Identity recognition method based on multi-part photoelectric volume pulse wave signals

Abstract

The invention discloses an identity recognition method based on multi-part photoelectric volume pulse wave signal feature combination. The identity recognition method is particularly suitable for dynamic environment identity recognition in different physiological states. According to the method, PPG signals are collected from six body parts (the left neck, the right neck, the left brachial, the right brachial, the left radial artery and the right radial artery), three physiological states of calmness, movement and concentration are covered, a model is trained through mixed state data, and performance is tested through independent state data. Signal processing comprises low-pass filtering and baseline drift correction and display, multi-part feature fusion significantly improves the recognition precision, and especially 100% accuracy is achieved under the combination of four parts of the left neck, the right neck, the left brachial part and the right brachial part. The invention provides a method for realizing high-precision identity recognition through multi-part photoelectric volume pulse wave signal feature combination under a multi-state test condition, and the method is suitable for high-precision identity recognition in a dynamic environment and has the application potential of popularization in the field of safety authentication of wearable equipment.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal processing, specifically relating to an identity recognition method based on multi-site photoplethysmography pulse wave signals. Background Technology

[0002] The rapid development of the Internet of Things (IoT) and smart home technologies has brought unprecedented convenience to daily life. However, at the same time, the increasing prevalence of identity spoofing and security threats poses a serious challenge to personal privacy and data security. Traditional identity verification methods, such as user ID-based systems or verification mechanisms relying on passwords and smart cards, are easily cracked, forged, or lost, making it difficult to meet the high security and convenience requirements of modern intelligent systems. Therefore, developing low-cost, non-invasive, and highly robust identity verification technologies has become a research hotspot. In recent years, biometric technologies have demonstrated superior security advantages due to their identification mechanisms based on unique individual physiological characteristics. Among them, photoplethysmography (PPG) technology has attracted widespread attention due to its uniqueness and low cost.

[0003] Compared to traditional biometric methods (such as fingerprint or facial recognition), PPG offers superior anti-counterfeiting capabilities. As a dynamic biosignal, it is difficult to steal or copy. Furthermore, PPG devices require only simple LEDs and photodiodes to achieve multi-site data collection (such as fingertips, neck, or earlobes), significantly improving system flexibility and versatility. The low cost of PPG devices gives them a significant advantage in practical applications. Their low cost and non-invasive nature have been widely validated in the field of health monitoring, such as wearable devices for heart rate and sleep monitoring. These advantages make PPG a promising candidate for identity verification.

[0004] Despite the immense potential of photoplethysmography (PPG) signals in identity recognition, several challenges remain in their practical application. First, signal acquisition is susceptible to motion artifacts, ambient lighting, and respiratory interference. These factors can alter the signal morphology, particularly during motion, where the spectral characteristics of PPG differ significantly from those at rest. Second, existing research largely focuses on single acquisition sites (such as fingertips or wrists) or resting states, lacking a systematic exploration of robustness to features from multiple sites and various physiological states. Therefore, a recognition technology capable of integrating PPG signal features from multiple sites and adapting to diverse physiological states is urgently needed. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and provide an identity recognition method based on multi-site photoplethysmography (PPG) signals. This method collects PPG signals from six sites: the left and right carotid arteries, left and right brachial arteries, left and right radial arteries, covering three physiological states: resting, active, and focused. The signals are preprocessed using low-pass filtering and moving average methods to extract waveform and time-frequency domain features. After filtering using a random forest algorithm, a K-nearest neighbor (KNN) classifier is used for identity recognition. Experimental results show that the four-site combination (left carotid, right carotid, left and right brachial, right brachial) achieves 100% accuracy under all test conditions, significantly improving recognition performance in dynamic environments. This invention provides an innovative technical solution for identity recognition of wearable devices under different physiological states.

[0006] Compared with the prior art, the innovative contribution of this invention lies in:

[0007] Multi-site signal acquisition captures unique physiological information from different sites, and the system analyzes the impact of site combinations on recognition accuracy, providing a more comprehensive feature set;

[0008] Multi-state testing verified the robustness of the model in dynamic environments;

[0009] By combining the random forest algorithm to optimize feature selection, the high-precision results demonstrate the potential of this method in the security authentication of wearable devices, laying the foundation for the application of PPG identity recognition technology in complex scenarios. Attached Figure Description

[0010] Figure 1 This is a block diagram of the experimental data acquisition system.

[0011] Figure 2 This is an example of a waveform after PPG signal preprocessing, showing the baseline drift correction effect.

[0012] Figure 3 The line graph shows the recognition accuracy achieved by the optimal combination of different numbers of sensors in the experiment, illustrating the trend of recognition accuracy changes in three different physiological states: calm, motion, and focus. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Example 1: Data Acquisition

[0015] Figure 1 This is a block diagram of the experimental data acquisition system. The data acquisition system uses an MSP430F5529 microcontroller (Texas Instruments, a 16-bit low-power microcontroller with a main frequency of 25MHz) combined with a reflective green PPG sensor to acquire high-quality photoplethysmography (PPG) signals. The system sampling frequency is 210Hz, and the single acquisition time is 1 minute, ensuring the temporal resolution and data integrity of the signal.

[0016] The experiment involved 24 healthy participants (aged 16-28 years, 16 males and 8 females), all of whom signed informed consent forms before participating in the experiment. PPG signals were acquired from the left and right carotid arteries, left and right brachial arteries, left and right radial arteries, covering the major arteries of the body to comprehensively assess pulse wave characteristics. The acquisition process consisted of three physiological states: resting, active, and focused. In the resting state, participants maintained a seated posture, avoiding swallowing, head movement, or other actions that might introduce artifacts to ensure signal quality. In the active state, participants performed 2 minutes of standardized exercise (including jumping jacks and squats at moderate intensity, with PPG signals acquired immediately afterward). In the focused state, participants focused their attention using the Stroop test; the task lasted 1 minute, and PPG signals were simultaneously acquired from six sites using six sensors.

[0017] Example 2: Signal Preprocessing

[0018] Figure 2 This is an example of the waveform after PPG signal preprocessing, demonstrating the baseline drift correction effect. The frequency of the PPG signal is mainly distributed in the range of 0.5-3Hz, reflecting physiological characteristics related to heart rate. To effectively remove high-frequency noise, such as electromyographic interference or environmental noise, a 12th-order Butterworth low-pass filter with a cutoff frequency of 4Hz is used on the acquired PPG signal. Baseline drift is often present in the PPG signal due to the influence of respiration, body movement, or other low-frequency interference, which may interfere with subsequent analysis. Baseline drift correction uses a moving average method with a window size of 210 sampling points (corresponding to 1 second). The low-frequency trend is calculated and subtracted from the original signal to generate the corrected PPG signal.

[0019] Example 3: Feature Extraction

[0020] Waveform characteristics:

[0021] A simple algorithm was used to manually filter normal waveforms, remove disordered waveforms, and retain complete periodic waveforms. Heart rate and heart rate variability were detected using an algorithm and used as two features. Then, a complete single heartbeat cycle was detected using an algorithm to obtain the start point, end point, and peak systolic blood pressure of a heartbeat cycle. To further analyze the PPG waveform, its first derivative VPG and second derivative APG were calculated using the following formulas:

[0022]

[0023]

[0024] Where PPG(t) represents the PPG signal at time t, and Δt is the sampling interval. VPG reflects the rate of change of the PPG signal, and APG reflects the acceleration of the PPG signal. By analyzing APG, local minima and maxima are detected to identify key feature points, such as the dicrotic notch and the diastolic peak. Eleven features are extracted from each cardiac cycle, specifically heart rate, heart rate variability, cardiac cycle area S, pulse map parameter K, and seven additional features based on the dicrotic notch and diastolic peak feature points.

[0025] Time-frequency domain characteristics:

[0026] The time-frequency domain features are decomposed into the signal using a 5th-order discrete wavelet transform, which is defined as follows:

[0027] W ψ (j,k)=∑ n x[n]·ψ j,k [n]

[0028] Where, ψ j,k [n] represents the wavelet basis function generated by scaling and translating the mother wavelet function ψ[n]. The wavelet basis is generated using the Daubechies mother wavelet function, and 18 features are extracted, including the approximation coefficients and detail coefficients of the five-level decomposition.

[0029] Example 4: Feature Filtering and Classification

[0030] The Random Forest method was used to evaluate feature importance in order to select the most discriminative features, and then the K-Nearest Neighbors (KNN) classifier was used for identity recognition.

[0031] Random Forest:

[0032] Random forest is a classifier based on ensemble learning that achieves robust feature selection and classification by constructing multiple decision trees and combining their predictions. Its core principle lies in generating multiple subsets from the training data through bootstrap aggregating (Bagging). Each decision tree is trained independently on these subsets, and feature subsets are randomly selected during node splits to enhance model diversity. Feature importance is evaluated by calculating the reduction in information gain or Gini impurity for each feature across all decision trees. Its mathematical expression can be approximated as:

[0033]

[0034] Where N is the number of decision trees, Gini parent (t) and Gini child (t, c) represent the Gini index before and after node splitting, respectively, and n c Let and n be the number of samples in the child and parent nodes, respectively, and f be the feature to be evaluated. The higher the importance score, the stronger the correlation between the feature and the identity label.

[0035] The importance of PPG signal features from different body parts was evaluated using the random forest method. Features highly relevant to the identity recognition task were extracted from the high-dimensional feature space, reducing redundant information. By selecting the features with the highest importance ranking for each body part, the feature dimensionality was reduced, and the feature quality was improved, thereby mitigating the risk of overfitting.

[0036] Example 5: Experimental Design and Methods

[0037] Data partitioning:

[0038] The PPG signal data of each subject was divided into training and testing sets to ensure that the model training and evaluation fully considered the multi-state characteristics. Data collection included 24 subjects (i.e., 24 identity tags), with data collected from each subject in three physiological states (resting state, post-exercise state, and focused state). The data collection time for each state was 1 minute, with a sampling frequency of 210 Hz, totaling 12,600 sampling points (60 seconds × 210 Hz). The specific division is as follows:

[0039] Training set: For each state of each subject, the data from the first 45 seconds (i.e., the first 9,450 sampling points) is used as training data. The training set consists of mixed data from three states, with each state containing 72 samples (24 subjects × 3 15-second segments), for a total of 216 samples (72 samples × 3 states), covering 24 identities.

[0040] Test Set: For each subject, the last 15 seconds of data (i.e., the last 3,150 sampling points) for each state are used as test data. The test set is divided independently by state, with each state containing 24 samples (24 subjects × 1 15-second segment), for a total of 72 samples (24 samples × 3 states), also covering 24 identities.

[0041] Experimental Design:

[0042] Single-path experiments: Experiments were conducted on each of the six body parts. The top six most important features were selected from the feature set for each part, with a step size of two features to test the effect of different feature counts. The recognition accuracy was calculated on the test set under three different states to observe the performance of single-part features under different conditions.

[0043] Multi-path experiments: Combinations of 2, 3, 4, 5, and 6 body parts were tested. In each combination, the top 6 most important features for each body part were selected, with a step size of 2 features, to test the effect of different numbers of features. Similar to the single-path experiments, the recognition accuracy was calculated on the test sets for the three states, and the impact of multi-body fusion on recognition performance was analyzed.

[0044] The experiment was divided into three scenarios: excluding the carotid artery, excluding the brachial artery, and excluding the radial artery. The aim was to systematically evaluate the impact of each artery category on the accuracy of identity recognition.

[0045] Example 6: Experimental Results

[0046] This study conducted real-time acquisition and analysis of PPG signals from multiple sites in 24 subjects under calm, post-exercise, and focused states, aiming to evaluate the impact of different sensor configurations on identity recognition accuracy. The experimental design covered configurations ranging from single-channel to six-channel; Tables 1-3 summarize the recognition accuracy achieved by the best-performing configuration under each experimental condition.

[0047] Table 1. Identification performance of non-carotid artery access combinations

[0048] Table 2 Identification performance of non-brachial artery access combinations

[0049] Table 3. Identification performance of non-radial artery access combinations

[0050] The results in Tables 1-3 show that the configuration excluding the radial artery in Table 3 performed best in single-path to four-path experiments, especially the four-path configuration (left carotid + right carotid + left brachial + right brachial) which achieved 100% accuracy in all states, demonstrating the superiority of the carotid and brachial artery combination. To further verify the saturation effect of multi-site fusion, Table 4 extends to five-path and six-path experiments.

[0051] Table 4. Identity Recognition Performance of Multi-Channel Combinations

[0052] To visually demonstrate the effect of merging multiple parts, Figure 3 The line graph presents the trend of recognition accuracy under different states for the optimal configuration of the six random combinations in Table 4. The accuracy is stable at 100% under all states, which is in stark contrast to the single path (right neck, 75% motion state), highlighting the advantages of the multi-part strategy in mitigating noise interference.

[0053] Through a systematic multi-state (calm, moving, focused) experimental design, the significant effect of multi-site photoplethysmography (PPG) signal fusion on improving the accuracy of identity recognition was demonstrated. (See Tables 1-4 and...) Figure 3 As shown, the experimental results indicate that the recognition accuracy significantly improves with the increase in the number of acquisition sites. Under the three experimental conditions, the single-path configuration performed worse than the calm state in both motion and focused states, reflecting the limitations of a single site. The dual-path configuration initially enhanced system stability, while the three-path configuration further improved robustness. The four-path configuration (left neck, right neck, left humerus, right humerus) achieved 100% accuracy in all three states—calm, motion, and focused—marking a substantial performance breakthrough. Based on this four-path combination, the five-path (left neck + right neck + left humerus + right humerus + left / right radial) and six-path (left neck + right neck + left humerus + right humerus + left radial + right radial) configurations continued to maintain 100% accuracy, suggesting that the performance may have reached its saturation threshold.

[0054] In contrast, the single-channel configuration in both cases results in unsatisfactory recognition accuracy under motion conditions, reflecting the significant impact of motion artifacts on signal quality. Our proposed multi-part fusion strategy effectively mitigates the effects of motion interference by integrating signals from six parts, significantly enhancing the stability and discriminative power of the features.

[0055] The experimental results also revealed the outstanding robustness of carotid artery signals in dynamic environments. The dual-carotid artery combination consistently outperformed the brachial and radial artery combination under all three experimental conditions. This may be related to the anatomical location of the carotid artery, which is close to the heart, allowing its PPG signal to more directly reflect cardiac pulsation characteristics, exhibiting clearer waveforms (such as significant biphasic waves and pulse amplitude). In contrast, traditional identification methods often focus on peripheral areas such as the fingertips or wrists, while this method significantly enriches the individual specificity of the feature set by incorporating neck signals. The experimental results in Tables 1-3 validate the necessity of this strategy.

[0056] In summary, this invention proposes a practical strategy for multi-part PPG signal fusion, which significantly improves the accuracy and robustness of identity recognition, especially in dynamic environments. The achievement of 100% accuracy with a four-way configuration highlights the potential of multi-part fusion in mitigating noise interference and enhancing feature stability, providing an important reference for the application of PPG technology in wearable devices and IoT security authentication.

[0057] Although embodiments of the invention have been shown and described in detail above, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An identity recognition method based on multi-site photoplethysmography (PPG) signals, comprising the following steps: Photoplethysmography (PPG) signals were collected from six body sites of the subject: the left carotid artery, right carotid artery, left brachial artery, right brachial artery, left radial artery, and right radial artery. The collection covered three physiological states: resting state, movement state, and focused state. The acquired PPG signals were preprocessed, including using a low-pass filter to remove noise and a moving average method to correct baseline drift. Eleven waveform features and 18 time-frequency domain features extracted by discrete wavelet transform were extracted from the preprocessed PPG signal. The random forest algorithm is used to filter the features, and redundant information is eliminated according to the importance of the features to achieve the integration of features from multiple parts. The K-nearest neighbor classifier is used for identity recognition based on the filtered features.

2. The identity recognition method based on multi-site photoplethysmography (PPG) signals according to claim 1, characterized in that: The PPG signal acquisition process involves six synchronously operating sensors continuously acquiring data for 60 seconds in each physiological state, with the acquisition sites covering major arteries to assess pulse wave characteristics.

3. The identity recognition method based on multi-site photoplethysmography (PPG) signals according to claim 1, characterized in that: During modeling, the raw PPG data of the three states (calm, moving, focused) were preprocessed using a 12th-order Butterworth low-pass filter. The cutoff frequency of the filter was set to 4Hz according to the state characteristics to remove noise. The moving average method uses a window size of 210 sampling points to calculate the low-frequency trend of the signal based on a sampling rate of 210Hz, and subtracts this trend from the filtered data to correct baseline drift caused by breathing and body movement.

4. The identity recognition method based on multi-site photoplethysmography (PPG) signals according to claim 1, characterized in that: The waveform features include 11 features extracted from each cardiac cycle, specifically heart rate, heart rate variability, area of ​​the heartbeat cycle, pulse map parameter K value, and 7 additional features based on diphtheria notch and diastolic peak feature points. The time-frequency domain features are decomposed by the signal using a 5th-order discrete wavelet transform, and a wavelet basis is generated using the Daubechies mother wavelet function to extract 18 features, including the approximation coefficients and detail coefficients of the five-level decomposition.

5. The identity recognition method based on multi-site photoplethysmography (PPG) signals according to claim 1, characterized in that: The random forest algorithm described therein selects the features most relevant to the identity recognition task from the high-dimensional feature space of PPG signals from multiple locations by calculating feature importance scores, thereby reducing the feature dimension to less than 50% and reducing the risk of overfitting.

6. The identity recognition method based on multi-site photoplethysmography (PPG) signals according to claim 1, characterized in that: The random forest is used to filter the features of each part, and the features that are highly relevant to the identity recognition task are selected. Then, the features of different parts are randomly combined to form a new feature vector, which is then input into the K-nearest neighbor classifier for identity recognition and classification.

Images