A continuous blood pressure signal anomaly detection method and system

Abstract

The application discloses a continuous blood pressure signal anomaly detection method and system, and relates to the technical field of signal anomaly detection.The method comprises the following steps: collecting a continuous blood pressure signal, segmenting the continuous blood pressure signal, and automatically classifying each signal segment based on the low-frequency band energy proportion and the preset high-frequency band energy proportion to obtain a signal segment set with a category label; sequentially performing CEEMDAN decomposition processing, IMF correlation screening processing, adaptive wavelet threshold processing and signal reconstruction processing on a single signal segment to obtain a denoised signal segment; performing adaptive feature extraction and fusion on the denoised signal segment based on the category label to obtain a fusion feature vector; calculating a health index of each signal segment based on the fusion feature vector to obtain a health index value of each signal segment; arranging the health index values in time sequence to form a continuous health index sequence; and performing anomaly detection and classification on the continuous health index sequence and outputting an anomaly result.

Description

Technical Field

[0001] This invention relates to the field of signal anomaly detection technology, and in particular to a method and system for continuous blood pressure signal anomaly detection. Background Technology

[0002] In recent years, with the rapid development of wearable devices and continuous physiological parameter monitoring technology, non-invasive and real-time monitoring of continuous blood pressure signals has become an important means of early warning and health management of cardiovascular diseases. By continuously analyzing blood pressure signals, it is possible to effectively capture their dynamic change patterns and promptly detect abnormal blood pressure events. This has important value in monitoring clinical conditions such as hypertension, hypotension, and excessive blood pressure fluctuations. However, since blood pressure signals are easily interfered with by various factors such as motion artifacts, device noise, and individual physiological variations, how to accurately and reliably detect abnormal segments from complex continuous signals and identify their specific types has always been a core challenge in this field.

[0003] Currently, most common methods for detecting abnormal blood pressure signals rely on preset fixed thresholds or statistical outlier detection. These methods do not adequately consider the non-stationary characteristics and noise sensitivity of the signals, making it difficult to adapt to the diversity of blood pressure patterns between and within individuals. In addition, existing methods usually adopt a uniform feature extraction and processing strategy, failing to perform differentiated analysis based on the time-frequency characteristics of the signals themselves, such as trend changes and high-frequency fluctuations. This results in limited feature representation capabilities and low distinguishability between transient and slowly changing abnormalities. Furthermore, most methods only focus on detecting abnormal segments and lack further discrimination of abnormality types, limiting their practical application value in clinical decision support. Summary of the Invention

[0004] The technical problem solved by this invention is that common methods for detecting abnormal blood pressure signals mostly rely on preset fixed thresholds or statistical outlier detection. These methods do not adequately consider the non-stationary characteristics and noise sensitivity of the signal, making it difficult to adapt to the diversity of blood pressure patterns between and within individuals. In addition, existing methods usually adopt a uniform feature extraction and processing strategy, failing to perform differentiated analysis based on the time-frequency characteristics of the signal itself, such as trend changes and high-frequency fluctuations. This results in limited feature representation capabilities and low distinguishability between transient and slowly changing abnormalities. Furthermore, most methods only focus on detecting abnormal segments and lack further discrimination of abnormality types, limiting their practical application value in clinical decision support.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for detecting abnormal continuous blood pressure signals, comprising the following steps, Step S1: Collect continuous blood pressure signals, segment the continuous blood pressure signals, and automatically classify each signal segment based on the energy ratio of the low frequency band and the preset energy ratio of the high frequency band to obtain a set of signal segments with category labels. Step S2: Perform CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction on a single signal segment in sequence to obtain the denoised signal segment. Step S3: Based on the category label, perform adaptive feature extraction and fusion on the denoised signal segment to obtain the fused feature vector; Step S4: Calculate the health index for the fused feature vector of each signal segment to obtain the health index value of each signal segment, and arrange all the health index values ​​in chronological order to form a continuous health index sequence. Step S5: Based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined abnormality type template, perform abnormality detection and classification, and output the abnormality results.

[0006] As a preferred embodiment of the continuous blood pressure signal abnormality detection method of the present invention, step S1 involves acquiring a continuous blood pressure signal, segmenting the continuous blood pressure signal, and automatically classifying each signal segment based on the energy ratio of the low-frequency band and the preset energy ratio of the high-frequency band, thereby obtaining a set of signal segments with category labels. Step S101: The continuous blood pressure signal is segmented using a sliding window of preset duration to obtain N signal segments; Step S102: Perform frequency domain transformation on each signal segment to obtain the power spectrum of each signal segment; Step S103: Based on the power spectrum, calculate the low-frequency band energy value of the signal segment in the preset low-frequency band and the high-frequency band energy value in the preset high-frequency band, respectively. The low-frequency band energy value is the sum of the power spectral density values ​​of the power spectrum at all frequency points within the preset low-frequency band, and the high-frequency band energy value is the sum of the power spectral density values ​​of the power spectrum at all frequency points within the preset high-frequency band. Step S104: Calculate the low-frequency band energy ratio and the high-frequency band energy ratio of the signal segment; The low-frequency band energy ratio is the ratio of the low-frequency band energy value to the total energy value, the high-frequency band energy ratio is the ratio of the high-frequency band energy value to the total energy value, and the total energy value is the sum of the low-frequency band and high-frequency band energy values; The sum of the energy proportions of the low-frequency band and the high-frequency band is 1.

[0007] In a preferred embodiment of the continuous blood pressure signal abnormality detection method of the present invention, in step S105, if the proportion of low frequency band energy is greater than the first determination threshold, the signal segment is determined to be a trend segment and associated with a trend category label. If the high-frequency band energy ratio is greater than the second determination threshold, the signal segment is determined to be a wave-type segment and associated with a wave category label; The signal segment collection consists of all signal segments with category labels, including trend category labels and volatility category labels.

[0008] As a preferred embodiment of the continuous blood pressure signal abnormality detection method of the present invention, in step S2, a single signal segment is sequentially subjected to CEEMDAN decomposition processing, IMF correlation screening processing, and adaptive wavelet threshold processing and signal reconstruction processing to obtain a denoised signal segment. Step S201, the CEEMDAN decomposition process includes performing CEEMDAN decomposition on a single signal segment to obtain k intrinsic mode components and residuals; Step S202, the IMF correlation screening process includes calculating the Pearson correlation coefficient for each intrinsic mode component and signal segment to obtain the correlation coefficient. When the correlation coefficient is greater than the correlation threshold, all corresponding intrinsic mode components are taken as the correlation component set. Step S203, the adaptive wavelet thresholding process includes performing 3-level db4 wavelet decomposition on each component in the relevant component set to obtain wavelet coefficients; The wavelet coefficients are processed using a piecewise adaptive threshold function to obtain the processed coefficients. The expression for the piecewise adaptive threshold function is as follows: ; in, This represents the coefficients after processing. Represents wavelet coefficients, Indicates the general threshold. , This represents an estimate of the noise standard deviation. Indicates the length of the wavelet coefficients. Indicates the preset adjustment parameters and If greater than 0, e represents the natural constant; Step S204, the signal reconstruction process includes reconstructing the processed coefficients through inverse wavelet transform to obtain the denoised components; The denoised signal segment is obtained by summing the denoised components and residuals after all components in the relevant component set have been processed by the piecewise adaptive threshold function and reconstructed.

[0009] As a preferred embodiment of the continuous blood pressure signal abnormality detection method of the present invention, in step S3, adaptive feature extraction and fusion are performed on the denoised signal segment based on the category label to obtain a fused feature vector; Step S301, the adaptive feature extraction process includes selecting the corresponding feature extraction path based on the category label; If the category label is a trend category label, then proceed to step S302; Step S302: For the denoised signal segment, a moving average smoothing is performed using a first fixed one-dimensional convolution kernel with coefficients [1 / 3, 1 / 3, 1 / 3] and length 3, and abrupt change point detection is performed using a second fixed one-dimensional convolution kernel with coefficients [1, -2, 1] and length 3 to obtain a trend feature vector; If the category label is a fluctuation category label, then proceed to step S303; Step S303: Perform CEEMDAN decomposition on the denoised signal segment to obtain multiple intrinsic mode components. Select the highest one-third of the components from the multiple intrinsic mode components as high-frequency components. Calculate the instantaneous frequency characteristics and energy characteristics of each high-frequency component to obtain the wave feature vector.

[0010] In a preferred embodiment of the continuous blood pressure signal anomaly detection method of the present invention, in step S304, the denoised signal segment is input to a predefined width learning system mapping node. The width learning system mapping node calculates the current output through the mapping function of the width learning system mapping node. The input of the mapping function includes the denoised signal segment and the output value generated by the width learning system mapping node when processing the previous signal segment. The current output is used as a temporal feature vector. Step S305: Select the corresponding feature combination method according to the category label and fuse them to obtain the fused feature vector; If the selected logical dimension is a trend category label, then the trend feature vector and the time series feature vector are concatenated to obtain a fused feature vector. If the label is a fluctuation category label, the fluctuation feature vector and the time series feature vector are concatenated to obtain a fused feature vector.

[0011] In a preferred embodiment of the continuous blood pressure signal abnormality detection method of the present invention, step S4 involves calculating health indicators for the fusion feature vector of each signal segment to obtain the health indicator value of each signal segment, and arranging all health indicator values ​​in chronological order to form a continuous health indicator sequence. Step S401: Based on a set of fused feature vectors as historical samples and corresponding physiological state reference values, the least squares method is used to fit and obtain a mapping weight vector. Based on the mapping weight vector, a linear mapping relationship from the fused feature vector to the health score is constructed. The fused feature vectors are mapped to a preliminary health score using a linear mapping relationship; Step S402: Normalize the preliminary health score to obtain the health indicator value; Step S403: Arrange the health indicator values ​​corresponding to all signal segments in chronological order to form a continuous health indicator sequence.

[0012] As a preferred embodiment of the continuous blood pressure signal abnormality detection method of the present invention, step S5 involves performing abnormality detection and classification based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined abnormality type template, and outputting the abnormality result. Step S501: Calculate local statistics in the health indicator sequence using a sliding window method, and dynamically generate anomaly detection thresholds based on the local statistics. Step S502: If the values ​​of M+1 consecutive segments traced back from the current segment in the health indicator sequence are all less than their respective dynamic thresholds, then the M+1 segments are determined to be abnormal segments. Step S503: Using the predefined anomaly type template, match the fusion feature vector corresponding to the anomaly fragment to obtain an anomaly type label; Step S504: Based on the judgment result of the abnormal segment and the abnormal type label, generate and output the detection result; The detection results include anomaly warning signals and anomaly type labels.

[0013] As a preferred embodiment of the continuous blood pressure signal anomaly detection method of the present invention, the process of constructing the predefined anomaly type template includes obtaining O blood pressure signal sample fragments with known anomaly types from historical data; For each of the sample segments, steps S1 to S4 are executed sequentially to obtain the corresponding fused feature vector; For sample segments belonging to the same anomaly type, calculate the statistical center vector of the fusion feature vector, and use the statistical center vector as the template feature vector representing the anomaly type. The template feature vectors of all exception types constitute a predefined exception type template.

[0014] A continuous blood pressure signal abnormality detection system includes a classification module, a decomposition module, a fusion module, a calculation module, and a detection module; The classification module collects continuous blood pressure signals, segments the continuous blood pressure signals, and automatically classifies them based on the energy ratio of each signal segment in the low-frequency band and the preset energy ratio in the high-frequency band, thus obtaining a set of signal segments with category labels. The decomposition module sequentially performs CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction on a single signal segment to obtain a denoised signal segment. The fusion module performs adaptive feature extraction and fusion on the denoised signal segments based on category labels to obtain a fused feature vector. The calculation module performs health index calculations on the fused feature vector of each signal segment to obtain the health index value of each signal segment, and arranges all health index values ​​in chronological order to form a continuous health index sequence. The detection module performs anomaly detection and classification based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined anomaly type template, and outputs the anomaly results.

[0015] The beneficial effects of this invention are as follows: By automatically classifying signal segments based on frequency domain energy proportion, it adaptively distinguishes between trend-type and fluctuation-type signal segments and matches them with differentiated feature extraction paths. This fundamentally overcomes the shortcomings of traditional one-size-fits-all processing strategies that fail to adequately consider the time-frequency characteristics of signals, achieving a precise characterization of the essential characteristics of signals. Secondly, by combining CEEMDAN decomposition, IMF correlation screening, and segmented adaptive wavelet thresholding technology, a cascaded denoising process is constructed, effectively removing complex noise while retaining key physiological information components. This significantly improves the signal quality and signal-to-noise ratio of subsequent analyses, solving the problems of traditional methods being sensitive to noise and prone to false alarms. Finally, by fusing multi-dimensional features and constructing a linear mapping from features to health indicators, combined with dynamic threshold detection and predefined abnormal template matching, this scheme not only achieves robust and continuous detection of abnormal segments but also further identifies abnormal types, thus providing a deeper level of analysis results from "whether it is abnormal" to "what kind of abnormality," greatly enhancing the clinical applicability and decision support value of the method. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a continuous blood pressure signal anomaly detection method according to an embodiment of the present invention.

[0017] Figure 2 This is a basic flowchart of a continuous blood pressure signal abnormality detection system provided in one embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for detecting abnormal continuous blood pressure signals is provided, comprising the following steps: Step S1: Collect continuous blood pressure signals, segment the continuous blood pressure signals, and automatically classify each signal segment based on the energy ratio of the low frequency band and the preset energy ratio of the high frequency band to obtain a set of signal segments with category labels. Step S2: Perform CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction on a single signal segment in sequence to obtain the denoised signal segment. Step S3: Based on the category label, perform adaptive feature extraction and fusion on the denoised signal segment to obtain the fused feature vector; Step S4: Calculate the health index for the fused feature vector of each signal segment to obtain the health index value of each signal segment, and arrange all the health index values ​​in chronological order to form a continuous health index sequence. Step S5: Based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined abnormality type template, perform abnormality detection and classification, and output the abnormality results.

[0020] In one embodiment, step S1 involves acquiring continuous blood pressure signals using a non-invasive blood pressure monitoring device based on the photoplethysmography (PPG) principle. The acquired continuous blood pressure signals are segmented and automatically classified according to the energy proportions of each segment in preset low-frequency and high-frequency bands, assigning trend or fluctuation category labels. This aims to identify the dominant pattern of the signal at its source, providing a basis for subsequent differentiated processing. Step S2 involves sequentially performing CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction for each signal segment. This series of steps constitutes a refined cascaded denoising process, effectively removing noise interference and preserving valuable physiological components in the signal to the maximum extent. Step S3 involves processing the signal according to the category labels. Step S4: For the denoised segments, the most relevant features, such as trend features or fluctuation features, are adaptively selected and extracted, and fused with a feature reflecting time-series dependence to form a high-dimensional fused feature vector that can comprehensively characterize the state of the segments. Step S5: Then, through a linear mapping relationship learned from historical data, the fused feature vector of each segment is transformed into an intuitive health indicator value, and arranged in chronological order to form a continuous health indicator sequence, thereby mapping the complex feature space to a one-dimensional traceable health trajectory. Step S6: Finally, based on the dynamic threshold detection of the health indicator sequence, abnormal time periods are located, and the fused feature vectors corresponding to the abnormal time periods are matched and classified using predefined abnormal type templates, and the final output includes the judgment result of the specific abnormal type. Through a progressive design—step S1 classification, step S2 denoising, step S3 adaptive feature extraction and fusion, step S4 health indexing, and step S5 anomaly detection and classification—end-to-end automated analysis from raw signals to final anomaly conclusions is achieved. This not only enhances the targeting and representational capabilities of the analysis through signal classification and adaptive feature extraction, but also improves the robustness and interpretability of the detection through cascaded denoising and the construction of health index sequences. Finally, template matching deepens the analysis from anomaly detection to anomaly classification, forming a complete and adaptive intelligent solution for continuous blood pressure signal anomaly detection.

[0021] Step S1: Collect continuous blood pressure signals, segment the continuous blood pressure signals, and automatically classify each signal segment based on the energy ratio of the low frequency band and the preset energy ratio of the high frequency band to obtain a set of signal segments with category labels. Step S101: The continuous blood pressure signal is segmented using a sliding window of preset duration to obtain N signal segments; Step S102: Perform frequency domain transformation on each signal segment to obtain the power spectrum of each signal segment; Step S103: Calculate the low-frequency energy value of the signal segment in the preset low-frequency band and the high-frequency energy value in the preset high-frequency band according to the power spectrum. The low-frequency band energy value is the sum of the power spectral density values ​​of the power spectrum at all frequency points within the preset low-frequency band, and the high-frequency band energy value is the sum of the power spectral density values ​​of the power spectrum at all frequency points within the preset high-frequency band. Step S104: Calculate the low-frequency band energy ratio and the high-frequency band energy ratio of the signal segment; The low-frequency band energy ratio is the ratio of the low-frequency band energy value to the total energy value, and the high-frequency band energy ratio is the ratio of the high-frequency band energy value to the total energy value. The total energy value is the sum of the low-frequency band and high-frequency band energy values. The sum of the energy proportions of the low-frequency band and the high-frequency band is 1.

[0022] Step S105: If the proportion of low-frequency band energy is greater than the first determination threshold, the signal segment is determined to be a trend segment and associated with a trend category label. If the proportion of high-frequency band energy is greater than the second judgment threshold, the signal segment is judged as a wave segment and associated with a wave category label; The signal segment collection consists of all signal segments with category labels, including trend category labels and volatility category labels.

[0023] In one embodiment, step S101 involves segmenting the acquired continuous blood pressure signal using a fixed-duration sliding window. To ensure both the capture of meaningful blood pressure fluctuation patterns (typically including several cardiac cycles and possible trend changes) and the maintenance of temporal resolution, the preset sliding window duration is set to 30 seconds. This setting is primarily based on a balance between the physiological characteristics and practicality of the blood pressure signal. On one hand, a 30-second window typically includes 25 to 60 cardiac cycles, which, at a typical heart rate of 60-100 beats per minute, is sufficient to stably capture the periodic fluctuation patterns and short-term trends of blood pressure. On the other hand, this duration provides sufficient frequency resolution to distinguish between preset low-frequency and high-frequency bands when calculating frequency domain features, while maintaining the response speed of the analysis in the time dimension, making it suitable for continuous monitoring scenarios. Simultaneously, to ensure the continuity of the analysis and avoid information loss, the sliding step size is preferably half the preset sliding window duration, i.e., 15 seconds, thereby obtaining N signal segments that may overlap. This segmentation operation transforms the continuous signal... The analysis unit is discrete but coherent; step S102, each signal segment obtained in step S101 is subjected to frequency domain transformation to obtain its frequency component information. Specifically, a Fast Fourier Transform (FFT) is applied to each signal segment to calculate its power spectrum. This step transforms the signal from the time domain to the frequency domain, laying the foundation for energy analysis; step S103, based on the power spectrum obtained in step S102, the energy value of each signal segment in two preset frequency bands is calculated. The preset low frequency band and preset high frequency band need to be determined based on blood pressure information. The frequencies are divided according to their physiological significance. For example, the low-frequency band is set to [0.04Hz, 0.15Hz), mainly corresponding to the slow regulation trend of blood vessels and the physiological activity of Mayer waves; the high-frequency band is set to [0.15Hz, 0.4Hz], mainly corresponding to respiratory waves and high-frequency fluctuations more related to heart rate. In calculation, the energy value of the low-frequency band is the sum of the power spectral density values ​​at all frequency points in the frequency range of [0.04Hz, 0.15Hz); similarly, the energy value of the high-frequency band is [0.15Hz, 0.4Hz].The sum of power spectral density values ​​at all frequency points within the 4Hz range; Step S104, calculate the energy proportion characteristics of each signal segment. First, calculate the total energy value of the signal segment, which is the sum of the low-frequency band energy value and the high-frequency band energy value, i.e., ignoring the higher-frequency noise band with extremely small energy; then, calculate the low-frequency band energy proportion and the high-frequency band energy proportion; where the low-frequency band energy proportion is the low-frequency band energy value divided by the total energy value, and the high-frequency band energy proportion is the high-frequency band energy value divided by the total energy value, and the sum of the low-frequency band energy proportion and the high-frequency band energy proportion is 1. The low-frequency band energy proportion and the high-frequency band energy proportion quantify the degree of concentration of signal energy in different frequency domains; Step S105, based on the energy proportion calculated in step S104, automatically classify the signal segments, set two judgment thresholds, and have The system consists of a first judgment threshold and a second judgment threshold. The first judgment threshold is used to determine whether a trend is dominant, and the second judgment threshold is used to determine whether a fluctuation is dominant. The values ​​of the first and second judgment thresholds are set to ensure the clarity of the category determination. For example, if the first judgment threshold = the second judgment threshold = 0.6, and the low-frequency band energy proportion of a signal segment is greater than 0.6, it indicates that its energy is mainly concentrated in the low frequency, and it is determined to be a trend segment and associated with a trend category label; if its high-frequency band energy proportion is greater than 0.6, it indicates that its energy is mainly concentrated in the high frequency, and it is determined to be a fluctuation segment and associated with a fluctuation category label. All signal segments assigned category labels constitute the input set for subsequent processing. Through objective frequency domain energy, the automatic and quantitative initial screening of signal patterns is achieved. Step S1 is the key starting point for the entire method to achieve adaptive processing. By using the inherent physical characteristics of the signal's frequency domain energy distribution, rather than fixed rules, the data stream is intelligently divided into trend stream and fluctuation stream. This data-driven classification enables the invention to flexibly cope with the diversity of blood pressure signal manifestations in different individuals and under different conditions, laying a solid first foundation for building a universal and precise anomaly detection system.

[0024] Step S2: Perform CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction on a single signal segment in sequence to obtain the denoised signal segment. Step S201, the CEEMDAN decomposition process includes performing CEEMDAN decomposition on a single signal segment to obtain k intrinsic mode components and residuals; Step S202, the IMF correlation screening process includes calculating the Pearson correlation coefficient for each intrinsic mode component and signal segment to obtain the correlation coefficient. When the correlation coefficient is greater than the correlation threshold, all corresponding intrinsic mode components are taken as the correlation component set. Step S203, the adaptive wavelet thresholding process includes performing 3-level db4 wavelet decomposition on each component in the relevant component set to obtain wavelet coefficients; The wavelet coefficients are processed using a piecewise adaptive threshold function to obtain the processed coefficients. The expression for the piecewise adaptive threshold function is as follows: ; in, This represents the coefficients after processing. Represents wavelet coefficients, Indicates the general threshold. , This represents an estimate of the noise standard deviation. Indicates the length of the wavelet coefficients. Indicates the preset adjustment parameters and If greater than 0, e represents the natural constant; Step S204, the signal reconstruction process includes reconstructing the processed coefficients through inverse wavelet transform to obtain the denoised components; The denoised signal segment is obtained by summing the denoised components and residuals after processing all components in the relevant component set through a piecewise adaptive threshold function and reconstructing them.

[0025] In one embodiment, step S2 involves refined denoising of a single blood pressure signal segment, constructing a cascaded denoising process of decomposition, screening, thresholding, and reconstruction. First, the signal is adaptively decomposed into oscillatory components of different scales using CEEMDAN. Then, components highly correlated with the original signal are screened out, ensuring that the denoising operation focuses on the dominant signal portion and avoids ineffective processing of irrelevant noise components or accidental deletion of weak signal components. Finally, the screened components undergo adaptive, nonlinear wavelet thresholding to achieve an optimal balance between noise elimination and signal detail preservation. This design ensures... The targeted and effective denoising provides high-quality input signals for subsequent feature extraction. Specifically, step S201 involves CEEMDAN decomposition, where a single signal segment is decomposed using CEEMDAN. To balance computational efficiency and decomposition completeness, a key parameter is set: the amplitude (noise standard deviation) of the added white noise is set to 0.2 times the standard deviation of the original signal segment. This setting is based on the premise that sufficiently strong white noise needs to be added to assist the CEEMDAN algorithm in effectively separating modes of different scales and avoiding mode aliasing, while ensuring that the added noise does not excessively distort the waveform of the original signal. The characteristic is that 0.2 times the standard deviation is a commonly used value that has been verified to be effective for blood pressure signals under this balance. The maximum allowed number of iterations is either overall decomposition or 500 iterations for each mode. The rationale is to ensure that the algorithm fully searches and extracts all possible oscillatory components in the signal to achieve complete decomposition. At the same time, to avoid getting bogged down in excessively long and ineffective calculations in the rare cases of slow convergence, 500 iterations is a robust upper limit that effectively controls the computational cost while ensuring that the vast majority of physiological signal segments can be fully decomposed. After performing the decomposition, k intrinsic mode components (IMFs) and one... The residual step adaptively deconstructs the non-stationary signal segment into a series of stationary oscillating components and trend terms arranged from high frequency to low frequency. Step S202 involves IMF correlation screening, calculating the Pearson correlation coefficient between each intrinsic mode component (IMF) and the original signal segment to obtain a set of correlation coefficients. A correlation threshold is set to distinguish between the dominant signal component and the dominant noise component. For example, this correlation threshold is set to 0.3, based on empirical values, and is usually used for the initial separation of effective signals and noise. Setting it to 0.3 ensures that all correlation coefficients are greater than 0.The intrinsic mode components (IMFs) of the 3-element signal are filtered out to form a correlation component set. This step achieves preliminary signal purification, retaining only the oscillation modes strongly correlated with the original signal and excluding obviously uncorrelated high-frequency noise or irrelevant oscillations in advance from the subsequent fine denoising process, thereby improving processing efficiency and reducing the risk of signal distortion. Step S203, adaptive wavelet thresholding, performs discrete wavelet decomposition on each IMF component in the correlation component set. Specifically, a 3-level decomposition can be adopted, using the db4 wavelet basis. After decomposition, the wavelet coefficients of each level are obtained. Subsequently, a piecewise adaptive nonlinear threshold function is used to process these wavelet coefficients. This function is defined as follows: when... At that time, the processed coefficients ;when At that time, the processed coefficients in, Represents the original wavelet coefficients. The coefficients represent the processed values, sgn(·) is the sign function, e is the natural constant, and ε is the universal threshold. The calculation formula is: Where N is the length of the wavelet coefficients and σ is an estimate of the noise standard deviation, which can be robustly estimated by dividing the median absolute value of the first-level detail coefficients (the finest scale in the wavelet decomposition) by 0.6745. The 0.6745 is set based on the scaling parameter correction factor of the standard normal distribution, which determines the median absolute value of the wavelet coefficients when the noise in the signal follows a Gaussian distribution. Dividing the deviation by 0.6745 yields a robust and nearly unbiased estimate of the noise standard deviation σ. This estimation method is insensitive to non-Gaussian components or impulse interference that may exist in the signal, making it particularly suitable for signals such as blood pressure signals that may include physiological spikes. It provides a more stable noise level assessment than directly calculating the standard deviation. p is a preset adjustment parameter; p > 0 controls the smoothness of the threshold function's transition around the threshold ε. A larger p value results in a function shape closer to a hard threshold, which is beneficial for preserving signal characteristics; a smaller p value results in a closer resemblance to a soft threshold, leading to smoother denoising. Empirically, p can be between 3 and 10; for example, an effective implementation sets p to 5. This step involves... This smooth and adaptive threshold function overcomes the shortcomings of traditional hard thresholding (which may introduce pseudo-Gibbs phenomenon) and soft thresholding (which may lead to excessive signal attenuation). It can remove noise remaining in the relevant IMF more precisely, while preserving the abrupt change points and waveform details of the blood pressure signal to the greatest extent. Step S204, signal reconstruction processing: For each IMF in the relevant component set, the wavelet coefficients of each layer after processing in step S203 are inversely transformed (wavelet reconstruction) through the corresponding wavelet basis (db4) to obtain the denoised IMF component. Finally, all the denoised IMF components are summed with the residual obtained from CEEMDAN decomposition to obtain the final denoised signal segment. Step S2 is not a simple denoising process, but a composite denoising framework that integrates adaptive signal decomposition (CEEMDAN), correlation-based component screening, and nonlinear wavelet threshold optimization. Its design logic is as follows: first, CEEMDAN's scale separation capability is used to decouple the mixed signals; then, correlation analysis is used for initial component selection to lock in the processing targets, avoiding resource waste and side effects; finally, refined wavelet thresholding with adjustable parameters and smooth transitions is applied to the selected key components. This divide-and-conquer, targeted, and refined processing strategy collectively achieves the core objective of effectively extracting and enhancing weak but crucial blood pressure physiological signal components against a strong noise background, providing a high-quality, high-fidelity signal foundation for the entire anomaly detection process.

[0026] Step S3: Based on the category label, perform adaptive feature extraction and fusion on the denoised signal segment to obtain the fused feature vector; Step S301, the adaptive feature extraction process includes selecting the corresponding feature extraction path based on the category label; If the category label is a trend category label, then proceed to step S302; Step S302: For the denoised signal segment, a moving average smoothing is performed using a first fixed one-dimensional convolution kernel with coefficients [1 / 3, 1 / 3, 1 / 3] and length 3, and abrupt change point detection is performed using a second fixed one-dimensional convolution kernel with coefficients [1, -2, 1] and length 3 to obtain a trend feature vector; If the category label is a fluctuation category label, then proceed to step S303; Step S303: Perform CEEMDAN decomposition on the denoised signal segment to obtain multiple intrinsic mode components. Select the highest one-third of the components from the multiple intrinsic mode components as high-frequency components. Calculate the instantaneous frequency characteristics and energy characteristics of each high-frequency component to obtain the wave feature vector.

[0027] Step S304: Input the denoised signal segment into a predefined width learning system mapping node. The width learning system mapping node calculates the current output through its mapping function. The input of the mapping function includes the denoised signal segment and the output value generated by the width learning system mapping node when processing the previous signal segment. The current output is used as the temporal feature vector. Step S305: Select the corresponding feature combination method according to the category label and fuse them to obtain the fused feature vector; If the selected logical dimension is a trend category label, then the trend feature vector and the time series feature vector are concatenated to obtain a fused feature vector. If the label is a fluctuation category label, the fluctuation feature vector and the time series feature vector are concatenated to obtain a fused feature vector.

[0028] In one embodiment, step S3 adaptively extracts and fuses the most discriminative features based on the inherent pattern category of the signal segments. This abandons the traditional approach of using the same feature extraction strategy for all signals. Instead, it intelligently selects and calculates features representing the dynamic characteristics of the signal type based on the trend category label or fluctuation category label determined in step S1, and uniformly fuses a feature reflecting the temporal dependency between signal segments. This constructs a high-dimensional representation that can both reflect the local characteristics within segments and characterize the global evolution between segments. This design ensures the high degree of specificity of feature engineering, laying a precise foundation for subsequent health indicator mapping. Based on this, the specific steps include: Step S301, initiating the adaptive feature extraction path selection logic, the core of which is to branch to different feature calculation processes based on the trend category label or fluctuation category label of the input signal segment; Step S302, for the signal segment determined to be of the trend type, its feature extraction is to capture its slowly changing trend and possible abrupt change points, such as the rapid rise or fall of blood pressure. The specific steps are as follows: First, a first fixed one-dimensional convolution kernel with coefficients of [1 / 3, 1 / 3, 1 / 3] and a length of 3 is used, i.e., a three-point moving average filter, to perform convolution operation on the denoised signal segment to achieve signal... The smoothed sequence is used to highlight the baseline trend and is taken as the first trend sub-feature. Next, a second fixed one-dimensional convolution kernel with coefficients [1,-2,1] and length 3 is used to convolve the same signal segment. This operation is equivalent to the second-order difference of the discrete approximation, the Laplacian operator, and inflection points or abrupt change regions in the highlighted signal. The resulting abrupt change detection sequence is taken as the second trend sub-feature. Finally, local statistics such as mean, variance, and maximum are typically extracted from these smoothed and abrupt change detection sequences to form a trend feature vector. For example, a simple implementation is to extract the mean of the smoothed sequence to reflect the average... The trend level is calculated, and the maximum absolute value of the mutation detection sequence is extracted to reflect the intensity of the most dramatic instantaneous change, thus forming a 2D trend feature vector. This is only an exemplary way to construct the feature vector. This step effectively extracts the macroscopic trend and key change point information of the signal in the time domain with low computational cost. In step S303, for the signal segment identified as a fluctuation, its feature extraction is to quantify the subtle patterns of its high-frequency oscillations, such as frequency modulation and energy distribution. Specifically, the segment is first decomposed again by CEEMDAN, with the parameter settings consistent with step S201, for example, the noise amplitude is 0.Using a standard deviation of 2 and a maximum iteration of 500, multiple intrinsic mode components (IMFs) are obtained. Then, based on the characteristic that IMF components are typically arranged in descending order of frequency, the top third of the IMF components with the highest frequency are selected as the high-frequency component set. For example, if a decomposition yields 6 IMFs, the first two, i.e., IMFs with indices 1 and 2, are selected as the high-frequency components. Next, for each IMF component in the high-frequency component set, instantaneous frequency characteristics and energy characteristics are calculated. Instantaneous frequency characteristics are typically obtained by performing a Hilbert transform on the IMF to obtain its analytic signal, then calculating the instantaneous phase, and finally differentiating the instantaneous phase to obtain the instantaneous frequency sequence. Statistical features, such as mean and standard deviation, are then extracted from this sequence. Energy characteristics are calculated by... The sum of squares or root mean square of the amplitudes of the IMF components is used as their energy. The instantaneous frequency and energy characteristics of all these high-frequency components are arranged in order to form the fluctuation feature vector. This step focuses on the high-frequency components through quadratic decomposition and extracts their instantaneous physical parameters, which can finely characterize the high-frequency fluctuation patterns in the blood pressure signal, such as respiratory modulation and short-term arrhythmia. Step S304, temporal feature vector extraction, this step injects the temporal correlation information between each signal segment and historical segments to make up for the lack of contextual memory of the features within the aforementioned segments. The specific steps are to maintain a mapping node of a wide learning system (BLS), which has a simple mapping function. The specific form of the mapping function can be a fully connected layer plus a nonlinear activation function, for example. ;in This represents the temporal feature vector of the current segment. This refers to the current time step, i.e., the denoised signal of the current signal segment. It usually needs to be converted into a fixed-dimensional representation first, for example, by taking several statistical measures such as its mean and standard deviation to form a low-dimensional input vector. This is the output value generated by the mapping node after processing the previous signal segment, i.e., the hidden state at the previous time step. W and U are weight matrices, b is the bias vector, and f is the activation function, such as Sigmoid or Tanh. The weights and biases are randomly generated and fixed during system initialization, for example, sampled from a uniform distribution [-0.1, 0.1]. The output of the mapping node... The dimensions can be preset, for example, 5 dimensions, which is the temporal feature vector of the current segment. This step uses a very simple recursive structure to encode the dynamic change information of each segment relative to the previous segment with negligible computational overhead, enhancing the continuity of features on the time axis. Step S305, feature fusion, according to the category label, combines the segment feature vector and the temporal feature vector obtained in the above steps. If it is a trend category label, the trend feature vector obtained in step S302 and the temporal feature vector obtained in step S304 are concatenated to form the final fused feature vector. If it is a fluctuation category label, the fluctuation feature vector obtained in step S303 and the temporal feature vector obtained in step S304 are concatenated to form the final fused feature vector. Vector concatenation means connecting all elements of two vectors in order to form a new longer one-dimensional vector. For example, if the trend feature vector is 2-dimensional and the temporal feature vector is 5-dimensional, then the fused feature vector is 7-dimensional. Step S3 is the core step in achieving differential analysis and context awareness. Through path selection (steps S301-S303), the most suitable feature descriptors are used for the two fundamentally different physiological patterns of trend and fluctuation, which greatly improves the discriminative power and physical interpretability of the features. At the same time, by introducing a lightweight memory unit, namely step S304, the temporal correlation information is supplemented for each isolated segment, enabling the features to reflect the dynamic evolution of the signal. This feature construction strategy, which combines category adaptation and temporal memory, ensures that the final fused feature vector can comprehensively and accurately represent the health status of blood pressure signal segments, thereby directly supporting the accuracy and reliability of subsequent health indicator calculations (step S4).

[0029] S4. Calculate the health index for the fused feature vector of each signal segment to obtain the health index value of each signal segment, and arrange all the health index values ​​in chronological order to form a continuous health index sequence. Step S401: Based on a set of fused feature vectors as historical samples and corresponding physiological state reference values, the least squares method is used to fit and obtain the mapping weight vector. Based on the mapping weight vector, a linear mapping relationship from the fused feature vector to the health score is constructed. The fused feature vectors are mapped to a preliminary health score using a linear mapping relationship; Step S402: Normalize the preliminary health score to obtain health indicator values; Step S403: Arrange the health indicator values ​​corresponding to all signal segments in chronological order to form a continuous health indicator sequence.

[0030] In one embodiment, step S4 transforms the high-dimensional, heterogeneous fused feature vector into an intuitive, single, and time-trackable health metric. This step establishes a stable mapping from a complex feature space to a one-dimensional health score using a linear model. This model is used to realize the linear mapping relationship and standardize the score, ultimately forming an indicator curve reflecting continuous changes in health status. This design achieves a key transformation from feature analysis to status assessment. Specifically, step S401 involves constructing and applying a linear mapping model from the fused feature vector to the health score. The purpose of this step is to utilize a set of historical sample data based on clinical standard annotations. The study aims to learn a mathematical model that comprehensively assesses the health status of blood pressure signal segments. Specifically, it involves first preparing a sufficiently large set of historical sample data to ensure the statistical robustness and generalization ability of the learned model. This dataset includes m samples, each consisting of two parts: a fused feature vector (generated in step S3) and a corresponding physiological state reference value. This reference value is set based on authoritative clinical guidelines or standards to ensure the medical validity of the model's learning objectives. For example, it can be calibrated according to the blood pressure level classifications specified in the American Heart Association (AHA) guidelines for hypertension management, specifically including those within the normal blood pressure range. Signal segments with systolic blood pressure less than 120 mmHg and diastolic blood pressure less than 80 mmHg are assigned a high reference value, such as 95; segments in the clearly defined hypertension range (e.g., systolic blood pressure greater than or equal to 140 mmHg or diastolic blood pressure greater than or equal to 90 mmHg) or hypotension range (systolic blood pressure less than 90 mmHg) are assigned a low reference value, such as 20; borderline states can be assigned an intermediate value, such as 60. These reference values ​​are directly derived from labeled data in publicly available clinical databases (e.g., MIMIC-III) that conform to these criteria. Then, a least squares method is used to perform a linear fit on this set of sample data, with the goal of finding a set of mapping weight vectors. This includes a bias term that minimizes the mean square error between the predicted score and the true reference value of the linear model. After successful fitting, a linear mapping relationship from any fused feature vector to the preliminary health score is established, which is a weighted summation formula. Finally, for any new signal segment to be evaluated, the fused feature vector output from step S3 is substituted into this linear relationship, and the calculated value is the preliminary health score of the segment. This step learns an interpretable model that can map multidimensional features to a one-dimensional score that conforms to the consensus of medical assessment through a data-driven approach based on clinical standard annotation, thus realizing the objectification, quantification and automation of assessment.Step S402 involves normalizing the preliminary health score to obtain a standardized health indicator value. The purpose of this step is to eliminate potential differences in scoring scales caused by different models or data batches, ensuring that the final health indicator value falls within a fixed and intuitive range. This facilitates setting a unified detection threshold and understanding. Specifically, based on the set of preliminary health scores from historical samples, the minimum and maximum values ​​of the scores are determined. Then, a minimum-maximum normalization method is used to map the preliminary health score of the current segment to a preset standard interval [0, 100]. The specific calculation method is to subtract the historical minimum score from the current score, divide by the historical score range (maximum minus minimum), and finally multiply by 100. The calculated value is the final health indicator value. After normalization, the health indicator value is limited to between 0 and 100. A higher value indicates a healthier blood pressure signal for that segment, while a lower value indicates a higher risk of abnormality. This step produces a standardized health indicator value. Standardized, dimensionless, and comparable health scalars with fixed ranges greatly simplify the setting and interpretation of subsequent global or individualized thresholds. Step S403 forms a continuous health index sequence. The purpose of this step is to organize the health assessment results of all discrete time segments according to their original time order to present the dynamic evolution of health status. Specifically, according to the time sequence of the signal segments in the original continuous blood pressure signal, the health index values ​​calculated by each segment in steps S401 and S402 are extracted sequentially, and these values ​​are stored in a list in order. This sequence is the continuous health index sequence, which is a one-dimensional curve that changes over time. It intuitively shows the continuous fluctuation of the health status of the monitored person's blood pressure signal. This step reconstructs the fragmented analysis results into a continuous monitoring trajectory, providing a direct data basis for detecting abnormal patterns of trend deterioration and sudden decline in health status in the time dimension. Step S4 acts as an evaluator and standardization interface, summarizing the results of all previous signal processing and feature engineering, i.e., the fused feature vectors, and outputting a unified health metric through a simple linear model pre-calibrated by clinical standards. This design ensures the objectivity, consistency, and medical rationality of the evaluation process. At the same time, the normalization operation standardizes the output to a fixed range, enabling the method to adapt to different individual baselines or different modeling data, enhancing the robustness and versatility of the system. The final generated continuous health indicator sequence is the core bridge connecting the complex signal analysis at the front end with the intuitive abnormality judgment strategy at the back end.

[0031] Step S5: Based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined abnormality type template, perform abnormality detection and classification, and output the abnormality results. Step S501: Calculate local statistics in the health indicator sequence using a sliding window method, and dynamically generate anomaly detection thresholds based on the local statistics. Step S502: If the values ​​of M+1 consecutive segments traced back from the current segment in the health indicator sequence are all less than their respective dynamic thresholds, then the M+1 segments are determined to be abnormal segments. Step S503: Using a predefined anomaly type template, match the fusion feature vector corresponding to the anomaly fragment to obtain the anomaly type label; Step S504: Based on the judgment result of the abnormal segment and the abnormal type label, generate and output the detection result; The detection results include anomaly warning signals and anomaly type labels.

[0032] The process of constructing a predefined abnormality type template includes obtaining O blood pressure signal sample fragments with known abnormality types from historical data; For each sample segment, steps S1 to S4 are executed sequentially to obtain the corresponding fused feature vector; For sample segments belonging to the same anomaly type, calculate the statistical center vector of the fusion feature vector, and use the statistical center vector as the template feature vector representing the anomaly type. The template feature vectors of all exception types constitute a predefined exception type template.

[0033] In one embodiment, step S5, based on a continuous health indicator sequence and its corresponding features, achieves robust detection and fine classification of blood pressure abnormalities. First, a dynamic threshold overcomes the insensitivity of fixed thresholds to individual differences and signal baseline drift, enabling adaptive initial screening of abnormalities. Second, by requiring multiple consecutive health indicator values ​​to be below their dynamic threshold, false alarms caused by transient interference are suppressed, improving detection stability. Finally, through matching with predefined templates, detected abnormalities are further classified into specific physiological or pathological patterns, achieving the transformation from abnormality detection to abnormality type identification, thus improving the clinical accuracy of the results. The design ensures the reliability, interpretability, and practicality of the test results, providing valuable guidance for clinical application. Specifically, step S501 involves generating a dynamic anomaly detection threshold. This step aims to set a time-varying anomaly judgment threshold for the health indicator sequence that adapts to its local fluctuations. Specifically, a sliding window containing W health indicator values ​​moves across the sequence. The window length W needs to balance sensitivity and stability; for example, setting it to 10 corresponds to approximately 2.5 minutes of data. This is based on the 30-second segment duration and 15-second step size settings in step S101. The rationale is that, from a physiological perspective, 2.5... The window duration of 10 minutes is sufficient to encompass clinically significant slow-moving trends or intermittent abnormalities in blood pressure, while avoiding detection delays caused by excessively long windows. From a signal processing perspective, the number of data points included in this window (10) provides sufficient statistical samples for calculating the local mean and standard deviation, thus ensuring the robustness of the estimated dynamic threshold. At each window location, the mean and standard deviation of all health indicator values ​​within the window are calculated as local statistics. Then, based on the local statistics, the abnormality detection threshold corresponding to the current window center point is dynamically generated. An effective threshold calculation method is to subtract the adjustment factor from the local mean. The adjustment coefficient is k times the standard deviation, where the adjustment coefficient k is a positive number used to control the sensitivity of the test. Its setting is mainly based on a balance between statistical principles and practical application effects. Under the assumption that the data approximately follows a normal distribution, the probability of a measured value falling outside the mean ± 2 standard deviations is less than 5%, and the probability of falling outside the mean ± 3 standard deviations is extremely small (approximately 0.3%). Therefore, it is usually set to a value between 2 and 3 based on this principle. For example, setting k=2.5 is a commonly used empirical value that strikes a balance between avoiding excessive false positives and false negatives, meaning it will be lower than the local mean of 2.Health indicator values ​​below 5 standard deviations are considered potential abnormalities. The threshold can automatically adjust to the local average level and fluctuation range of the health indicator sequence, thus adapting to changes in the signal baseline over different time periods and significantly reducing false negatives and false negatives caused by individual differences or slow changes in physiological state. Step S502, determination of continuous abnormal segments: The purpose of this step is to distinguish true pathological events from transient noise or artifact interference through the logic of continuous abnormalities. Specifically, it includes setting a number M of continuous abnormal segments. The setting of M is based on ensuring the persistence of abnormal events. Typically, M is not less than 1. Considering the step size of the sliding window, for example, 15 seconds, to capture abnormal events lasting at least about 1 minute, M=3, i.e., 4 consecutive segments. Starting from the current segment to be judged, trace back and check M+1 consecutive segments, including the current segment. The checking condition is that the health indicator value of each of these M+1 segments must be less than its corresponding dynamic threshold, which is generated by step S501 at each segment's time point. If this condition is met, the time period consisting of these M+1 consecutive segments is determined to be an abnormal segment set. This step, by requiring the abnormal state to be continuous in time, effectively filters out isolated abnormal points caused by instantaneous measurement errors, motion artifacts, or other random interferences, significantly improving the specificity and reliability of abnormal detection. Step S503, abnormal classification based on template matching, aims to identify pathological or physiological patterns in the detected abnormal segments. Specifically, it includes pre-constructing an abnormal type template library. After a set of abnormal segments is determined in step S502, the fusion feature vectors corresponding to these segments are extracted. For the abnormal time period, the average vector of the fusion feature vectors of all segments is usually calculated as the comprehensive feature representation of the abnormal event. Then, the similarity of this average feature vector with each template feature vector in the template library is calculated. The similarity calculation uses cosine similarity. For example, the cosine similarity with each template vector is calculated, and the one with the highest similarity and exceeding a preset minimum matching threshold, such as 0, is selected.7. The threshold is set based on the fact that, during model training and validation, this value can effectively distinguish between different categories of abnormal patterns while ensuring sufficient recall for similar abnormal patterns. It is a verified balance point. The type corresponding to this template serves as the abnormal type label for the abnormal event. If the highest similarity is lower than the lowest matching threshold, it can be marked as an unclassified abnormality. This step elevates abnormality detection from binary judgment to multi-class recognition, enabling the differentiation of different types such as persistent hypertension, paroxysmal sudden drops in blood pressure, and periodic large fluctuations, providing more targeted information for subsequent clinical analysis and intervention. Step S504, generation and output of detection results, aims to format and deliver the results. The final detection conclusion specifically includes summarizing the abnormal time period determination results from step S502 (usually including the abnormal start time and end time) and the corresponding abnormal type labels obtained from step S503, generating a structured detection result. This result consists of two parts: an abnormality occurrence warning signal and an abnormality type label. The abnormality occurrence warning signal indicates that an abnormality was detected within a specific time period. The abnormality type label describes which predefined type the abnormality is most likely to belong to, and the output format is a visual report, database record, or real-time warning message. This transforms the complex analysis results in the above process into clear, actionable clinical or monitoring information, completing the final closed loop of the entire automated detection process. The process of constructing predefined abnormality type templates includes: First, collecting O blood pressure signal sample fragments of known abnormality types from historical data. The number O should be sufficient to ensure the statistical significance of each template type, for example, each type includes 20-50 sample fragments. For each sample fragment, steps S1 to S4 of this patent method are executed independently and completely to obtain its corresponding fusion feature vector. Then, the fusion feature vectors of all sample fragments belonging to the same abnormality type are aggregated. The statistical center vector of these vectors is calculated. The most common and effective method is to calculate their arithmetic mean vector, that is, to calculate the average of each dimension of all vectors of the same type to obtain a new vector of the same dimension. This vector is the template feature vector representing this type of abnormality. This process is repeated for each abnormality type that needs to be identified. Finally, the template feature vectors of all types constitute the predefined abnormality type template library. Step S5 is the final step in realizing the value of this method. It designs a three-level progressive discrimination pipeline of dynamic initial screening, continuous confirmation, and fine classification. The dynamic threshold (step S501) ensures the individualized adaptability of the detection; the continuous judgment (step S502) ensures the robustness of the detection; and the template matching (step S503) realizes the automated and data-driven identification of abnormality types, avoiding the limitations of traditional methods that only alarm but do not differentiate symptoms. This design can not only reliably detect abnormalities, but also provide abnormality type information with clinical discrimination, thereby realizing the functional sublimation from simple abnormality monitoring to preliminary abnormality auxiliary diagnosis.

[0034] Example 2, refer to Figure 2 In another embodiment of the present invention, which differs from the first embodiment, a continuous blood pressure signal abnormality detection system is provided, including a classification module, a decomposition module, a fusion module, a calculation module, and a detection module; The classification module collects continuous blood pressure signals, segments the continuous blood pressure signals, and automatically classifies them based on the energy ratio of each signal segment in the low-frequency band and the preset energy ratio in the high-frequency band, thus obtaining a set of signal segments with category labels. The decomposition module sequentially performs CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction on a single signal segment to obtain a denoised signal segment. The fusion module performs adaptive feature extraction and fusion on the denoised signal segments based on category labels to obtain a fused feature vector. The calculation module performs health index calculations on the fused feature vector of each signal segment to obtain the health index value of each signal segment, and arranges all health index values ​​in chronological order to form a continuous health index sequence. The detection module performs anomaly detection and classification based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined anomaly type template, and outputs the anomaly results.

[0035] In one embodiment, the classification module acquires continuous blood pressure signals using a non-invasive blood pressure monitoring device based on the photoplethysmography (PPG) principle. The acquired continuous blood pressure signals are segmented and automatically classified according to the energy proportions of each segment in a preset low-frequency and high-frequency band, assigning trend or fluctuation category labels. This aims to identify the dominant pattern of the signal at its source, providing a basis for subsequent differentiated processing. The decomposition module, for a single signal segment, sequentially performs CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction. This series of steps constitutes a refined cascaded denoising process, effectively removing noise interference and preserving valuable physiological components in the signal to the maximum extent. The fusion module, based on the category labels... The system adaptively selects and extracts the most relevant features, such as trend or fluctuation features, for the denoised segments and fuses them with a feature reflecting time-series dependence to form a high-dimensional fused feature vector that can comprehensively characterize the state of the segments. The calculation module then transforms the fused feature vector of each segment into an intuitive health indicator value through a linear mapping relationship learned from historical data, and arranges them in chronological order to form a continuous health indicator sequence, thereby mapping the complex feature space to a one-dimensional traceable health trajectory. Finally, the detection module locates abnormal periods based on the dynamic threshold detection of the health indicator sequence, and uses a predefined abnormality type template to match and classify the fused feature vectors corresponding to the abnormal periods, ultimately outputting a judgment result including the specific abnormality type. This invention overcomes the shortcomings of traditional one-size-fits-all processing strategies that fail to adequately consider the time-frequency characteristics of signals by automatically classifying signal segments based on frequency domain energy proportions and matching them with differentiated feature extraction paths. This achieves a precise characterization of the essential characteristics of signals. Secondly, by combining CEEMDAN decomposition, IMF correlation screening, and segmented adaptive wavelet thresholding technology, a cascaded denoising process is constructed, effectively removing complex noise while retaining key physiological information components. This significantly improves the signal quality and signal-to-noise ratio of subsequent analyses and solves the problems of traditional methods being sensitive to noise and prone to false alarms. Finally, by fusing multi-dimensional features and constructing a linear mapping from features to health indicators, combined with dynamic threshold detection and predefined abnormal template matching, this solution not only achieves robust and continuous detection of abnormal segments but also further identifies the type of abnormality. This provides a deeper level of analysis results, from "whether it is abnormal" to "what kind of abnormality," greatly enhancing the clinical applicability and decision support value of the method.

[0036] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0037] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for detecting abnormal continuous blood pressure signals, characterized in that, Includes the following steps, Step S1: Collect continuous blood pressure signals, segment the continuous blood pressure signals, and automatically classify each signal segment based on the energy ratio of the low frequency band and the preset energy ratio of the high frequency band to obtain a set of signal segments with category labels. Step S2: Perform CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction on a single signal segment in sequence to obtain the denoised signal segment. Step S3: Based on the category label, perform adaptive feature extraction and fusion on the denoised signal segment to obtain the fused feature vector; Step S4: Calculate the health index for the fused feature vector of each signal segment to obtain the health index value of each signal segment, and arrange all the health index values ​​in chronological order to form a continuous health index sequence. Step S5: Based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined abnormality type template, perform abnormality detection and classification, and output the abnormality results.

2. The method for detecting abnormal continuous blood pressure signals as described in claim 1, characterized in that, Step S1: Collect continuous blood pressure signals, segment the continuous blood pressure signals, and automatically classify each signal segment based on the energy ratio of the low frequency band and the preset energy ratio of the high frequency band to obtain a set of signal segments with category labels. Step S101: The continuous blood pressure signal is segmented using a sliding window of preset duration to obtain N signal segments; Step S102: Perform frequency domain transformation on each signal segment to obtain the power spectrum of each signal segment; Step S103: Based on the power spectrum, calculate the low-frequency band energy value of the signal segment in the preset low-frequency band and the high-frequency band energy value in the preset high-frequency band, respectively. The low-frequency band energy value is the sum of the power spectral density values ​​of the power spectrum at all frequency points within the preset low-frequency band, and the high-frequency band energy value is the sum of the power spectral density values ​​of the power spectrum at all frequency points within the preset high-frequency band. Step S104: Calculate the low-frequency band energy ratio and the high-frequency band energy ratio of the signal segment; The low-frequency band energy ratio is the ratio of the low-frequency band energy value to the total energy value, the high-frequency band energy ratio is the ratio of the high-frequency band energy value to the total energy value, and the total energy value is the sum of the low-frequency band and high-frequency band energy values; The sum of the energy proportions of the low-frequency band and the high-frequency band is 1.

3. The method for detecting abnormal continuous blood pressure signals as described in claim 2, characterized in that, Step S105: If the proportion of low-frequency band energy is greater than the first determination threshold, the signal segment is determined to be a trend segment and associated with a trend category label. If the high-frequency band energy ratio is greater than the second determination threshold, the signal segment is determined to be a wave-type segment and associated with a wave category label; The signal segment collection consists of all signal segments with category labels, including trend category labels and volatility category labels.

4. The method for detecting abnormal continuous blood pressure signals as described in claim 3, characterized in that, Step S2: Perform CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction on a single signal segment in sequence to obtain the denoised signal segment. Step S201, the CEEMDAN decomposition process includes performing CEEMDAN decomposition on a single signal segment to obtain k intrinsic mode components and residuals; Step S202, the IMF correlation screening process includes calculating the Pearson correlation coefficient for each intrinsic mode component and signal segment to obtain the correlation coefficient. When the correlation coefficient is greater than the correlation threshold, all corresponding intrinsic mode components are taken as the correlation component set. Step S203, the adaptive wavelet thresholding process includes performing 3-level db4 wavelet decomposition on each component in the relevant component set to obtain wavelet coefficients; The wavelet coefficients are processed using a piecewise adaptive threshold function to obtain the processed coefficients. The expression for the piecewise adaptive threshold function is as follows: ; in, This represents the coefficients after processing. Represents wavelet coefficients, Indicates the general threshold. , This represents an estimate of the noise standard deviation. Indicates the length of the wavelet coefficients. Indicates the preset adjustment parameters and If the value is greater than 0, then e represents the natural constant; Step S204, the signal reconstruction process includes reconstructing the processed coefficients through inverse wavelet transform to obtain the denoised components; The denoised signal segment is obtained by summing the denoised components and residuals after all components in the relevant component set have been processed by the piecewise adaptive threshold function and reconstructed.

5. The method for detecting abnormal continuous blood pressure signals as described in claim 4, characterized in that, Step S3: Based on the category label, perform adaptive feature extraction and fusion on the denoised signal segment to obtain a fused feature vector; Step S301, the adaptive feature extraction process includes selecting the corresponding feature extraction path based on the category label; If the category label is a trend category label, then proceed to step S302; Step S302: For the denoised signal segment, a moving average smoothing is performed using a first fixed one-dimensional convolution kernel with coefficients [1 / 3, 1 / 3, 1 / 3] and length 3, and abrupt change point detection is performed using a second fixed one-dimensional convolution kernel with coefficients [1, -2, 1] and length 3 to obtain a trend feature vector; If the category label is a fluctuation category label, then proceed to step S303; Step S303: Perform CEEMDAN decomposition on the denoised signal segment to obtain multiple intrinsic mode components. Select the highest one-third of the components from the multiple intrinsic mode components as high-frequency components. Calculate the instantaneous frequency characteristics and energy characteristics of each high-frequency component to obtain the wave feature vector.

6. The method for detecting abnormal continuous blood pressure signals as described in claim 5, characterized in that, Step S304: Input the denoised signal segment into a predefined width learning system mapping node. The width learning system mapping node calculates the current output through its mapping function. The input of the mapping function includes the denoised signal segment and the output value generated by the width learning system mapping node when processing the previous signal segment. The current output is used as a temporal feature vector. Step S305: Select the corresponding feature combination method according to the category label and fuse them to obtain the fused feature vector; If the selected logical dimension is a trend category label, then the trend feature vector and the time series feature vector are concatenated to obtain a fused feature vector. If the label is a fluctuation category label, the fluctuation feature vector and the time series feature vector are concatenated to obtain a fused feature vector.

7. The method for detecting abnormal continuous blood pressure signals as described in claim 6, characterized in that, Step S4: Calculate the health index for the fused feature vector of each signal segment to obtain the health index value of each signal segment, and arrange all the health index values ​​in chronological order to form a continuous health index sequence. Step S401: Based on a set of fused feature vectors as historical samples and corresponding physiological state reference values, the least squares method is used to fit and obtain a mapping weight vector. Based on the mapping weight vector, a linear mapping relationship from the fused feature vector to the health score is constructed. The fused feature vectors are mapped to a preliminary health score using a linear mapping relationship; Step S402: Normalize the preliminary health score to obtain the health indicator value; Step S403: Arrange the health indicator values ​​corresponding to all signal segments in chronological order to form a continuous health indicator sequence.

8. The method for detecting abnormal continuous blood pressure signals as described in claim 7, characterized in that, Step S5: Based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined abnormality type template, perform abnormality detection and classification, and output the abnormality results. Step S501: Calculate local statistics in the health indicator sequence using a sliding window method, and dynamically generate anomaly detection thresholds based on the local statistics. Step S502: If the values ​​of M+1 consecutive segments traced back from the current segment in the health indicator sequence are all less than their respective dynamic thresholds, then the M+1 segments are determined to be abnormal segments. Step S503: Using the predefined anomaly type template, match the fusion feature vector corresponding to the anomaly fragment to obtain an anomaly type label; Step S504: Based on the judgment result of the abnormal segment and the abnormal type label, generate and output the detection result; The detection results include anomaly warning signals and anomaly type labels.

9. The method for detecting abnormal continuous blood pressure signals as described in claim 8, characterized in that, The process of constructing the predefined abnormality type template includes obtaining O blood pressure signal sample fragments with known abnormality types from historical data; For each of the sample segments, steps S1 to S4 are executed sequentially to obtain the corresponding fused feature vector; For sample segments belonging to the same anomaly type, calculate the statistical center vector of the fusion feature vector, and use the statistical center vector as the template feature vector representing the anomaly type. The template feature vectors of all exception types constitute a predefined exception type template.

10. A continuous blood pressure signal anomaly detection system, applied to a continuous blood pressure signal anomaly detection method as described in any one of claims 1-9, characterized in that, It includes a classification module, a decomposition module, a fusion module, a calculation module, and a detection module; The classification module collects continuous blood pressure signals, segments the continuous blood pressure signals, and automatically classifies them based on the energy ratio of each signal segment in the low-frequency band and the preset energy ratio in the high-frequency band, thus obtaining a set of signal segments with category labels. The decomposition module sequentially performs CEEMDAN decomposition, IMF correlation screening, adaptive wavelet thresholding, and signal reconstruction on a single signal segment to obtain a denoised signal segment. The fusion module performs adaptive feature extraction and fusion on the denoised signal segments based on category labels to obtain a fused feature vector. The calculation module performs health index calculations on the fused feature vector of each signal segment to obtain the health index value of each signal segment, and arranges all health index values ​​in chronological order to form a continuous health index sequence. The detection module performs anomaly detection and classification based on the continuous health indicator sequence, the corresponding fusion feature vector, and the predefined anomaly type template, and outputs the anomaly results.

Images