BiLSTM-FCN-based ballistocardiogram signal multi-class classification method and system
Through the local adaptive transform denoising and feature fusion of BiLSTM-FCN model, the classification problems brought about by time-degeneration and individual differences in cardiac impact signals are solved, and higher accuracy and stability are achieved, adapting to individual differences between different patients, and supporting real-time monitoring and intelligent diagnosis is enhanced.
Patent Information
- Application Number
- CN202510313678.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-17
- Publication Date
- 2025-08-08
AI Technical Summary
The existing core impact signal classification methods are difficult to achieve accurate classification when facing time-varying, non-stationarity and individual differences. The traditional methods are limited by manual feature engineering and lack of noise resistance.
BiLSTM-FCN model is used to classify the core impact signal. Through local adaptive transform denoising, timing encoding, feature fusion and Softmax classification, combined with BiLSTM timing feature extraction and FCN spatial feature extraction, the influence of human factors is reduced and the robustness and accuracy of the model is improved.
It improves the accuracy and stability of cardiac impact signal classification, adapts to individual differences between different patients, enhances noise anti-noise ability, and improves the support capabilities of real-time monitoring and intelligent diagnosis.
Smart Images
Figure CN120448998A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical signal processing technology, and in particular to a multi-category classification method and system for cardiac ballistic signals based on BiLSTM-FCN. Background Art
[0002] Ballistocardiographic signals are important biosignals used to describe cardiac activity and have broad applications in medical diagnosis, health monitoring, and intelligent medical devices. Accurate classification of ballistocardiographic signals is a crucial component of early screening and intelligent diagnostic systems for cardiovascular disease. However, due to the time-varying, non-stationary, and individual variability of ballistocardiographic signals, traditional signal processing and classification methods have limitations in practical applications.
[0003] Currently, classification methods for cardiac ballistic signals mainly include those based on statistical features, those based on time-frequency analysis, and those based on deep learning. Statistical methods extract the time-domain or frequency-domain features of cardiac ballistic signals and combine them with traditional machine learning algorithms (such as support vector machines and decision trees) for classification. However, these methods are limited by manual feature engineering and are difficult to adapt to the complex variations of cardiac ballistic signals. Summary of the Invention
[0004] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a multi-category classification method and system for ballistocardial signals based on BiLSTM-FCN.
[0005] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:
[0006] In a first aspect, an embodiment of the present invention provides a multi-class classification method for cardiac ballistic signals based on BiLSTM-FCN, the method comprising:
[0007] S1, collecting cardiac shock signals;
[0008] S2. Preprocessing the ballistocardiographic signal to obtain a preprocessed signal;
[0009] S3, performing time series coding on the preprocessed signal to obtain a time coding matrix;
[0010] S4. Input the time coding matrix into the trained BiLSTM-FCN model to obtain the category prediction label corresponding to the heartbeat signal;
[0011] The BiLSTM-FCN model includes:
[0012] Input layer, used to receive the time encoding matrix in the input BiLSTM-FCN model;
[0013] The BiLSTM temporal feature extraction layer is used to obtain the temporal feature matrix based on the time coding matrix input by the input layer;
[0014] The FCN spatial feature extraction layer is used to process the temporal feature matrix output by the BiLSTM temporal feature extraction layer to obtain a spatial feature matrix;
[0015] The feature fusion layer is used to fuse the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix;
[0016] The classification decision layer is used to perform feature mapping and classification calculation based on the fusion feature matrix, ultimately obtain the category probability distribution of the heartbeat signal, and output the category prediction label;
[0017] The class prediction label is the label corresponding to the class with the highest class probability.
[0018] Preferably, the S2 specifically includes:
[0019] The signal preprocessing step adopts a local adaptive transformation denoising method to remove local noise and non-stationary baseline drift of the heart ballistic signal. The calculation formula of the local adaptive transformation denoising method is:
[0020]
[0021] S pre (t) is the preprocessed signal;
[0022] S raw (t) is the heartbeat signal;
[0023] N is the total number of components of local noise;
[0024] a k is the local amplitude weight corresponding to the kth local noise component;
[0025] β k is the time-varying attenuation factor corresponding to the kth local noise component;
[0026] ω k is the dominant frequency of the signal corresponding to the kth local noise component;
[0027] is the local phase offset corresponding to the kth local noise component;
[0028] t is time.
[0029] Preferably, the S3 specifically includes:
[0030] S31, determining the short-time spectral component of the ballistocardi signal and its corresponding dominant frequency according to the preprocessed signal, and extracting phase information;
[0031] S32. Construct a time coding matrix based on the short-time spectral components of the ballistocardial signal and its corresponding dominant frequency and extracting phase information.
[0032] Preferably, the S32 specifically includes:
[0033] According to the short-time spectrum components of the heartbeat signal and its corresponding dominant frequency, and extracting the phase information, the time coding matrix is constructed using formula (1);
[0034] The formula (1) is:
[0035] T c (i, j) = |F k |·sin(2πf k j+θ k )
[0036] Among them, T c (i, j) is the element with index i, j in the time coding matrix;
[0037] F k is the complex value of the short-time spectral component;
[0038] f k is the dominant frequency of the short-time spectral component;
[0039] θ k is the phase information of the short-time spectral component.
[0040] Preferably, the BiLSTM time series feature extraction layer obtains a time series feature matrix based on the time coding matrix input by the input layer, specifically including:
[0041] The BiLSTM time series feature extraction layer is used to perform bidirectional time series feature extraction on the time coding matrix input by the input layer, and calculate the time series feature matrix;
[0042] Formula (2) is used to extract the time series features;
[0043] The formula (2) is:
[0044]
[0045] H BiLSTM =[h1,…,h t ,…,h T ] is the time series feature matrix calculated by BiLSTM;
[0046] h tis the hidden state vector calculated by BiLSTM at time step t;
[0047] x t is the input of the time encoding matrix at time step t;
[0048] W h 、U h 、b h is the trainable weight matrix and bias term of BiLSTM;
[0049] σ( ) is the activation function;
[0050] is the phase change of the time coding matrix at time step t;
[0051] γ is the gated memory decay coefficient.
[0052] Preferably, the FCN spatial feature extraction layer is used to process the temporal feature matrix output by the BiLSTM temporal feature extraction layer to obtain a spatial feature matrix, specifically including:
[0053] The FCN spatial feature extraction layer calculates the temporal feature matrix output by the BiLSTM temporal feature extraction layer using formula (3) to obtain the spatial feature matrix;
[0054] The formula (3) is:
[0055]
[0056] H FAN (i, j) is the i-th, j-th element of the spatial feature matrix calculated by the FCN spatial feature extraction layer;
[0057] H BiLSTM (i, j) is H BiLSTM The i-th and j-th elements of the time series feature matrix;
[0058] W m is the weight parameter of the convolution kernel, m is the convolution window index, and M is the convolution kernel size;
[0059] μ is the transformation factor based on the phase shift;
[0060] a1 is the nonlinear enhancement coefficient;
[0061] a2 is the gradient modulation coefficient;
[0062] Indicates H BiLSTM Output the local gradient of the matrix.
[0063] Preferably, the feature fusion layer fuses the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix, which specifically includes:
[0064] The feature fusion layer uses formula (4) to fuse the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix;
[0065] The formula (4) is:
[0066] H fusion =c·H BiLSTM +(1-c)·H FCN ;
[0067] H fusion is the fusion feature matrix;
[0068] H BiLSTM is the time series feature matrix;
[0069] H FCN is the spatial feature matrix;
[0070] c is the dynamic fusion weight;
[0071] in,
[0072]
[0073] Preferably, the method further comprises, before S1:
[0074] S0. Use the training data set to train the BiLSTM-FCN model until the pre-set training conditions are met, and then a trained BiLSTM-FCN model is obtained;
[0075] The training data set includes a plurality of samples of cardiac shock signals, and each sample includes: a time coding matrix for training, and a true category label corresponding to the time coding matrix for training.
[0076] Preferably,
[0077] The pre-set training condition is that during the training process, the cross entropy loss function of the BiLSTM-FCN model changes less than a preset threshold in several consecutive iterations.
[0078] On the other hand, this embodiment also provides a multi-class classification system for ballistocardial signals based on BiLSTM-FCN, including:
[0079] at least one processor; and
[0080] At least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the multi-category classification method for cardiac ballistic signals based on BiLSTM-FCN as described in the first aspect.
[0081] The beneficial effects of the present invention are:
[0082] The present invention discloses a multi-category classification method for cardiac ballistogram signals based on BiLSTM-FCN. By adopting a BiLSTM-FCN model, in which BiLSTM is used to extract temporal features and FCN is used to extract spatial features, and optimization is performed through a feature fusion layer, the method can more accurately extract the temporal and spatial features of cardiac ballistogram signals compared to the existing technology, thereby improving the robustness and accuracy of classification and achieving the effect of improving the classification performance of cardiac ballistogram signals.
[0083] The present invention adopts a BiLSTM-FCN-based multi-category classification method for cardiac ballistic signals. Due to the use of automatic feature extraction, it avoids the reliance on manual feature engineering. Compared with the existing technology, it can reduce the influence of human factors on the classification results, improve the generalization ability of the model, and achieve the effect of adapting to the individual differences of different patients.
[0084] The present invention provides a multi-category classification method for cardiac ballistometry signals based on BiLSTM-FCN. By adopting a local noise removal method, the method effectively reduces external interference in the signal preprocessing stage. Compared with the existing technology, the method can enhance the noise resistance and improve the classification stability, thereby achieving a more reliable cardiac ballistometry signal classification effect.
[0085] The present invention proposes a multi-category classification method for cardiac ballistometry signals based on BiLSTM-FCN. By adopting the BiLSTM structure instead of the traditional RNN, the computational efficiency is improved. Compared with the existing technology, the method can process time series signals faster, thereby achieving the effect of improving the real-time cardiac ballistometry signal monitoring capability.
[0086] The present invention provides a multi-category classification method for cardiac ballistic signals based on BiLSTM-FCN. Since the Softmax function is used to calculate the category probability and the final category label is output based on the maximum probability principle, compared with the existing technology, the method can improve the accuracy of cardiac ballistic signal classification and provide reliable data support for intelligent diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0087] Figure 1 This is a flow chart of a multi-class classification method for cardiac ballistic signals based on BiLSTM-FCN of the present invention;
[0088] Figure 2This is a flow chart of a multi-category classification method for cardiac ballistic signals based on BiLSTM-FCN in Example 2 of the present invention. DETAILED DESCRIPTION
[0089] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.
[0090] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. Instead, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.
[0091] Example 1
[0092] See also Figure 1 This embodiment provides a multi-class classification method for cardiac ballistic signals based on BiLSTM-FCN, the method comprising:
[0093] S1, collecting cardiac shock signals;
[0094] S2. Preprocessing the ballistocardiographic signal to obtain a preprocessed signal;
[0095] S3, performing time series coding on the preprocessed signal to obtain a time coding matrix;
[0096] S4. Input the time coding matrix into the trained BiLSTM-FCN model to obtain the category prediction label corresponding to the heartbeat signal;
[0097] The BiLSTM-FCN model includes:
[0098] Input layer, used to receive the time encoding matrix in the input BiLSTM-FCN model;
[0099] The BiLSTM temporal feature extraction layer is used to obtain the temporal feature matrix based on the time coding matrix input by the input layer;
[0100] The FCN spatial feature extraction layer is used to process the temporal feature matrix output by the BiLSTM temporal feature extraction layer to obtain a spatial feature matrix;
[0101] The feature fusion layer is used to fuse the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix;
[0102] The classification decision layer is used to perform feature mapping and classification calculation based on the fusion feature matrix, ultimately obtain the category probability distribution of the heartbeat signal, and output the category prediction label;
[0103] The class prediction label is the label corresponding to the class with the highest class probability.
[0104] For example, suppose you want to classify different types of cardiac impulse signals (such as normal cardiac impulses, premature beats, atrial fibrillation, ventricular fibrillation, etc.). The following is an application example of this method:
[0105] Collect cardiac shock signals from an electrocardiogram (ECG) or other physiological monitoring equipment, such as 5-second cardiac shock data.
[0106] Perform filtering and denoising, such as using wavelet transform or bandpass filtering to remove power frequency interference, baseline drift, and myoelectric noise. Normalize the signal to reduce amplitude differences between individuals.
[0107] A sliding window method is used to extract time segments, and each segment is time-coded, such as by using short-time Fourier transform (STFT) or wavelet transform to convert the one-dimensional signal into a two-dimensional time-coded matrix.
[0108] The BiLSTM temporal feature extraction layer extracts time series features, such as the short-term and long-term dependencies of the cardiac signal.
[0109] The FCN spatial feature extraction layer further learns the spatial patterns in the temporal features, such as waveform shape, amplitude changes, etc.
[0110] The feature fusion layer fuses the features extracted by BiLSTM and FCN to improve information integrity.
[0111] The classification decision layer uses Softmax to calculate the category probability distribution and outputs the category label with the highest category probability, such as "atrial fibrillation" or "ventricular fibrillation".
[0112] In a specific embodiment, the S2 specifically includes:
[0113] The signal preprocessing step adopts a local adaptive transformation denoising method to remove local noise and non-stationary baseline drift of the heart ballistic signal. The calculation formula of the local adaptive transformation denoising method is:
[0114]
[0115] S pre (t) is the preprocessing signal; S raw (t) is the heartbeat signal; N is the total number of local noise components; a k is the local amplitude weight corresponding to the kth local noise component; β kis the time-varying attenuation factor corresponding to the kth local noise component; ω k is the dominant frequency of the signal corresponding to the kth local noise component; is the local phase offset corresponding to the kth local noise component; t is time.
[0116] This formula models the local noise in the cardiac signal as multiple sinusoidal components with decaying amplitudes and uses the exponential decay term This method gradually suppresses noise, thereby removing local interference. This method adapts to noise characteristics over time, improving denoising effectiveness, unlike traditional filtering methods that uniformly process the entire signal. Because it uses a locally adaptive approach, the denoising process preserves key features of the ballistocardial signal, such as amplitude and timing information, ensuring that the subsequent BiLSTM-FCN classification model can accurately identify the class.
[0117] In a specific application, the S3 specifically includes:
[0118] S31, determining the short-time spectral component of the ballistocardi signal and its corresponding dominant frequency according to the preprocessed signal, and extracting phase information;
[0119] S32. Construct a time coding matrix based on the short-time spectral components of the ballistocardial signal and its corresponding dominant frequency and extracting phase information.
[0120] The S32 specifically includes:
[0121] According to the short-time spectrum components of the heartbeat signal and its corresponding dominant frequency, and extracting the phase information, the time coding matrix is constructed using formula (1);
[0122] The formula (1) is:
[0123] T c (i, j) = |F k |·sin(2πf k j+θ k )
[0124] Among them, T c (i, j) is the element with index i, j in the time coding matrix;
[0125] F k is the complex value of the short-time spectral component;
[0126] f k is the dominant frequency of the short-time spectral component; this value reflects the primary oscillation characteristics of the ballistocardia signal within different time windows. By incorporating the dominant frequency into the encoding matrix, we can better capture the pattern of the ballistocardia signal and improve the BiLSTM's ability to identify signal categories.
[0127] θ k is the phase information of the short-time spectral component. Since BiLSTM is primarily used to learn temporal features, introducing phase information can improve the model's ability to understand the correlation between different time windows, making the classification results more accurate.
[0128] In formula (1), |F k | represents the complex modulus value (i.e., amplitude) of the short-time spectral component. Since different types of cardiac ballistic signals may have different amplitude distributions in the frequency domain, this information helps to distinguish different types of cardiac ballistic signals. sin(2πf k j+θ k ) forms a time-varying oscillating signal on the time axis, effectively reflecting the signal's periodicity. This enables the BiLSTM to better learn the timing characteristics of the heartbeat signal and improve classification robustness.
[0129] In this embodiment, the BiLSTM time series feature extraction layer obtains a time series feature matrix based on the time coding matrix input by the input layer, specifically including:
[0130] The BiLSTM time series feature extraction layer is used to perform bidirectional time series feature extraction on the time coding matrix input by the input layer, and calculate the time series feature matrix;
[0131] Formula (2) is used to extract the time series features;
[0132] The formula (2) is:
[0133]
[0134] H BiLSTM =[h1,…,h t ,…,h T ] is the time series feature matrix calculated by BiLSTM;
[0135] h t is the hidden state vector calculated by BiLSTM at time step t; x t is the input of the time encoding matrix at time step t; W h 、U h 、b h is the trainable weight matrix and bias term of BiLSTM; W h is the input weight matrix, U h is the hidden state weight matrix, b h is the bias term; σ( ) is the activation function; is the phase change of the time coding matrix at time step t; γ is the gated memory decay coefficient.
[0136] In this embodiment, the FCN spatial feature extraction layer is used to process the temporal feature matrix output by the BiLSTM temporal feature extraction layer to obtain a spatial feature matrix, specifically including:
[0137] The FCN spatial feature extraction layer calculates the temporal feature matrix output by the BiLSTM temporal feature extraction layer using formula (3) to obtain the spatial feature matrix;
[0138] The formula (3) is:
[0139]
[0140] H FAN (i, j) is the i-th, j-th element of the spatial feature matrix calculated by the FCN spatial feature extraction layer;
[0141] H BiLSTM (i, j) is H BiLSTM The i-th and j-th elements of the time series feature matrix;
[0142] W m is the weight parameter of the convolution kernel, m is the convolution window index, and M is the convolution kernel size;
[0143] μ is the transformation factor based on the phase shift;
[0144] a1 is the nonlinear enhancement coefficient;
[0145] a2 is the gradient modulation coefficient;
[0146] Indicates H BiLSTM Output the local gradient of the matrix.
[0147] This example uses BiLSTM to extract temporal features, then utilizes FCN for spatial feature extraction, and combines nonlinear transformations and gradient modulation to enhance the model's ability to handle complex spatiotemporal data. This allows for more accurate modeling of spatiotemporal dependencies, improving performance in various spatiotemporal prediction and classification tasks.
[0148] The feature fusion layer fuses the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix, which specifically includes:
[0149] The feature fusion layer uses formula (4) to fuse the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix;
[0150] The formula (4) is:
[0151] H fusion =c·H BiLSTM +(1-c)·H FCN ;
[0152] H fusion is the fusion feature matrix;
[0153] H BiLSTM is the time series feature matrix;
[0154] H FCN is the spatial feature matrix;
[0155] c is the dynamic fusion weight;
[0156] in,
[0157] This formula indicates that c is an adaptive dynamic weight calculated based on the modulus (or absolute value sum) of the feature matrix. BILSTM With stronger time series characteristics (larger values), c is larger, indicating a stronger tendency towards time series characteristics. FCN The greater the contribution (the spatial features are more significant), the smaller c is, indicating a greater bias towards spatial features.
[0158] In this embodiment, before S1, the following steps are further included: S0, using a training data set to train the BiLSTM-FCN model until a preset training condition is met, thereby obtaining a trained BiLSTM-FCN model;
[0159] The training data set includes a plurality of samples of cardiac shock signals, and each sample includes: a time coding matrix for training, and a true category label corresponding to the time coding matrix for training.
[0160] The pre-set training condition is that during the training process, the cross entropy loss function of the BiLSTM-FCN model changes less than a preset threshold in several consecutive iterations.
[0161] Example 2
[0162] This second embodiment provides a multi-class classification method for ballistocardial signals based on BiLSTM-FCN, including the following steps:
[0163] A1. Obtain the original cardiac signal and perform preprocessing to remove local noise;
[0164] A2. Extract features from the preprocessed signal and construct a time coding matrix;
[0165] A3. Use the training dataset to train the BiLSTM-FCN model to obtain a trained classification model.
[0166] A4. Use the trained BiLSTM-FCN model to classify the newly input cardiac arrest signal and output the corresponding category prediction label.
[0167] The preprocessing process in step A1 includes local noise removal, which is mathematically expressed as follows:
[0168]
[0169] in:
[0170] S pre (t) is the denoised cardiac signal;
[0171] S raw (t) is the original cardiac signal;
[0172] N is the total number of local noise components;
[0173] a k is the local amplitude weight, which indicates the intensity of the noise component;
[0174] β k is the time-varying attenuation factor, which indicates the speed at which the noise decays over time;
[0175] ω k is the dominant frequency extracted by short-time Fourier transform (STFT);
[0176] is the local phase shift;
[0177] t is the time variable.
[0178] In this embodiment, the time coding matrix T in step A2 i，j The construction method is:
[0179]
[0180] in:
[0181] T i，j is the element in the time coding matrix;
[0182] F i The complex spectral components extracted by short-time Fourier transform (STFT) or continuous wavelet transform (CWT);
[0183] θ i is the phase information of the spectral component;
[0184] i represents the time window index;
[0185] j represents the frequency index.
[0186] The training data set in step A3 includes multiple cardiac shock signal samples, each of which includes:
[0187] Temporal encoding matrix for training; true category labels, stored in one-hot encoding form.
[0188] The sources of training data sets include: real patient data or simulated signal data.
[0189] The BiLSTM-FCN model consists of a BiLSTM temporal feature extraction layer, which extracts temporal features and takes an input of shape (T, D), where T is the number of time steps and D is the feature dimension; an FCN spatial feature extraction layer, which extracts spatial features and uses a convolution kernel of size (3, 3); a feature fusion layer, which uses a concatenation operation to fuse the BiLSTM and FCN features; and a classification decision layer, which uses a softmax activation function to output class labels. If the training loss does not reach the set threshold, additional training is performed using the Adam optimizer with an adjusted learning rate.
[0190] In this embodiment, the BiLSTM (bidirectional long short-term memory) network captures the temporal dependencies of cardiac signal impulses and is particularly suitable for long-term trend analysis, such as rhythm patterns and abnormal impulses in cardiac signal impulses. The FCN (fully convolutional network) extracts local spatial features, such as spectral distribution and transient signal characteristics, and is suitable for analyzing short-term variations in cardiac signal impulses. The feature fusion layer combines the features extracted by the BiLSTM and FCN, enabling the model to simultaneously consider global temporal information and local spatial information, thereby improving classification accuracy.
[0191] The training dataset contains both real patient data and simulated signal data, improving the model's adaptability and generalization capabilities and preventing overfitting to certain data distributions. The use of one-hot encoding for the true category labels makes the model's training objectives more stable.
[0192] Optimization strategy: The Adam optimizer is used to adjust the learning rate to achieve faster model convergence. When the training loss does not reach the set threshold, the learning rate is dynamically adjusted to prevent overfitting or gradient vanishing, thereby improving training stability.
[0193] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.
[0194] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; and internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.
[0195] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is at a lower level than the second feature.
[0196] In the description of this specification, the terms "one embodiment", "some embodiments", "embodiments", "examples", "specific examples" or "some examples" refer to the specific features, structures, materials or characteristics described in conjunction with the embodiment or example and included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and features of different embodiments or examples, unless they are mutually inconsistent.
[0197] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.
Claims
1. A multi-class classification method for cardiac ballistic signals based on BiLSTM-FCN, characterized in that: The method comprises: S1, collecting cardiac shock signals; S2. Preprocessing the ballistocardiographic signal to obtain a preprocessed signal; S3, performing time series coding on the preprocessed signal to obtain a time coding matrix; S4. Input the time coding matrix into the trained BiLSTM-FCN model to obtain the category prediction label corresponding to the heart ballistic signal; The BiLSTM-FCN model includes: Input layer, used to receive the time encoding matrix in the input BiLSTM-FCN model; The BiLSTM temporal feature extraction layer is used to obtain the temporal feature matrix based on the time coding matrix input by the input layer; The FCN spatial feature extraction layer is used to process the temporal feature matrix output by the BiLSTM temporal feature extraction layer to obtain a spatial feature matrix; The feature fusion layer is used to fuse the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix; The classification decision layer is used to perform feature mapping and classification calculation based on the fusion feature matrix, ultimately obtain the category probability distribution of the heartbeat signal, and output the category prediction label; The class prediction label is the label corresponding to the class with the highest class probability.
2. The multi-class classification method for ballistic heart signals based on BiLSTM-FCN according to claim 1, characterized in that The S2 specifically includes: The signal preprocessing step adopts a local adaptive transformation denoising method to remove local noise and non-stationary baseline drift of the heart ballistic signal. The calculation formula of the local adaptive transformation denoising method is: S pre (t) is the preprocessed signal; S raw (t) is the heartbeat signal; N is the total number of components of local noise; a k is the local amplitude weight corresponding to the kth local noise component; β k is the time-varying attenuation factor corresponding to the kth local noise component; ω k is the dominant frequency of the signal corresponding to the kth local noise component; is the local phase offset corresponding to the kth local noise component; t is time.
3. The multi-class classification method for ballistic heart signals based on BiLSTM-FCN according to claim 1, characterized in that The S3 specifically includes: S31, determining the short-time spectral component of the ballistocardi signal and its corresponding dominant frequency according to the preprocessed signal, and extracting phase information; S32. Construct a time coding matrix based on the short-time spectral components of the ballistocardial signal and its corresponding dominant frequency and extracting phase information.
4. The multi-class classification method for cardiac ballistic signals based on BiLSTM-FCN according to claim 3, characterized in that: The S32 specifically includes: According to the short-time spectrum components of the heartbeat signal and its corresponding dominant frequency, and extracting the phase information, the time coding matrix is constructed using formula (1); The formula (1) is: T c (i,j)=|F k |·sin(2πf k j+θ k ) Among them, T c (i, j) is the element with index i, j in the time coding matrix; F k is the complex value of the short-time spectral component; f k is the dominant frequency of the short-time spectral component; θ k is the phase information of the short-time spectral component.
5. The multi-class classification method for ballistic heart signals based on BiLSTM-FCN according to claim 1, characterized in that The BiLSTM temporal feature extraction layer obtains the temporal feature matrix based on the time coding matrix input by the input layer, which specifically includes: The BiLSTM time series feature extraction layer is used to perform bidirectional time series feature extraction on the time coding matrix input by the input layer, and calculate the time series feature matrix; Formula (2) is used to extract the time series features; The formula (2) is: H BiLSTM =[h1,…,h t ,…,h T ] is the time series feature matrix calculated by BiLSTM; h t is the hidden state vector calculated by BiLSTM at time step t; x t is the input of the time encoding matrix at time step t; W h 、U h 、b h is the trainable weight matrix and bias term of BiLSTM; σ() is the activation function; is the phase change of the time coding matrix at time step t; γ is the gated memory decay coefficient.
6. The multi-class classification method for ballistic heart signals based on BiLSTM-FCN according to claim 5, characterized in that: The FCN spatial feature extraction layer is used to process the temporal feature matrix output by the BiLSTM temporal feature extraction layer to obtain a spatial feature matrix. Specifically, it includes: The FCN spatial feature extraction layer calculates the temporal feature matrix output by the BiLSTM temporal feature extraction layer using formula (3) to obtain the spatial feature matrix; The formula (3) is: H FAN (i, j) is the i-th, j-th element of the spatial feature matrix calculated by the FCN spatial feature extraction layer; H BiLSTM (i, j) is H BiLSTM The i-th and j-th elements of the time series feature matrix; W m is the weight parameter of the convolution kernel, m is the convolution window index, and M is the convolution kernel size; μ is the transformation factor based on the phase shift; a1 is the nonlinear enhancement coefficient; a2 is the gradient modulation coefficient; Indicates H BiLSTM Output the local gradient of the matrix.
7. The multi-class classification method for ballistic heart signals based on BiLSTM-FCN according to claim 1, characterized in that: The feature fusion layer fuses the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix, which specifically includes: The feature fusion layer uses formula (4) to fuse the temporal feature matrix and the spatial feature matrix to obtain a fused feature matrix; The formula (4) is: H fusion =c·H BiLSTM +(1-c)·H FCN ; H fusion is the fusion feature matrix; H BiLSTM is the time series feature matrix; H FCN is the spatial feature matrix; c is the dynamic fusion weight; in, 8. The multi-class classification method for ballistic heart signals based on BiLSTM-FCN according to claim 7, characterized in that: The method further includes, before S1: S0. Use the training data set to train the BiLSTM-FCN model until the pre-set training conditions are met, and then a trained BiLSTM-FCN model is obtained; The training data set includes a plurality of samples of cardiac shock signals, and each sample includes: a time coding matrix for training, and a true category label corresponding to the time coding matrix for training.
9. The multi-class classification method for ballistic heart signals based on BiLSTM-FCN according to claim 8, characterized in that: in, The pre-set training condition is that during the training process, the cross entropy loss function of the BiLSTM-FCN model changes less than a preset threshold in several consecutive iterations.
10. A multi-class classification system for cardiac ballistic signals based on BiLSTM-FCN, characterized in that: include: at least one processor; as well as At least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the multi-category classification method for cardiac ballistic signals based on BiLSTM-FCN as described in any one of claims 1-9.