A millimeter wave sign detection method based on swarm intelligence and improved wavelet threshold

Abstract

The application discloses a kind of millimeter wave sign detection methods based on swarm intelligence and improved wavelet threshold value, wherein the method includes: obtaining the original vital sign signal collected by millimeter wave radar;Construct the multi-objective fitness function of fusion sample entropy, pearson correlation coefficient and kurtosis;The multi-objective fitness function is optimized using the improved Harris eagle optimization algorithm, and the optimal parameter combination of CEEMDAN decomposition algorithm is adaptively solved;Based on the optimal parameter combination, the original vital sign signal is decomposed by CEEMDAN, and a plurality of intrinsic mode function components are obtained.The application constructs the cascade processing framework of "parameter optimization-signal separation-noise suppression", solves the problem that CEEMDAN parameter depends on artificial experience, respiratory and heartbeat signal band aliasing and the problem of insufficient performance of traditional wavelet threshold denoising.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing and non-contact sensing technology, and in particular to a millimeter-wave radar vital sign detection method based on swarm intelligence optimization and improved wavelet threshold. Background Technology

[0002] Millimeter-wave radar, with its advantages of being non-contact, having strong penetration, and protecting privacy, has shown great potential in vital sign detection scenarios such as sleep monitoring, health monitoring, and disaster relief. The signals received by radar are a mixture of respiratory and heartbeat signals caused by subtle movements of the chest cavity and body surface, as well as environmental noise, circuit noise, and random body movement interference. Respiratory signals (approximately 0.1-0.4 Hz) and heartbeat signals (approximately 0.8-2.0 Hz) have similar frequencies but vastly different energies, and are often submerged in noise, resulting in severe signal band aliasing. Therefore, how to robustly and accurately separate weak respiratory and heartbeat signals from complex noise backgrounds is a core challenge currently facing the technology. Adaptive signal decomposition methods, such as Fully Adaptive Noise Ensemble Empirical Mode Decomposition (CEEMDAN), are widely used to process non-stationary and nonlinear signals. However, the decomposition effect of CEEMDAN is heavily dependent on the settings of its key parameters (such as the amplitude of added noise and the number of ensembles). Traditional methods rely heavily on empirical trial and error, lacking adaptability and easily leading to mode aliasing or component distortion, affecting the accuracy of subsequent separation. Furthermore, even after decomposition and reconstruction, residual noise still exists in the separated vital sign signals. Traditional wavelet thresholding denoising methods suffer from discontinuities and constant biases, making it difficult to effectively filter noise while preserving signal details, especially under low signal-to-noise ratio conditions where performance degrades significantly. Therefore, there is an urgent need for an adaptive signal processing method that can automatically optimize decomposition parameters and achieve refined post-processing. Summary of the Invention

[0003] The purpose of this invention is to provide a millimeter-wave radar vital sign detection method based on swarm intelligence optimization and improved wavelet threshold. By constructing a cascaded processing framework of "parameter optimization-signal separation-noise suppression", it solves the problems of CEEMDAN parameters relying on human experience, frequency band mixing of respiratory and heartbeat signals, and insufficient denoising performance of traditional wavelet threshold.

[0004] According to one aspect of the present invention, a millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet thresholding is provided, the method comprising: S1. Acquire raw vital signs signals collected by millimeter-wave radar; S2. Construct a multi-objective fitness function that integrates sample entropy, Pearson correlation coefficient, and kurtosis; S3. The multi-objective fitness function is optimized using the improved Harris Eagle optimization algorithm, and the optimal parameter combination of the CEEMDAN decomposition algorithm is adaptively solved. S4. Based on the optimal parameter combination, perform CEEMDAN decomposition on the original vital signs signal to obtain several intrinsic mode function components. S5. Based on the three-dimensional evaluation criteria of dominant frequency proportion, Pearson correlation coefficient and frequency band energy distribution, the intrinsic mode function components are screened and reconstructed to obtain the initially separated respiratory signal and heartbeat signal. S6. An improved wavelet threshold denoising algorithm is used to perform secondary denoising on the reconstructed respiratory signal and heartbeat signal respectively; S7. Perform spectral analysis on the denoised breathing and heartbeat signals to extract the respiratory rate and heart rate.

[0005] In the above technical solution, considering the technical problems existing in the background technology and the technical features adopted in the solution, the advantages of this method are mainly reflected in the following aspects: (1) To address the problem that CEEMDAN decomposition parameters in the background technology rely on manual trial and error and lack adaptability, this scheme constructs a multi-objective fitness function that integrates sample entropy, Pearson correlation coefficient, and kurtosis, and uses an improved Harris Eagle optimization algorithm to optimize the multi-objective fitness function, adaptively solving for the optimal parameter combination of the CEEMDAN decomposition algorithm. The decomposition effect of CEEMDAN heavily depends on the setting of key parameters (such as the amplitude of added noise and the number of ensembles). Traditional methods mostly rely on trial and error and lack adaptability. This scheme transforms the parameter selection problem into a multi-objective optimization problem by constructing a multi-objective function that comprehensively considers signal complexity (sample entropy), correlation (Pearson correlation coefficient), and statistical characteristics (kurtosis). Furthermore, the improved Harris Eagle optimization algorithm is used to optimize this function, which can automatically search for the parameter combination that makes the decomposition quality optimal (i.e., minimizes the fitness function). This process avoids the subjectivity and inefficiency of manual parameter tuning, realizes the adaptive determination of decomposition parameters, and provides accurate and adaptive input conditions for subsequent signal separation.

[0006] (2) To address the problem of severe frequency band aliasing and difficulty in accurate separation of respiratory and heartbeat signals in the background technology, this solution proposes a three-dimensional evaluation criterion based on the CEEMDAN decomposition, using the dominant frequency ratio, Pearson correlation coefficient, and frequency band energy distribution to screen and reconstruct the intrinsic mode function components. Respiratory and heartbeat signals have similar frequencies but vastly different energies, often being submerged in noise, leading to frequency band aliasing. This solution does not simply superimpose all the intrinsic mode function components obtained from the decomposition, but introduces a multi-dimensional screening mechanism. The dominant frequency ratio in this criterion measures the energy concentration of the component within a preset physiological frequency band (respiratory or heartbeat band); the Pearson correlation coefficient measures the correlation strength between the component and the original signal; and the frequency band energy distribution measures the sharpness of the component's spectrum. By weightedly combining these three dimensions, effective components belonging to respiration and heartbeat can be accurately identified from numerous components, solving the problem of single-indicator interference and inaccurate separation, and achieving accurate reconstruction of respiratory and heartbeat signals.

[0007] (3) To address the shortcomings of traditional wavelet thresholding denoising methods in the background technology, such as discontinuity and constant bias, this scheme employs an improved wavelet thresholding denoising algorithm to perform secondary denoising on the reconstructed respiratory and heartbeat signals. Traditional wavelet thresholding denoising methods (such as discontinuous hard thresholding functions and constant bias in soft thresholding functions) exhibit significant performance degradation under low signal-to-noise ratio conditions. The improved wavelet thresholding denoising algorithm used in this scheme includes a continuous and differentiable thresholding function (such as a hyperbolic tangent function). This function is continuous at the threshold point, avoiding the oscillation problem caused by the reconstructed signal from the hard thresholding function. At the same time, when the wavelet coefficients are large, the processed coefficients approach the original coefficients, reducing the constant bias caused by the soft thresholding function and better preserving signal details. Furthermore, this algorithm adopts an adaptive thresholding rule based on subband characteristics, dynamically adjusting the threshold according to the noise level of each decomposition layer, achieving refined and adaptive noise filtering, suppressing residual noise while protecting the weak features of the respiratory and heartbeat signals.

[0008] In summary, this solution constructs a multi-objective function to guide parameter optimization, designs multi-dimensional criteria for accurate signal separation, and introduces an improved threshold function for fine noise reduction, forming a complete "parameter optimization-signal separation-noise suppression" processing flow. This solution comprehensively addresses the problems of parameter dependence on experience, inaccurate signal separation, and insufficient noise reduction performance in the background technologies.

[0009] In some embodiments, the method for constructing the multi-objective fitness function includes: Calculate the mean sample entropy of all IMF components after CEEMDAN decomposition. Mean absolute value of Pearson correlation coefficient with original vital signs signals and the mean absolute value of kurtosis ; We sum the three factors by weight to construct the fitness function, whose expression is:

[0010] in, As preset weights, and .

[0011] In the above technical solution, in order to address the problem that the CEEMDAN decomposition parameters in the background technology rely on human experience and trial and error and lack adaptability, this solution constructs a multi-objective fitness function that integrates sample entropy, Pearson correlation coefficient and kurtosis, and transforms the parameter optimization problem into a minimization problem of a multi-objective function.

[0012] Specifically, the method for constructing the fitness function includes: calculating the mean sample entropy of all IMF components after CEEMDAN decomposition, the mean absolute value of the Pearson correlation coefficient with the original signal, and the mean absolute value of the kurtosis; weighted summing of the three values, and using a specific functional form, the expression of which is: The function's form has the following technical implications: the mean sample entropy is used to quantify the complexity of each component; ideal vital sign signals should have low complexity, so the smaller the expected value of this term, the better. The mean absolute value of the correlation coefficient is used to measure the correlation between each component and the original signal; a higher correlation indicates more effective information contained in the component, so 1 is subtracted from this term, making the overall function value smaller, corresponding to higher correlation. The mean absolute value of kurtosis is used to capture the impact characteristics in the signal; this term helps to preserve transient components such as heartbeats. Through the weighted combination of the above three indicators, this function can comprehensively evaluate the quality of CEEMDAN decomposition from three dimensions: complexity, correlation, and impact characteristics. The construction of this fitness function transforms the problem of evaluating decomposition performance, which is difficult to quantify directly, into a computable objective function, providing a quantitative basis for subsequent automatic search of optimal decomposition parameters using optimization algorithms. Furthermore, the three indicators in this function complement each other: sample entropy focuses on the regularity of the signal, the correlation coefficient focuses on the fidelity of information, and kurtosis focuses on the preservation of transient characteristics; the combination of these three can more comprehensively evaluate the adaptability of the decomposition results to different signal components. In addition, the function uses... The form of this approach ensures that the three metrics have a unified direction within the minimization framework, facilitating iterative optimization of subsequent algorithms. This construction method lays the foundation for adaptive selection of CEEMDAN parameters, avoiding the limitations of traditional methods that rely on trial and error based on human experience.

[0013] In some embodiments, the improved Harris Hawk optimization algorithm includes the following steps: Initialize the population, and set the search dimension, boundary, and maximum number of iterations; Calculate the fitness of each individual to determine the current optimal solution; Update escape energy And introduce a dynamic adaptive escape energy factor. To balance exploration and development; The Laplacian crossover operator and dynamic adaptive weighting factor are introduced during the development phase. This enhances population diversity and local search capabilities; Based on escape energy and random number The combination of these strategies employs soft encirclement, hard encirclement, rapid dive soft encirclement, or rapid dive hard encirclement to update individual positions. Iterate to the maximum number of times and output the optimal parameter combination.

[0014] In the above technical solution, addressing the problem that the CEEMDAN decomposition parameters rely on manual trial and error and lack adaptability in the background technology, this solution introduces an improved Harris Eagle optimization algorithm to optimize the multi-objective fitness function, thereby achieving automatic solution of decomposition parameters. This improved algorithm, based on the standard Harris Eagle algorithm, optimizes its position update mechanism through multiple strategies.

[0015] Specifically, the improved algorithm's process includes: initializing the population and setting the search dimension, boundary, and maximum number of iterations; calculating individual fitness to determine the current optimal solution; updating the escape energy E and introducing a dynamic adaptive escape energy factor. To balance global exploration and local development capabilities, the Laplace crossover operator and dynamic adaptive weighting factor are introduced during the development phase. To enhance population diversity and local search accuracy; based on escape energy and random numbers Different combinations of parameters are used to update the individual position using soft encirclement, hard encirclement, rapid dive soft encirclement, or rapid dive hard encirclement strategies; after iterating to the maximum number of times, the optimal parameter combination is output.

[0016] The innovation of this algorithm lies in several improvements over the standard Harris Eagle optimizer. Firstly, the dynamic adaptive escape energy factor... The introduction of [a specific feature] improves the algorithm's ability to balance the exploration and development phases during iteration, enabling it to broadly search the potential solution space in the early stages and finely mine around the optimal solution in the later stages. Secondly, the Laplace crossover operator strategy generates new offspring individuals by performing a Laplace distribution-based crossover operation on individuals with better fitness, allowing them to compete with their parents. This mechanism increases population diversity and helps the algorithm escape local optima. Thirdly, [a specific feature] dynamically adaptive weighting factors [is used]. The introduction of this approach improves attack behavior during the development phase. By adaptively adjusting weights with the number of iterations, it enhances the ability to approximate the prey's location and improves the convergence accuracy of local searches. Furthermore, the algorithm retains the escape energy-based approach found in standard HHO algorithms. and random numbers The four-quadrant siege strategy (soft siege, hard siege, rapid dive soft siege, and rapid dive hard siege) was proposed, and the relevant calculation formulas were adapted and adjusted according to the improved dynamic factors, making the siege behavior more adaptable to complex parameter optimization tasks.

[0017] The organic combination of the aforementioned improvement strategies enables this algorithm to exhibit stronger global optimization capabilities, higher convergence accuracy, and better population diversity compared to traditional empirical trial-and-error methods or standard swarm intelligence algorithms when solving the CEEMDAN parameter optimization problem. The optimal parameter combination obtained through the algorithm's adaptive solution allows CEEMDAN decomposition to better adapt to the characteristics of the original signal, providing better input conditions for the subsequent accurate separation of respiratory and heartbeat signals, thereby enhancing the adaptability and robustness of the entire vital sign detection method.

[0018] In some embodiments, the dynamic adaptive escape energy factor The update formula is:

[0019] in, A random number within (0,1) This represents the current iteration number. This represents the maximum number of iterations.

[0020] In the above technical solution, in the standard Harris Eagle optimization algorithm, the escape energy A linear decreasing strategy is usually adopted. The escape energy factor E1 is used to control the transition from the exploration phase to the development phase during the iteration process. However, the linear decreasing strategy is prone to causing a rapid loss of population diversity in the later stages of iteration, leading the algorithm into a local optimum. To address this problem, this scheme designs a dynamic adaptive escape energy factor E1 update formula, the expression of which is: The ingenuity of this formula lies in the nonlinear adjustment characteristics brought about by its composite function structure. Specifically, the formula consists of three parts: the first part... It provides a basic trend of decreasing with the number of iterations, consistent with standard algorithms; Part Two Random perturbations are introduced to enhance the algorithm's random search capability, helping to avoid premature convergence; Part Three It is a nonlinear decay factor based on an exponential function, which includes the natural constant. The power function form. This nonlinear factor in the early stage of iteration ( The smaller value takes a larger value, making The overall value is relatively high, which is beneficial for the algorithm to maintain strong exploration ability in the early stage; in the later stage of iteration ( (Approaching 1), the value of this factor is close to 0.9 - 0.5 = 0.4, which makes the overall value of E1 decrease but still maintain a certain non-zero range, avoiding the escape energy from decaying to zero too early, and thus retaining a certain ability to jump out of the local optimum in the later stage of iteration.

[0021] Compared to the linear decreasing strategy in the standard algorithm, this dynamic adaptive factor has the following technical advantages: First, the nonlinear decay characteristic allows the algorithm to adaptively adjust the balance between exploration and development according to the current progress during iteration. It maintains a high exploration probability in the early stages to fully search the solution space, and maintains appropriate escape energy in the later stages to prevent the population from getting trapped in local optima. Second, the formula includes a random term... This allows the escape energy value to fluctuate randomly in each iteration, increasing the diversity of population behavior and helping the algorithm maintain its vitality in complex multimodal function optimization. Furthermore, the nonlinear design based on the exponential function causes the escape energy decay rate to exhibit a pattern of initially slow, then rapid, and finally slowing down during iteration. This non-monotonic decay curve, compared to linear strategies, can more precisely match the search needs of intelligent optimization algorithms at different stages.

[0022] The dynamic adaptive escape energy factor The introduction of this algorithm improves upon the shortcomings of the standard Harris Eagle algorithm in exploring and developing balance control, giving the improved algorithm stronger global search capabilities and higher convergence accuracy when solving the CEEMDAN parameter optimization problem, and providing better parameter inputs for the subsequent accurate separation of respiratory and heartbeat signals.

[0023] In some embodiments, the Laplacian crossover operator is generated as follows: Calculate the cross coefficient , which follows a Laplace distribution; Generate two offspring individuals:

[0024] in, The coordinates of the two parent nodes; Compare the fitness of offspring with that of parents and retain the better individuals.

[0025] In the aforementioned technical solutions, the standard Harris Eagle optimization algorithm primarily relies on a predator-besieging strategy to update individual positions, lacking an explicit crossover mechanism. This can easily lead to a decrease in population diversity in the later stages of iteration, resulting in local optima. To address this issue, this solution introduces a Laplace crossover operator during the development phase of the improved Harris Eagle optimization algorithm. Its generation method is as follows: first, calculate the crossover coefficient β, which follows a Laplace distribution; then, based on two parent individuals... and Two offspring individuals are generated; finally, the fitness of the offspring and the parent is compared, and the individual with better fitness is retained to enter the next generation.

[0026] The innovation of this Laplace crossover operator lies in its distribution characteristics, generation mechanism, and integration with selection strategies. First, the crossover coefficient β follows a Laplace distribution, which has a longer tail than the Gaussian distribution, enabling it to generate larger random step sizes with a certain probability, thus introducing stronger perturbations in individual updates. When β is positive, offspring individuals will explore in directions away from their parents; when β is negative, offspring individuals will be generated between two parents or in the opposite direction. This coefficient generation method based on the Laplace distribution, compared to uniform or Gaussian distributions, can maintain local search capabilities while retaining a certain probability of large-step exploration, contributing to increased population diversity. Second, this operator utilizes the difference information between two parent individuals. As a scaling benchmark, the generation range of offspring individuals is adaptively correlated with the distance between parents. When two parent individuals are far apart, the exploration range of the offspring individuals expands accordingly; when parent individuals converge to a similar region, the perturbation range of the offspring individuals naturally decreases. This adaptive characteristic allows the operator to automatically adjust the search step size at different stages of algorithm iteration, which is beneficial for exploration in the early stages and for fine-grained search in the later stages. Furthermore, this operator is combined with a selection strategy to compare the fitness of offspring with that of their parents after generation, retaining only the better individuals. This competitive mechanism ensures that the overall quality of the population does not decrease due to crossover operations, while allowing a certain probability of deterioration to maintain diversity, but ultimately ensuring the convergence direction through fitness selection. In addition, this Laplace crossover operator does not replace the original Harris Eagle siege strategy, but rather serves as a supplementary mechanism, applying additional perturbation to individuals with better fitness during the development phase. This dual update mechanism of "siege + crossover" allows the algorithm to generate new candidate solutions near the current better solution through crossover operations while conducting local development around the current optimal solution, further exploring potential better solutions in the local region.

[0027] The introduction of this Laplace crossover operator effectively enhances the population diversity of the improved Harris Eagle optimization algorithm during the iteration process, reduces the risk of the algorithm getting trapped in local optima, and maintains convergence efficiency. By generating diverse candidate solutions and competing with the parent algorithm, this operator improves the algorithm's global optimization ability for complex multimodal functions (such as the CEEMDAN parameter optimization problem), providing a more reliable optimization foundation for adaptively solving the optimal decomposition parameters.

[0028] In some embodiments, the construction of the three-dimensional evaluation criteria includes: For each IMF component, calculate its energy percentage within the preset respiratory and cardiac frequency bands. , Calculate the absolute value of its Pearson correlation coefficient with the original vital signs signal. ; Calculate its spectral energy concentration ; Construct respiratory signal scoring functions and heart rate signal scoring functions:

[0029]

[0030] in, These are preset weighting coefficients; The IMF components with scores higher than the threshold are superimposed and reconstructed to obtain the initially separated respiratory and heartbeat signals.

[0031] In the aforementioned technical solutions, respiratory and heartbeat signals have similar frequencies but vastly different energies, and are often submerged in noise, resulting in severe signal aliasing. Accurately separating respiratory and heartbeat signals from the decomposed intrinsic mode function (IMF) components is a core technical challenge. To address this issue, this solution constructs a three-dimensional evaluation criterion based on the dominant frequency proportion, Pearson correlation coefficient, and frequency band energy distribution, used to screen and reconstruct the IMF components obtained from CEEMDAN decomposition.

[0032] The method for constructing this criterion includes: for each IMF component, firstly, calculating its energy proportion within the preset respiratory and cardiac frequency bands; secondly, calculating the absolute value of its Pearson correlation coefficient with the original vital signs signal; thirdly, calculating its spectral energy concentration; then constructing a respiratory signal scoring function and a cardiac signal scoring function; finally, selecting IMF components with scores higher than the corresponding thresholds for superposition and reconstruction to obtain the initially separated respiratory and cardiac signals.

[0033] The innovation of this criterion lies in the design of a multi-dimensional comprehensive evaluation mechanism. Firstly, the proportion of frequency band energy... , The energy concentration of a component is used to measure the energy concentration of the component within the target physiological frequency band. This indicator directly reflects the frequency band matching degree between the component and the respiratory or heartbeat signal, and is the basis for signal attribution judgment. Secondly, the absolute value of the Pearson correlation coefficient is used to measure the strength of the correlation between the component and the original signal. This indicator can identify components that contain more effective components of the original signal, avoiding the loss of effective components with high correlation to the original signal but slightly offset frequency bands due to over-reliance on frequency band characteristics. Thirdly, the spectral energy concentration is used to measure the sharpness of the component's spectrum. This indicator can distinguish between periodic signals with a clear dominant frequency and broadband noise, helping to screen out respiratory or heartbeat components with pure spectrum and strong periodicity. The above three dimensions evaluate IMF components from different perspectives: the frequency band energy proportion focuses on the frequency domain matching degree, the correlation coefficient focuses on the time domain correlation, and the energy concentration focuses on the spectral purity. The three complement each other to form a three-dimensional evaluation system. A single frequency band energy proportion may lead to misjudgment due to frequency band aliasing; a single correlation coefficient may be biased due to noise interference; and a single energy concentration may ignore the overall correlation between the signal and the original signal. By weighting and combining the three factors, this criterion can comprehensively weigh information from each dimension, improving the accuracy and robustness of component selection. Scoring Function and The construction uses the same weighting coefficients α, θ, and γ, but substitutes different frequency band energy proportions. , This design allows the same component to obtain two different score values, representing its suitability as a respiratory component and as a heartbeat component, respectively. This design allows a component to simultaneously compete for scores in both respiratory and heartbeat metrics, ultimately determining its assignment based on the score, thus adapting to potential frequency band overlap in real-world signals. IMF components with scores above a threshold are then superimposed and reconstructed, translating the quantitative evaluation results into a concrete signal reconstruction operation. By setting an appropriate threshold, components with low scores, potentially containing noise or interference, can be eliminated, retaining only high-quality, effective components for reconstruction, thereby improving the purity of the reconstructed signal.

[0034] The introduction of this three-dimensional evaluation criterion solves the problem of inaccurate separation caused by relying solely on frequency band division or subjective experience to select components in traditional methods. Through multi-dimensional comprehensive scoring, effective components of respiratory and heartbeat signals can be accurately identified from complex IMF sets, providing high-quality input signals for subsequent denoising and frequency extraction, thereby improving the separation accuracy and robustness of the entire vital sign detection method.

[0035] In some embodiments, the improved wavelet thresholding denoising algorithm includes: Use a continuously differentiable threshold function:

[0036] in, These are the original wavelet coefficients. These are the wavelet coefficients after thresholding. For the threshold, The adjustment parameter is greater than 0.

[0037] In the aforementioned technical solutions, traditional wavelet thresholding denoising methods have inherent defects: hard thresholding functions are discontinuous at the threshold point, leading to oscillations and pseudo-Gibbs phenomena in the reconstructed signal; while soft thresholding functions are continuous, there is a constant deviation between the processed wavelet coefficients and the original coefficients, resulting in excessive smoothing of signal details. To address these problems, this solution proposes an improved wavelet thresholding denoising algorithm. This algorithm employs a continuously differentiable threshold function, the innovation of which lies in its functional form design and the resulting mathematical properties. First, this function... Continuous: When Approaching from the right hour, Approaching 0, ,therefore Approaching 0, it seamlessly connects with the 0 value on the left, overcoming the discontinuity problem of hard thresholding functions at the threshold point. Secondly, the function is continuously differentiable within its domain, ensuring the smoothness of the reconstructed signal and avoiding spurious oscillations introduced by thresholding. This function also exhibits advantages in the region of large amplitude coefficients. hour, The value is relatively large. The function approaches 1, therefore Approaching This means that for signal coefficients with large amplitudes (typically corresponding to the main components of the signal), this threshold function can preserve the original coefficients with almost no attenuation, overcoming the constant bias problem caused by the soft threshold function's uniform contraction of all coefficients greater than the threshold. Adjustment parameters The introduction of threshold functions provides flexibility. The value of controls Steepness of the function: when When the value is large, the function changes drastically near the threshold point, approaching the characteristics of a hard threshold function; when When the value of 'a' is small, the function changes smoothly near the threshold point, exhibiting a certain smoothing characteristic. By adjusting the value of 'a', a trade-off can be struck between preserving signal details and suppressing noise, adapting to the needs of different signal characteristics and noise levels. The difference between this threshold function and traditional threshold functions lies in the following: hard threshold functions are piecewise constants that jump at the threshold point; soft threshold functions are piecewise linear and exhibit constant deviation; while the method used in this scheme... The morphological threshold function is a smooth, nonlinear contraction function that combines continuity and asymptotic unbiasedness. Compared to other nonlinear threshold functions (such as the Garrote function), it exhibits... The function has a clear mathematical expression and adjustable steep parameters, making it simpler to implement and providing clear physical meaning for the parameters. Applying this improved threshold function to the secondary denoising in step S6 can effectively filter out residual noise in respiratory and heartbeat signals while better preserving the detailed features of the signals. For respiratory signals, the function can maintain the integrity of their periodic waveform; for heartbeat signals with weak amplitudes, the asymptotically unbiased nature of the function can reduce excessive attenuation of detail coefficients, helping to preserve the impact characteristics of the heartbeat.

[0038] In some embodiments, the threshold For adaptive threshold The calculation formula is as follows:

[0039] in, For the first The noise standard deviation estimate of the layer wavelet coefficients. This is the signal length.

[0040] In the aforementioned technical solutions, traditional wavelet thresholding denoising methods exhibit significant performance degradation under low signal-to-noise ratio conditions, and fixed thresholds struggle to adapt to differences in noise levels across different decomposition layers. To address this issue, this solution employs an adaptive thresholding rule based on sub-band characteristics in its improved wavelet thresholding denoising algorithm. The innovation of this threshold calculation formula lies in its hierarchical adaptive mechanism design. Firstly, the formula includes a basic term. This approach originates from the universal threshold proposed by Donoho, which is theoretically based on the premise that the maximum amplitude of Gaussian white noise has a high probability of not exceeding the threshold, thus effectively removing noise components. However, the universal threshold applies the same threshold to all decomposition layers, failing to consider the differences in signal and noise distribution at different scales. This scheme, based on this, introduces a layered attenuation factor. This factor varies with the number of decomposition layers. The value decreases as the number of decomposition levels increases. Due to the multi-resolution nature of wavelet decomposition, the value decreases as the number of decomposition levels increases. As the scale increases, the frequency band corresponding to the detail coefficients gradually decreases, the energy proportion of the signal components gradually increases, while the energy of the noise components typically attenuates with increasing scale. Therefore, in the high-frequency layer ( Smaller frequencies require a larger threshold to suppress stronger noise, especially in the low-frequency layer. Larger values ​​require a smaller threshold to preserve signal details. Attenuation factor This adaptive adjustment was precisely achieved: when When the factor is 0.707, the threshold decay is relatively small; when At this point, the factor is approximately 0.177, significantly reducing the threshold and making it easier to preserve low-frequency detail coefficients. The base of this layered attenuation factor is designed to be... That is, approximately 0.707. This value is chosen because it relates to the frequency harmonic characteristics of wavelet decomposition. With each additional layer of decomposition, the bandwidth is halved, and noise energy exhibits a corresponding attenuation pattern between layers. This can approximately match the trend of noise energy decay with scale, achieving a reasonable match between the threshold and the number of layers. In the formula... For the first The noise standard deviation estimate of the layer wavelet coefficients is typically based on the median absolute deviation of the detail coefficients for that layer. This estimation method is insensitive to outliers and robustly reflects the actual noise level of each layer. Combined with a layer threshold factor, this formula achieves layer-by-layer adaptation: it considers the actual differences in noise levels between layers (through...). Furthermore, a priori decay law based on the number of layers was introduced (through ( The combination of these two factors allows for a more precise and reasonable threshold setting. This is in contrast to the improved threshold function. When used in conjunction with this adaptive threshold This constitutes a complete denoising scheme. The threshold function provides a continuously differentiable nonlinear shrinkage mechanism, while the adaptive threshold provides the function with input parameters that vary with the number of layers. For high-frequency layers, a larger threshold... This allows more wavelet coefficients to be set to zero or contracted, effectively filtering out noise; for the low-frequency layer, smaller... This allows the coefficients corresponding to the principal components of the signal to be preserved, while the asymptotically unbiased nature of the threshold function ensures that these coefficients are not excessively attenuated. The introduction of this adaptive threshold rule solves the problem of under-denoising or over-smoothing caused by the traditional fixed threshold method's uniform processing across different decomposition levels. By combining the attenuation factor related to the level and the actual noise level estimation, this rule can dynamically adjust the denoising intensity according to the characteristics of the wavelet coefficients at each level, achieving refined and adaptive noise filtering. This mechanism, together with the improved threshold function, provides a higher quality input signal for the spectral analysis in step S7, thereby improving the accuracy of respiratory rate and heart rate extraction.

[0041] In some embodiments, the spectrum analysis employs Fast Fourier Transform to search for the peak values ​​of the denoised signal spectrum within preset respiratory and heart rate bands, respectively, and extracts the respiratory rate and heart rate.

[0042] In the above technical solution, the signal received by the millimeter-wave radar is a mixture of respiratory and heartbeat signals, as well as environmental noise, circuit noise, and random body motion interference. After parameter optimization, signal separation, and secondary denoising in the aforementioned steps, it is necessary to accurately extract the respiratory rate and heart rate from the processed signal. To address this requirement, this solution uses Fast Fourier Transform (FFT) for spectrum analysis. Specifically, it searches for the peak values ​​of the denoised signal spectrum in step S6 within preset physiological respiratory and heartbeat frequency bands, and converts the peak frequencies into respiratory and heart rates. The innovation of this spectrum analysis method lies in its connection mechanism with the aforementioned steps and the application of a frequency band-limited search strategy. First, this method is closely integrated with the output of step S6, directly processing the respiratory and heartbeat signals after improved wavelet threshold denoising. The improved wavelet threshold denoising algorithm in step S5 effectively suppresses residual noise and preserves signal details through a continuously differentiable threshold function and adaptive threshold rules, providing a cleaner input signal with a higher signal-to-noise ratio for spectrum analysis, making subsequent peak detection more accurate and reliable. Second, this method uses preset physiological frequency bands to limit the search range. The respiratory physiological frequency band is typically set to 0.1-0.4 Hz, corresponding to a respiratory rate of 6-24 breaths / minute; the cardiac physiological frequency band is typically set to 0.8-2.0 Hz, corresponding to a heart rate of 48-120 beats / minute. When searching for spectral peaks, the maximum value is sought only within the corresponding physiological frequency band, rather than searching the entire frequency band. This limitation has the following technical significance: on the one hand, it eliminates the influence of noise interference, harmonic components, or body motion artifacts that may exist outside the physiological frequency band on peak detection, improving the accuracy of frequency extraction; on the other hand, it reduces the search range, lowering the probability of false detection caused by occasional large amplitude values ​​outside the frequency band, and enhancing the robustness of the extraction results. Furthermore, this method uses Fast Fourier Transform (FFT) as a spectral analysis tool. The FFT algorithm has high computational efficiency and can meet the application requirements of real-time or near-real-time vital sign monitoring. Combined with the signal quality after denoising in step S5, the spectrum obtained by FFT has clear peak characteristics, facilitating accurate location of the main peak frequency through peak detection algorithms. This spectral analysis method, together with the aforementioned steps, constitutes a complete detection process: steps S2-S3 achieve adaptive parameter optimization, steps S4-S5 achieve precise signal separation, step S6 achieves residual noise suppression, and step S7 transforms the above processing results into specific physiological parameters. The output of each step provides quality assurance for the input of the next step. Spectral analysis, as the final output, directly verifies the effectiveness of the aforementioned steps and quantifies the processing results into readable respiratory and heart rates. The technical contribution of this method lies in its use of a bandwidth-limited search strategy to fully utilize prior physiological knowledge of respiratory and heartbeat signals, avoiding misjudgments that may result from blind searches; and in its integration with an improved wavelet threshold denoising algorithm, it ensures the signal quality of the input spectrum, resulting in higher reliability of peak detection.This spectrum analysis method is simple and effective, complementing the aforementioned complex processing steps, and together they achieve high-precision, robust, and adaptive detection of vital signs signals from millimeter-wave radar.

[0043] According to another aspect of the present invention, a millimeter-wave vital sign detection device based on swarm intelligence and improved wavelet thresholding is provided, wherein the device, based on the above-described method, comprises: The signal acquisition module is used to acquire raw vital sign signals collected by millimeter-wave radar. The fitness function module is used to construct a multi-objective fitness function that integrates sample entropy, Pearson correlation coefficient, and kurtosis. The solution module is used to optimize the multi-objective fitness function using the improved Harris Eagle optimization algorithm and adaptively solve for the optimal parameter combination of the CEEMDAN decomposition algorithm. The decomposition module is used to perform CEEMDAN decomposition on the original vital sign signal based on the optimal parameter combination to obtain several intrinsic mode function components. The separation module is used to propose a three-dimensional evaluation criterion based on the main frequency ratio, Pearson correlation coefficient and frequency band energy distribution, and to screen and reconstruct the intrinsic mode function components to obtain the initially separated respiratory signal and heartbeat signal. The denoising module is used to perform secondary denoising on the reconstructed respiratory signal and heartbeat signal respectively using an improved wavelet threshold denoising algorithm; The extraction module is used to perform spectral analysis on the denoised breathing and heartbeat signals to extract the respiratory rate and heart rate.

[0044] In order to better utilize the above method, this application proposes a millimeter-wave vital sign detection device based on swarm intelligence and improved wavelet threshold. Each module corresponds to a step of the above method, and its specific principle has been described above and will not be repeated here. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating an embodiment of a millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold according to the present invention. Figure 2 This is an overall flowchart of an embodiment of a millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold of the present invention; Figure 3 This is a comparison of the convergence performance of the improved Harris Eagle Optimization Algorithm (IHHO) and the traditional algorithm in an embodiment of a millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold according to the present invention. Figure 4 This is a three-dimensional evaluation criterion analysis diagram of the intrinsic mode function (IMF) components of an embodiment of a millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet thresholding according to the present invention. Figure 5 This is a time-domain waveform diagram of respiratory and heartbeat signals reconstructed based on three-dimensional evaluation criteria, representing an embodiment of a millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold of the present invention. Figure 6 This is a comparison of the respiratory / heartbeat signal spectra before and after processing by the improved wavelet threshold denoising algorithm in an embodiment of a millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold according to the present invention. Figure 7 This is a schematic diagram of an embodiment of a millimeter-wave vital sign detection device based on swarm intelligence and improved wavelet threshold according to the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1 Please see Figure 1 , Figure 2 A millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet thresholding, the method comprising: S1. Acquire raw vital signs signals collected by millimeter-wave radar; For example, using a millimeter-wave radar evaluation kit (such as IWR6843ISK with DCA1000EVM acquisition board) or an integrated radar acquisition system, raw echo data containing thoracic displacement information is acquired with the test subject sitting 0.5m-1m in front of the radar, and stored as a digital signal file as input for subsequent processing.

[0049] S2. Construct a multi-objective fitness function that integrates sample entropy, Pearson correlation coefficient, and kurtosis; In this embodiment, the multi-objective fitness function The calculation formula is:

[0050] in, Let represent the set of sample entropy of all IMF components obtained after one CEEMDAN decomposition, the set of absolute values ​​of correlation coefficients with the original signal, and the set of absolute values ​​of kurtosis, respectively. This indicates an average value operation; The preset positive weighting coefficients, and satisfy the following conditions: . In this embodiment, the formulas for calculating the sample entropy set, correlation coefficient, and kurtosis index are as follows:

[0051] For example, the multi-objective fitness function is used to quantitatively evaluate the quality of the CEEMDAN decomposition results. The construction and calculation process of this function is as follows: Step 1: Perform a CEEMDAN decomposition once. Suppose that the location of a certain individual in the current Harris Eagle population corresponds to a set of CEEMDAN parameters (e.g., noise standard deviation ratio). average number of sets Using this set of parameters to analyze the original vital signs signals. Perform CEEMDAN decomposition to obtain The intrinsic mode function components are denoted as: and a residual component (The residual components are usually not included in subsequent calculations.) Step 2: Calculate the three indicators for each IMF component. For each component (in Calculate the following three values ​​respectively: (1) Sample entropy ( ): Purpose of the calculation: To measure the complexity of the components. The lower the sample entropy, the more regular and periodic the signal (ideal vital sign signals should have strong periodicity).

[0052] Example calculation: Assume that for components Let the embedding dimension be... Similarity tolerance By statistically matching the logarithm of the template, the sample entropy value of this component is finally calculated.

[0053] (2) Absolute value of Pearson correlation coefficient ( ): Calculation objective: To measure the relationship between this component and the original signal. The degree of correlation. The larger the absolute value, the more effective components of the original signal are contained in that component.

[0054] Example calculation: Assume the components With the original signal Perform a correlation analysis. The calculation formula is:

[0055] (3) Absolute value of kurtosis ( ): Purpose of the calculation: To measure the impact characteristics of the components. The greater the kurtosis, the more pronounced the impact components (such as the instantaneous impact of a heartbeat) or transient characteristics are in the signal.

[0056] Example calculation:

[0057] Step 3: Constructing the indicator set The calculation results of all the above components are summarized into three sets: Sample entropy set:

[0058] Correlation coefficient set:

[0059] Kurtosis set:

[0060] Step 4: Calculate the objective function value F Set preset weighting coefficients and substitute them into the formula for calculation:

[0061] Step 5: Results Feedback Calculated This refers to the fitness value corresponding to the set of CEEMDAN parameters. The smaller this value, the lower the overall complexity of the decomposed IMF components, the stronger their correlation with the original signal, and the more pronounced their impact characteristics; in other words, the better the decomposition effect. The goal of the IHHO optimization algorithm is to find the optimal solution through iterative optimization. The set of CEEMDAN parameters with the smallest value.

[0062] S3. The multi-objective fitness function is optimized using the improved Harris Eagle optimization algorithm, and the optimal parameter combination of the CEEMDAN decomposition algorithm is adaptively solved. In this embodiment, the improved Harris Hawk optimization algorithm specifically includes the following steps: S31. Randomly generate a parameter of size in the parameter search space. The initial population of individuals is set, and the maximum number of iterations is set. Search dimensionality The lower bound of the search range is set according to the number of parameters to be optimized in CEEMDAN. and the Upper Realm ; S32. Decode the position of the individual in the population into CEEMDAN parameters, run CEEMDAN decomposition and calculate the multi-objective fitness function value described in step S2, and find the position of the individual with the smallest fitness value as the position of the current optimal parameter combination. S33, Update Escape Energy ;when At this point, the algorithm enters the exploration phase. At this point, the algorithm enters the development phase; at different phases, the positions of individuals in the population are updated according to the improvement strategy; S34. If the maximum number of iterations is reached... If the condition is met, the globally optimal individual position is output as the optimal parameter combination for CEEMDAN; otherwise, return to step S32. In this embodiment, in S33, the escape energy The expression is:

[0063] in The initial escape energy is a random number within the interval (-1, 1) in each iteration. This represents the current iteration number. This represents the maximum number of iterations.

[0064] In this embodiment, in S33, when During the exploration phase, the position update method is as follows:

[0065] in, It is the position of an individual randomly selected in the current population. It is the current average position of the population. It is a random number within (0,1). and These are the lower and upper bounds of the search space, respectively.

[0066] In this embodiment, in S33, when When entering the development phase, its location update includes the following three improvement strategies: (1) Dynamic adaptive escape energy: The escape energy factor controlling development behavior is improved to... Its update formula is:

[0067] in, It is a random number within (0,1).

[0068] (2) Laplace crossover operator strategy: Introduce a crossover operation based on the Laplace distribution for individuals with better current fitness, and the crossover coefficients are... for:

[0069] in, Scaling factor It is a very small positive number, ensuring that the parameter is strictly greater than 0. Then, two new offspring individuals are generated:

[0070]

[0071] Then, the fitness of offspring and parents is compared, and the better solution is retained.

[0072] (3) Dynamic adaptive weights: Introducing dynamic weight factors into attack behaviors during the development phase. Its update formula is:

[0073] In this embodiment, in S33, random numbers are introduced during the development phase. Indicates the prey's chance of escape, according to and Different combinations of values ​​are used to update the position in the following four ways: (1) Soft encirclement strategy: when and At this time, the prey still has relatively sufficient ability to escape. The eagle flock then adopts a soft encirclement strategy, gradually approaching the prey by moving slowly. The position update formula is:

[0074] in, The difference between the prey's current position and its position. The random jump intensity of the prey, ∈(0,1) is a random number.

[0075] (2) Hard siege strategy: when and At this time, prey has low escape energy and is difficult to escape. The eagle flock adopts a direct and rapid hard-hit strategy, with the position update formula being:

[0076] (3) Rapid dive-type soft encirclement strategy: when and At this time, the prey has a higher probability of escaping. The flock of eagles first calculates a potential location according to strategy 1. The formula is:

[0077] in, For improved dynamic weighting factors, if the new position The fitness is better than the current position (Right now If the position is updated, then the position is... Otherwise, the encirclement is considered a failure, and the flock of eagles will perform a random walk based on the Lévy flight function to calculate another candidate location. :

[0078] in, A random vector of size 1×D. Let be the Lévy flight function, which is a 1×D random vector whose components are generated by a random walk model that follows a Lévy stable distribution. Specifically, the step size of each dimension is... It can be calculated using the following formula:

[0079] in, and It is a random variable that follows a standard normal distribution. It is a constant (usually taken as 1.5).

[0080] (4) Rapid dive-and-attack strategy: When and At that time, the prey was low on energy but still trying to escape. The flock of eagles first calculated candidate positions for an attack based on average position. :

[0081] in, This represents the current average position of the population. If... Then the updated position is Otherwise, perform the Lévy fly-through random walk again to compute another candidate location. And decide whether to update the position based on fitness. .

[0082] Exemplary, S3 example: Improved Harris Hawk algorithm to optimize CEEMDAN parameters In this embodiment, the improved Harris Hawk optimization algorithm (IHHO) is used to optimize two key parameters of CEEMDAN—the noise standard deviation ratio. with set average degree Adaptive optimization is performed to enable the multi-objective fitness function constructed by S2. Minimize. The specific process of the IHHO algorithm is as follows: S31. Parameter Setting and Population Initialization Parameter dimensions to be optimized Corresponding decision variables .

[0083] Search scope: , (All are integers, but are treated as continuous variables in the algorithm and rounded down during decoding).

[0084] Population size That is, 20 Harris Eagle individuals are randomly generated, and the position of each eagle represents a set of candidate parameters. .

[0085] Maximum number of iterations .

[0086] The initial population is generated uniformly and randomly within the search space. For example, the position of the first randomly generated individual is... The second individual And so on.

[0087] S32: Calculate fitness and determine the current best individual. Decode the position of each individual in the population into CEEMDAN parameters, run CEEMDAN decomposition, and calculate the fitness value according to S2. Assuming the current number is... In the generation, the individual The position with the minimum fitness value (i.e. optimal) is recorded.

[0088] Step S33: Update escape energy Calculate the escape energy using the formula:

[0089] in, Let be a number randomly generated in the range (-1, 1) for each iteration. Assume the current iteration... , Randomly generated ,but:

[0090] because The algorithm has entered the development stage.

[0091] S33 (Development Phase): Update individual locations based on improvement strategies. During the development phase, random numbers are introduced. (Represents the prey's chance of escaping, randomly within the range (0,1)) and current escape energy. The decision is made regarding which of the four siege strategies to employ. Simultaneously, the escape energy factor used during the development phase utilizes an improved dynamic adaptive escape energy. Dynamic weighting factor And apply the Laplace crossover operator.

[0092] Suppose we are currently processing the th ... individual Its current location is .

[0093] First, calculate the improved dynamic adaptive factor:

[0094] For example, let Then the calculation yields .

[0095]

[0096] Substitution Calculated .

[0097] Then, according to and Choose a specific strategy: Assuming random generation ,and This aligns with the "rapid dive-and-surround soft encirclement strategy" ( and ).

[0098] First, calculate the potential locations. :

[0099] in, , Given a random number within (0,1), assume ,but Substitute each vector into the calculation (note that this is vector operation):

[0100] Calculation yields specific Value (e.g.) ).

[0101] calculate fitness .like Then use replace Otherwise, perform a Lévy flight random walk to generate another candidate location. :

[0102] in It is a 1×2 random vector. For Lévy's flight stride, the stride length in each dimension. It is generated by the following formula:

[0103] Hypothesis generation , Take a random vector ,but .like Then use replace Otherwise, keep the original position.

[0104] Application of the Laplace crossover operator: Every certain number of generations (e.g., every 5 generations), perform the Laplace crossover operation on a subset of the best-fit individuals in the current population (e.g., the top 10%). Randomly select two best-fit individuals. Generate cross coefficients :

[0105] Generate two offspring:

[0106] If the offspring are more fit than the parents, they are replaced. This increases population diversity and avoids premature maturation.

[0107] S34: The iteration terminates. Repeat the above process until the maximum number of iterations is reached. The position of the individual with the lowest fitness throughout the process is recorded as the optimal parameter combination. After 20 iterations, the globally optimal individual position output by the IHHO algorithm is:

[0108] The fitness value corresponding to this set of parameters is the smallest. That is, under these parameters, the IMF component obtained by CEEMDAN decomposition has low overall complexity, strong correlation with the original signal, and obvious kurtosis characteristics, which is beneficial for the subsequent separation of respiratory and heartbeat signals.

[0109] S4. Based on the optimal parameter combination, perform CEEMDAN decomposition on the original vital signs signal to obtain several intrinsic mode function components. For example, S4 performs CEEMDAN decomposition on the original vital signs signal based on the optimal parameter combination to obtain several intrinsic mode function components.

[0110] In this embodiment, the optimal parameter combination obtained based on S3 optimization ( , The original vital signs signals were decomposed using CEEMDAN, and respiratory and heartbeat signals were reconstructed using three-dimensional evaluation criteria.

[0111] 1. CEEMDAN decomposition process Input: Raw vital signs signals sampling frequency Hz, duration 51.2 seconds, total number of sampling points .

[0112] Optimal parameter: Proportion of noise standard deviation average number of sets .

[0113] Decomposition steps: First-level decomposition: Towards the original signal Sixty different adaptive white noises (with an amplitude of 0.12 times the standard deviation of the original signal) were added, and EMD decomposition was performed. The average value of the first-order IMF components of each decomposition was taken as the first intrinsic mode function. At the same time, the first-order residual is obtained. .

[0114] Second-level decomposition: For the residual signal Add 60 more adaptive white noise iterations, perform EMD decomposition, and take the average of the second-order IMF components from each decomposition as... To obtain the residual .

[0115] Repeat the iteration: Continue the above process until the residual signal can no longer be decomposed (the number of extreme points does not exceed 2).

[0116] Decomposition results: Original signal Decomposed into One intrinsic mode function component and one residual component:

[0117] in, It mainly contains high-frequency noise. Corresponding heart rate frequency band (0.8-2.0Hz). Corresponding respiratory frequency band (0.1-0.4 Hz). This is the signal trend term.

[0118] S5. Based on the three-dimensional evaluation criteria of dominant frequency proportion, Pearson correlation coefficient and frequency band energy distribution, the intrinsic mode function components are screened and reconstructed to obtain the initially separated respiratory signal and heartbeat signal. In this embodiment, the comprehensive scoring function of the three-dimensional evaluation criterion described in S5 is: for the intrinsic mode function components It belongs to the respiratory signal scoring And scores belonging to heartbeat signals Calculate according to the following formulas:

[0119]

[0120] in, Components The energy percentage within the preset breathing and heart rate bands; For components The absolute value of the Pearson correlation coefficient with the original signal; For components Spectral energy concentration; These are the preset weighting coefficients.

[0121] In this embodiment, the formula for the band power ratio is as follows:

[0122] In this embodiment, S5 filters out and IMF components exceeding their corresponding thresholds are superimposed and reconstructed to obtain a preliminary separated respiratory signal. and heartbeat signals It is important to note the residual components. Items lacking valid vital signs information are directly rejected and will not participate in the IMF's screening and reconstruction process.

[0123] For example, S5 uses three-dimensional evaluation criteria for IMF screening and signal reconstruction. 1. Calculation of three-dimensional evaluation criteria For each IMF component ( ), calculate the following three indicators: (1) Frequency band energy ratio and

[0124] Assume respiratory frequency band: Hz Assuming a heart rate frequency band: Hz For components Perform a Fast Fourier Transform (FFT) to obtain the spectrum. Calculate the total energy:

[0125] Calculate the energy within the respiratory band:

[0126] Calculate the energy within the heartbeat frequency band:

[0127] The frequency band energy ratio is:

[0128] Example calculation (assuming...) (For example) Assuming the calculation yields , ,

[0129] but ,

[0130] (2) Absolute value of Pearson correlation coefficient

[0131] Calculate components With the original signal Correlation coefficient:

[0132] Take absolute value .

[0133] Example calculation: Assume Correlation coefficient with the original signal ,but

[0134] (3) Spectral energy concentration

[0135] Spectral energy concentration reflects the sharpness of the spectral peaks of a component. The calculation formula is:

[0136] Example calculation: Assumption peak frequency of the spectrum Hz (belonging to the breathing frequency band), with a peak energy of 0.85 and a total energy of 1.25, then

[0137] (4) Summary of three-dimensional indicators The above indicators were calculated for all nine components, and the results are assumed to be as shown in the table below:

[0138] 2. Comprehensive scoring calculation and screening Set the weight coefficients for the three-dimensional evaluation criteria (in this embodiment, we take...). , , ,satisfy ).

[0139] For each component Calculate respiratory signal scores respectively Heartbeat signal score :

[0140] Example calculation (assuming...) (For example)

[0141]

[0142] Calculate scores for all components, assuming the results are as follows:

[0143] Assuming the filtering threshold is set as follows: Respiratory signal scoring threshold:

[0144] Heartbeat signal scoring threshold:

[0145] Filtering results: Respiratory signal components: (0.772) (0.665) is above the threshold and is selected.

[0146] Heartbeat signal components: (0.394) (0.416) In (0.336), and Above the threshold Slightly below the threshold (can be retained or removed depending on actual requirements; this embodiment retains). and ).

[0147] 3. Signal Reconstruction The selected respiratory signal components are then superimposed and reconstructed.

[0148] The selected heartbeat signal components are then superimposed and reconstructed.

[0149] S6. An improved wavelet threshold denoising algorithm is used to perform secondary denoising on the reconstructed respiratory signal and heartbeat signal respectively; In this embodiment, in step S6, the improved wavelet thresholding denoising algorithm includes a continuously differentiable threshold function, which is defined as:

[0150] in, These are the original wavelet coefficients. These are the wavelet coefficients after thresholding. For the threshold, The adjustment parameter is greater than 0.

[0151] In this embodiment, in step S6, the improved wavelet thresholding denoising algorithm adopts an adaptive thresholding rule based on sub-band characteristics. Threshold of layer detail factor The calculation formula is:

[0152] in, The noise standard deviation is estimated based on the wavelet coefficients of this layer. This is the signal length.

[0153] In this embodiment, σ is estimated by the median absolute deviation of the wavelet coefficients of this layer.

[0154] For example, the improved wavelet thresholding denoising algorithm in S6 performs secondary denoising on the reconstructed signal. 1. Assume the input signal is as follows: respiratory signals : Obtained by reconstruction in step S5, length Sampling rate Hz, with a main frequency of approximately 0.23 Hz.

[0155] Heartbeat signal : Obtained by reconstruction in step S5, length Sampling rate Hz, with a main frequency of approximately 1.2 Hz.

[0156] Some noise remains in both signals, which needs to be refined using an improved wavelet threshold denoising algorithm.

[0157] 2. Algorithm Parameter Settings Wavelet basis function: Daubechies 6th order wavelet (db6) is selected because it has good regularity and tight support properties, making it suitable for processing local features of vital signs signals.

[0158] Number of decomposition layers: For a signal with a sampling rate of 20 Hz, the 5-layer decomposition can divide the frequency into: detail layer D1 (5-10 Hz), D2 (2.5-5 Hz), D3 (1.25-2.5 Hz), D4 (0.625-1.25 Hz), D5 (0.3125-0.625 Hz), and approximation layer A5 (0-0.3125 Hz), which can cover the respiratory and heartbeat frequency bands.

[0159] Improve the threshold function parameters: Adjust the parameters (Empirical value, controlling the steepness of the threshold function).

[0160] Adaptive threshold rule: according to the formula Calculate the threshold for each layer, where the noise standard deviation is... Estimated from the median absolute deviation of the wavelet coefficients at this level: , This is the signal length.

[0161] 3. Improve the definition of the threshold function For wavelet coefficients Coefficients after thresholding for:

[0162] This function is It is continuously differentiable at [location], and when [condition] hour, , This avoids the constant bias of traditional soft threshold functions.

[0163] 4. Respiratory signals noise reduction processing (1) Wavelet decomposition, for Perform a 5-level wavelet decomposition to obtain the detail coefficients of each level. ( ) and approximation coefficients .

[0164] (2) Estimate the standard deviation of noise at each floor Level 1 detail factor Calculate the absolute deviation of the median, and get ,but .

[0165] Level 2: , .

[0166] Level 3: , .

[0167] Level 4: , .

[0168] Level 5: , .

[0169] (3) Threshold processing For each level of detail coefficients, apply the improved threshold function ( Processing is performed using a certain coefficient in the first layer. (greater than) For example:

[0170] (Note: In actual calculations) when (When the value is very small, this is just an example value) Approximation coefficient The low-frequency components representing the signal are not subject to thresholding and are retained directly.

[0171] (4) Wavelet reconstruction The processed detail coefficients and the retained approximation coefficients are then subjected to inverse wavelet transform to obtain the denoised breathing signal. .

[0172] (5) Regarding heartbeat signals noise reduction processing Similarly to Perform a 5-level wavelet decomposition and estimate the noise standard deviation of each level: Level 1 detail factor: , .

[0173] Level 2: , .

[0174] Level 3: , .

[0175] Level 4: , .

[0176] Level 5: , .

[0177] (6) Threshold processing Similarly, the improved threshold function is applied ( The main energy of the heartbeat signal is concentrated in the detail coefficients of the 3rd and 4th layers (corresponding to the frequency bands of 1.25-5 Hz and 0.625-1.25 Hz), which are well preserved, while the noise in the high-frequency layers (the 1st and 2nd layers) is effectively suppressed.

[0178] (7) Wavelet reconstruction, the reconstructed heartbeat signal is obtained. .

[0179] 6. Output of denoising results After improved wavelet threshold denoising, the final high-quality respiratory signal was obtained. and heartbeat signals The data is then used in step S7 for spectral analysis to extract respiratory rate and heart rate.

[0180] S7. Perform spectral analysis on the denoised breathing and heartbeat signals to extract the respiratory rate and heart rate.

[0181] In this embodiment, in step S7, the spectrum analysis adopts the fast Fourier transform method; within the preset respiratory physiological frequency band and cardiac physiological frequency band, the peak frequency of the signal spectrum after noise reduction in step S6 is searched and used as the final extracted respiratory rate and heart rate.

[0182] For example, S7 spectral analysis extracts respiratory rate and heart rate. 1. Assume the input signal Noise-reduced breathing signal : Obtained from step S6, length sampling frequency Hz, duration is 51.2 seconds.

[0183] Denoising the heartbeat signal : Obtained from step S6, length sampling frequency Hz, duration is 51.2 seconds.

[0184] 2. Assuming the spectrum analysis parameters are set... Analysis method: Fast Fourier Transform (FFT) FFT points: (Double zero padding to improve spectral resolution) Window function: Hanning window, used to reduce spectral leakage. Frequency resolution: Hz Respiratory physiological frequency bands: Hz (corresponding to a respiratory rate of 6-24 breaths per minute) Physiological frequency bands of heartbeat: Hz (corresponding to a heart rate of 48-120 beats per minute) 3. Respiratory signal spectrum analysis and respiratory rate extraction (1) Windowing processing for respiratory signals Apply Hanning window:

[0185] The Hanning window function is:

[0186] (2) FFT calculation Perform a 2048-point FFT on the windowed signal to obtain the spectrum. , The corresponding frequency value is:

[0187] (3) Calculate the amplitude spectrum Take the absolute value of the spectral amplitude. It is then converted into a single-sided spectrum (only the positive frequency portion is retained).

[0188] (4) Search for peak values ​​within the respiratory frequency band In the respiratory band The index range corresponding to Hz Within, search for the location of the maximum value in the amplitude spectrum:

[0189] (5) Calculate respiratory rate Convert peak index to frequency value:

[0190] Example calculation: Assuming the search yields the peak index Corresponding frequency: Hz Converted to respiratory rate (breaths / minute): times / minute (6) Peak value verification Check if there is a distinct single main peak near the peak frequency to ensure there are no false detections caused by body movement or other interference.

[0191] 4. Heartbeat signal spectrum analysis and heart rate extraction (1) Windowing processing for heartbeat signals Apply the same Hanning window:

[0192] (2) FFT calculation: Perform a 2048-point FFT on the windowed heartbeat signal to obtain the spectrum. .

[0193] (3) Calculate the amplitude spectrum and take the absolute value. , which is converted into a one-sided spectrum.

[0194] (4) Search for peak values ​​within the heartbeat frequency band. The index range corresponding to Hz Within, search for the location of the maximum value in the amplitude spectrum:

[0195] (5) Calculate the heart rate and convert the peak index to a frequency value:

[0196] Example calculation: Assuming the search yields the peak index Corresponding frequency: Hz Converted to heart rate (beats / minute): times / minute (6) Peak value verification Check if there is a distinct single main peak within the heart rate frequency band. In this example, such as... Figure 5 As shown, the main peak in the heartbeat frequency band is clear, the harmonic components are weak, and there is no obvious interference.

[0197] To verify the effectiveness of the method proposed in this invention, real millimeter-wave radar vital sign data were used for testing. First, multiple 51.2-second segments of resting state data from healthy adults were directly selected from the vital sign radar database as the raw test signals. . Regarding parameter settings, the population size for the Improved Harris Eagle Optimization Algorithm (IHHO) is set. =20, Maximum number of iterations =20, and set two key parameters of CEEMDAN—the proportion of noise standard deviation. with set average degree As two-dimensional decision variables to be optimized, their search ranges are respectively set as follows: and A multi-objective fitness function used to evaluate decomposition quality. The weighting coefficient is set to Run IHHO to adjust parameters. Perform adaptive optimization, fitness The convergence process is as follows Figure 3 As shown. The final result is the fitness function. The smallest optimal parameter combination is . Combine the optimal parameters Input CEEMDAN to the original signal Decompose to obtain Each intrinsic mode IMF component is denoted as . Subsequently, based on the proportion of main frequency... absolute value of Pearson correlation coefficient With spectral energy concentration The three-dimensional evaluation criteria constituted for each IMF component. A comprehensive evaluation was conducted. The weighting coefficients for the three dimensions were set as follows: Based on this criterion, IMF components meeting the criteria were selected, and these components were then superimposed to reconstruct the initially separated respiratory signals. With heartbeat signals The radar chart of the three-dimensional evaluation indicators for each IMF component is shown below. Figure 4 As shown, the time-domain waveform of the reconstructed signal is as follows: Figure 5 As shown. Finally, the reconstructed signal and The improved wavelet thresholding denoising algorithm proposed in this invention is applied for refined post-processing. Specific settings include: selecting the db6 wavelet basis and performing a 5-level decomposition; and improving the adjustment parameters in the threshold function. Adaptive threshold noise standard deviation According to the Median estimation of layer wavelet coefficients. The spectral result after denoising is as follows: Figure 6 As shown.

[0198] Based on the above embodiments, it can be seen that the present invention has the following beneficial effects: This invention designs a novel evaluation criterion to guide the CEEMDAN decomposition and screening process. First, a multi-objective function considering signal complexity, correlation, and statistical characteristics is constructed to find the optimal decomposition parameters. Second, a set of explicit rules based on bandwidth energy proportion, correlation, and bandwidth energy ratio is proposed to accurately select signal components belonging to respiration and heartbeat from the decomposition results. This method avoids manual adjustment based on experience, making the signal separation process more objective and accurate. This invention introduces three key improvements to the standard Harris Eagle Optimization (HHO) algorithm, resulting in the IHHO algorithm. These improvements include: better balancing "broad exploration" and "fine-grained search" during the search process; and increasing the diversity of solutions through a new strategy to avoid prematurely getting trapped in local optima. The improved algorithm finds better parameter combinations and achieves more stable results when searching for parameters, thus laying a solid foundation for subsequent signal decomposition. The improved wavelet thresholding denoising method proposed in this invention uses a novel threshold function, which solves the problem that traditional methods easily lead to signal discontinuities or deviations when removing noise. It can better preserve the details of the original vital signs signal while filtering out noise. Furthermore, this method can automatically adjust the denoising intensity according to the noise levels of different frequency components of the signal, achieving more refined and adaptive noise filtering.

[0199] Example 2 Please see Figure 7 A millimeter-wave vital sign detection device based on swarm intelligence and improved wavelet thresholding, based on the method described in one embodiment, the device comprising: The signal acquisition module is used to acquire raw vital sign signals collected by millimeter-wave radar. The fitness function module is used to construct a multi-objective fitness function that integrates sample entropy, Pearson correlation coefficient, and kurtosis. The solution module is used to optimize the multi-objective fitness function using the improved Harris Eagle optimization algorithm and adaptively solve for the optimal parameter combination of the CEEMDAN decomposition algorithm. The decomposition module is used to perform CEEMDAN decomposition on the original vital sign signal based on the optimal parameter combination to obtain several intrinsic mode function components. The separation module is used to propose a three-dimensional evaluation criterion based on the main frequency ratio, Pearson correlation coefficient and frequency band energy distribution, and to screen and reconstruct the intrinsic mode function components to obtain the initially separated respiratory signal and heartbeat signal. The denoising module is used to perform secondary denoising on the reconstructed respiratory signal and heartbeat signal respectively using an improved wavelet threshold denoising algorithm; The extraction module is used to perform spectral analysis on the denoised breathing and heartbeat signals to extract the respiratory rate and heart rate.

[0200] In this embodiment, in order to better utilize the method described in one of the embodiments, this application proposes a millimeter-wave vital sign detection device based on swarm intelligence and improved wavelet threshold. Each module corresponds to each step of the above method, and its specific principle has been described above and will not be repeated here.

[0201] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet thresholding, characterized in that, The method includes: S1. Acquire raw vital signs signals collected by millimeter-wave radar; S2. Construct a multi-objective fitness function that integrates sample entropy, Pearson correlation coefficient, and kurtosis; S3. Optimize the multi-objective fitness function using the improved Harris Eagle optimization algorithm, and adaptively solve for the optimal parameter combination of the CEEMDAN decomposition algorithm; the improved Harris Eagle optimization algorithm includes the following steps: Initialize the population, and set the search dimension, boundary, and maximum number of iterations; Calculate the fitness of each individual to determine the current optimal solution; Update escape energy And introduce a dynamic adaptive escape energy factor. To balance exploration and development; The Laplacian crossover operator and dynamic adaptive weighting factor are introduced during the development phase. This enhances population diversity and local search capabilities; Based on escape energy and random number The combination of these strategies employs soft encirclement, hard encirclement, rapid dive soft encirclement, or rapid dive hard encirclement to update individual positions. Iterate to the maximum number of times and output the optimal parameter combination; S4. Based on the optimal parameter combination, perform CEEMDAN decomposition on the original vital signs signal to obtain several intrinsic mode function components. S5. Based on the three-dimensional evaluation criteria of dominant frequency proportion, Pearson correlation coefficient and frequency band energy distribution, the intrinsic mode function components are screened and reconstructed to obtain the initially separated respiratory signal and heartbeat signal. S6. An improved wavelet threshold denoising algorithm is used to perform secondary denoising on the reconstructed respiratory signal and heartbeat signal respectively; S7. Perform spectral analysis on the denoised breathing and heartbeat signals to extract the respiratory rate and heart rate.

2. The millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold as described in claim 1, characterized in that, The method for constructing the multi-objective fitness function includes: Calculate the mean sample entropy of all IMF components after CEEMDAN decomposition. Mean absolute value of Pearson correlation coefficient with original vital signs signals and the mean absolute value of kurtosis ; We sum the three factors by weight to construct the fitness function, whose expression is: in, For preset weights, and .

3. The millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold as described in claim 1, characterized in that, The dynamic adaptive escape energy factor The update formula is: in, A random number within (0,1) This represents the current iteration number. This represents the maximum number of iterations.

4. The millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold as described in claim 1, characterized in that, The Laplace crossover operator is generated as follows: Calculate the cross coefficient , which follows a Laplace distribution; Generate two offspring individuals: in, The coordinates of the two parent nodes; Compare the fitness of offspring with that of parents and retain the better individuals.

5. The millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold as described in claim 1, characterized in that, The construction of the three-dimensional evaluation criteria includes: For each IMF component, calculate its energy percentage within the preset respiratory and cardiac frequency bands. , Calculate the absolute value of its Pearson correlation coefficient with the original vital signs signal. ; Calculate its spectral energy concentration ; Construct respiratory signal scoring functions and heart rate signal scoring functions: in, These are preset weighting coefficients; The IMF components with scores higher than the threshold were superimposed and reconstructed to obtain the initially separated respiratory and heartbeat signals.

6. The millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold as described in claim 1, characterized in that, The improved wavelet threshold denoising algorithm includes: Use a continuously differentiable threshold function: in, These are the original wavelet coefficients. These are the wavelet coefficients after thresholding. For the threshold, The adjustment parameter is greater than 0.

7. The millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold as described in claim 6, characterized in that, The threshold For adaptive threshold The calculation formula is as follows: in, For the first The noise standard deviation estimate of the layer wavelet coefficients. This is the signal length.

8. The millimeter-wave vital sign detection method based on swarm intelligence and improved wavelet threshold as described in claim 1, characterized in that, The spectrum analysis employs Fast Fourier Transform to search for the peak values ​​of the denoised signal spectrum within preset respiratory and heart rate bands, respectively, and extracts the respiratory rate and heart rate.

9. A millimeter-wave radar vital sign detection system based on swarm intelligence optimization and improved wavelet threshold, characterized in that, The method according to any one of claims 1 to 8 includes: The signal acquisition module is used to acquire raw vital sign signals collected by millimeter-wave radar. The fitness function module is used to construct a multi-objective fitness function that integrates sample entropy, Pearson correlation coefficient, and kurtosis. The solution module is used to optimize the multi-objective fitness function using an improved Harris Eagle optimization algorithm, adaptively solving for the optimal parameter combination of the CEEMDAN decomposition algorithm; the improved Harris Eagle optimization algorithm includes the following steps: Initialize the population, and set the search dimension, boundary, and maximum number of iterations; Calculate the fitness of each individual to determine the current optimal solution; Update escape energy And introduce a dynamic adaptive escape energy factor. To balance exploration and development; The Laplacian crossover operator and dynamic adaptive weighting factor are introduced during the development phase. This enhances population diversity and local search capabilities; Based on escape energy and random number The combination of these strategies employs soft encirclement, hard encirclement, rapid dive soft encirclement, or rapid dive hard encirclement to update individual positions. Iterate to the maximum number of times and output the optimal parameter combination; The decomposition module is used to perform CEEMDAN decomposition on the original vital sign signal based on the optimal parameter combination to obtain several intrinsic mode function components. The separation module is used to propose a three-dimensional evaluation criterion based on the main frequency ratio, Pearson correlation coefficient and frequency band energy distribution, and to screen and reconstruct the intrinsic mode function components to obtain the initially separated respiratory signal and heartbeat signal. The denoising module is used to perform secondary denoising on the reconstructed respiratory signal and heartbeat signal respectively using an improved wavelet threshold denoising algorithm; The extraction module is used to perform spectral analysis on the denoised breathing and heartbeat signals to extract the respiratory rate and heart rate.

Images