Algorithm and system for eliminating motion artifacts of multi-wavelength PPG signals

Abstract

The invention relates to the technical field of biomedical signal processing, and discloses a multi-wavelength PPG signal motion artifact elimination algorithm and system. The method comprises the following steps: acquiring a multi-wavelength PPG signal by using acquisition equipment; inputting the acquired noise-containing multi-wavelength PPG signal into a filter; the filter processes an input signal and then generates a denoised PPG signal; extracting a period of the denoised signal in real time by using an autocorrelation function; inputting the denoised PPG signal and a period extracted in real time into a BSS algorithm to find an optimal noise reference signal; taking the denoised PPG signal and the optimal noise signal as a first input and a second input of adaptive filtering to eliminate motion artifacts; the motion artifacts are eliminated by using the algorithm, and finally the PPG signals with the motion artifacts eliminated are output; the method overcomes the problems of dependence on additional hardware, low correlation of noise reference signals, complex algorithm and the like in the prior art.

Description

Technical Field

[0001] The present invention belongs to the technical field of biomedical signal processing, and particularly relates to an algorithm and system for eliminating motion artifacts based on multi-wavelength photoplethysmogram signals. Background Art

[0002] In recent years, with the popularization of intelligent wearable devices and the growth of personalized medical needs, non-invasive physiological monitoring technology based on photoplethysmogram (PPG) has attracted much attention due to its convenience and low cost. PPG irradiates the skin surface with red and infrared light, and detects key physiological parameters such as heart rate and blood oxygen saturation by using the light absorption difference caused by blood flow changes. However, in dynamic scenarios (such as exercise), the PPG signal is easily interfered by motion artifacts (MA). The root cause is the change in the contact state between the sensor and the skin caused by non-cardiac human motion (such as limb swing), which makes the signal mixed with noise overlapping the target physiological feature frequency band, seriously affecting the accuracy of parameter detection.

[0003] Currently, the suppression techniques for motion artifacts are mainly divided into two categories: signal processing algorithm optimization and multi-sensor fusion solutions. At the algorithm level, methods such as wavelet denoising, independent component analysis (ICA), and empirical mode decomposition (EMD) achieve noise suppression through time-frequency domain separation or signal decomposition. However, their performance significantly decreases when the motion artifact and the physiological signal frequency bands overlap. Adaptive filtering technology has become a research hotspot due to its universality for frequency-domain independent noise, but its core depends on a highly correlated noise reference signal. In existing research, hardware-assisted solutions collect motion noise as a reference signal through an accelerometer or gyroscope (such as the accelerometer adaptive noise cancellation method proposed by Mohan et al. in 2016 and the improved gyroscope-based scheme by Lee et al. in 2018). Although the noise correlation is improved, it introduces additional hardware costs and device volume, making it difficult to meet the requirements of low cost and lightweight of wearable devices.

[0004] To solve the problem of hardware dependence, the method of synthesizing a reference signal from multi-channel PPG has emerged. For example, noise components are extracted from the corrupted PPG signal based on fast Fourier transform (FFT), singular value decomposition (SVD), or ICA. However, the reference signal synthesized by such methods has insufficient correlation with the real motion artifact, and is limited by the different responses of red and infrared light wavelengths to motion interference (such as the reference signal misalignment caused by wavelength characteristics in the two-stage adaptive algorithm proposed by Yousef et al. in 2014), making it difficult to achieve stable noise cancellation. In addition, the signal acquisition synchronization problem (such as the timing deviation in the research by Park et al. in 2023) further reduces the correlation between the noise component and the reference signal.

[0005] In summary, the existing technologies face the following core challenges: 1) Relying on additional sensors increases the system complexity and cost; 2) The correlation between the synthesized reference signal and the real motion artifacts is low; 3) The difference in the motion noise response among multi-wavelength signals limits the adaptability of the algorithm; 4) Complex algorithms are difficult to meet the real-time requirements of wearable devices. Therefore, there is an urgent need to develop a motion artifact cancellation algorithm that does not require external hardware, has high noise correlation, and low complexity, so as to improve the quality of PPG signals and the detection accuracy of physiological parameters in dynamic scenarios. Summary of the Invention

[0006] To solve the above technical problems in the existing technologies, the present invention provides an algorithm for canceling motion artifacts of multi-wavelength PPG signals, including the following steps:

[0007] S1. Synchronously collect multi-wavelength PPG signals using a collection device;

[0008] S2. Input the noisy multi-wavelength PPG signals into a band-pass filter for preprocessing to filter out high-frequency noise and baseline signals, and generate denoised PPG signals;

[0009] S3. The filter processes the input signals to generate denoised PPG signals, and generates denoised PPG signals;

[0010] S4. Use the autocorrelation function to extract the period of the denoised signals in real time;

[0011] S5. Input the denoised PPG signals and the period extracted in real time into the BSS algorithm to find the optimal noise reference signal;

[0012] S6. Use the denoised PPG signals and the optimal noise signals as the first input and the second input of the adaptive filter to cancel motion artifacts;

[0013] S7. Output the PPG signals with motion artifacts canceled, and calculate physiological parameters based on these signals.

[0014] The filter is a fourth-order Butterworth filter, and the transfer function is expressed as:

[0015]

[0016] Where H BP represents the transfer function of the band-pass filter, n represents the order of the filter, w LP represents the low cut-off frequency, w HP represents the high cut-off frequency, and s represents the complex frequency variable.

[0017] Frame the denoised PPG signals using the autocorrelation function, and extract the signal period in real time through a sliding window and an adaptive threshold, and adaptively adjust the threshold by calculating the average value and the standard deviation to extract the period.

[0018] The BSS algorithm generates a noise reference signal through the following steps:

[0019] (1) Establish a linear equation for the dual-wavelength PPG signal:

[0020] x1 = PPG R = AC R + MA R = aAC IR + cMA IR

[0021] x2 = PPG IR = AC IR + MA IR

[0022] where x1 and x2 represent the red-light PPG signal and the infrared-light PPG signal, IR and R represent the subscripts of the infrared light and the red light respectively, MA is the motion artifact component, and AC is the arterial component; a represents the ratio of AC R to AC IR , and c represents the ratio of MA R to MA IR .

[0023] (2) Establish a linear relationship between the dual wavelengths by using a β coefficient as:

[0024] N(n) = x1 - βx2 = (a - β)AC IR + (c - β)MA IR

[0025] where N(n) represents the noise reference signal, AC is a periodic component with a time structure, and the correlation signal contains an AC component (β = c), an MA component (β = a), or a linear combination of these two sources.

[0026] (3) When using the BSS algorithm to find a β that is updated in real time, first minimize the mean square error of the error signal e1(n) at the output, and extract the AC component estimation in the x2 signal. The mathematical expression of the e1(n) error signal:

[0027] e1(n) = S(n) - bS(n + T a )

[0028] where S(n) = x1 - cx2, b is a simple FIR filter coefficient vector with a single delay , and T a is the period of the discrete-time arterial source signal.

[0029] The filter adopts the least mean square algorithm, and its weight coefficient update rule is:

[0030] The output of the filter is:

[0031] y(n) = W T (n)N(n)

[0032] The error signal is:

[0033] e(n) = x(n) - y(n)

[0034] Weight coefficient iteration:

[0035] W(n + 1) = W(n) + mue(n)N(n)

[0036] Where y(n) is the output of the filter, W is the weight coefficient, N(n) is the noise reference signal, e(n) is the error signal, x(n) is the desired signal, mu is the learning rate, and the superscript T represents the transpose.

[0037] The learning rate mu and the filter order N are dynamically adjusted through the entropy value. The calculation formula for the entropy of the desired signal is:

[0038] (1) Calculate the power spectrum entropy of the desired signal:

[0039]

[0040] Where H represents the power spectrum entropy, P xxi is the power spectral density of the i-th frequency component, ε is a very small value, and ∑Pxx is the sum of the power spectral densities of all frequency components;

[0041] (2) Adjust the ranges of mu and N according to the entropy threshold: If the entropy value is less than 2, then set mu ∈ [0.0003, 0.0007] and N ∈ [25, 50]; otherwise, set mu ∈ [0.0008, 0.0015] and N ∈ [51, 120].

[0042] The physiological parameters include heart rate, and its accuracy is evaluated by the Pearson correlation coefficient and the mean absolute error. Among them,

[0043] Pearson correlation coefficient R: It is used to measure the linear relationship between two variables. Its value ranges from -1 to 1. When it is 1, the two are linearly correlated; when it is 0, the two are independent of each other; when it is negative, the two are negatively correlated. The calculation formula is as follows:

[0044]

[0045] Where Cov(HR est .HR ref ) represents the covariance between the estimated heart rate and the reference heart rate, and σ represents the standard deviation;

[0046] Mean Absolute Error (MAE): It represents the average of the absolute values of the deviations between all estimated values and the reference values. By taking the absolute value of the deviation for calculation, the situation where positive and negative deviations cancel each other out will not occur, and it can reflect the actual error better than the average error. The calculation formula is as follows:

[0047]

[0048] where N is the length of the heart rate sequence, and HR est (i) is the estimated heart rate after eliminating motion artifacts through the BSS-LMS adaptive filtering algorithm at the i-th moment, and HR ref (i) is the reference heart rate at the i-th moment.

[0049] On the other hand, the present invention provides a system based on the above algorithm, including:

[0050] (1) A multi-wavelength PPG sensor, which includes closely arranged red and infrared light LEDs and a photodiode covered with a filter;

[0051] (2) A signal acquisition module that synchronously acquires dual-wavelength PPG signals at a sampling rate of 1000 Hz;

[0052] (3) A processing module for performing filtering, period extraction, BSS algorithm, and adaptive filtering;

[0053] (4) A parameter dynamic adjustment module that adjusts the learning rate and order of the filter based on the entropy value.

[0054] The PPG sensor is fixed to the wrist of the subject through a wristband and synchronously acquires the reference signal with a four-limb electrocardiogram device.

[0055] The present invention provides a computer-readable storage medium that adopts the above algorithm.

[0056] The present invention has the following beneficial effects compared with the prior art:

[0057] Through multi-wavelength signal cooperation, BSS-LMS algorithm optimization, and dynamic entropy adjustment mechanism, the present invention realizes the efficient elimination of motion artifacts without external hardware, and has the advantages of low cost, low complexity, and high precision, providing a reliable solution for the dynamic physiological monitoring of wearable devices. Description of the Drawings

[0058] Figure 1 is a flowchart of a motion artifact elimination algorithm provided by the present invention;

[0059] Figure 2 is the system architecture for acquiring signals and the self-designed and fabricated PPG sensor covered with a filter provided by the present invention;

[0060] Figure 3 is the experimental setup for data collection provided by the present invention;

[0061] Figure 4 is the indication diagram of the detailed movements of each sequence provided by the present invention;

[0062] Figure 5 is the overall framework diagram of the BSS-LMS adaptive filtering algorithm of the present invention;

[0063] Figure 6 is the effect diagram of the reconstructed PPG signal waveform using the BSS-LMS adaptive filtering algorithm to eliminate MA provided by the present invention;

[0064] Figure 7 is the Bland-Altman plot of the HR estimated value and the reference value of the PPG signal after eliminating MA provided by the present invention.

[0065] Figure 8 is the Bland-Altman plot of the HR estimated value and the reference value of the PPG signal containing MA provided by the present invention. Detailed implementation manners

[0066] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be described below through specific embodiments shown in the drawings. However, it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the descriptions of known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present invention.

[0067] Referring to Figure 1 , the present invention provides an algorithm for eliminating motion artifacts of multi-wavelength PPG signals, including the following steps: S1. Synchronously collect multi-wavelength PPG signals using a collection device; S2. Input the noisy multi-wavelength PPG signals into a band-pass filter for preprocessing to filter out high-frequency noise and baseline signals and generate denoised PPG signals; S3. The filter processes the input signals to generate denoised PPG signals; S4. Use the autocorrelation function to extract the period in real time from the denoised signals; S5. Input the denoised PPG signals and the period extracted in real time into the BSS algorithm to find the optimal noise reference signal; S6. Use the denoised PPG signals and the optimal noise signal as the first input and the second input of the adaptive filtering to eliminate motion artifacts; S7. Output the PPG signals with motion artifacts eliminated and calculate physiological parameters based on these signals.

[0068] Specifically, this embodiment provides an algorithm for eliminating motion artifacts, including the following steps: synchronously collecting multi-wavelength PPG signals using a collection device, and eliminating ambient light interference through a photodiode covered with a filter; inputting the noisy multi-wavelength PPG signals into a band-pass filter for preprocessing to filter out high-frequency noise and baseline signals, generating denoised PPG signals; performing frame segmentation on the denoised PPG signals using the autocorrelation function, and extracting signal periods in real time through a sliding window and an adaptive threshold; inputting the denoised PPG signals and the extracted periods into a blind source separation (BSS) algorithm, establishing a dual-wavelength linear relationship model, and generating a best noise reference signal updated in real time by minimizing the mean square error of the error signal; using the denoised PPG signals and the best noise reference signal as inputs to an adaptive filter, dynamically adjusting filtering parameters in combination with entropy values to eliminate motion artifacts; outputting PPG signals with motion artifacts eliminated, and calculating physiological parameters based on these signals.

[0069] In one implementation, multi-wavelength PPG signals are synchronously collected using a collection device, and ambient light interference is eliminated through a photodiode covered with a filter. To improve the correlation of the MA components in the red and infrared light PPG signals and eliminate the influence of ambient light, a multi-wavelength PPG sensor collection device is used to synchronously collect PPG signals, and the photodiode (PD) is covered with a filter to eliminate the influence of ambient light, as Figure 2 shown. A total of 13 adult subjects participated in the experiment, their ages ranging from 18 to 27 years old, including 8 males and 5 females, covering different ethnic groups.

[0070] The experimental device for data collection is as Figure 3 shown. A multi-wavelength PPG sensor is worn on the left wrist of the subject and fixed using a wristband, and signals are synchronously collected using the II lead of a four-limb electrocardiogram (ECG) collection device for reference. During the entire experiment, the subjects were required to remain stationary except for their left hands, and then the left hands were operated according to the instructions, including continuous movement and rest.

[0071] Two sets of experiments were conducted for each subject. The two sets of experiments involved horizontal movement of the left hand and gait arm swing respectively, and each set of experiments included 4 sequences. The subjects moved at arbitrary amplitudes and frequencies according to the instructions. All data were collected at a sampling rate of 1000 Hz. The detailed instructions for each sequence are as Figure 4 shown.

[0072] The collected noisy multi-wavelength PPG signals are input into a filter. The filter processes the input signals to generate denoised PPG signals;

[0073] In the present invention, the frequency range of the AC component is between 0.5 Hz and 5 Hz. A fourth-order Butterworth band-pass filter is used to preprocess the collected data. The transfer function of the Butterworth band-pass filter is:

[0074]

[0075] where H BP represents the transfer function of the band-pass filter, n represents the order of the filter, w LP represents the low cut-off frequency, w HP represents the high cut-off frequency, and s represents the complex frequency variable.

[0076] During preprocessing, the high cut-off frequency is set to 5 Hz and the low cut-off frequency is set to 0.5 Hz to filter out high-frequency noise, smooth the signal, filter out the PPG baseline signal, and improve the quality of the PPG signal.

[0077] S4: Use the autocorrelation function to extract the period of the denoised red light signal in real time; when using the autocorrelation function to extract the period, the present invention extracts the non-stationary PPG signal containing motion artifacts. The heart rate of a normal person is between 60 and 100. Here, the frame length is set to 900 sampling points. Each autocorrelation calculation only processes 2 times the frame length, that is, 1800 sampling points. Therefore, the data is sliced and framed at a sliding interval of 1.8 seconds, and the overlapping size is 0.8 seconds to achieve the effect of time-varying autocorrelation and perform more detailed local period analysis on the non-stationary signal. When calculating the autocorrelation frame by frame, an adaptive threshold method is adopted, and the threshold is adjusted in real time by calculating the average value and the standard deviation to more accurately extract the period.

[0078] S5: Input the denoised PPG signal and the period extracted in real time into the BSS algorithm to find the best noise reference signal (that is, obtain the real-time updated β);

[0079] To obtain the noise reference signal of the adaptive filter, the BSS algorithm is used to obtain the best noise reference signal. The BSS algorithm blindly finds a linear combination of the mixed signals, and this combination can recover the original source signals, which may be rescaled and randomly permuted in the output. First, since the MA between the two wavelengths is highly correlated, the infrared light (x2) and red light (x1) signals are modeled. The linear equation of the PPG signal is:

[0080] x2 = PPG IR = AC IR + MA IR (2)

[0081] x1 = PPG R = AC R + MA R = aACIR +cMA IR (3)

[0082] Among them, x1 and x2 represent the red light PPG signal and the infrared light PPG signal, and the subscripts IR and R represent infrared light and red light respectively. a and c represent a = AC R / AC IR 、c = MA R / MA IR 。

[0083] The linear relationship between the two wavelengths is established by using a β coefficient as:

[0084] N(n) = x1 - βx2 = (a - β)AC IR +(c - β)MA IR (4)

[0085] Among them, N(n) represents the noise reference signal, AC is the periodic component with a time structure, and the correlation signal contains the AC component (β = c), the MA component (β = a), or a linear combination of these two sources.

[0086] When using the BSS algorithm to find the real-time updated β, first minimize the mean square error of the error signal e1(n) at the output, and extract the AC component estimation in the x2 signal. The mathematical expression of the e1(n) error signal:

[0087] e1(n) = S(n) - bS(n + T a ) (5)

[0088] Among them, S(n) = x1 - cx2, b is the coefficient vector of a simple FIR filter with a single delay ; T a is the period of the discrete-time arterial source signal and is extracted using the autocorrelation function.

[0089] c and b are completed by minimizing the mean square error defined as the cost function. The cost function is:

[0090] J(c,b) = E[e1 2 (n)] (6)

[0091] When the gradients of the cost function with respect to c and b are zero, this function reaches the minimum value. Therefore, by setting the gradients of the cost function with respect to c and b equal to zero, the update rules for c and b are obtained as follows:

[0092]

[0093] Among them, E[] represents the expectation.

[0094] Extract the noise reference signal by eliminating the arterial signal in x1. This process is calculated by the variance of the error signal e2(n), and its mathematical expression is:

[0095] e2(n) = x1 - αS(n) = x1 - α(x1 - cx2) (9)

[0096] The update rule of α is derived as above:

[0097]

[0098] Actually, e2(n) is the noise reference signal of the adaptive filter. The update rule of the β coefficient can be obtained according to formulas (4) and (9):

[0099]

[0100] Use the denoised PPG signal and the optimal noise signal as the first input and the second input of the adaptive filter to eliminate motion artifacts;

[0101] In the present invention, a (Least Mean Squares, LMS) adaptive filter is used to eliminate motion artifacts. The equation of the conventional LMS algorithm is as follows:

[0102] The output of the LMS adaptive filter is:

[0103] y(n) = W T (n)N(n) (12)

[0104] The error signal is:

[0105] e(n) = x(n) - y(n) (13)

[0106] Weight coefficient iteration:

[0107] W(n + 1) = W(n) + mue(n)N(n) (14)

[0108] Where y(n) is the output of the adaptive filter, W is the weight coefficient, N(n) is the noise reference signal, e(n) is the error signal, x(n) is the desired signal, mu is the learning rate, and the superscript T represents transpose.

[0109] Use the proposed algorithm to eliminate motion artifacts, and finally output the PPG signal with motion artifacts eliminated.

[0110] In order to find the appropriate learning rate and filtering order (N) to eliminate motion artifacts, we use entropy to judge the regularity and complexity of the desired signal in each frame, dynamically adjust the parameters of the filter, and eliminate the motion artifacts of the PPG signal. The calculation formula of the entropy of the desired signal is:

[0111]

[0112] where H represents the power spectrum entropy, and P xxi is the power spectral density of the i-th frequency component, ∑Pxx is the sum of the power spectral densities of all frequency components, the power spectral density is calculated using the Welch method, and ε is a very small value used to prevent the logarithmic function from being undefined at zero.

[0113] To distinguish between noisy frames and noise-free frames, the PPG signals in the moving state and the stationary state are measured and the power spectrum entropy is analyzed. The entropy of the PPG signal during movement is easily distinguishable from the entropy in the stationary state. In the stationary state, the entropy is very low and exhibits very small fluctuations in different frames. On the contrary, the entropy value is relatively high in the moving state. Although in this case, there are large fluctuations in different frames, the entropy of each frame is significantly higher than that in the stationary state. Therefore, the empirically determined threshold is H T = 2, and this threshold provides a sufficient margin between the stationary state and the moving state. By comparing the entropy values in the stationary state and the moving state, when N = 0.0005 and mu = 132, the two states can be clearly distinguished. Therefore, initially, the initial values of the filter parameters are set as mu = 0.0005 and N = 132, and then the entropy of the infrared light after denoising the slices is calculated. The threshold of the entropy is set equal to 2. When H≥2, the range of mu is set to 0.0003 - 0.0007, and the range of N is set to 25 - 50. Otherwise, the range of mu is set to 0.0008 - 0.0015, and the range of N is set to 51 - 150. Different mus and Ns are traversed within the parameter range, and the combination that minimizes the entropy value of the error signal e(n) is selected.

[0114] The overall framework diagram of the proposed algorithm is as Figure 5 shown.

[0115] To prove the beneficial effects of this elimination method, in this embodiment, the Pearson correlation coefficient (PCC), the mean absolute error (MAE), and the Bland-Altman plot are selected to evaluate the quality of the PPG signal generated by the model.

[0116] Pearson correlation coefficient R: It is used to measure the linear relationship between two variables. Its value ranges from -1 to 1. When it is 1, the two are linearly correlated. When it is 0, the two are independent of each other. When it is negative, the two are negatively correlated. The calculation formula is as follows:

[0117]

[0118] where Cov(HR est.HR ref ) represents the covariance between the estimated heart rate and the reference heart rate, and σ represents the standard deviation.

[0119] Mean Absolute Error MAE: It represents the average of the absolute values of the deviations between all estimated values and the reference values. Taking the absolute value of the deviation for calculation, the situation of positive and negative deviations canceling each other out will not occur, and it can better reflect the actual error compared with the average error. The calculation formula is as follows:

[0120]

[0121] where N is the length of the heart rate sequence, HR est (i) is the heart rate estimated by eliminating motion artifacts through the BSS-LMS adaptive filtering algorithm at the i-th moment, and HR ref (i) is the reference heart rate at the i-th moment.

[0122] Bland-Altman plot: It combines the quantitative analysis of the consistency limit with the qualitative description of the scatter plot distribution to accurately evaluate the consistency between the estimated value and the reference value. Its horizontal axis is the mean of the estimated and reference values with the same coordinates, and the vertical axis is the difference between the two. Let μ be the average of the differences between the two groups of data, and σ be its standard deviation. The confidence interval is taken as [μ - 1.96σ, μ + 1.96σ].

[0123] To show the denoising effect of the proposed algorithm, Table 1 summarizes the comparison of various indexes of the PPG signal before and after denoising and the ECG signal used for reference. It can be seen from the data in the table that after the BSS-LMS adaptive filtering proposed in the present invention eliminates motion artifacts, the estimated values have significantly improved compared with the original noisy signal in terms of various indexes, and the denoising effect is remarkable. Specifically, the Pearson correlation coefficient has increased by 97.23% relatively, and the mean absolute error has decreased by 94.94% relatively.

[0124] The noisy PPG signal is processed by the BSS-LMS adaptive filtering algorithm to eliminate motion artifacts, and the reconstructed PPG signal obtained is as Figure 6 shown. It can be seen from the figure that the algorithm makes the reconstructed PPG waveform have obvious periodicity while restoring some important details, making the waveform more valuable for reference.

[0125] Table 1. Performance evaluation

[0126]

[0127] To further evaluate the effect of the proposed BSS-LMS adaptive filtering algorithm in eliminating MA, Figure 7 and Figure 8Bland-Altman plots of heart rate values before and after eliminating motion artifacts respectively against the reference heart rate value. As can be seen from the figure, there is a relatively low consistency between the heart rate estimated from the original noisy PPG signal and the reference heart rate value, while the consistency between the estimated value and the reference value after being processed by the BSS-LMS adaptive algorithm is relatively high.

[0128] The motion artifact elimination algorithm for multi-wavelength PPG signals provided by the present invention has the following significant advantages compared with the prior art:

[0129] Firstly, the present invention reduces the hardware dependence and system cost. By synchronously collecting dual-wavelength PPG signals and utilizing the high correlation of their motion artifact (MA) components, a noise reference signal is directly synthesized without relying on additional hardware such as accelerometers and gyroscopes, significantly reducing the system volume and manufacturing cost.

[0130] Secondly, the present invention enhances the correlation of the noise reference signal. By combining the blind source separation (BSS) algorithm to optimize the noise reference signal in real time, the problem of high randomness and low correlation of the noise reference signal in traditional methods is solved, thereby improving the denoising performance of the adaptive filter.

[0131] Thirdly, the algorithm of the present invention is efficient and has low complexity. The present invention extracts the signal period in real time through the autocorrelation function and combines dynamic parameter adjustment (such as the optimization of the learning rate and filtering order based on the entropy criterion), simplifies the algorithm process, reduces the consumption of computing resources, and is suitable for real-time processing in embedded systems.

[0132] On the other hand, the present invention significantly improves the signal quality. Experimental data shows that after being processed by this algorithm, the Pearson correlation coefficient (PCC) of the PPG signal increases from 0.869 to 0.996, and the mean absolute error (MAE) decreases from 15.129 bpm to 0.765 bpm, and the consistency between the heart rate estimation and the reference value is significantly improved (verified by the Bland-Altman plot). The present invention introduces the entropy criterion to analyze the spectral characteristics of the signal and dynamically adjusts the filtering parameters according to the degree of motion artifact interference, enhancing the robustness and adaptability of the algorithm in complex motion scenarios.

[0133] At the same time, by optimizing the signal acquisition and processing process, the algorithm of the present invention is more suitable for fields such as smart wearable devices and mobile medical monitoring that have strict requirements for volume, cost, and real-time performance, providing a reliable solution for the accurate detection of physiological parameters in dynamic environments.

[0134] In summary, through the collaborative innovation of multiple technologies, the present invention effectively solves the pain points in the prior art such as strong hardware dependence, poor quality of the noise reference signal, and complex algorithm, providing an efficient, low-cost, and reliable solution for the motion artifact elimination of PPG signals.

[0135] The above has described an exemplary flowchart for implementing the elimination of motion artifacts in PPG signals according to an embodiment of the present invention with reference to the accompanying drawings. It should be noted that the large number of details included in the above description are only exemplary illustrations of the present invention, rather than limitations on the present invention. In other embodiments of the present invention, the method may have more, fewer, or different steps, and the relationships such as the sequence, inclusion, and functions between the steps may be different from those described and illustrated.

Claims

1. An algorithm for eliminating motion artifacts in multi-wavelength PPG signals, characterized in that, It includes the following steps: S1. Use a collection device to synchronously collect multi-wavelength PPG signals; S2. Input the noisy multi-wavelength PPG signals into a band-pass filter for preprocessing to filter out high-frequency noise and baseline signals, generating denoised PPG signals; S3. The filter processes the input signals to generate denoised PPG signals, generating denoised PPG signals; S4. Use the autocorrelation function to extract the period of the denoised signals in real time; S5. Input the denoised PPG signals and the period extracted in real time into the BSS algorithm to find the optimal noise reference signal; S6. Use the denoised PPG signals and the optimal noise signals as the first input and the second input of the adaptive filter to eliminate motion artifacts; S7. Output the PPG signals with motion artifacts eliminated and calculate physiological parameters based on these signals.

2. The algorithm according to claim 1, characterized in that, The filter is a fourth-order Butterworth filter with a low cut-off frequency of 0.5 Hz and a high cut-off frequency of 5 Hz, and its transfer function is expressed as: Where H BP represents the transfer function of the band - pass filter, n represents the order of the filter, w LP represents the low - cut - off frequency, w HP represents the high - cut - off frequency, and s represents the complex - frequency variable.

3. The algorithm according to claim 1, characterized in that, Perform frame processing on the denoised PPG signals using the autocorrelation function, and extract the signal period in real time through a sliding window and an adaptive threshold; the frame processing slides at an interval of 1.8 seconds with an overlap size of 0.8 seconds, and adaptively adjusts the threshold by calculating the average value and the standard deviation to extract the period.

4. The algorithm according to claim 1, characterized in that, The BSS algorithm generates a noise reference signal through the following steps: (1) Establish a linear equation for the dual-wavelength PPG signals: x1 = PPG R = AC R + MA R = aAC IR + cMA IR x2 = PPG IR = AC IR + MA IR Among them, x1 and x2 represent the red-light PPG signal and the infrared-light PPG signal, and the subscripts IR and R represent infrared light and red light respectively. MA is the motion artifact component, and AC is the arterial component; a represents AC R The ratio with AC IR , and c represents MA R The ratio with MA IR . (2) Establish a linear relationship between the dual wavelengths by using a β coefficient as: N(n) = x1 - βx2 = (a - β)AC IR +(c - β)MA IR where N(n) represents the noise reference signal, AC is a periodic component with a time structure, and the correlation signal contains the AC component (β = c), the MA component (β = a), or a linear combination of these two components. (3) When using the BSS algorithm to find the β updated in real time, first minimize the mean square error of the error signal e1(n) at the output to extract the AC estimation component in the x2 signal. The mathematical expression of the e1(n) error signal: e1(n) = S(n) - bS(n + T a ) where S(n) = x1 - cx2, and b is a simple FIR filter coefficient vector with a single delay Z-T a where T a is the period of the discrete-time arterial source signal.

5. The algorithm according to claim 1, characterized in that, The filter adopts the least mean square algorithm, and its weight coefficient update rule is: The output of the adaptive filter is: y(n) = W T (n)N(n) The error signal is: e(n) = x(n) - y(n) Weight coefficient iteration: W(n + 1) = W(n) + mue(n)N(n) where y(n) is the output of the filter, W is the weight coefficient, N(n) is the noise reference signal, e(n) is the error signal, x(n) is the desired signal, mu is the learning rate, and the superscript T represents the transpose.

6. The algorithm according to claim 5, wherein The learning rate mu and the filter order N are dynamically adjusted through the entropy value. The calculation formula for the entropy of the desired signal is: (1) Calculate the power spectrum entropy of the desired signal: Among them, H represents the power spectrum entropy, and P xxi is the power spectrum density of the i-th frequency component, ε is a very small value, and ∑Pxx is the sum of the power spectrum densities of all frequency components; (2) Adjust the ranges of mu and N according to the entropy value: If the entropy value is less than 2, then set mu ∈ [0.0003, 0.0007] and N ∈ [25, 50]; otherwise, set mu ∈ [0.0008, 0.0015] and N ∈ [51, 120].

7. The algorithm according to claim 1, characterized in that The physiological parameters include heart rate, and its accuracy is evaluated through the Pearson correlation coefficient and the mean absolute error. Among them, Pearson correlation coefficient R: It is used to measure the linear relationship between two variables. Its value ranges from -1 to 1. When it is 1, the two variables are linearly correlated; when it is 0, the two variables are independent of each other; when it is negative, the two variables are negatively correlated. The calculation formula is as follows: where Cov(HR est .HR ref ) represents the covariance between the estimated heart rate HR est and the reference heart rate HR ref , and σ represents the standard deviation. Mean Absolute Error MAE: It represents the average of the absolute values of the deviations between all estimated values and the reference value. Taking the absolute value of the deviation for calculation, there will be no situation where positive and negative deviations cancel each other out, and it can better reflect the actual error compared to the mean error. The calculation formula is as follows: where N is the length of the heart rate sequence, and HR est (i) is the estimated heart rate after eliminating motion artifacts by the BSS-LMS adaptive filtering algorithm at time i, and HR ref (i) is the reference heart rate at time i.

8. A system for implementing the algorithm according to any one of claims 1-7, characterized in that, Including: (1) A multi-wavelength PPG sensor, which includes closely arranged red and infrared LEDs and a photodiode covered with a filter; (2) A signal acquisition module, which synchronously acquires dual-wavelength PPG signals at a sampling rate of 1000 Hz; (3) A processing module, which is used to perform filtering, period extraction, BSS algorithm, and adaptive filtering; (4) A parameter dynamic adjustment module, which adjusts the learning rate and order of the filter based on the entropy value.

9. The system according to claim 8, wherein The PPG sensor is fixed to the subject's wrist through a wristband and synchronously acquires the reference signal with a four-limb electrocardiogram device.

10. A computer-readable storage medium storing a computer program, characterized in that, When the program is executed, it implements the steps of the algorithm described in any one of claims 1-7.

Images