Non-invasive blood pressure simulator fault intelligent diagnosis method

Abstract

The invention discloses a non-invasive blood pressure simulator fault intelligent diagnosis method, and relates to the technical field of medical equipment fault detection, and the method comprises the steps: preprocessing the operation data of a non-invasive blood pressure simulator to obtain a data matrix; performing principal component analysis on the data matrix to obtain a plurality of principal components, constructing a principal component projection matrix based on the principal components, and performing dimension reduction on the data matrix based on the principal component projection matrix to obtain principal component data; obtaining score data based on the principal component data and the principal component projection matrix, and weighting the score data to obtain an abnormal score; if the abnormal score is larger than or equal to a preset abnormal threshold value, it is judged that the non-invasive blood pressure simulator is abnormal, and the problem that an existing intelligent fault diagnosis method for the non-invasive blood pressure simulator has high-dimensional data redundancy and feature coupling and cannot adapt to abnormal mode diversity, and consequently fault recognition accuracy is low is solved.

Description

Technical Field

[0001] This invention relates to the field of medical equipment fault detection technology, and more specifically, to an intelligent diagnostic method for non-invasive blood pressure simulator faults. Background Technology

[0002] Non-invasive blood pressure simulators are crucial tools for the calibration and testing of medical devices such as electronic blood pressure monitors and patient monitors, and their accuracy and stability are paramount. Existing intelligent fault diagnosis methods for non-invasive blood pressure simulators typically employ univariate thresholding or isolated forest models. However, the data generated during the operation of non-invasive blood pressure simulators (such as pressure control signals, flow sensor readings, valve action sequences, and pressure decay curves) are characterized by high dimensionality, strong temporal sequence, and strong correlation among features. Univariate thresholding methods struggle to fully exploit the data structure and effectively handle complex nonlinear relationships between features, leading to information loss or redundancy and affecting detection sensitivity. Isolated forest models are typically only sensitive to a specific type of anomaly (such as point anomalies or contextual anomalies), exhibiting limited adaptability to different anomaly types, resulting in high false alarm and false negative rates and low fault identification accuracy. Summary of the Invention

[0003] To address the problems of high-dimensional data redundancy, feature coupling, and inability to adapt to diverse anomaly patterns in existing intelligent fault diagnosis methods for non-invasive blood pressure simulators, resulting in low fault identification accuracy, this invention provides an intelligent fault diagnosis method for non-invasive blood pressure simulators, the method comprising:

[0004] Preprocess the operating data of the non-invasive blood pressure simulator to obtain a data matrix;

[0005] Principal component analysis is performed on the data matrix to obtain several principal components. A principal component projection matrix is ​​constructed based on the principal components. The data matrix is ​​then dimensionality-reduced based on the principal component projection matrix to obtain principal component data.

[0006] Score data is obtained based on the principal component data and the principal component projection matrix, and abnormal scores are obtained by weighting the score data.

[0007] If the abnormal score is greater than or equal to the preset abnormal threshold, the non-invasive blood pressure simulator is determined to be abnormal.

[0008] Potential drivers of failure modes: Equipment failures are often driven by a few core factors that are masked by noise and irrelevant features in the original high-dimensional data.

[0009] This method introduces Principal Component Analysis (PCA) and dimensionality reduction, utilizing the enhancement mechanisms of PCA: 1) Variance maximization to preserve key information: PCA prioritizes retaining the principal components with the largest variance (i.e., containing the most significant changes in the data). These principal components often correspond to the core driving patterns of faults (such as the trend of pressure changes), thus capturing the potential driving patterns of faults; 2) Noise filtering: Principal components with low variance usually correspond to noise or minor disturbances. By discarding these components, irrelevant fluctuations can be suppressed, making fault characteristics more prominent, thereby improving the separation between fault types (such as normal and abnormal states being easier to distinguish in low-dimensional space), thus improving fault separation and overcoming interference resistance.

[0010] Collinearity is essentially a phenomenon where multiple features in high-dimensional data may be highly linearly correlated due to their physical relationship, leading to data redundancy and model overfitting.

[0011] This method utilizes the solution mechanism of PCA: 1) Covariance matrix decomposition: PCA calculates the covariance matrix of the data and decomposes it into eigenvalues ​​(variance) and eigenvectors (principal component directions). The variance of collinear features will be integrated into a few principal components, reducing redundant correlations between features; 2) Orthogonal principal components: The principal components are orthogonal to each other (correlation is 0), which mathematically eliminates collinearity and avoids the multicollinearity trap during model training (such as unstable regression coefficients), thereby solving the problem of collinearity in high-dimensional data.

[0012] The essence of the curse of dimensionality is that data sparsity increases dramatically in high-dimensional spaces, requiring a large number of samples for models to fit patterns, which easily leads to overfitting; at the same time, high-dimensional computation has high complexity (such as distance calculation and parameter training costs increase exponentially).

[0013] This method utilizes the optimization mechanisms of PCA: 1) Dimensionality reduction to reduce redundant features: By retaining a small number of core principal components (e.g., using 2-3 principal components to explain more than 95% of the variance), the data dimensionality is reduced, the number of parameters that the model needs to learn is reduced, and the risk of overfitting is reduced; 2) Increased data density: Data samples are more densely distributed in low-dimensional space, and the model can capture general rules more efficiently, rather than remembering noise or special cases, thereby enhancing the generalization ability to new samples; 3) Improved computational efficiency: Low-dimensional data significantly reduces the time cost of model training and inference, making it easier to implement complex algorithms (such as SVM, deep learning) in fault detection, thereby improving the model's generalization ability and breaking the curse of dimensionality.

[0014] In summary, PCA addresses the issues of collinearity and sparsity in high-dimensional data by preserving key information and eliminating redundant interference. Furthermore, it enhances the distinguishability of fault features by focusing on core driving patterns, ultimately achieving a simultaneous improvement in model accuracy, robustness, and generalization ability. This characteristic makes it particularly suitable for high-dimensional data scenarios such as medical device fault detection (e.g., non-invasive blood pressure simulators, multi-parameter monitors).

[0015] The reconstruction error score maps multiple parameters to a low-dimensional principal component space using PCA, and then reconstructs the original data. Under normal conditions, the multi-parameter coordination follows a fixed pattern (such as the systolic-diastolic characteristics of blood pressure waveforms), resulting in extremely low reconstruction error. However, if there are multi-parameter coupling anomalies (such as a pressure control module malfunction causing asynchrony between pressure and flow signals), the difference between the original data and the reconstructed data will increase significantly (the reconstruction error score will jump), thereby capturing coordination anomalies that traditional thresholding methods cannot identify, achieving global capture of coupling anomalies.

[0016] Physiological disturbances such as patient tremors are localized, transient fluctuations, and their variance is much smaller than the systematic bias caused by the fault. By retaining the principal components (corresponding to the core laws of equipment operation) and discarding noise components (corresponding to interference), the reconstruction error score filters out physiological disturbances during the reconstruction process. The reconstruction error score does not increase significantly due to local fluctuations, thereby achieving noise filtering of physiological disturbances and reducing the false alarm rate.

[0017] In summary, obtaining the reconstruction error score overcomes the fundamental shortcomings of traditional threshold methods, such as the inability of a single parameter threshold to capture multi-parameter coupling anomalies, leading to an increased false alarm rate; the fixed fluctuation range, ignoring patient physiological interference (such as tremors), leading to an increased false alarm rate; and the independent judgment of each sensor, without utilizing the correlation between subsystems, leading to an increased false alarm rate. This effectively improves the sensitivity of fault detection and the robustness against interference.

[0018] Mahalanobis distance scores correct for correlations between parameters (such as the positive correlation between pressure and flow) using the covariance matrix. They calculate the weighted distance from sample points to normal data centers, capturing a spatial measure of coupling correlation. If multiple parameters deviate from their normal distribution due to coupling anomalies (such as flow not increasing proportionally while pressure increases), the Mahalanobis distance score will increase significantly, thus identifying correlation anomalies that are ignored by traditional methods.

[0019] Mahalanobis distance scores dynamically calculate the normal range (rather than a fixed threshold) based on the distribution of normal samples, enabling adaptive correction of individualized baselines. For example, for patients in high-altitude areas, Mahalanobis distance scores will be centered on the blood pressure baseline of this group, allowing for higher baseline values; for older devices, the covariance matrix will be adjusted based on their historical normal data to reduce false alarms caused by baseline differences and improve cross-scenario adaptability.

[0020] In summary, obtaining Mahalanobis distance scores can resolve issues such as strong coupling of multiple parameters (e.g., the mutual influence between pressure control, signal acquisition, and circuit systems in blood pressure monitoring) and individualized baseline differences (e.g., different baselines between new / old equipment and high-altitude / plain areas), thereby improving fault separation, enhancing resistance to environmental interference, and improving early fault detection capabilities.

[0021] Local Outlier Factor (LOF) identifies anomalies by comparing the local density of a sample point with that of its neighboring samples: the local density of normal samples is close to that of their neighbors (LOF≈1); the local density of anomaly samples (such as fault points) is significantly lower than that of their neighbors (LOF>>1), thus achieving density quantification of outliers at multiple scales. Whether it is a sudden and severe fault (a sharp drop in local density) or a slow drift (a gradual decrease in local density), LOF can accurately capture anomalies through density differences, reducing the misjudgment of normal samples in sparse regions by traditional distance methods.

[0022] Local outlier factors are insensitive to local noise. By comparing the relative anomalies of samples in the neighborhood density, noise-resistant identification of complex faults can be achieved. For example, in multi-sensor data, the noise of a single sensor is diluted by normal samples in the neighborhood (local density is close to normal), while the multi-sensor collaborative anomaly caused by a fault will significantly reduce the local density (increase the LOF value), thereby improving the identification rate of complex faults (such as multi-module collaborative failure).

[0023] In summary, obtaining local outlier factor scores addresses multi-scale outlier characteristics and overcomes the density blind zone of traditional outlier detection, thereby improving the recognition rate of complex faults and noise resistance.

[0024] Reconstruction error scores can effectively detect sudden faults, but are susceptible to transient interference. A fusion compensation mechanism is used: stability-weighted suppression of Mahalanobis distance scores and local outlier scores. Mahalanobis distance scores can effectively detect faults coupled with related parameters, but have low sensitivity to progressive faults. A fusion compensation mechanism is used: local outlier scores enhance progressive pattern recognition. Local outlier scores can effectively detect faults with local density anomalies, but miss global sparse faults. A fusion compensation mechanism is used: reconstruction error scores provide a global anomaly baseline. By fusing these three anomaly scores, the advantages of each score are absorbed while overcoming the limitations of a single score. This comprehensive approach to anomaly scores adapts to various types of equipment faults, significantly enhancing anti-interference capabilities and improving fault identification performance.

[0025] Combining PCA dimensionality reduction with multi-score fusion, principal component analysis can effectively reduce dimensionality while retaining key information, and anomaly score fusion can integrate multiple detection indicators, thereby improving the comprehensiveness and robustness of anomaly detection, and enhancing detection accuracy and adaptability.

[0026] Furthermore, the specific steps for obtaining several principal components include:

[0027] Based on the data matrix, a covariance matrix is ​​obtained; based on the covariance matrix, an eigenvalue diagonal matrix is ​​obtained; and based on the eigenvalue diagonal matrix, the principal components are obtained.

[0028] The first formula for obtaining the covariance matrix is:

[0029] ;

[0030] in, Represents the covariance matrix. This represents the number of samples in the data matrix. Represents a data matrix. Indicates transpose;

[0031] The second formula for obtaining the eigenvalue diagonal matrix is:

[0032] ;

[0033] in, Represents the eigenvector matrix, Represents an eigenvalue diagonal matrix;

[0034] The third calculation formula for obtaining the principal components is:

[0035] ;

[0036] in, Indicates the first The eigenvalues ​​of the principal components Represents an integer greater than or equal to 1. Indicates the number of features. Indicates the number of principal components. This represents the threshold for the cumulative variance contribution rate;

[0037] The fourth calculation formula for obtaining the principal component data is:

[0038] ;

[0039] in, Represents principal component data. This represents the principal component projection matrix.

[0040] Furthermore, the score data includes the reconstruction error score and the Mahalanobis distance score, and the fifth calculation formula for obtaining the reconstruction error score is:

[0041] ;

[0042] ;

[0043] in, This indicates data reconstruction. This represents the mean of the principal component data. Represents the reconstruction error fraction. Indicates runtime data, Represents principal component data. Represents the principal component projection matrix. Indicates transpose;

[0044] The sixth formula for obtaining the Mahalanobis distance score is:

[0045] ;

[0046] in, Represents the Mahalanobis distance fraction. and These represent the mean and covariance of the principal component data, respectively.

[0047] Furthermore, the score data also includes local outlier scores, and the specific steps for obtaining the local outlier scores include:

[0048] Based on the principal component data, obtain any two samples to obtain a first sample and a second sample, and obtain the Euclidean distance between the first sample and the second sample;

[0049] A preset number of neighbors is used to obtain the nearest neighbor set of the first sample and the nearest neighbor distance of the second sample based on the Euclidean distance and the number of neighbors.

[0050] Based on the Euclidean distance and the nearest neighbor distance, the reachability distance between the first sample and the second sample is obtained;

[0051] Based on the reachability distance and the nearest neighbor set, the local reachability density is obtained;

[0052] The local outlier factor score is obtained based on the nearest neighbor set and the local reachability density.

[0053] Furthermore, the seventh formula for calculating the Euclidean distance is as follows:

[0054] ;

[0055] in, This represents the Euclidean distance between the first and second samples. Indicates the first sample. Indicates the second sample. and Both represent integers greater than or equal to 1;

[0056] The eighth formula for obtaining the locally reachable density is:

[0057] ;

[0058] in, This represents the local reachability density of the first sample. Represents the nearest neighbor set. This represents the reachability distance between the first sample and the second sample.

[0059] The ninth formula for obtaining the local outlier score is:

[0060] ;

[0061] in, This represents the local outlier score of the first sample.

[0062] Furthermore, the tenth formula for obtaining the abnormal score is:

[0063] ;

[0064] ;

[0065] in, Represents the standardized first Category score data, Indicates the first Category score data, express standard deviation express The mean, Indicates abnormal scores. Indicates the first The weighting coefficients of the category score data, This indicates the number of categories in the score data.

[0066] Considering a single anomaly detection algorithm, such as Isolation Forest, the following shortcomings still exist:

[0067] Incomplete coverage of abnormal patterns: A single model is usually only sensitive to specific types of anomalies (such as point anomalies and context anomalies), while the root causes of failures in non-invasive blood pressure simulators are diverse, including mechanical wear, solenoid valve sticking, sensor zero drift and micro-leakage in the gas path, etc. These failures manifest as completely different patterns in the data, which are difficult for a single model to capture comprehensively.

[0068] Poor interpretability: Most existing methods output a binary label of normal or abnormal, which cannot indicate the specific type, severity or possible component of the fault, resulting in a lack of targeted maintenance decisions and prolonged equipment downtime.

[0069] Therefore, this method abandons the single anomaly detection approach and adopts a strategy of multimodal temporal feature decoupling and collaborative analysis. It decomposes the working cycle of the entire non-invasive blood pressure simulator system into multiple functional sub-processes (rapid inflation, pressure stabilization, linear pressure relief, and pulse simulation). For each sub-process, it extracts the feature set that best reflects the health status of its core components from the raw data and uses a fault diagnosis model to achieve refined identification and location of multiple fault modes.

[0070] Furthermore, the method also includes:

[0071] Based on the abnormal running data, multi-channel time series data is obtained. Based on key event points, the multi-channel time series data is segmented to obtain several sub-sequences. The key feature set corresponding to each sub-sequence is obtained. Based on the key feature set, the total feature vector is obtained.

[0072] An output vector is obtained based on the pre-trained fault diagnosis model and the total feature vector, and fault information is obtained based on the output vector.

[0073] Furthermore, the multi-channel timing data includes the main pressure sensor signal and the exhaust valve control signal, and the key event points include the cycle start point, inflation end point, depressurization start point, and cycle end point. The specific steps for obtaining several sub-sequences include:

[0074] A pressure curve is obtained based on the main pressure sensor signal, and the slope and pressure value at each time point are obtained based on the pressure curve; a pressure relief curve is obtained based on the exhaust valve control signal.

[0075] If the slope continues to increase and the pressure value is greater than a preset pressure threshold, then the time point at which the pressure value first exceeds the preset pressure threshold is obtained to obtain the starting point of the cycle.

[0076] If the pressure value is within a preset contraction range, then the pressure derivative is obtained based on the pressure value, and the rate of decrease of the pressure derivative is obtained. If the rate of decrease is greater than the decrease threshold and the pressure derivative is lower than the preset contraction threshold, then the inflation end point is obtained based on the pressure derivative.

[0077] Obtain the rising edge of the pressure relief curve, and obtain the pressure relief start point based on the rising edge;

[0078] If the pressure value remains below the preset pressure threshold within a preset time range, the end point of the cycle is obtained based on the pressure value.

[0079] The multi-channel time-series data is segmented based on the key event points to obtain the sub-sequences.

[0080] By identifying key event points in each calibration cycle (such as the start of inflation, reaching the target pressure, and the start of depressurization), and according to their inherent physical processes and functional logic, the entire time-series data is segmented into subsequences corresponding to different physiological simulation stages. By decoupling the complex system into multiple functional subprocesses and extracting features that can sensitively reflect the health status of specific components, it is possible to capture diverse fault modes caused by different root causes and exhibited at different stages, significantly improving the coverage of fault detection.

[0081] Furthermore, the sub-sequences include an inflation stage sub-sequence, a pressure stabilization stage sub-sequence, a pressure relief stage sub-sequence, and a pulse wave stage sub-sequence. The specific steps for obtaining the key feature set corresponding to each sub-sequence include:

[0082] The inflation stage subsequence is obtained by acquiring the stage curve, the average slope and pressure feedback value of the stage curve, and the inflation rate is obtained based on the average slope; the inflation time is obtained based on the cycle start point and the inflation end point; the maximum pressure value is obtained based on the stage curve, and the pressure overshoot is obtained based on the maximum pressure value and the target pressure; the high-pressure source control signal is obtained based on the inflation stage subsequence, the error between the set value of the high-pressure source control signal and the pressure feedback value is acquired, and the error is integrated to obtain the control signal integral error;

[0083] Based on the inflation rate, the inflation time, the pressure overshoot, and the control signal integral error, the key feature set of the inflation stage subsequence is obtained;

[0084] The stabilizing pressure value of the stabilizing stage subsequence is obtained, the standard deviation of the stabilizing pressure value is obtained, and the standard deviation of pressure fluctuation is obtained based on the standard deviation; the stabilizing pressure value is linearly fitted to obtain the pressure drift slope; the micro-pressure control signal of the stabilizing stage subsequence is frequency domain analyzed to obtain the control signal oscillation frequency;

[0085] Based on the pressure fluctuation standard deviation, the pressure drift slope, and the control signal oscillation frequency, the key feature set of the subsequence of the voltage stabilization stage is obtained;

[0086] Obtain the pressure relief curve of the pressure relief stage subsequence, fit the pressure relief curve to obtain the pressure decay time constant and fitting error, obtain the nonlinearity of the decay curve based on the fitting error, and obtain the average pressure relief rate based on the pressure relief curve;

[0087] Based on the pressure decay time constant, the nonlinearity of the decay curve, and the average pressure relief rate, the key feature set of the pressure relief stage subsequence is obtained;

[0088] The pulse wave pressure curve of the pulse wave stage subsequence is obtained, the peak value and trough value of the pulse wave pressure curve are obtained, and the pulse wave amplitude is obtained based on the peak value and the trough value; the average time interval between two adjacent peak values ​​is obtained to obtain the pulse wave period; the pulse wave is obtained based on the pulse wave pressure curve, and the pulse wave is compared with the reference wave to obtain the waveform distortion.

[0089] Based on the pulse wave amplitude, the pulse wave period, and the waveform distortion, the key feature set of the pulse wave stage subsequence is obtained.

[0090] For each stage of the subsequence, a multi-dimensional feature set related to the core physical process of that stage is extracted to decouple the features. This transforms a complex time-series signal that is difficult to analyze directly into a structured, characteristic, and physically meaningful data representation, laying a solid foundation for subsequent high-precision and interpretable fault diagnosis.

[0091] Furthermore, the specific steps for obtaining the fault diagnosis model include:

[0092] Acquire historical data from a non-invasive blood pressure simulator. The historical data includes several historical samples in normal and fault states. Each historical sample corresponds to a total feature vector and a multi-label vector. The multi-label vector includes several fault labels for different fault types.

[0093] Based on the historical samples and the fault relationship table, fault relationships are obtained, and the statistical correlation between any two fault labels is obtained to obtain a correlation matrix. A fault relationship graph is constructed based on the fault labels, the fault relationships, and the correlation matrix.

[0094] Based on the historical samples and the fault labels, single fault samples and composite fault samples are obtained. A preset model is trained based on the single fault samples to obtain a first model. The first model is trained based on the composite fault samples and the fault relationship graph to obtain the fault diagnosis model.

[0095] Fault diagnosis is restructured from a multiple-choice classification problem into a multiple-choice probabilistic inference problem with correlations. It explicitly models the dependencies between labels, which is more consistent with the physical fact that fault mechanisms are interconnected. This enables the model to learn the complex coexistence, mutual exclusion, and causal relationships between different fault modes, thereby outputting a probabilistic multi-label vector with high reliability and strong decision support capabilities. Using a supervised multi-label classification model, it can not only determine whether there is an anomaly, but also give an accurate judgment on where the anomaly is and what kind of anomaly it is.

[0096] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0097] 1. This method introduces Principal Component Analysis (PCA) and dimensionality reduction. PCA captures the potential driving patterns of faults, improves fault separation, overcomes interference resistance, and solves the problem of collinearity in high-dimensional data. Dimensionality reduction improves the model's generalization ability and overcomes the curse of dimensionality. By combining PCA dimensionality reduction with multi-score fusion, PCA can effectively reduce dimensionality while retaining key information, and anomaly score fusion can integrate multiple detection indicators, improving the comprehensiveness and robustness of anomaly detection, and enhancing detection accuracy and adaptability.

[0098] 2. This method abandons the single anomaly detection approach and adopts a strategy of multimodal temporal feature decoupling and collaborative analysis. It decomposes the working cycle of the entire non-invasive blood pressure simulator system into multiple functional sub-processes (rapid inflation, pressure stabilization, linear pressure relief, and pulse simulation). For each sub-process, it extracts the feature set that best reflects the health status of its core components from the raw data and uses a fault diagnosis model to achieve refined identification and location of multiple fault modes.

[0099] 3. Identify key event points in each calibration cycle (such as the start of inflation, reaching the target pressure, and the start of depressurization). According to its inherent physical process and functional logic, the entire time series data is segmented into subsequences corresponding to different physiological simulation stages. By decoupling the complex system into multiple functional subprocesses and extracting features that can sensitively reflect the health status of specific components, it is possible to capture diverse fault modes caused by different fundamental reasons and exhibited at different stages, significantly improving the coverage of fault detection.

[0100] 4. For each stage of the subsequence, extract the multi-dimensional feature set related to the core physical process of that stage to achieve feature decoupling. This transforms a complex time-series signal that is difficult to analyze directly into a structured, characteristic, and physically meaningful data representation, laying a solid foundation for subsequent high-precision and interpretable fault diagnosis.

[0101] 5. Fault diagnosis is restructured from a multiple-choice classification problem into a multiple-choice probabilistic inference problem, explicitly modeling the dependencies between labels. This better reflects the physical fact that fault mechanisms are interconnected, enabling the model to learn the complex coexistence, mutual exclusion, and causal relationships between different fault modes. As a result, it outputs a probabilistic multi-label vector with high reliability and strong decision support capabilities. Using a supervised multi-label classification model, it can not only determine whether there is an anomaly, but also provide accurate judgments on where and what kind of anomaly it is. Attached Figure Description

[0102] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0103] Figure 1This is a flowchart illustrating the intelligent fault diagnosis method for non-invasive blood pressure simulators in this invention. Detailed implementation details...

[0104] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0105] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0106] Example 1

[0107] refer to Figure 1 This embodiment provides an intelligent diagnostic method for non-invasive blood pressure simulator malfunctions, the method comprising:

[0108] Preprocessing the operating data of the non-invasive blood pressure simulator to obtain a data matrix Each row represents a data sample, and each column represents a feature. The sample size represents the number of running cycles or detection records contained in the device dataset. The feature number represents the number of device features contained in each sample; the data matrix includes several features, which may include pressure, flow rate, temperature, current, voltage, sensor signals, control commands, alarm signals, and my historical fault records, etc.

[0109] In this example, the preprocessing methods include: initial screening of features based on expert experience, removal of redundant features through statistical methods such as analysis of variance, correlation analysis, and information gain, and Z-score standardization of features to reduce the influence of dimensions.

[0110] In this embodiment, after Z-score standardization, the data matrix is ​​further centered, and the calculation method can be as follows:

[0111] ;

[0112] This represents the data matrix after centralization. This represents the data matrix after Z-score standardization. This represents the mean of the data matrix after Z-score standardization;

[0113] Principal component analysis is performed on the data matrix to obtain several principal components; the specific steps for obtaining the several principal components include:

[0114] Based on the data matrix, a covariance matrix is ​​obtained; based on the covariance matrix, an eigenvalue diagonal matrix is ​​obtained; and based on the eigenvalue diagonal matrix, the principal components are obtained.

[0115] The first formula for obtaining the covariance matrix is:

[0116] ;

[0117] in, Represents the covariance matrix. This represents the number of samples in the data matrix. Represents a data matrix. Indicates transpose;

[0118] The second formula for obtaining the eigenvalue diagonal matrix is:

[0119] ;

[0120] in, Represents the eigenvector matrix, Represents an eigenvalue diagonal matrix;

[0121] The third calculation formula for obtaining the principal components is:

[0122] ;

[0123] Using the cumulative variance contribution rate method, the top... One principal component makes the above formula hold;

[0124] in, Indicates the first The eigenvalues ​​of the principal components Represents an integer greater than or equal to 1. Indicates the number of features. This indicates the number of principal components, i.e., the number of principal components selected. This represents the cumulative variance contribution rate threshold, used to determine... The threshold;

[0125] Construct a principal component projection matrix based on the principal components. Principal component data are obtained by dimensionality reduction of the data matrix based on the principal component projection matrix.

[0126] The fourth calculation formula for obtaining the principal component data is:

[0127] ;

[0128] in, Represents principal component data. This represents the principal component projection matrix.

[0129] Score data are obtained based on the principal component data and the principal component projection matrix;

[0130] The score data includes reconstruction error score and Mahalanobis distance score. The fifth calculation formula for obtaining the reconstruction error score is as follows:

[0131] ;

[0132] ;

[0133] in, This represents the reconstructed data, specifically the sample data restored using principal components. This represents the mean of the principal component data. Represents the reconstruction error fraction. Indicates runtime data, Represents principal component data. Represents the principal component projection matrix. Indicates transpose;

[0134] The sixth formula for obtaining the Mahalanobis distance score is:

[0135] ;

[0136] in, Represents the Mahalanobis distance fraction. and These represent the mean and covariance of the principal component data, respectively.

[0137] The score data also includes local outlier scores, and the specific steps for obtaining the local outlier scores include:

[0138] Based on the principal component data, obtain any two samples to obtain a first sample and a second sample, and obtain the Euclidean distance between the first sample and the second sample;

[0139] A preset number of neighbors is used to obtain the nearest neighbor set of the first sample and the nearest neighbor distance of the second sample based on the Euclidean distance and the number of neighbors. Specifically, the Euclidean distances between other samples and the second sample are obtained and sorted in descending order. The Euclidean distances of the first sample are taken from the first sample and the smallest Euclidean distance is taken as the nearest neighbor distance of the second sample. The Euclidean distances between other samples and the first sample are obtained and sorted in descending order. The Euclidean distances of the first sample are taken from the first sample and the nearest neighbor set is obtained.

[0140] Based on the Euclidean distance and the nearest neighbor distance, the reachable distance between the first sample and the second sample is obtained, that is, the distance with the largest value between the Euclidean distance and the nearest neighbor distance is selected as the reachable distance; when the first sample is very close to the second sample, the lower limit of the nearest neighbor distance is used to support it, and the magnification of the local scale by points with extremely high density is suppressed.

[0141] Based on the reachability distance and the nearest neighbor set, the local reachability density is obtained;

[0142] The local outlier factor score is obtained based on the nearest neighbor set and the local reachability density.

[0143] The seventh formula for obtaining the Euclidean distance is:

[0144] ;

[0145] in, This represents the Euclidean distance between the first and second samples. Indicates the first sample. Indicates the second sample, i.e. and They are respectively The The first sample and the first One sample, and Both represent integers greater than or equal to 1;

[0146] The eighth formula for obtaining the locally reachable density is:

[0147] ;

[0148] in, This represents the local reachability density of the first sample. Represents the nearest neighbor set. This represents the reachability distance between the first sample and the second sample.

[0149] The ninth formula for obtaining the local outlier score is:

[0150] ;

[0151] in, This represents the local outlier score of the first sample.

[0152] ≈1 means It is comparable to the density of its neighborhood (normal); >1 indicates The density is lower than that of the neighborhood (the larger the density, the more abnormal); <1 indicates The density is higher than that of the neighborhood (more internally).

[0153] Abnormal scores are obtained by weighting the aforementioned score data;

[0154] If the abnormal score is greater than or equal to a preset abnormal threshold (which can be determined by expert experience or historical data), the non-invasive blood pressure simulator is deemed abnormal.

[0155] The tenth formula for obtaining the abnormal score is:

[0156] For each abnormal score Perform Z-score standardization:

[0157] ;

[0158] Abnormal score fusion:

[0159] ;

[0160] in, Represents the standardized first Category score data, Indicates the first Category score data, express standard deviation express The mean, Indicates abnormal scores. Indicates the first The weighting coefficients for the category score data can be determined using expert experience. This indicates the number of categories in the score data.

[0161] Example 2

[0162] Based on Embodiment 1, in this embodiment, the method further includes:

[0163] Based on the abnormal operating data, multi-channel timing data is obtained. The multi-channel timing data may include high-pressure source control signal, micro-pressure control signal, fast exhaust valve action signal, precision flow sensor reading, and main pressure sensor reading.

[0164] The multi-channel time series data is segmented based on key event points to obtain several sub-sequences. The key feature set corresponding to each sub-sequence is obtained. The total feature vector is obtained based on the key feature set. All key feature sets are concatenated into a total feature vector.

[0165] The multi-channel timing data includes main pressure sensor signals and exhaust valve control signals. The key event points include the cycle start point, inflation end point, depressurization start point, and cycle end point. The specific steps to obtain several sub-sequences include:

[0166] A pressure curve is obtained based on the main pressure sensor signal, and the slope and pressure value at each time point are obtained based on the pressure curve; a pressure relief curve is obtained based on the exhaust valve control signal; if the slope continues to increase and the pressure value is greater than a preset pressure threshold, the time point at which the pressure value first exceeds the preset pressure threshold is obtained to obtain the cycle start point;

[0167] That is, a pressure threshold is set on the pressure signal p(t). When the pressure p(t) rises continuously from the ambient pressure and exceeds the pressure threshold for the first time, the time point is marked as the start of the cycle t0.

[0168] If the pressure value is within a preset contraction range, then the pressure derivative is obtained based on the pressure value, and the rate of decrease of the pressure derivative is obtained. If the rate of decrease is greater than the decrease threshold and the pressure derivative is lower than the preset contraction threshold, then the inflation end point is obtained based on the pressure derivative.

[0169] That is, when the pressure signal p(t) reaches the preset contraction range, the derivative of the pressure signal p(t) is calculated. When the derivative drops significantly from a positive value and approaches zero (below the drop threshold), it indicates that the inflation process is over and the system enters the pressure stabilization stage. This inflection point time is marked as the inflation end point t1.

[0170] Obtain the rising edge of the pressure relief curve, and obtain the pressure relief start point based on the rising edge; this point is usually explicitly given by the system control logic, or it can be accurately determined by detecting the rising or falling edge of the exhaust valve control signal (depending on the valve logic), and this time point is marked as the pressure relief start point t2;

[0171] If the pressure value remains below the preset pressure threshold within a preset time range, the end point of the cycle is obtained based on the pressure value.

[0172] That is, when the pressure signal p(t) decreases and stabilizes near the ambient pressure (continuously below the preset pressure threshold), this time point is marked as the end point t3 of the cycle.

[0173] The multi-channel time-series data is segmented based on the key event points to obtain the sub-sequences.

[0174] The subsequences include an inflation stage subsequence, a pressure stabilization stage subsequence, a pressure relief stage subsequence, and a pulse wave stage subsequence. For the inflation stage, the following are extracted: inflation rate, time to reach target pressure, pressure overshoot during inflation, and integral error of the control signal. For the pressure stabilization stage, the following are extracted: pressure fluctuation variance, pressure drift slope, and oscillation frequency of the pressure stabilization control signal. For the pressure relief stage, the following are extracted: pressure decay time constant, nonlinearity of the decay curve, and fitting error with the ideal decay curve. For the pulse wave stage, the following are extracted: amplitude, period, and waveform distortion of the simulated pulse wave (e.g., through harmonic analysis).

[0175] The specific steps for obtaining the key feature set corresponding to each of the sub-sequences include:

[0176] Obtain the stage curve of the inflation stage subsequence, obtain the average slope and pressure feedback value of the stage curve, and obtain the inflation rate based on the average slope; that is, average slope = (p(t1) - p(t0)) / t1 - t0, which is sensitive to the performance of the air pump, the air inlet valve and the unobstructedness of the pipeline.

[0177] The inflation time is obtained based on the cycle start point and the inflation end point; t1-t0 is taken directly, and an abnormal extension may indicate a decrease in air pump power or the presence of a leak.

[0178] The maximum pressure value is obtained based on the stage curve, and the pressure overshoot is obtained based on the maximum pressure value and the target pressure; the pressure overshoot = maximum pressure value - target pressure. A large overshoot may indicate PID parameter misalignment of the pressure controller or sensor response delay.

[0179] The high-pressure source control signal is obtained based on the inflation stage subsequence. The error between the set value of the high-pressure source control signal and the pressure feedback value is obtained. The error is integrated to obtain the integral error of the control signal, which reflects the stability and efficiency of the control system.

[0180] Based on the inflation rate, the inflation time, the pressure overshoot, and the control signal integral error, the key feature set of the inflation stage subsequence is obtained;

[0181] Obtain the stabilizing pressure value of the subsequence of the stabilizing stage, obtain the standard deviation of the stabilizing pressure value, and obtain the standard deviation of pressure fluctuation based on the standard deviation, that is, calculate the standard deviation of the pressure signal p(t) in this stage, which directly reflects the sealing performance and pressure stability of the system and is sensitive to micro-leakage; perform linear fitting on the stabilizing pressure value to obtain the pressure drift slope, that is, perform linear fitting on the pressure signal p(t) to obtain its slope, a significant negative slope is a typical indicator of slow leakage; perform frequency domain analysis on the micro-pressure control signal of the subsequence of the stabilizing stage to obtain the oscillation frequency of the control signal; perform frequency domain analysis (such as Fast Fourier Transform, FFT) on the micro-pressure control signal to extract its dominant oscillation frequency, abnormal oscillation mode may indicate controller instability or malfunction of solenoid valve.

[0182] Based on the pressure fluctuation standard deviation, the pressure drift slope, and the control signal oscillation frequency, the key feature set of the subsequence of the voltage stabilization stage is obtained;

[0183] Obtain the pressure relief curve of the subsequence of the pressure relief stage, fit the pressure relief curve to obtain the pressure decay time constant and fitting error, and obtain the nonlinearity of the decay curve based on the fitting error; approximate the pressure curve p(t) of the pressure relief stage as an exponential decay model: p(t) = p0 * exp(-t / τ), where p0 represents the initial pressure at the beginning of the pressure relief stage, t represents time, τ represents the time constant, and exp() represents the natural exponential function. The time constant is obtained by fitting, and an increase in the time constant may indicate insufficient opening or blockage of the exhaust valve; calculate the fitting error (such as root mean square error) between the actual pressure curve and the ideal exponential decay model. An ideal pressure relief valve should produce smooth exponential decay, and an increase in nonlinearity may indicate that the valve action is nonlinear or there is foreign object interference.

[0184] The average pressure relief rate is obtained based on the pressure relief curve, i.e., = (p(t2) - p(t3)) / t2 - t3;

[0185] Based on the pressure decay time constant, the nonlinearity of the decay curve, and the average pressure relief rate, the key feature set of the pressure relief stage subsequence is obtained;

[0186] The pulse wave pressure curve of the pulse wave stage subsequence is obtained, the peak and trough values ​​of the pulse wave pressure curve are obtained, the pulse wave amplitude is obtained based on the peak and trough values, and the peak-trough amplitude of the pulse wave is extracted from the pressure signal; the average time interval between two adjacent peak values ​​is obtained to obtain the pulse wave period; the pulse wave is obtained based on the pulse wave pressure curve, and the pulse wave is compared with a reference wave to obtain the waveform distortion; the extracted pulse wave is compared with a standard waveform (such as a sine wave, a more accurate physiological waveform template) to calculate its distortion, for example, by calculating the total harmonic distortion to obtain the waveform distortion.

[0187] Based on the pulse wave amplitude, the pulse wave period, and the waveform distortion, the key feature set of the pulse wave stage subsequence is obtained.

[0188] An output vector is obtained based on the pre-trained fault diagnosis model and the total feature vector. The output vector is a multi-label output vector, such as [solenoid valve failure, pressure sensor drift, micro-leakage, ...]. Each label represents a fault mode or component, and the output value is the probability of the fault occurring. Fault information is obtained based on the output vector.

[0189] The specific steps for obtaining the fault diagnosis model include:

[0190] Historical data of a non-invasive blood pressure simulator is acquired. The historical data includes several historical samples in normal and fault states. Each historical sample corresponds to a total feature vector and a multi-label vector. The multi-label vector includes fault labels for several different fault types. The multi-label vector is the target of model training and is a binary vector of length K, where K is the number of all known fault modes and faulty components.

[0191] Assuming K=5, the fault modes are defined as [solenoid valve jamming, air pump power attenuation, main sensor zero drift, air circuit micro-leakage, controller parameter misalignment]. A sample can be labeled [1, 0, 1, 0, 0], indicating that the sample simultaneously exhibits both solenoid valve jamming and main sensor zero drift. This breaks through the traditional single fault assumption and truly reflects the coexistence of multiple faults in reality.

[0192] Based on the historical samples and the fault relationship table, fault relationships are obtained. In this embodiment, the fault relationship table may include correlations between known fault modes, between faulty components, and between fault modes and faulty components. Specifically, it may include:

[0193] Coexistence: Micro-leakage and pump power reduction may occur simultaneously because the pump needs to work harder to compensate for the leakage;

[0194] Mutual exclusion: Some faults are physically almost impossible to occur simultaneously;

[0195] Causal relationship: Solenoid valve jamming may cause the illusion of controller parameter misalignment.

[0196] Obtain the statistical correlation between any two fault labels to obtain a correlation matrix. Construct a fault relationship graph based on the fault labels, the fault relationships, and the correlation matrix, which can be represented as G = (V, E, A), where each node V in the graph represents a single fault mode or faulty component, the total number of nodes V N is equal to the length of the multi-label vector, and the edges E represent the existence of some kind of association between them. An N×N adjacency matrix A can be used to quantify the strength of these relationships.

[0197] In this embodiment, the adjacency matrix A uses point mutual information to calculate the statistical correlation between all pairs of fault labels. The calculated correlation matrix is ​​then thresholded (for example, only edges with an absolute value greater than 0.1 are retained) and normalized to the interval [0,1] or [-1, 1] to form the adjacency matrix A.

[0198] In this embodiment, domain experts can also directly define the elements in the adjacency matrix A based on system physical principles and maintenance experience. :

[0199] = 1.0: Indicates a strong coexistence or causal relationship. For example, there is a strong correlation between micro-leakage in the gas path and power reduction in the air pump, because the leak causes the air pump to continue operating to maintain pressure; = 0.5: indicates a moderate or common accompaniment relationship; = 0: indicates that there is no known direct relationship; = -0.5: indicates a weak mutual exclusion relationship.

[0200] Based on the historical samples and fault labels, single fault samples and compound fault samples are obtained. A preset model is trained based on the single fault samples to obtain a first model, allowing the model to firmly grasp the characteristic patterns of each single fault. The first model is then trained based on the compound fault samples and the fault relationship graph to obtain the fault diagnosis model. All compound fault samples are then introduced, and the model trained in the first stage is fine-tuned. At this point, special attention is paid to compound fault samples that are incorrectly predicted by the model to enhance the model's ability to distinguish complex situations.

[0201] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0202] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent diagnosis of faults in a non-invasive blood pressure simulator, characterized in that: The method includes: Preprocess the operating data of the non-invasive blood pressure simulator to obtain a data matrix; Principal component analysis is performed on the data matrix to obtain several principal components. A principal component projection matrix is ​​constructed based on the principal components. The data matrix is ​​then dimensionality-reduced based on the principal component projection matrix to obtain principal component data. Score data is obtained based on the principal component data and the principal component projection matrix, and abnormal scores are obtained by weighting the score data. If the abnormal score is greater than or equal to the preset abnormal threshold, the non-invasive blood pressure simulator is determined to be abnormal.

2. The intelligent fault diagnosis method for non-invasive blood pressure simulator according to claim 1, characterized in that, The specific steps for obtaining several principal components include: Based on the data matrix, a covariance matrix is ​​obtained; based on the covariance matrix, an eigenvalue diagonal matrix is ​​obtained; and based on the eigenvalue diagonal matrix, the principal components are obtained. The first formula for obtaining the covariance matrix is: ； in, Represents the covariance matrix. This represents the number of samples in the data matrix. Represents a data matrix. Indicates transpose; The second formula for obtaining the eigenvalue diagonal matrix is: ； in, Represents the eigenvector matrix, Represents an eigenvalue diagonal matrix; The third calculation formula for obtaining the principal components is: ； in, Indicates the first The eigenvalues ​​of the principal components Represents an integer greater than or equal to 1. Indicates the number of features. Indicates the number of principal components. This represents the threshold for the cumulative variance contribution rate; The fourth calculation formula for obtaining the principal component data is: ； in, Represents principal component data. This represents the principal component projection matrix.

3. The intelligent fault diagnosis method for non-invasive blood pressure simulator according to claim 1, characterized in that, The score data includes reconstruction error score and Mahalanobis distance score. The fifth calculation formula for obtaining the reconstruction error score is: ； ； in, This indicates data reconstruction. This represents the mean of the principal component data. Represents the reconstruction error fraction. Indicates runtime data, Represents principal component data. Represents the principal component projection matrix. Indicates transpose; The sixth formula for obtaining the Mahalanobis distance score is: ； in, Represents the Mahalanobis distance fraction. and These represent the mean and covariance of the principal component data, respectively.

4. The intelligent fault diagnosis method for non-invasive blood pressure simulator according to claim 3, characterized in that, The score data also includes local outlier scores, and the specific steps to obtain the local outlier scores include: Based on the principal component data, obtain any two samples to obtain a first sample and a second sample, and obtain the Euclidean distance between the first sample and the second sample; A preset number of neighbors is used to obtain the nearest neighbor set of the first sample and the nearest neighbor distance of the second sample based on the Euclidean distance and the number of neighbors. Based on the Euclidean distance and the nearest neighbor distance, the reachability distance between the first sample and the second sample is obtained; Based on the reachability distance and the nearest neighbor set, the local reachability density is obtained; The local outlier factor score is obtained based on the nearest neighbor set and the local reachability density.

5. The intelligent fault diagnosis method for non-invasive blood pressure simulator according to claim 4, characterized in that, The seventh formula for calculating the Euclidean distance is: ； in, This represents the Euclidean distance between the first and second samples. Indicates the first sample. Indicates the second sample. and Both represent integers greater than or equal to 1; The eighth formula for obtaining the locally reachable density is: ； in, This represents the local reachability density of the first sample. Represents the nearest neighbor set. This represents the reachability distance between the first sample and the second sample. The ninth formula for obtaining the local outlier score is: ； in, This represents the local outlier score of the first sample.

6. The intelligent fault diagnosis method for non-invasive blood pressure simulator according to claim 5, characterized in that, The tenth formula for obtaining the abnormal score is: ； ； in, Represents the standardized first Category score data, Indicates the first Category score data, express standard deviation express The mean, Indicates abnormal scores. Indicates the first The weighting coefficients of the category score data, This indicates the number of categories in the score data.

7. The intelligent fault diagnosis method for non-invasive blood pressure simulator according to claim 1, characterized in that, The method further includes: Based on the abnormal running data, multi-channel time series data is obtained. Based on key event points, the multi-channel time series data is segmented to obtain several sub-sequences. The key feature set corresponding to each sub-sequence is obtained. Based on the key feature set, the total feature vector is obtained. An output vector is obtained based on the pre-trained fault diagnosis model and the total feature vector, and fault information is obtained based on the output vector.

8. The intelligent fault diagnosis method for non-invasive blood pressure simulator according to claim 7, characterized in that, The multi-channel timing data includes main pressure sensor signals and exhaust valve control signals. The key event points include the cycle start point, inflation end point, depressurization start point, and cycle end point. The specific steps to obtain several sub-sequences include: A pressure curve is obtained based on the main pressure sensor signal, and the slope and pressure value at each time point are obtained based on the pressure curve; a pressure relief curve is obtained based on the exhaust valve control signal. If the slope continues to increase and the pressure value is greater than a preset pressure threshold, then the time point at which the pressure value first exceeds the preset pressure threshold is obtained to obtain the starting point of the cycle. If the pressure value is within a preset contraction range, then the pressure derivative is obtained based on the pressure value, and the rate of decrease of the pressure derivative is obtained. If the rate of decrease is greater than the decrease threshold and the pressure derivative is lower than the preset contraction threshold, then the inflation end point is obtained based on the pressure derivative. Obtain the rising edge of the pressure relief curve, and obtain the pressure relief start point based on the rising edge; If the pressure value remains below the preset pressure threshold within a preset time range, the end point of the cycle is obtained based on the pressure value. The multi-channel time-series data is segmented based on the key event points to obtain the sub-sequences.

9. The intelligent fault diagnosis method for a non-invasive blood pressure simulator according to claim 8, characterized in that, The subsequence includes an inflation phase subsequence, a pressure stabilization phase subsequence, a pressure relief phase subsequence, and a pulse wave phase subsequence. The specific steps for obtaining the key feature set corresponding to each subsequence include: The inflation stage subsequence is obtained by acquiring the stage curve, the average slope and pressure feedback value of the stage curve, and the inflation rate is obtained based on the average slope; the inflation time is obtained based on the cycle start point and the inflation end point; the maximum pressure value is obtained based on the stage curve, and the pressure overshoot is obtained based on the maximum pressure value and the target pressure; the high-pressure source control signal is obtained based on the inflation stage subsequence, the error between the set value of the high-pressure source control signal and the pressure feedback value is acquired, and the error is integrated to obtain the control signal integral error; Based on the inflation rate, the inflation time, the pressure overshoot, and the control signal integral error, the key feature set of the inflation stage subsequence is obtained; The stabilizing pressure value of the stabilizing stage subsequence is obtained, the standard deviation of the stabilizing pressure value is obtained, and the standard deviation of pressure fluctuation is obtained based on the standard deviation; the stabilizing pressure value is linearly fitted to obtain the pressure drift slope; the micro-pressure control signal of the stabilizing stage subsequence is frequency domain analyzed to obtain the control signal oscillation frequency; Based on the pressure fluctuation standard deviation, the pressure drift slope, and the control signal oscillation frequency, the key feature set of the subsequence of the voltage stabilization stage is obtained; Obtain the pressure relief curve of the pressure relief stage subsequence, fit the pressure relief curve to obtain the pressure decay time constant and fitting error, obtain the nonlinearity of the decay curve based on the fitting error, and obtain the average pressure relief rate based on the pressure relief curve; Based on the pressure decay time constant, the nonlinearity of the decay curve, and the average pressure relief rate, the key feature set of the pressure relief stage subsequence is obtained; The pulse wave pressure curve of the pulse wave stage subsequence is obtained, the peak value and trough value of the pulse wave pressure curve are obtained, and the pulse wave amplitude is obtained based on the peak value and the trough value; the average time interval between two adjacent peak values ​​is obtained to obtain the pulse wave period; the pulse wave is obtained based on the pulse wave pressure curve, and the pulse wave is compared with the reference wave to obtain the waveform distortion. Based on the pulse wave amplitude, the pulse wave period, and the waveform distortion, the key feature set of the pulse wave stage subsequence is obtained.

10. The intelligent fault diagnosis method for a non-invasive blood pressure simulator according to claim 9, characterized in that, The specific steps for obtaining the fault diagnosis model include: Acquire historical data from a non-invasive blood pressure simulator. The historical data includes several historical samples in normal and fault states. Each historical sample corresponds to a total feature vector and a multi-label vector. The multi-label vector includes several fault labels for different fault types. Based on the historical samples and the fault relationship table, fault relationships are obtained, and the statistical correlation between any two fault labels is obtained to obtain a correlation matrix. A fault relationship graph is constructed based on the fault labels, the fault relationships, and the correlation matrix. Based on the historical samples and the fault labels, single fault samples and composite fault samples are obtained. A preset model is trained based on the single fault samples to obtain a first model. The first model is trained based on the composite fault samples and the fault relationship graph to obtain the fault diagnosis model.

Images