A non-contact abnormal event prompting method and system based on heart impact signal rhythm-morphology collaborative modeling

Abstract

This invention discloses a non-contact abnormal event alerting method and system based on rhythm-morphology co-modeling of cardiac impact signals. The method includes: acquiring cardiac impact signals from the tested object using a non-contact sensor; constructing rhythmic signal segments based on the cardiac impact signals, and constructing corresponding morphological signal segments using the rhythmic signal segments as basic units, assigning both to the same unique identifier; extracting rhythmic and morphological features using a dual-branch feature extraction module; performing homologous pairing of the rhythmic and morphological features based on the unique identifier, splicing them to form joint features, and generating channel weights based on channel responses to obtain weighted fusion features; inputting the weighted fusion features into a discriminant model, and outputting cardiac impact signal abnormal event alert information. This invention can improve the accuracy and robustness of non-contact BCG signal abnormal event alerting.

Description

Technical Field

[0001] This invention belongs to the field of cardiac impact signal processing technology, specifically relating to a non-contact abnormal event alerting method and system based on cardiac impact signal rhythm-morphology co-modeling. Background Technology

[0002] Premature ventricular contractions (PVCs) are a common type of ventricular arrhythmia in clinical practice, and their occurrence is directly related to abnormalities in cardiac electrophysiological activity and mechanical contraction processes. Frequent PVCs often indicate an underlying risk of organic heart disease, and in severe cases, may induce more complex arrhythmias such as ventricular tachycardia and ventricular fibrillation. Therefore, long-term, continuous monitoring and identification of abnormal patterns related to PVCs are of great significance for cardiovascular disease risk screening, status observation, and subsequent manual review.

[0003] Currently, the identification of premature ventricular contractions (PVCs) primarily relies on electrocardiogram (ECG) acquisition technology. This technology uses electrodes attached to the skin to collect potential changes generated during myocardial depolarization and repolarization, which can accurately reflect the state of cardiac electrical activity. However, this contact-based detection method has significant limitations: direct contact between the electrodes and the skin can easily cause discomfort, skin allergies, or irritation. In long-term monitoring or sleep monitoring scenarios, the electrodes are also prone to falling off due to changes in body position, affecting the stability and continuity of signal acquisition and limiting its application in non-invasive monitoring scenarios.

[0004] To overcome the aforementioned drawbacks of contact-based ECG acquisition, non-contact monitoring technologies based on cardiac mechanical activity have gradually gained attention. Among these, cardiac impulse signal (BCG) indirectly reflects the state of cardiac mechanical activity by sensing the weak mechanical vibrations generated in the human body during the heart's ejection process. It can acquire heartbeat-related signal information without electrodes contacting the skin, and has application advantages in the fields of sleep monitoring and long-term health monitoring.

[0005] Unlike ECG signals, which primarily reflect cardiac electrical activity, BCG signals characterize the mechanical vibrations of the human body caused by cardiac ejection and hemodynamic changes. They are more significantly affected by respiratory modulation, body motion artifacts, changes in mattress coupling, and individual differences, exhibiting strong non-stationarity and a low signal-to-noise ratio. For abnormal events related to premature ventricular contractions (PVCs), the abnormal manifestations in BCG signals are not limited to a single level: on the one hand, they manifest as changes in the cardiac interval and overall rhythm structure; on the other hand, they are accompanied by distortions in the mechanical waveforms of single or adjacent cardiac beats. Because these abnormalities differ significantly in time scale and representation, traditional single-scale analysis methods that borrow from ECG processing approaches struggle to reliably and comprehensively identify and indicate related abnormal patterns.

[0006] Furthermore, in non-contact BCG acquisition scenarios, when constructing local morphological fragments from rhythmic fragments, inconsistencies in local fragment quality, incomplete feature extraction, and unreliable sample correspondences may arise between samples of different scales. Directly fusing the bi-branch features can easily introduce mispairing and redundant noise, affecting the stability of the abnormal event alert results.

[0007] Therefore, there is an urgent need to develop a signal processing technology solution suitable for non-contact BCG signal scenarios, which can effectively model cardiac rhythm information and cardiac morphology information under complex interference conditions, and achieve reliable fusion based on the correspondence of homologous samples, thereby providing stable indication of abnormal patterns related to premature ventricular contractions in cardiac impact signal segments, so as to meet the engineering application needs in long-term, continuous, and non-intrusive monitoring scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide a non-contact abnormal event alerting method and system based on rhythm-morphology co-modeling of cardiac impact signals, in order to solve the problems that existing contact-based ECG acquisition is difficult to apply to long-term non-intrusive monitoring scenarios, and that existing cardiac impact signal processing methods are difficult to stably characterize rhythm abnormalities and local mechanical waveform distortions simultaneously under complex interference conditions, and lack reliable correspondence and fusion mechanisms between features of different scales.

[0009] In a first aspect, the present invention provides a non-contact abnormal event alerting method based on cardiac impulse signal rhythm-morphology co-modeling, comprising:

[0010] The cardiac impact signal of the tested object is collected using a non-contact sensor.

[0011] Based on the cardiac impact signal, a rhythmic signal segment is constructed, and a corresponding morphological signal segment is constructed using the rhythmic signal segment as the basic unit. The same unique identifier is assigned to the rhythmic signal segment and its corresponding morphological signal segment.

[0012] The rhythm signal segment is input into the rhythm feature extraction branch. The heart rate interval fluctuation and rhythm instability in the rhythm signal segment are modeled by time-series coding to obtain multiple rhythm sub-features and their corresponding abnormal response intensities. The multiple rhythm sub-features are then aggregated according to the abnormal response intensities to obtain the rhythm features.

[0013] The morphological signal segment is input into the morphological feature extraction branch. The morphological information of the local cardiac mechanical waveform is extracted by local waveform encoding. The temporal evolution relationship between adjacent cardiac beats is modeled, and the key time steps are weighted and aggregated to obtain the morphological features.

[0014] The rhythmic features and morphological features are paired based on the unique identifier; only rhythmic features and morphological features with the same unique identifier are fused into a joint feature; when the rhythmic features and morphological features do not have the same unique identifier, or when either feature is not a valid feature, no dual-branch fusion is performed on the corresponding sample.

[0015] Channel weights are generated based on the channel responses of the joint features, and the joint features are then subjected to channel-by-channel weighted modulation using the channel weights to obtain weighted fused features.

[0016] The weighted fusion features are input into the discrimination model to obtain abnormal event prompts for cardiac impact signals.

[0017] Preferably, the rhythm feature extraction branch uses a temporal convolutional network to perform temporal encoding on the rhythm signal segment, and the temporal convolutional network includes multiple dilated convolutional layers.

[0018] Aggregating the multiple rhythmic sub-features based on the abnormal response intensity includes: selecting several rhythmic sub-features with higher abnormal response intensity from the multiple rhythmic sub-features according to the abnormal response intensity, and performing weighted aggregation on the selected rhythmic sub-features to obtain the rhythmic features.

[0019] Preferably, the morphological feature extraction branch includes a one-dimensional convolutional network, a bidirectional long short-term memory network, and an attention mechanism; the one-dimensional convolutional network is used to perform local waveform encoding on the morphological signal segment; the bidirectional long short-term memory network is used to model the temporal evolution relationship between adjacent heartbeats; the attention mechanism is used to calculate the attention weight of each time step, and to perform weighted aggregation of the temporal features of each time step according to the attention weight to obtain the morphological features.

[0020] Preferably, before fusing the rhythmic features and the morphological features that have the same unique identifier into a joint feature, the method further includes:

[0021] Determine whether both the rhythmic features and the morphological features have been successfully extracted.

[0022] If both the rhythmic feature and the morphological feature are successfully extracted, then the rhythmic feature and the morphological feature are concatenated into the joint feature; if either the rhythmic feature or the morphological feature is not successfully extracted, then the rhythmic feature and the morphological feature of the corresponding sample are not concatenated into the joint feature.

[0023] Preferably, the process of obtaining the weighted fusion features includes: calculating the sum of squared eigenvalues ​​of each feature channel in the joint features, and obtaining the joint feature energy based on the sum of squared eigenvalues; normalizing the joint features according to the joint feature energy to obtain normalized joint features; generating the channel weights based on the normalized joint features and learnable parameters; and multiplying the channel weights by the joint features channel by channel to obtain the weighted fusion features.

[0024] Preferably, the cardiac impact signal acquired by the non-contact sensor is preprocessed. The preprocessing includes: performing baseline drift removal, spike noise removal, bandpass filtering, and normalization on the cardiac impact signal in sequence to obtain the preprocessed cardiac impact signal.

[0025] Preferably, the baseline drift removal process includes linear detrending processing, the spike noise removal process includes Hampel filtering, the bandpass filtering process includes Butterworth bandpass filtering, and the normalization process includes Z-score normalization.

[0026] Preferably, constructing rhythmic signal segments based on the cardiac impact signal includes: dividing the preprocessed cardiac impact signal into non-overlapping segments according to a fixed time window length to obtain multiple rhythmic signal segments.

[0027] Constructing a corresponding morphological signal segment using the rhythmic signal segment as the basic unit includes: detecting the main J wave of the heartbeat from the rhythmic signal segment, extracting a local heartbeat sequence containing multiple consecutive heartbeats based on the main J wave of the heartbeat, and using the local heartbeat sequence as the morphological signal segment.

[0028] Preferably, the fixed time window length is 8s to 20s; the local heartbeat sequence contains 3 to 5 consecutive heartbeats; the unique identifier includes the subject number, the rhythm signal segment number, and the sampling start time.

[0029] Secondly, the present invention provides a non-contact abnormal event alerting system based on cardiac impulse signal rhythm-morphology co-modeling, used to execute the aforementioned non-contact abnormal event alerting method; the non-contact abnormal event alerting system includes:

[0030] The signal acquisition module is used to acquire cardiac impact signals of the tested object through a non-contact sensor.

[0031] The signal preprocessing module is used to preprocess the cardiac impact signal to obtain a preprocessed cardiac impact signal.

[0032] The dual-branch feature extraction module is used to construct rhythmic signal segments based on the preprocessed cardiac impact signals, construct corresponding morphological signal segments using the rhythmic signal segments as basic units, and assign the same unique identifier to the rhythmic signal segments and their corresponding morphological signal segments. Rhythmic features are obtained through the rhythmic feature extraction branch, and morphological features are obtained through the morphological feature extraction branch.

[0033] The feature fusion module is used to perform homologous pairing of the rhythmic features and the morphological features based on the unique identifier, and to concatenate only the rhythmic features and morphological features that have the same unique identifier and have been successfully extracted into a joint feature. The module generates channel weights based on the channel responses of the joint features, and uses the channel weights to perform channel-by-channel weighted modulation on the joint features to obtain a weighted fusion feature.

[0034] The classification decision module is used to input the weighted fusion features into the discrimination model to obtain abnormal event prompt information of cardiac impact signal.

[0035] Compared with the prior art, the present invention has at least the following beneficial effects.

[0036] 1. This invention assigns the same unique identifier to rhythmic signal segments and their corresponding morphological signal segments, and performs homologous pairing and selective fusion based on the unique identifier. It can adapt to the characteristics of inconsistent quality of local segments of BCG signals and the easy distortion of sample correspondence, and reduce the interference of incorrect pairing, missing features and low-quality segments on the fusion results.

[0037] 2. This invention splices rhythmic and morphological features after homologous pairing into joint features, and generates channel weights based on the channel responses of the joint features. The joint features are then subjected to channel-by-channel weighted modulation, which can adapt to the characteristics of low signal-to-noise ratio and a lot of redundant interference in BCG signals, enhance the discriminative information related to ventricular premature beats, and suppress irrelevant noise.

[0038] 3. This invention constructs rhythmic signal segments and corresponding morphological signal segments, and sets rhythmic feature extraction branches and morphological feature extraction branches respectively. It can take advantage of the fact that BCG signals simultaneously contain information on changes in cardiac rhythm and mechanical waveform distortion, and can collaboratively characterize abnormal patterns related to premature ventricular contractions from two scales: long-term rhythmic structure and local cardiac morphology.

[0039] 4. This invention performs time-series modeling of cardiac interval fluctuations and rhythm instability through rhythm feature extraction branches, and aggregates multiple rhythm sub-features according to the intensity of abnormal response. It can highlight segments with strong abnormal responses, taking into account the sparse distribution of abnormal rhythm segments in long-term BCG signals, and reduce the dilution of abnormal rhythm features by a large number of normal cardiac segments.

[0040] 5. This invention extracts the morphological information of local cardiac mechanical waveforms through morphological feature extraction branches, models the temporal evolution relationship between adjacent cardiac beats, and performs weighted aggregation on key time steps. This can enhance the ability of key local waveforms to characterize abnormal patterns, given that ventricular premature beats in BCG signals are characterized by local mechanical vibration waveform distortion. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of non-contact cardiac impact signal acquisition in Embodiment 1 of the present invention.

[0042] Figure 2 This is a flowchart of the signal preprocessing process in Embodiment 1 of the present invention.

[0043] Figure 3 This is a flowchart of the rhythm feature extraction branch in Embodiment 1 of the present invention.

[0044] Figure 4 This is a flowchart of the morphological feature extraction branch in Embodiment 1 of the present invention.

[0045] Figure 5 This is a flowchart of the feature fusion process in Embodiment 1 of the present invention.

[0046] Figure 6 A system block diagram of a non-contact abnormal event notification system is provided for Embodiment 2 of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be described below with reference to the accompanying drawings and embodiments. The described embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention.

[0048] This invention provides a non-contact abnormal event alerting method based on rhythm-morphology co-modeling of cardiac impulse signals. The method and system process non-contact acquired cardiac impulse signals, employing a dual-branch co-modeling approach combining rhythm and morphological features, and selectively fusing data based on the correspondence of homologous samples to output abnormal event alerting information within cardiac impulse signal segments. This abnormal event alerting information characterizes the presence of signal abnormalities related to premature ventricular contractions (PVCs) within the cardiac impulse signal segments and is used for subsequent monitoring, screening, or manual review; it is not directly used as a clinical diagnostic conclusion.

[0049] Example 1

[0050] A non-contact abnormal event indication method based on rhythm-morphology co-modeling of cardiac impulse signals is proposed to indicate abnormal signal patterns related to ventricular premature beats (PVCs) based on non-contactly acquired cardiac impulse signals. The method uses a cardiac impulse signal abnormal event identification model to predict cardiac impulse signal anomalies. This model includes a signal preprocessing module, a dual-branch feature extraction module, a feature fusion module, and a classification decision module. Specifically, this embodiment addresses the mechanical vibration signal characteristics of BCG, which differ from ECG, by constructing a rhythm feature extraction branch and a morphological feature extraction branch to co-model long-term rhythm abnormalities and local mechanical waveform distortions caused by PVCs, respectively. Simultaneously, by assigning the same unique identifier to dual-branch samples generated from the same long-term sample and performing selective fusion based on this identifier, only samples from which effective features have been successfully extracted in both branches are jointly judged, thereby improving the accuracy, stability, and robustness of PVC-related abnormal pattern indication under complex interference conditions.

[0051] The non-contact abnormal event notification method includes the following steps:

[0052] Step S100: Acquire raw cardiac impact signals and perform preprocessing.

[0053] S110: As Figure 1 As shown, a non-contact sensor array is used to collect cardiac mechanical vibration signals of the monitored subject during sleep. This non-contact sensor array can be deployed in a smart mattress, sleep monitoring device, or other non-contact physiological signal monitoring device.

[0054] The cardiac mechanical vibration signal is converted into a voltage signal by a signal conditioning circuit, with an amplitude range of -5V to +5V. The data processing unit digitizes the voltage signal at a sampling rate of 125Hz to obtain the raw cardiac impact signal. Among them, the original cardiac impact signal This indicates the sequence number of the cardiac impact signal obtained after digital sampling. Amplitude at the location; sampling point number This indicates the discrete sampling point number in the original cardiac impact signal. Total number of sampling points This indicates the total number of sampling points contained in the original cardiac impact signal.

[0055] S120: As Figure 2 As shown, the original cardiac impact signal is preprocessed as follows:

[0056] S121: Linear detrending processing is performed on the raw cardiac impact signal. The trend term is fitted using the least squares method.

[0057]

[0058] in, This represents the estimated value of the trend term, specifically the first value obtained through linear fitting. The low-frequency trend term corresponding to each sampling point; , These represent the trend slope parameter and the trend intercept parameter in the linear trend term, respectively.

[0059] Trend slope parameter and trend intercept parameter The following least squares constraints must be satisfied:

[0060]

[0061] The detrending signal is:

[0062]

[0063] The linear detrending process described above can reduce low-frequency baseline drift caused by sensor drift, slow changes in body position, or changes in mattress coupling state, allowing subsequent feature extraction to focus more on heartbeat-related mechanical vibration information.

[0064] S122: Perform Hampel filtering to denoise the detrended signal. Set the sliding window length to... , where half window length Calculate the window value:

[0065]

[0066] in, Indicates the first The sliding window corresponding to each sampling point is specifically based on the sampling point number. A local signal window centered on the signal; This represents the value in a window, specifically a sliding window. The median of the amplitude of each signal within the range. This is for median extraction.

[0067] Sliding window Represented as:

[0068]

[0069] Calculate the absolute median:

[0070]

[0071] in, This refers to the index of the sampling points within the window, specifically representing the sliding window. The sampling point number within.

[0072] Calculate the robust standard deviation based on the absolute median:

[0073]

[0074] Among them, robust standard deviation Used to identify sudden spike noise; coefficient 1.4826 is the scaling factor for converting the absolute median deviation into an estimate of the standard deviation of the Gaussian distribution.

[0075] When the following conditions are met:

[0076]

[0077] When an outlier is identified, the current sampling point is determined to be an outlier and replaced with a window value. Output denoised signal .

[0078] Denoising signal This is the cardiac impulse signal obtained after replacing outliers with Hampel filtering. Hampel filtering denoising can reduce the impact of sudden spike noise on cardiac waveform recognition and local morphological feature extraction.

[0079] S123: For the denoised signal Perform Butterworth bandpass filtering. A fourth-order Butterworth bandpass filter with a passband range of 1Hz to 12Hz is used to filter out breathing components in the range of 0.1Hz to 0.5Hz and high-frequency noise above 12Hz. The output filtered signal... Filtered signal The cardiac impulse signal obtained after Butterworth bandpass filtering is at time step number Amplitude at the specified point; Time step number This indicates the discrete-time number in the filtered signal.

[0080] S124: For the filtered signal Perform Z-score normalization. Calculate the mean of the filtered signal:

[0081]

[0082] in, This represents the total number of time steps, specifically the total number of time steps involved in the standardization calculation.

[0083] Calculate the standard deviation of the filtered signal:

[0084]

[0085] Calculate the standardized cardiac impact signal:

[0086]

[0087] This embodiment reduces amplitude differences between different channels, individuals, and acquisition time periods through standardization, ensuring that the model input is on a uniform scale.

[0088] Step S200: Construct rhythmic signal segments and morphological signal segments

[0089] S210: The preprocessed cardiac impulse signal obtained in step S100 is divided into non-overlapping segments according to a fixed time window length to construct rhythm signal segments.

[0090] In this embodiment, the preprocessed cardiac impact signal is divided into non-overlapping 12-second time windows. Since the sampling rate is 125Hz, each 12-second time window contains 1500 sampling points.

[0091] This yields a long-term rhythmic signal segment sequence. .in, Indicates the first The first monitored object A long time-history signal segment, specifically from the first... The first segment obtained from the preprocessed cardiac impact signal of the monitored object Each rhythmic signal segment at time step number The amplitude at that point; Indicates the number of the monitored object; Indicates the segment number; Indicates the total number of segments.

[0092] S220: For each rhythmic signal segment As the basic unit, from the rhythm signal segment Extract the corresponding local heartbeat waveforms to construct morphological signal segments. Specifically, from each rhythm signal segment... The J-wave, the dominant peak of the heartbeat, is detected, and a local heartbeat sequence containing four consecutive heartbeats is extracted based on the detected J-wave. Extracting the local heartbeat sequence based on the dominant J-wave allows different heartbeat waveforms to be aligned in time, thereby improving the consistency of morphological feature extraction.

[0093] The same unique identifier is assigned to the same rhythmic signal segment and the morphological signal segments generated from it, establishing a one-to-one correspondence between them at the sample level. In some embodiments, the unique identifier may consist of the monitored object number, the long-term signal segment number, the sampling start time, or a combination thereof. Through this unique identifier mechanism, it is possible to determine whether the rhythmic features and morphological features originate from the same long-term cardiac impulse signal segment in the subsequent feature fusion stage, avoiding erroneous fusion between samples from different sources.

[0094] Step S300: Extract rhythm features through the rhythm feature extraction branch

[0095] rhythmic signal segments The input rhythm feature extraction branch further divides each rhythm signal segment into multiple rhythm sub-segments; these multiple rhythm sub-segments all originate from the rhythm signal segment corresponding to the same unique identifier. The rhythm feature extraction branch uses a temporal convolutional network (TCN) to encode features for each rhythm sub-segment, expands the receptive field through dilated convolution, and extracts features related to cardiac interval fluctuations and rhythm instability.

[0096] In this embodiment, a 3-layer TCN is used to encode the features of each rhythm signal segment. The length of the first layer convolution kernel is... Inflation factor ; Length of the second convolutional kernel Inflation factor ; Length of the third convolutional kernel Inflation factor The first two layers have 64 output channels, and the third layer has 128 output channels. After TCN encoding, each rhythm signal segment outputs a 128-dimensional feature vector. and anomaly confidence score The anomaly confidence score Reflecting the The abnormal response intensity corresponding to each rhythmic signal segment.

[0097] No. The layer dilation convolution operation is defined as follows:

[0098]

[0099] in, This indicates the temporal convolutional network layer currently participating in the dilated convolution calculation; Indicates the first Layer output features; Indicates the first Layer input features; Indicates the first The kernel parameters of the convolutional layer are specifically the parameters of the first layer. In the layer dilated convolution, the first Each convolutional kernel weight; Indicates the kernel position number; Indicates the kernel length; Indicates the first Layer expansion factor.

[0100] After multi-layer TCN encoding, each rhythmic sub-segment outputs a corresponding feature vector. and anomaly confidence score Using a Top-k aggregation strategy, the segment with the highest confidence level is selected from multiple rhythmic sub-segments corresponding to the same unique identifier. Each rhythmic segment is aggregated to generate a rhythmic feature that corresponds to this unique identifier. and rhythm discrimination score The aggregated rhythmic features still correspond to the unique identifier and are used to splice with morphological features that have the same unique identifier.

[0101]

[0102]

[0103]

[0104] in, Indicates the number of aggregated fragments.

[0105] In this embodiment, the following settings are provided: The three rhythmic segments with the highest confidence levels are selected, and rhythmic features and rhythm discrimination scores are calculated. This Top-k aggregation strategy can highlight rhythmic segments with strong abnormal responses in long-term cardiac impulse signals, reducing the dilution effect of numerous normal cardiac beat segments on abnormal pattern recognition.

[0106] Step S400: Extract morphological features through the morphological feature extraction branch

[0107] The morphological signal segment is input into the morphological feature extraction branch. This branch, targeting the characteristics of ventricular premature beats (PVCs) in BCG signals as local mechanical waveform distortions and the dynamic evolution of adjacent heartbeats, employs 1D-CNN to encode the morphological structure of the local heartbeat sequence and combines it with bidirectional LSTM to model the temporal evolution relationship between adjacent heartbeats. Furthermore, an attention mechanism is used to weight and enhance key time steps corresponding to abnormal patterns related to PVCs in the local heartbeat sequence, generating morphological features.

[0108] In this embodiment, the 1D-CNN includes two one-dimensional convolutional layers with kernel lengths of 7 and 5, respectively, and 16 output channels for each. After extracting local morphological texture features using the 1D-CNN, the resulting feature sequence is input into a bidirectional LSTM to capture the temporal evolution of local morphological signals between adjacent time steps. The unidirectional hidden layer of the bidirectional LSTM has a dimension of 64, and after bidirectional concatenation, it outputs a 128-dimensional temporal feature representation at each time step. Among these features, the local heartbeat time step number is included. . This indicates the time step length of the local cardiac rhythm sequence.

[0109] Calculate the attention score using a fully connected layer:

[0110]

[0111] in, This indicates that the attention mechanism is used to process temporal feature vectors. Learnable attention weight matrix that undergoes linear transformation; This represents the learnable attention bias vector used to calculate the attention score in the attention mechanism; attention score Specifically, in the partial cardiac sequence, the [number]th ... The response strength of each time step to the anomaly pattern characterization.

[0112] Softmax normalization is applied to the attention scores to obtain the time step weights:

[0113]

[0114] in, Indicates the time step index.

[0115] The weighted morphological characteristics are calculated as follows:

[0116]

[0117] in, This represents the morphological features obtained after attention-weighted aggregation of the temporal features at each time step of the local heartbeat sequence. In this embodiment, the local heartbeat sequence contains 4 consecutive heartbeats, therefore the time step length of the local heartbeat sequence is... .

[0118] Through the above morphological feature extraction process, we can model the mechanical waveform distortion of a single or adjacent heartbeat caused by abnormal patterns related to premature ventricular contractions, and perform weighted enhancement on key time steps, thereby improving the ability to express local morphological abnormalities.

[0119] Step S500: Selective fusion based on unique identifier

[0120] S510: Perform homology pairing of rhythmic and morphological features based on unique identifiers. If rhythmic features... and morphological characteristics If the rhythmic features and morphological features have the same unique identifier and both branches successfully extract valid features, then the fusion operation is performed on the two. If the rhythmic features and morphological features do not have the same unique identifier, or if either branch fails to extract valid features, then the two-branch fusion operation is not performed on the corresponding samples.

[0121] Successful extraction means that the corresponding input segment meets the preset data quality conditions and, after processing by the corresponding feature extraction branch, can generate a feature vector with complete dimensions and valid values. Unsuccessful extraction means that a feature vector that meets the preset dimension and validity requirements cannot be generated, including but not limited to: missing, non-numerical, or invalid timestamps or signal values ​​of the input segment; insufficient signal length to form a preset analysis window; insufficient number of valid signal segments or valid heartbeats; signal quality lower than a preset threshold; failure to detect the main heartbeat peak or morphological positioning point that meets the preset conditions; or the generated feature vector has at least one of the following: missing values, infinite values, non-numerical values, or dimension mismatch.

[0122] S520: Entering the feature splicing stage, combining rhythmic features and morphological characteristics Perform splicing. Rhythmic characteristics. It has 128 dimensions and morphological features. The 128-dimensional features are concatenated to obtain a 256-dimensional joint feature vector:

[0123]

[0124] S530: Calculate joint eigenenergy:

[0125]

[0126] in, The total number of feature channels is represented by the joint feature vector. The number of feature channels contained therein; The joint eigenvector is represented by the first... The characteristic values ​​of each channel; This refers to a zero-prevention constant, specifically a small positive number used to prevent division by zero. In this embodiment, the zero-prevention constant... You can take 1×10 -6 .

[0127] S540: Normalize the joint eigenvectors:

[0128]

[0129] in, This represents the normalized joint eigenvector, specifically the joint eigenvector. The eigenvectors obtained after global energy normalization.

[0130] S550: Through learnable parameters , , Constructing channel weights:

[0131]

[0132] in, This represents the channel scaling parameter, specifically a learnable parameter used to adjust the channel response amplitude during the channel weight generation process; This represents the channel modulation parameters, specifically the learnable parameters used to adjust the contributions of the normalized joint feature channels during the channel weight generation process; This represents the channel bias parameter, specifically a learnable parameter used to adjust the channel baseline response during the channel weight generation process; This represents the channel weight vector, specifically the channel weights generated based on the joint feature vector. This represents the element-wise multiplication operator; This represents the Sigmoid function, used to generate channel weights.

[0133] S560: Perform weighted fusion to obtain weighted fusion features. :

[0134]

[0135] In this embodiment, the channel weight vector Each dimension of the weights acts on the joint feature vector. The corresponding characteristic channels are used to enhance the channel response related to abnormal event prompts and suppress redundant channel responses.

[0136] The channel weight generation process can adaptively adjust the contribution of different channels based on the global channel response of the joint features, enhance the discrimination information corresponding to abnormal patterns related to premature ventricular contractions, and suppress redundant noise.

[0137] Step S600: Output exception event message

[0138] Fusion features Input the discriminant model to obtain the abnormal probability corresponding to the cardiac impact signal segment. The discriminant model can be a logistic regression model, and the probability prediction calculation process is as follows:

[0139]

[0140] in, This indicates the probability of an anomaly, specifically the probability of identifying an abnormal pattern related to premature ventricular contractions in a cardiac impulse signal segment. Represents the discriminant model weight vector; weight vector transpose Represents the discriminant model weight vector Transpose; Discriminant model bias term The sigmoid activation function represents the bias parameter in a logistic regression discriminant model. This represents the method used to map the linear output of a logistic regression model to anomaly probabilities. Nonlinear functions.

[0141] Set preset threshold This is a probability threshold used to determine whether an abnormal event notification message is triggered. In this embodiment, a preset threshold is used. When the probability of an anomaly Greater than the preset threshold When the cardiac impulse signal segment is suspected of having an abnormal pattern related to premature ventricular contractions, an abnormal event prompt message is output; when the abnormal probability... Less than or equal to the preset threshold When this happens, output a message indicating that no exception event was triggered.

[0142] The abnormal event notification information may include at least one of the following: abnormal probability, abnormal event notification label, corresponding time window position, corresponding unique identifier, and signal segment index for manual review. Through the above method, the output of this embodiment is a notification of abnormal patterns in cardiac impact signal segments, rather than a direct diagnosis of disease in the monitored subject.

[0143] In some embodiments, the fixed time window length is not limited to 12s, and can be set to 8s, 10s, 15s, 20s, or other time lengths depending on the sampling rate, the state of the monitored object, and the model input requirements. The morphological signal segment is not limited to containing 4 consecutive heartbeats, but can also contain 3, 5, or more consecutive heartbeats to adapt to the local morphological modeling needs under different heart rate states. The rhythm feature extraction branch is not limited to a TCN structure, and can also employ a one-dimensional convolutional network, a Transformer encoder, a gated recurrent unit network, or a combination thereof. The morphological feature extraction branch is not limited to a combination of 1D-CNN and bidirectional LSTM, and can also employ a one-dimensional convolutional network and an attention network, a convolutional gated recurrent network, a temporal Transformer network, or a combination thereof. The discriminant model is not limited to a logistic regression model, and can also employ a fully connected neural network, a support vector machine, a gradient boosting tree, or other classification models to output the abnormal probability corresponding to the cardiac impact signal segment based on the fused features.

[0144] In this embodiment, a synchronous BCG-ECG dataset was used for training and validation. The dataset contained 53 subjects with abnormal cardiac impulse signals, specifically those exhibiting paroxysmal premature ventricular contractions (PVCs); 31 were male and 22 were female, aged 45 to 72 years. ECG signals were acquired using a clinical-grade Holter monitor at a sampling rate of 200 Hz, serving as a reference source for PVC-related abnormal events; BCG signals were acquired using the system described in this embodiment. Data was annotated by two senior physiological signal analysts, and discrepancies were arbitrated by a third expert.

[0145] The dataset was partitioned using a participant-independent splitting strategy, divided into training, validation, and test sets in a 7:2:1 ratio. The training set contained 37 participants, the validation set contained 11 participants, and the test set contained 5 participants. The AdamW optimizer was used for training, with an initial learning rate of 5 × 10⁻⁶. -5 The weight decays to 1×10 -3 The loss function used is the binary cross-entropy loss function, combined with Reduce-on-Plateau learning rate scheduling and early stopping mechanism. The maximum number of iterations is 60, and the batch size is 32. During training, overfitting is suppressed through gradient clipping and Dropout regularization, with a gradient clipping threshold of 1.0 and a Dropout ratio of 0.35.

[0146] The results of this embodiment and the comparison method are shown in the table below.

[0147]

[0148] The experimental results show that this embodiment can provide stable alerts for abnormal patterns related to premature ventricular contractions in cardiac impulse signal segments, and is suitable for abnormal event alerting tasks in long-term non-intrusive monitoring scenarios.

[0149] Example 2

[0150] A non-contact abnormal event alerting system based on rhythm-morphology co-modeling of cardiac impact signals is provided for executing the non-contact signal processing method described in Example 1. The system includes a signal acquisition module and a feature extraction and processing module. The feature extraction and processing module includes a signal preprocessing module, a dual-branch feature extraction module, a feature fusion module, and a classification decision module.

[0151] The signal acquisition module includes a non-contact sensor array, a signal conditioning circuit, and a data processing unit.

[0152] The non-contact sensor array is used to collect weak mechanical vibration signals caused by cardiac ejection during the subject's sleep state. In this embodiment, the non-contact sensor array is a PVDF piezoelectric film sensor array. Specifically, the sensor array can be an Emfit I series PVDF piezoelectric film sensor array with a size of 290×600mm. The sensor array is laid under the mattress or sheet in a 3-channel configuration, with channel 1 corresponding to the area directly below the chest cavity of the subject, and channels 2 and 3 corresponding to the upper and lower back areas, respectively.

[0153] By employing the aforementioned multi-channel arrangement, the signal acquisition module can expand its sensing range, improve the stability of cardiac impact signal acquisition when the monitored object's position changes, and reduce the impact of changes in the contact state of a single channel on subsequent abnormal event alerts.

[0154] The signal conditioning circuit includes a charge amplifier, a multi-stage amplifier circuit, and an anti-aliasing filter. The charge amplifier uses an OPA128 low-bias current operational amplifier to convert the high-impedance charge signal output from the sensor array into a low-impedance voltage signal and reduce the impact of cable capacitance on signal acquisition. The multi-stage amplifier circuit uses an INA128 instrumentation amplifier with a total amplification factor of 1000. The anti-aliasing filter uses an 8th-order elliptic filter with a cutoff frequency of 20Hz to filter out power frequency interference and high-frequency electromyographic noise.

[0155] The data processing unit uses an STM32H743 microcontroller with a built-in 16-bit ADC module and a sampling rate set to 125Hz. The data processing unit digitizes the conditioned analog signal and transmits the raw cardiac impact signal to the feature extraction and processing module via an SPI interface.

[0156] In some embodiments, the non-contact sensor array is not limited to a PVDF piezoelectric film sensor array, but may also employ a piezoelectric ceramic sensor, a flexible piezoresistive sensor, a mattress-type pressure sensor array, or a combination thereof, as long as it can acquire signals related to cardiac mechanical vibration.

[0157] The signal preprocessing module is used to receive the raw cardiac impact signal output by the signal acquisition module, and to perform baseline drift removal, spike noise removal, frequency band filtering and standardization on the raw cardiac impact signal to obtain the preprocessed cardiac impact signal.

[0158] During operation, the signal preprocessing module first uses linear detrending to reduce low-frequency baseline drift caused by sensor drift, slow changes in body position, or changes in mattress coupling state; then, it uses Hampel filtering to identify and replace sudden spike noise points; next, it uses a fourth-order Butterworth bandpass filter to retain heart rate-related components in the range of 1Hz to 12Hz and suppress respiratory components and high-frequency noise; finally, it uses Z-score normalization to unify the signal amplitude scale between different channels, different individuals, and different acquisition time periods.

[0159] Through the above processing, the signal preprocessing module can improve the stability of the input signal, providing high-quality cardiac impact signal segments for subsequent rhythm feature extraction and morphological feature extraction.

[0160] In some embodiments, the signal preprocessing module may include an analog preprocessing unit, an analog-to-digital conversion unit, and a digital preprocessing unit. The analog preprocessing unit is connected between the non-contact sensor array and the analog-to-digital conversion unit, and is used to perform impedance matching, pre-amplification, analog filtering, and level adaptation on the weak charge or voltage signals output by the sensor array. The analog-to-digital conversion unit is used to convert the analog preprocessed BCG signal into a digital BCG signal. The digital preprocessing unit is used to perform baseline drift removal, spike noise removal, bandpass filtering, and normalization on the digital BCG signal.

[0161] In some embodiments, the analog preprocessing unit may include an input protection circuit, a charge amplification circuit, a low-noise amplification circuit, an anti-aliasing filter circuit, and a bias adjustment circuit. The charge amplification circuit converts the high-impedance charge signal output from the PVDF piezoelectric film sensor into a low-impedance voltage signal; the low-noise amplification circuit amplifies the voltage signal; the anti-aliasing filter circuit limits the signal bandwidth before analog-to-digital conversion; and the bias adjustment circuit ensures the signal amplitude falls within the input range of the analog-to-digital conversion unit.

[0162] In some embodiments, the analog-to-digital conversion unit can be implemented using a microcontroller-embedded ADC, a separate multi-channel ADC chip, or an analog front-end chip-integrated ADC. The digital preprocessing unit can be implemented by a microcontroller, digital signal processor, embedded processor, field-programmable gate array, or system-on-a-chip, and transmits the preprocessed BCG signal to the feature extraction processing module via SPI, I2C, UART, USB, Bluetooth, Wi-Fi, or Ethernet interfaces.

[0163] The dual-branch feature extraction module includes a rhythm feature extraction branch and a morphological feature extraction branch. This module is used to extract abnormal pattern-related features from both a longer time scale and a local cardiac morphological scale, taking into account the mechanical vibration characteristics that distinguish cardiac impact signals from electrocardiogram signals.

[0164] The rhythm feature extraction branch is used for long-term modeling of the preprocessed cardiac impulse signal. Specifically, the rhythm feature extraction branch divides the preprocessed cardiac impulse signal into non-overlapping segments according to a fixed 12-second time window to obtain rhythm signal segments. For each rhythm signal segment, the rhythm feature extraction branch uses a 3-layer TCN for encoding. The first TCN has a kernel length of 3 and a dilation factor of 1; the second TCN has a kernel length of 3 and a dilation factor of 2; the third TCN has a kernel length of 3 and a dilation factor of 4; the first two TCNs have 64 output channels, and the third TCN has 128 output channels, and a 128-dimensional rhythm feature representation is obtained through adaptive average pooling.

[0165] After TCN encoding, the rhythm feature extraction branch outputs a 128-dimensional rhythm feature vector and anomaly confidence score for each rhythm signal segment. Subsequently, the rhythm feature extraction branch uses a Top-k aggregation strategy to select several rhythm sub-segments with high confidence, and generates overall rhythm features and rhythm discrimination scores based on the selected rhythm sub-segments. In this embodiment, the number of aggregated segments in the Top-k aggregation strategy is 3. Through the above processing, the rhythm feature extraction branch can highlight rhythm segments with strong abnormal responses in long-term cardiac impulse signals, reducing the dilution effect of a large number of normal cardiac beat segments on abnormal pattern recognition.

[0166] The morphological feature extraction branch is used to model the local heartbeat waveform and the dynamic evolution relationship between adjacent heartbeats. Specifically, the morphological feature extraction branch uses rhythm signal segments as basic units, detects the main peak J wave of the heartbeat from each rhythm signal segment, and extracts a local heartbeat sequence containing 4 consecutive heartbeats based on the detected J wave.

[0167] The morphological feature extraction branch first employs a 1D-CNN to encode the morphological structure of the local heartbeat sequence. This 1D-CNN consists of two one-dimensional convolutional layers with kernel lengths of 7 and 5, each with 16 output channels, used to extract local waveform edges, peak-valley variations, and short-term morphological patterns. Subsequently, the morphological feature extraction branch uses a bidirectional LSTM to capture the temporal evolution relationship between adjacent heartbeats. The unidirectional hidden layer dimension of the bidirectional LSTM is 64, resulting in an output feature dimension of 128. Further, the morphological feature extraction branch introduces an attention mechanism to adaptively weight and aggregate key time steps in the local heartbeat sequence, ultimately generating a 128-dimensional weighted morphological feature representation.

[0168] Through the above processing, the morphological feature extraction branch can model the mechanical waveform distortion of single or adjacent heartbeats caused by abnormal patterns related to ventricular premature beats, and enhance the contribution of key local morphological segments to the indication of abnormal events.

[0169] During sample construction, the dual-branch feature extraction module assigns the same unique identifier to the same rhythmic signal segment and the morphological signal segments generated from it. This unique identifier can consist of the monitored object number, the rhythmic signal segment sequence number, the sampling start time, or a combination thereof. Using this unique identifier, the system can determine whether the rhythmic features and morphological features originate from the same cardiac impulse signal segment during the feature fusion stage.

[0170] The feature fusion module is used to perform homologous pairing of rhythmic features and morphological features based on unique identifiers, and selectively fusion is performed only on samples where valid features are successfully extracted from both branches. Specifically, when rhythmic features and morphological features have the same unique identifier and both are valid features, the feature fusion module concatenates the 128-dimensional rhythmic features and the 128-dimensional morphological features to obtain a 256-dimensional joint feature vector. Then, the feature fusion module generates channel weights based on the global channel response of the joint feature vector, and uses the channel weights to perform weighted modulation on the joint feature vector to obtain a 256-dimensional weighted fused feature.

[0171] When rhythmic features and morphological features do not have the same unique identifier, or when any branch fails to extract a valid feature, the feature fusion module does not perform a two-branch fusion operation on the corresponding sample. Through the above selective fusion mechanism, the feature fusion module can reduce the interference of mismatched pairs, incomplete samples, and low-quality local fragments on the fusion results.

[0172] The channel weights are used to characterize the contribution of different channels in the joint features to the abnormal event indication task. By channel-weighting the joint features, the feature fusion module can enhance the discriminative information corresponding to abnormal patterns related to premature ventricular contractions and suppress redundant noise.

[0173] In some embodiments, the feature fusion module may be implemented by a processor, graphics processor, neural network accelerator, digital signal processor, field-programmable gate array, or a combination thereof in the feature extraction processing module. The feature fusion module receives rhythmic and morphological features output from the dual-branch feature extraction module and performs homology matching, feature concatenation, channel weight generation, and weighted fusion processing based on the unique identifier of the sample.

[0174] In some embodiments, the feature fusion module includes a feature caching unit, an identifier matching unit, a vector concatenation unit, a gated weight calculation unit, and a weighted output unit. The feature caching unit temporarily stores rhythmic features, morphological features, and their corresponding unique identifiers; the identifier matching unit determines whether the rhythmic features and morphological features originate from the same rhythmic signal segment; the vector concatenation unit concatenates homologous rhythmic features and morphological features into a joint feature vector; the gated weight calculation unit generates channel weights based on the joint feature vector; and the weighted output unit performs channel weighting on the joint feature vector based on the channel weights to obtain a weighted fused feature.

[0175] The classification decision module generates abnormal event alert information based on the weighted fusion features output by the feature fusion module. In this embodiment, the classification decision module uses a logistic regression model as the discriminant model. The discriminant model receives 128-dimensional weighted fusion features and outputs the abnormal probability corresponding to the cardiac impulse signal segment. The classification decision module compares the abnormal probability with a preset threshold. When the abnormal probability is greater than the preset threshold, the classification decision module outputs abnormal event alert information indicating the presence of a premature ventricular contraction-related abnormal pattern in the cardiac impulse signal segment; when the abnormal probability is not greater than the preset threshold, the classification decision module outputs information indicating that no abnormal event alert has been triggered.

[0176] The abnormal event notification information may include at least one of the following: abnormal probability, abnormal event notification label, corresponding time window position, corresponding unique identifier, and signal segment index for manual review.

[0177] In this embodiment, the feature extraction processing module adopts a hardware platform consisting of an Intel Core Ultra9 275HX CPU, an NVIDIA GeForce RTX 5070Ti GPU, and 32GB of memory, and runs the Python 3.12.7 framework to implement signal preprocessing, dual-branch feature extraction, selective fusion, and abnormal event prompting operations.

[0178] In some embodiments, the feature extraction and processing module may also be an edge computing device, server, embedded intelligent processing platform, or cloud computing platform with model inference capabilities. The discriminant model is not limited to logistic regression models, but may also employ fully connected neural networks, support vector machines, gradient boosting trees, or other classification models to output the anomaly probability corresponding to the cardiac impulse signal fragment based on the fused features.

[0179] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; these modifications or substitutions should not be construed as departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A non-contact abnormal event alerting method based on cardiac impulse signal rhythm-morphology co-modeling, characterized in that, include: The cardiac impact signal of the tested object is acquired through a non-contact sensor. Based on the cardiac impact signal, a rhythmic signal segment is constructed, and a corresponding morphological signal segment is constructed using the rhythmic signal segment as the basic unit. The same unique identifier is assigned to the rhythmic signal segment and its corresponding morphological signal segment. The rhythm signal segment is input into the rhythm feature extraction branch. The heart rate interval fluctuation and rhythm instability in the rhythm signal segment are modeled by time-series coding to obtain multiple rhythm sub-features and their corresponding abnormal response intensities. The multiple rhythm sub-features are then aggregated according to the abnormal response intensities to obtain rhythm features. The morphological signal segment is input into the morphological feature extraction branch. The morphological information of the local cardiac mechanical waveform is extracted by local waveform encoding. The temporal evolution relationship between adjacent cardiac beats is modeled, and the key time steps are weighted and aggregated to obtain the morphological features. Based on the unique identifier, the rhythmic features and the morphological features are paired according to their common origin; only rhythmic features and morphological features with the same unique identifier are fused into a joint feature; Channel weights are generated based on the channel responses of the joint features, and the joint features are then subjected to channel-by-channel weighted modulation using the channel weights to obtain weighted fused features. The weighted fusion features are input into the discrimination model to obtain abnormal event prompts for cardiac impact signals.

2. The non-contact abnormal event notification method according to claim 1, characterized in that, The rhythm feature extraction branch uses a temporal convolutional network to perform temporal encoding on the rhythm signal segment, and the temporal convolutional network includes multiple dilated convolutional layers. Aggregating the multiple rhythmic sub-features based on the abnormal response intensity includes: selecting several rhythmic sub-features with higher abnormal response intensity from the multiple rhythmic sub-features according to the abnormal response intensity, and performing weighted aggregation on the selected rhythmic sub-features to obtain the rhythmic features.

3. The non-contact abnormal event notification method according to claim 1, characterized in that, The morphological feature extraction branch includes a one-dimensional convolutional network, a bidirectional long short-term memory network, and an attention mechanism; the one-dimensional convolutional network is used to perform local waveform encoding on the morphological signal segments; The bidirectional long short-term memory network is used to model the temporal evolution relationship between adjacent heartbeats; the attention mechanism is used to calculate the attention weight of each time step, and to perform weighted aggregation of the temporal features of each time step according to the attention weight to obtain the morphological features.

4. The non-contact abnormal event notification method according to claim 1, characterized in that, Before fusing rhythmic features and morphological features with the same unique identifier into a joint feature, the method further includes: Determine whether both the rhythmic features and the morphological features have been successfully extracted; If both the rhythmic feature and the morphological feature are successfully extracted, then the rhythmic feature and the morphological feature are concatenated into the joint feature; if either the rhythmic feature or the morphological feature is not successfully extracted, then the rhythmic feature and the morphological feature of the corresponding sample are not concatenated into the joint feature.

5. The non-contact abnormal event notification method according to claim 1, characterized in that, The process of obtaining weighted fusion features includes: calculating the sum of squared eigenvalues ​​of each feature channel in the joint features, and obtaining the joint feature energy based on the sum of squared eigenvalues; normalizing the joint features according to the joint feature energy to obtain normalized joint features; generating the channel weights based on the normalized joint features and learnable parameters; and multiplying the channel weights by the joint features channel by channel to obtain the weighted fusion features.

6. The non-contact abnormal event notification method according to claim 1, characterized in that, The cardiac impact signal acquired by the non-contact sensor is preprocessed; the preprocessing includes: performing baseline drift removal, spike noise removal, bandpass filtering, and normalization on the cardiac impact signal in sequence to obtain the preprocessed cardiac impact signal.

7. The non-contact abnormal event notification method according to claim 6, characterized in that, The baseline drift removal process includes linear detrending processing, the spike noise removal process includes Hampel filtering, the bandpass filtering process includes Butterworth bandpass filtering, and the normalization process includes Z-score normalization.

8. The non-contact abnormal event notification method according to claim 1, characterized in that, Constructing rhythmic signal segments based on the cardiac impact signal includes: dividing the cardiac impact signal into non-overlapping segments according to a fixed time window length to obtain multiple rhythmic signal segments; Constructing a corresponding morphological signal segment using the rhythmic signal segment as the basic unit includes: detecting the main J wave of the heartbeat from the rhythmic signal segment, extracting a local heartbeat sequence containing multiple consecutive heartbeats based on the main J wave of the heartbeat, and using the local heartbeat sequence as the morphological signal segment.

9. The non-contact abnormal event notification method according to claim 8, characterized in that, The fixed time window length is 8s to 20s; the local heartbeat sequence contains 3 to 5 consecutive heartbeats; the unique identifier includes the subject number, the rhythm signal segment number, and the sampling start time.

10. A non-contact abnormal event alerting system based on cardiac impulse signal rhythm-morphology co-modeling, characterized in that, Used to perform the non-contact abnormal event notification method as described in any one of claims 1-9; The non-contact abnormal event notification system includes: The signal acquisition module is used to acquire cardiac impact signals of the tested object through a non-contact sensor; The signal preprocessing module is used to preprocess the cardiac impact signal to obtain a preprocessed cardiac impact signal. The dual-branch feature extraction module is used to construct rhythmic signal segments based on the preprocessed cardiac impact signal, construct corresponding morphological signal segments using the rhythmic signal segments as basic units, and assign the same unique identifier to the rhythmic signal segments and their corresponding morphological signal segments. Rhythmic features are obtained through the rhythmic feature extraction branch, and morphological features are obtained through the morphological feature extraction branch. The feature fusion module is used to perform homologous pairing of the rhythmic features and the morphological features according to the unique identifier, and only concatenate the rhythmic features and morphological features that have the same unique identifier and have been successfully extracted into a joint feature. The module generates channel weights based on the channel response of the joint feature, and uses the channel weights to perform channel-wise weighted modulation on the joint feature to obtain a weighted fusion feature. The classification decision module is used to input the weighted fusion features into the discrimination model to obtain abnormal event prompt information of cardiac impact signal.

Images