the permission dimension in its storage structure:

[0134] Master Template: Stores the polymorphic baseline template of the master user, has the highest priority, and serves as the trigger. The sole criterion for the threshold.

[0135] Sub-template: Stores trusted family member characteristics authorized by the primary user (only single-state sampling is required), used for driving... Permission range.

[0136] Blacklist Cache: Temporary storage identification failed. Anomaly snapshots are taken for subsequent security audits and risk warnings.

[0137] Step S6: During continuous device operation, steps S1 to S5 are executed cyclically in a sliding window manner to achieve streaming monitoring of user identities and dynamic adjustment of permissions.

[0138] Unlike existing technologies that use intermittent verification by only invoking the recognition module when a specific function is triggered, this solution employs a silent verification mechanism based on a sliding window. The system continuously acquires ECG feature segments of the current wearer in the background at fixed time windows. User identity verification runs as a low-priority background process, resident and not interfering with the user's real-time heart rate display. When the model determines that the current signal quality is better than a threshold, the system automatically completes the identity comparison. If the score S remains within the primary user range, the current high-level privileges are maintained, and the user does not need to perform any manual intervention.

[0139] If the similarity score S changes abruptly between two consecutive sampling periods during continuous wear, the system determines that the wearer has changed. At this point, the permission redirection engine will intervene within one second, closing the sensitive medical record interface and encrypting the currently running data. This "second-level response" capability ensures that even if the device is illegally accessed by others while logged in, there will be no substantial leakage of privacy data.

[0140] Example 1

[0141] This invention provides a multi-level access control method based on electrocardiogram biometrics, applicable to wearable devices, comprising the following steps:

[0142] Step S1: Collect the user's electrocardiogram (ECG) signal;

[0143] Step S2: Perform signal quality assessment on the electrocardiogram signal to obtain the signal quality assessment result;

[0144] Step S3: Process the ECG signal using a multi-scale residual network model to extract the identity feature vector. The multi-scale residual network model includes parallel multi-scale convolutional branches and residual connections. The parallel multi-scale convolutional branches are used to capture the morphological features of the ECG signal at different time scales. The residual connections are used to fuse the input signal and the multi-scale feature output to suppress the interference of physiological state fluctuations on identity feature extraction.

[0145] Step S4: Compare the identity feature vector with multiple benchmark templates in the pre-stored multi-level feature template library, and calculate the similarity score. The multiple benchmark templates correspond to the electrocardiogram features of the same user under different physiological states.

[0146] Step S5: Based on the comparison results of the similarity score and multiple preset thresholds, map the current user to the corresponding permission level. The permission level is a series of levels divided into continuous intervals, and each level is associated with differentiated device operation permissions and data access permissions.

[0147] Step S6: During continuous device operation, steps S1 to S5 are executed cyclically in a sliding window manner to achieve streaming monitoring of user identities and dynamic adjustment of permissions.

[0148] In the above scheme, the acquisition of the user's electrocardiogram signal includes the following steps:

[0149] Step 1.1: Receive the user's analog electrocardiogram signal in real time through the analog front-end circuit integrated into the wearable device;

[0150] Step 1.2: Perform analog-to-digital conversion on the ECG analog signal at a sampling rate of 250 Hz to 500 Hz to generate the original ECG signal;

[0151] The sampling rate is configured to preserve the detailed morphological features of the QRS complex in the original ECG signal to support subsequent ECG fingerprint construction.

[0152] In the above scheme, step S2 specifically includes:

[0153] To address the challenges of morphological deception caused by transient noise generated by wearable devices in dynamic environments, and the inability of traditional evaluation methods to distinguish between physiological fluctuations and physical interference due to reliance on past information, this paper employs a bidirectional gated cyclic unit combined with a high-overlap temporal window. By utilizing bidirectional hidden layer features, it simultaneously extracts the forward causality and backward evolution of the signal, achieving high-frequency scanning. This solves the problem of misjudgment caused by the difficulty in decoupling physiological signals and environmental noise in dynamic environments, and overcomes the computational waste caused by the blind operation of high-power backend recognition algorithms on low-quality data. While ensuring high confidence in the recognition results, it effectively reduces the system's ineffective computational overhead and significantly extends the battery life of the wearable terminal.

[0154] Step 2.1: Input the ECG signal into the signal quality assessment model, which is a timing assessment model based on a bidirectional gated cyclic unit;

[0155] Step 2.2: In the signal quality assessment model, the input ECG signal is sliced ​​with a window length of 2 seconds and a step size of 0.5 seconds to generate multiple ECG segment sequences. This high-overlapping temporal window slicing design, combined with the context awareness capability of the bidirectional gating unit, constitutes the key physical implementation method for achieving contactless verification and instantaneous permission isolation in this scheme.

[0156] Step 2.3: Perform time-series analysis on the multiple ECG segment sequences using the bidirectional gated loop unit, and output signal quality scores. The It represents a continuous value within the interval [0,1].

[0157] Step 2.4: According to the above The numerical value executes the corresponding signal processing strategy, including:

[0158] when At that time, the electrocardiogram signal is determined to be a high-quality signal and marked as a valid signal;

[0159] when If the ECG signal is determined to be a medium-quality signal, the system will only maintain the basic heart rate monitoring function and temporarily suspend the identity recognition process.

[0160] when If the ECG signal is deemed a poor quality signal, the current data segment is discarded, and the device interface prompts the user to adjust the wearing posture.

[0161] In the above scheme, step 3 includes the following steps:

[0162] Step 3.1: Input the preprocessed ECG segment into the multi-scale residual network model (multi-scale residual block);

[0163] Step 3.2: Using the initial feature mapping layer of the multi-scale residual network model, A one-dimensional convolutional kernel performs spatial mapping on the input ECG sequence;

[0164] Step 3.3: Process the output of step 3.2 through the multi-scale residual feature extraction layer of the multi-scale residual network model. The multi-scale residual feature extraction layer consists of cascaded multi-scale residual blocks, and each multi-scale residual block performs the following operations:

[0165] Step 3.3.1: By setting up in parallel Convolutional branching and The convolutional branches process the input signal respectively, wherein... Convolutional branches are used to capture fine morphological details of electrocardiogram signals under high heart rate conditions. Convolutional branches are used to capture the global waveform envelope of ECG signals in low heart rate states;

[0166] Step 3.3.2: [The sentence is incomplete and requires more context to translate accurately.] Convolutional branches and the The outputs of the convolutional branches are concatenated along the channel dimension;

[0167] Step 3.3.3: Add the input signal of the multi-scale residual block to the splicing result of step 3.3.2 through residual connection;

[0168] Step 3.4: Perform global average pooling dimensionality reduction on the output of Step 3.3 through the global feature compression layer of the multi-scale residual network model;

[0169] Step 3.5: Through the identity vector output layer of the multi-scale residual network model, the output of Step 3.4 is mapped to a fixed-length identity feature vector through the embedding layer.

[0170] In the above scheme, step 4 includes the following steps:

[0171] Step 4.1: During the device registration phase, ECG segments of the user under low and high heart rate conditions are collected. Feature vectors are extracted using a multi-scale residual network model, and multiple feature vectors under the same physiological state are averaged to generate a baseline template for the main user. and The reference template is stored in encrypted form in the multi-level feature template library of the local security chip;

[0172] Step 4.2: During the device verification phase, obtain the identity feature vector extracted in real time by the multi-scale residual network model. ;

[0173] Step 4.3: Calculate the above Compared with the base template of the master user in the multi-level feature template library and The cosine similarity is calculated, and the maximum value is taken as the similarity score. The calculation formula is:

[0174]

[0175] in, This represents the identity feature vector extracted in real time. This represents the baseline template for the primary user in a low heart rate resting state. This represents the baseline template for the primary user in a high heart rate state after exercise. Represents the cosine similarity function. This indicates the similarity score;

[0176] Step 4.4: When the similarity score is... At that time, an incremental update is performed on the baseline template, and the update formula is:

[0177]

[0178] in, This represents the updated baseline template. This represents the baseline template before the update. This represents the identity feature vector that meets the high-quality criteria in the current verification (i.e., the one in step 4.2). ), To smoothly update the coefficients and take values ​​of .

[0179] In the above scheme, the construction steps of the multi-level feature template library include the following steps:

[0180] Addressing the industry pain point that existing technologies using a single static template fail to recognize ECG waveforms due to nonlinear compression during high heart rate states such as exercise or anxiety, this solution abandons single-sample input and adopts a dual-state collaborative sampling mode. This mode guides users to input features at rest and after exercise, extracts vectors of different dimensions, and generates benchmark templates for resting and exercising states in the feature space. This fundamentally solves the problem of ECG waveform feature shift caused by large fluctuations in heart rate, overcomes the limitation of poor adaptability of traditional solutions in non-resting states across all scenarios, significantly reduces the false rejection rate after exercise, and ensures high security and high response speed for identity verification by relying on local secure chip encrypted storage.

[0181] Step a.1: During the initial registration phase of the device, guide the user to collect ECG segments under low heart rate conditions;

[0182] Step a.2: Guide the user to collect ECG segments under high heart rate conditions after exercise;

[0183] Step a.3: Extract features from the ECG segments under the low heart rate state using a multi-scale residual network model to generate a low heart rate feature vector;

[0184] Step a.4: Extract features from the ECG segments under the high heart rate state after exercise using a multi-scale residual network model to generate a high heart rate feature vector;

[0185] Step a.5: Averaging the low heart rate feature vector and the high heart rate feature vector to generate a resting state baseline template. and motion state reference template ;

[0186] Step a.6: The resting state reference template and motion state reference template It is stored in encrypted form on a local security chip.

[0187] In the above scheme, step 5 includes the following steps:

[0188] Step 5.1: Calculate the similarity score. The user's current permission level is determined by comparing the data with multiple preset thresholds; wherein the multiple preset thresholds include a first threshold. Second threshold and the third threshold ;

[0189] Step 5.2: Perform the corresponding access control operation according to the access level:

[0190] when When the user is identified as the primary user, the highest level of privileges are granted, allowing access to complete historical medical records, export of raw electrocardiogram data, and modification of medical warning thresholds.

[0191] when When the system is deemed to be a trusted family member, it enters the family sharing mode, allowing the user to view real-time ECG waveforms and daily health summaries, but automatically hides the main user's past medical history, sensitive diagnostic conclusions, and personal privacy settings.

[0192] when If the user is identified as a temporary visitor, the system will switch to anonymous measurement mode, displaying only instantaneous heart rate and blood oxygen saturation. The measurement data will not be recorded in the main user's health record and the local cache will be automatically cleared after the measurement is completed.

[0193] when If the user is identified as an unknown user, the core interactive interface will be locked and a "Permissions Restricted" message will be displayed.

[0194] Step 5.3: Monitor the similarity score of consecutive sampling periods during continuous equipment operation. When similarity scores from two consecutive sampling periods are detected When a mutation occurs, it is determined that the wearer has changed and a permission redirection operation is triggered. The permission redirection operation includes closing the sensitive data interface and encrypting the currently running data within 1 second.

[0195] In the above scheme, step 6 includes the following steps:

[0196] Step 6.1: During continuous operation of the device, steps S1 to S4 are executed cyclically using a sliding window mechanism, wherein the sliding window continuously acquires ECG signal segments at fixed time windows;

[0197] Step 6.2: Configure the identity verification module as a low-priority background process so that it runs in the background and does not interfere with the real-time heart rate display function in the foreground of the device;

[0198] Step 6.3: In each loop, when the signal quality evaluation result meets the preset quality conditions, execute steps S3 and S4 to obtain the current similarity score. ,in This represents the score obtained by calculating the cosine similarity between the real-time identity feature vector and the pre-stored benchmark template;

[0199] Step 6.4: Monitor the similarity score of consecutive sampling periods When similarity scores from two consecutive sampling periods are detected When a mutation occurs, it is determined that the wearer has changed;

[0200] Step 6.5: Trigger permission redirection operation, which includes closing the sensitive data access interface and encrypting the currently running data within 1 second.

[0201] Example 2

[0202] The present invention also provides a wearable device for implementing multi-level access control based on electrocardiogram biometrics, including:

[0203] The analog front-end circuit is configured to acquire the user's electrocardiogram signal at a sampling rate of 250 Hz to 500 Hz;

[0204] The local security chip is configured to encrypt and store a multi-level feature template library, which contains baseline templates of the same user at rest under low heart rate and high heart rate conditions after exercise. and motion state reference template ;

[0205] The processor is connected to the analog front-end circuit and the local security chip, and is configured to execute the multi-level access control method based on electrocardiogram biometrics.

[0206] The display interface is configured to dynamically present differentiated operation interfaces and data access content based on the permission levels determined by the method.

[0207] The processor is further configured to run the identity verification process as a low-priority background process, and when it is determined that the wearer's identity has changed, to trigger a permission redirection operation, close the sensitive data access interface, and encrypt the currently running data.

[0208] Example 3

[0209] As one possible implementation, a wearable device for implementing multi-level access control based on electrocardiogram biometrics includes:

[0210] The analog front-end circuit is configured to acquire the user's electrocardiogram signal at a sampling rate of 250 Hz to 500 Hz;

[0211] The signal quality assessment module is configured to slice the ECG signal with a window length of 2 seconds and a step size of 0.5 seconds, and output a signal quality score through a time-series assessment model based on a bidirectional gated cyclic unit. The For the range of values ​​within Continuous values ​​within;

[0212] The identity feature extraction module is configured to process the electrocardiogram signal using a multi-scale residual network model to extract identity feature vectors. The multi-scale residual network model includes an initial feature mapping layer, a multi-scale residual feature extraction layer composed of cascaded multi-scale residual blocks, a global feature compression layer, and an identity vector output layer. Each multi-scale residual block contains parallel first and second convolutional branches. The first convolutional branch uses a kernel size of [missing information]. The one-dimensional convolution operation is used to capture fine morphological details under high heart rate conditions, and the second convolution branch uses a convolution kernel size of... One-dimensional convolution operations are used to capture the global waveform envelope under low heart rate conditions, and the input features are accumulated with the concatenated multi-scale features through residual connections;

[0213] A secure storage chip is configured to store a multi-level feature template library in encrypted form, the multi-level feature template library containing resting state baseline templates generated by the same user in low heart rate scenarios. and the motion state baseline template generated in high heart rate scenarios after exercise. ;

[0214] The template comparison module is configured to compare the identity feature vector respectively with the above and Perform cosine similarity calculation and take the maximum value as the similarity score. ;

[0215] The permission mapping module is configured to map the similarity score. With preset threshold , , The system compares the permissions of the current user with those of the primary user, trusted family member, temporary visitor, or unknown user, and implements differentiated interface display and data access control policies for different permission levels.

[0216] The streaming monitoring and control module is configured to treat the identity verification process as a low-priority background process, periodically triggering the cyclic execution of the analog front-end circuit, signal quality assessment module, identity feature extraction module, and template comparison module in a sliding window manner, and determining the similarity score between two consecutive sampling periods. When a mutation occurs, a permission redirection operation is triggered within 1 second to close the sensitive data access interface and encrypt the currently running data;

[0217] in, This indicates the signal quality score; the higher the value, the lower the degree of interference from motion artifacts or poor contact in the ECG signal. Represents a real-time identity feature vector; Represents a baseline template for the resting state; Represents a reference template for motion states; The similarity score represents the degree of matching between the current user's identity features and the pre-stored template.

[0218] This invention has the following characteristics:

[0219] First, it possesses extremely strong robustness to physiological states and deep feature extraction capabilities. This invention abandons the crude logic of traditional patents that rely on matching single morphological features, instead employing a high-order feature extraction mechanism integrating multi-scale residual networks (multi-scale residual blocks). This mechanism simultaneously captures microscopic notches and macroscopic envelopes in ECG signals through parallel convolutional paths, enabling the system to automatically filter out physiological deviation noise caused by heart rate fluctuations and changes in exercise state. This means that whether the user is at rest, jogging, or even experiencing a high heart rate due to anxiety, the system can accurately extract the unique "ECG fingerprint" reflecting the heart's anatomical structure, achieving truly stable all-weather, multi-state recognition and significantly reducing the risk of "false rejection after exercise" common in existing technologies.

[0220] Secondly, it achieves a shift in security mode from "passive triggering" to "active continuous monitoring." This invention establishes a seamless, silent background verification mechanism, ensuring that the system only initiates high-energy-consuming verification when signal quality meets standards. This extends device battery life and fills security vulnerabilities during non-interactive periods. Once a shift in the wearer's identity characteristics is detected, the system can respond within seconds and automatically switch permission status.

[0221] Finally, a four-dimensional tiered permission decision-making system suitable for social scenarios was constructed. This invention breaks through the binary permission limitation of existing technologies, which is "black and white," and proposes for the first time a four-dimensional mapping model based on similarity score intervals, namely "main user, trusted family member, temporary visitor, and unknown user." This refined design precisely addresses the pain point of multiple family members sharing medical devices: by dynamically adjusting the UI interface, hiding medical history records, and isolating measurement data, the system protects the core privacy of the main user while taking into account the convenience of health monitoring for family members and the functional needs of temporary users. This tiered authorization logic, combined with the incremental self-evolution mechanism of the feature library, transforms the device from a static electronic tool into an intelligent health management terminal that can continuously adapt to the user's physiological evolution and has a high awareness of privacy boundaries.

Claims

1. A multi-level access control method based on electrocardiogram biometrics, applied to wearable devices, characterized in that, Includes the following steps: Step S1: Collect the user's electrocardiogram (ECG) signal; Step S2: Perform signal quality assessment on the electrocardiogram signal to obtain the signal quality assessment result; Step S3: Process the ECG signal using a multi-scale residual network model to extract the identity feature vector. The multi-scale residual network model includes parallel multi-scale convolutional branches and residual connections. The parallel multi-scale convolutional branches are used to capture the morphological features of the ECG signal at different time scales. The residual connections are used to fuse the input signal and the multi-scale feature output to suppress the interference of physiological state fluctuations on identity feature extraction. Step S4: Compare the identity feature vector with multiple benchmark templates in the pre-stored multi-level feature template library, and calculate the similarity score. The multiple benchmark templates correspond to the electrocardiogram features of the same user under different physiological states. Step S5: Based on the comparison results of the similarity score and multiple preset thresholds, map the current user to the corresponding permission level. The permission level is a series of levels divided into continuous intervals, and each level is associated with differentiated device operation permissions and data access permissions. Step S6: During continuous device operation, steps S1 to S5 are executed cyclically in a sliding window manner to achieve streaming monitoring of user identities and dynamic adjustment of permissions.

2. The method according to claim 1, characterized in that, The process of collecting the user's electrocardiogram (ECG) signal includes the following sub-steps: Step 1.1: Receive the user's analog electrocardiogram signal in real time through the analog front-end circuit integrated into the wearable device; Step 1.2: Perform analog-to-digital conversion on the ECG analog signal at a sampling rate of 250 Hz to 500 Hz to generate the original ECG signal; The sampling rate is configured to preserve the detailed morphological features of the QRS complex in the original ECG signal to support subsequent ECG fingerprint construction.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: Step 2.1: Input the ECG signal into the signal quality assessment model, which is a timing assessment model based on a bidirectional gated cyclic unit; Step 2.2: In the signal quality assessment model, the input ECG signal is sliced ​​with a window length of 2 seconds and a step size of 0.5 seconds to generate multiple ECG segment sequences; Step 2.3: Perform time-series analysis on the multiple ECG segment sequences using the bidirectional gated loop unit, and output signal quality scores. The It represents a continuous value within the interval [0,1]. Step 2.4: According to the above The numerical value executes the corresponding signal processing strategy, including: when At that time, the electrocardiogram signal is determined to be a high-quality signal and marked as a valid signal; when If the ECG signal is determined to be a medium-quality signal, the system will only maintain the basic heart rate monitoring function and temporarily suspend the identity recognition process. when If the ECG signal is deemed a poor quality signal, the current data segment is discarded, and the device interface prompts the user to adjust the wearing posture.

4. The method according to claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Input the preprocessed ECG segments into the multi-scale residual network model; Step 3.2: Using the initial feature mapping layer of the multi-scale residual network model, A one-dimensional convolutional kernel performs spatial mapping on the input ECG sequence; Step 3.3: Process the output of step 3.2 through the multi-scale residual feature extraction layer of the multi-scale residual network model. The multi-scale residual feature extraction layer consists of cascaded multi-scale residual blocks, and each multi-scale residual block performs the following operations: Step 3.3.1: By setting up in parallel Convolutional branching and The convolutional branches process the input signal respectively, wherein... Convolutional branches are used to capture fine morphological details of electrocardiogram signals under high heart rate conditions. Convolutional branches are used to capture the global waveform envelope of ECG signals in low heart rate states; Step 3.3.2: [The sentence is incomplete and requires more context to translate accurately.] Convolutional branches and the The outputs of the convolutional branches are concatenated along the channel dimension; Step 3.3.3: Add the input signal of the multi-scale residual block to the splicing result of step 3.3.2 through residual connection; Step 3.4: Perform global average pooling dimensionality reduction on the output of Step 3.3 through the global feature compression layer of the multi-scale residual network model; Step 3.5: Through the identity vector output layer of the multi-scale residual network model, the output of Step 3.4 is mapped to a fixed-length identity feature vector through the embedding layer.

5. The method according to claim 1, characterized in that, Step 4 includes the following steps: Step 4.1: During the device registration phase, ECG segments of the user under low and high heart rate conditions are collected. Feature vectors are extracted using a multi-scale residual network model, and multiple feature vectors under the same physiological state are averaged to generate a baseline template for the main user. and The reference template is stored in encrypted form in the multi-level feature template library of the local security chip; Step 4.2: During the device verification phase, obtain the identity feature vector extracted in real time by the multi-scale residual network model. ; Step 4.3: Calculate the above Compared with the base template of the master user in the multi-level feature template library and The cosine similarity is calculated, and the maximum value is taken as the similarity score. The calculation formula is: in, This represents the identity feature vector extracted in real time. This represents the baseline template for the primary user in a low heart rate resting state. This represents the baseline template for the primary user in a high heart rate state after exercise. Represents the cosine similarity function. This indicates the similarity score; Step 4.4: When the similarity score is... At that time, an incremental update is performed on the baseline template, and the update formula is: in, This represents the updated baseline template. This represents the baseline template before the update. This represents the identity feature vector that meets the high-quality criteria in the current verification (i.e., the one in step 4.2). ), To smoothly update the coefficients and take values ​​of .

6. The method according to claim 4, characterized in that, The construction steps of the multi-level feature template library include the following steps: Step a.1: During the initial registration phase of the device, guide the user to collect ECG segments under low heart rate conditions; Step a.2: Guide the user to collect ECG segments under high heart rate conditions after exercise; Step a.3: Extract features from the ECG segments under the low heart rate state using a multi-scale residual network model to generate a low heart rate feature vector; Step a.4: Extract features from the ECG segments under the high heart rate state after exercise using a multi-scale residual network model to generate a high heart rate feature vector; Step a.5: Averaging the low heart rate feature vector and the high heart rate feature vector to generate a resting state baseline template. and motion state reference template ; Step a.6: The resting state reference template and motion state reference template It is stored in encrypted form on a local security chip.

7. The method according to claim 1, characterized in that, Step 5 includes the following steps: Step 5.1: Calculate the similarity score. The user's current permission level is determined by comparing the data with multiple preset thresholds; wherein the multiple preset thresholds include a first threshold. Second threshold and the third threshold ; Step 5.2: Perform the corresponding access control operation according to the access level: when When the user is identified as the primary user, the highest level of privileges are granted, allowing access to complete historical medical records, export of raw electrocardiogram data, and modification of medical warning thresholds. when When the system is deemed to be a trusted family member, it enters the family sharing mode, allowing the user to view real-time ECG waveforms and daily health summaries, but automatically hides the main user's past medical history, sensitive diagnostic conclusions, and personal privacy settings. when If the user is identified as a temporary visitor, the system will switch to anonymous measurement mode, displaying only instantaneous heart rate and blood oxygen saturation. The measurement data will not be recorded in the main user's health record and the local cache will be automatically cleared after the measurement is completed. when If the user is identified as an unknown user, the core interactive interface will be locked and a "Restricted Permissions" message will be displayed. Step 5.3: Monitor the similarity score of consecutive sampling periods during continuous equipment operation. When similarity scores from two consecutive sampling periods are detected When a mutation occurs, it is determined that the wearer has changed and a permission redirection operation is triggered. The permission redirection operation includes closing the sensitive data interface and encrypting the currently running data.

8. The method according to claim 1, characterized in that, Step 6 includes the following steps: Step 6.1: During continuous operation of the device, steps S1 to S4 are executed cyclically using a sliding window mechanism, wherein the sliding window continuously acquires ECG signal segments at fixed time windows; Step 6.2: Configure the identity verification module as a low-priority background process so that it runs in the background and does not interfere with the real-time heart rate display function in the foreground of the device; Step 6.3: In each loop, when the signal quality evaluation result meets the preset quality conditions, execute steps S3 and S4 to obtain the current similarity score. ,in This represents the score obtained by calculating the cosine similarity between the real-time identity feature vector and the pre-stored benchmark template; Step 6.4: Monitor the similarity score of consecutive sampling periods When similarity scores from two consecutive sampling periods are detected When a mutation occurs, it is determined that the wearer has changed; Step 6.5: Trigger permission redirection operation, which includes closing the sensitive data access interface and encrypting the currently running data within 1 second.

9. A wearable device for implementing multi-level access control based on electrocardiogram biometrics, characterized in that, include: The analog front-end circuit is configured to acquire the user's electrocardiogram signal at a sampling rate of 250 Hz to 500 Hz; The local security chip is configured to encrypt and store a multi-level feature template library, which contains baseline templates of the same user at rest under low heart rate and high heart rate conditions after exercise. and motion state reference template ; The processor, connected to the analog front-end circuit and the local security chip, is configured to execute the multi-level access control method based on electrocardiogram biometrics as described in any one of claims 1 to 8; The display interface is configured to dynamically present differentiated operation interfaces and data access content based on the permission levels determined by the method. The processor is further configured to run the identity verification process as a low-priority background process, and when it is determined that the wearer's identity has changed, to trigger a permission redirection operation, close the sensitive data access interface, and encrypt the currently running data.

Images