Waveform morphological feature extraction method and device of biological time sequence signal and medium

Abstract

The invention discloses a waveform morphological feature extraction method and device of a biological time sequence signal, and a medium. The method comprises the following steps: inputting the biological time sequence signal into a pre-trained feature extraction model; global features of the biological time sequence signals are extracted through a TCN convolution module, and morphological features of the biological time sequence signals are extracted through a Shapelets module; and inputting the global features and the morphological features into a fusion encoder for fusion, and outputting a feature signal. In the task of global and local dynamic combination, the global and local features in the biological signal can be effectively captured, so that the overall understanding and classification performance of the model on time series data is improved, and the classification accuracy and robustness of the model are ensured.

Description

Technical Field

[0001] The present application relates to the field of computers, and in particular to a method, device and medium for extracting waveform morphological features of biological time series signals. Background Art

[0002] For the time series waveform data of physiological parameter monitoring (for example, signals such as continuous glucose monitoring), the morphological changes of the waveforms contain a lot of physical or clinical information. These morphological information are of great significance in realizing personalized analysis of physiological signs and pattern analysis of sensor signals.

[0003] However, when processing these morphological representation information, existing deep learning models are limited to the learning of global temporal patterns, or limited to the extraction of local morphological features of waveforms, and do not comprehensively analyze global and local morphological features, which affects the performance of the model for specific monitoring tasks. Summary of the invention

[0004] In order to solve the above problems, the present application proposes a method for extracting waveform morphological features of biological time series signals, including:

[0005] Inputting the biological time series signal into a pre-trained feature extraction model; the feature extraction model includes a TCN convolution module and a Shapelets module;

[0006] Extracting the global features of the biological time series signal through the TCN convolution module, and extracting the morphological features of the biological time series signal through the Shapelets module;

[0007] The global features and the morphological features are input into a fusion encoder for fusion, and a feature signal is output.

[0008] Optionally, extracting the global features of the biological time series signal through the TCN convolution module specifically includes:

[0009] Determine that the architecture of the TCN convolution module includes multiple convolution layers, the multiple convolution layers include an input layer, multiple hidden layers and an output layer; wherein the dilation factors corresponding to the input layer and the multiple hidden layers are increased in sequence according to the data forward processing order;

[0010] Performing dilated causal convolution, weight normalization, activation function, and regularization in sequence through the input layer and at least one of the multiple hidden layers;

[0011] The fully connected layer is used as the output layer to output the global features of the biological time series signal.

[0012] Optionally, the method further includes:

[0013] Adding a residual connection in each of the convolutional layers;

[0014] Among them, if the input dimension is the same as the output dimension, the residual connection is: F l [t,c]=Z l [t,c]+F l-1 [t,c], where F l [t,c] is the convolution value of the lth layer, time step t, and channel c after the dilated causal convolution, Z l [t,c] is the convolution value of layer l, time step t, channel c;

[0015] If the input dimension is different from the output dimension, F l [t,c]=Z l [t,c]+F xonv [t,c],F vonv [t,c]=W l1×c ·F l-1 ,in, is the weight of the one-dimensional convolution, is the output of one-dimensional convolution.

[0016] Optionally, extracting the morphological features of the biological time series signal through the Shapelets module specifically includes:

[0017] Based on the multivariate time series, obtaining the turning point of the biological timing signal;

[0018] Performing multiple rounds of selection of important perception points at the turning point, and after each selection, generating a first subsequence based on the important perception points selected in the current round and the important perception points adjacent to the important perception points selected in the current round and selected in the previous round;

[0019] For each label category corresponding to the biological time series signal, based on the perceptual subsequence distance between each first subsequence and each biological time series signal contained in the label category, determine a plurality of second subsequences having the highest information gain of each first subsequence relative to the label category, and add the second subsequences to the morphological subsequence pool corresponding to the label category;

[0020] For the morphological subsequence pool, the best fitting subsequence corresponding to each second subsequence is selected through a filter, and the morphological features of the biological time series signal are generated through the best fitting subsequence.

[0021] Optionally, generating a first subsequence specifically includes:

[0022] Among the turning points, the first turning point and the last turning point are taken as important perception points;

[0023] Among the remaining turning points, multiple rounds of important perception points are selected until the number of selected important perception points reaches a preset upper limit; wherein, in each round of selection, the turning point with the largest reconstruction distance is selected as the important perception point selected in the current round; the reconstruction distance refers to the vertical distance between the target turning point and the straight line reconstructed between the two nearest selected important perception points;

[0024] In each round, at least one first subsequence is generated according to three consecutive important perception points including the important perception points selected in the current round.

[0025] Optionally, generating the morphological features of the biological time series signal specifically includes:

[0026] Restrict the search for the best-fitting subsequence to a neighborhood within the hyperparameter window size;

[0027] For each second subsequence in the morphological subsequence pool, calculating the complexity invariant distance between it and the subsequence of the biological time series signal, and selecting the subsequence of the biological time series signal with the shortest complexity invariant distance as the best fitting subsequence of the biological time series signal;

[0028] Obtaining a difference feature corresponding to the best fitting subsequence based on a linear projection of the best fitting subsequence and the morphological subsequence pool in a feature space;

[0029] The differential features of all the best fitting subsequences are merged to generate the morphological features of the biological time series signal.

[0030] Optionally, for each second subsequence in the morphological subsequence pool, when calculating the corresponding complexity invariant distance, an upper limit of the number of calculations is preset, and the upper limit of the calculation is positively linearly related to the hyperparameter window size;

[0031] The difference features of all the best fitting subsequences are merged to generate the morphological features of the biological time series signal, specifically including:

[0032] The difference features of the best fitting subsequence are used as column vectors, and all column vectors are merged to generate the morphological features of the biological time series signal.

[0033] Optionally, the global feature and the morphological feature are input into a fusion encoder for fusion, and a feature signal is output, which specifically includes:

[0034] Inputting the global features and the morphological features into a fusion encoder, fusing the global features and the morphological features through one-dimensional convolution fusion to obtain a fusion feature;

[0035] The fused features are input into an encoder, and a feature signal is output.

[0036] Optionally, the global features and the morphological features are fused by one-dimensional convolution fusion to obtain fused features, which specifically includes:

[0037] Aligning the global features and the morphological features, and filling the morphological features;

[0038] The global features and the morphological features are concatenated, and a one-dimensional convolution is used to perform a convolution operation on the concatenated features to extract corresponding interactive features;

[0039] The interactive features are processed by an activation function to obtain fusion features;

[0040] The fusion feature is input into an encoder, and a feature signal is output, specifically including:

[0041] The fused features are processed through a self-attention mechanism to output feature signals.

[0042] Optionally, the biological time series signal is input into a pre-trained feature extraction model, specifically including:

[0043] Acquire a data set consisting of multiple biological time-series signals;

[0044] The biological time series signals in the data set are input into a pre-trained feature extraction model.

[0045] Optionally, the biological timing signal includes at least one of a current value obtained based on continuous blood glucose monitoring and a glucose value obtained based on the current value.

[0046] On the other hand, the present application also proposes a waveform morphology feature extraction device for biological time series signals, comprising:

[0047] at least one processor; and,

[0048] a memory communicatively connected to the at least one processor; wherein,

[0049] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the waveform morphology feature extraction method of the biological timing signal as described in the above example.

[0050] On the other hand, the present application also proposes a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured as: the waveform morphology feature extraction method of the biological timing signal described in the above example.

[0051] The waveform morphology feature extraction method of biological time series signals proposed in this application can bring the following beneficial effects:

[0052] The present application extracts global time series features through the TCN convolution module, processes the input biological signal time series waveform, extracts the long-term dependence features of the entire time series, extracts local morphological features through the Shapelets module, extracts local morphological features from the same signal, and then introduces the global features and local morphological features extracted by the TCN convolution module and the Shapelets module respectively into the fusion encoder for fusion to obtain feature signals. In the task of dynamically combining the global and local, the present application can effectively capture and fully utilize the global and local features in the biological signal for data analysis, and at the same time solves the problem that local morphological features are difficult to organically combine with global features, thereby improving the model's overall understanding and classification performance of time series data, and ensuring the classification accuracy and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0054] Figure 1 Schematic diagram of the process of extracting waveform morphology features of biological time series signals in an embodiment of the present application;

[0055] Figure 2 This is a schematic diagram of the architecture of a feature extraction model in an embodiment of the present application;

[0056] Figure 3 This is a schematic diagram of the architecture of the TCN convolution module in an embodiment of the present application;

[0057] Figure 4 This is a schematic diagram of important sensing points in an embodiment of the present application;

[0058] Figure 5 This is a schematic diagram of extracting the first subsequence in an embodiment of the present application;

[0059] Figure 6 Schematic diagram of a waveform morphology feature extraction device for biological timing signals in an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0061] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0062] At present, when processing time series waveform data for physiological parameter monitoring, existing deep learning models mainly focus on the learning of global time series patterns and ignore the extraction of local morphological features of the waveform. Therefore, it is difficult to effectively capture the specific morphological features in the signal, thus affecting the performance of the model for specific monitoring tasks.

[0063] For this reason, Figure 1 As shown, the embodiment of the present application provides a method for extracting waveform morphological features of biological time series signals, comprising:

[0064] S101: Inputting biological time series signals into a pre-trained feature extraction model; the feature extraction model includes a TCN convolution module and a Shapelets module.

[0065] Bio-time series signals refer to signals that describe the dynamic changes and activities of biological systems through time series data. They are derived from changes in certain physiological, behavioral or ecological phenomena of organisms or ecosystems over a period of time, and can be collected and analyzed through a variety of instruments and methods. For example, in the field of blood glucose monitoring, it can be a current value obtained based on continuous blood glucose monitoring (GCM), or a glucose value obtained based on the current value. In other fields, bio-time series signals can also be electrocardiogram (ECG) signals, electroencephalogram (EEG) signals, heart rate variability (HRV) signals, etc.

[0066] In actual acquisition, multiple biological time series signals are usually collected and formed into a data set. Each biological time series signal can be composed of segments of fixed time length, and each data set can correspond to a single user or a single user group (for example, a healthy user group, a user group with a specific disease, etc.).

[0067] At this point, the biological timing signals in the data set are input into the pre-trained feature extraction model. The feature extraction model extracts features for each biological timing signal in turn, and finally extracts the feature signal of the biological timing signal. Next, based on the needs, the feature signals extracted from all the biological timing signals in the data set can be processed in the next step, such as splicing, fusion, and analysis of the feature signals.

[0068] like Figure 2 As shown in FIG, the feature extraction model includes a TCN convolution module and a Shapelets module, which are used to extract the global features and morphological features of biological time series signals, respectively, and finally fuse them to output corresponding feature signals.

[0069] Assume that the input biological timing signal Where X is a data set consisting of biological time series signals, X1~X N is the biological time series signal, M is the length of the time series, N is the number of variables, and each column vector Represents the time series of variable i.

[0070] When training the model, the training data set Data = {(X i ,Y i )|1≤i≤z},X i , Y i They represent the i-th training sample (that is, a single biological time series signal) and the corresponding label, respectively, where z represents the number of time series instances.

[0071] The goal is to extract the overall and morphological features from the training data set as the input of the TCNSFormer model, the subsequence column vector of the time series is a column vector X i A continuous subsequence of , where start and end represent the start index and end index respectively.

[0072] S102: extracting the global features of the biological time series signal through the TCN convolution module, and extracting the morphological features of the biological time series signal through the Shapelets module.

[0073] TCN is the abbreviation of Temporal Convolutional Network, which is referred to as TCN convolution module here. It is a convolutional network structure for time series, which can extract time-dependent features without destroying the time order. Compared with traditional recurrent neural networks (RNN) or long short-term memory recursive neural networks (LSTM), TCN has a longer receptive field, can effectively capture long-term dependencies, identify local change trends of waveforms, and maintain the continuity of signals in different time periods.

[0074] Specifically, Figure 2 and Figure 3 As shown in FIG. 1 , the architecture of the TCN convolution module includes an input layer, multiple hidden layers, and an output layer; wherein the dilation factors corresponding to the input layer and multiple hidden layers increase in sequence according to the order of forward data processing.

[0075] After the biological time series signal is input into the TCN convolution module, it passes through the input layer and at least one of the multiple hidden layers, and performs dilated causal convolution, weight normalization, activation function, and regularization in sequence. Then, the fully connected layer is used as the output layer to output the global features of the biological time series signal.

[0076] Furthermore, the dilated causal convolution combines the dilated convolution with the causal convolution to ensure that the output depends only on the current and previous time steps, ensuring that there is no leakage of future information; by inserting holes between the convolution kernels, the receptive field is significantly increased to capture the dependencies between long time steps. This convolution enables the network to capture information over a larger time range without increasing the computational cost.

[0077] Perform dilated causal convolution, specifically the convolution output matrix of the previous layer and the weight matrix cyclic convolution: Perform dilated causal convolution. In which, is the input data, X is the input biological time series signal, is the real number space, M is the length of the time series, N is the number of variables, is the convolution value of channel c at layer l, time step t, C l is the number of output channels of layer l, is the weight of the convolution kernel, d l is the expansion rate of the lth layer of convolution, k is the size of the convolution kernel, c ′ is the input channel index, and c is the output channel index.

[0078] Weight normalization is used to normalize the weights of the convolution kernels to avoid gradient vanishing or explosion and ensure the stability of the weights during training. Perform weight normalization, where is the normalized weight, g l is the scalar parameter that can be learned by the lth layer, v l is the original weight of the convolution kernel of the lth layer, ‖v l ‖ is v l The L2 norm of .

[0079] The ReLU activation function is used to introduce nonlinearity after each convolution layer, so that the model can capture complex temporal patterns. The ReLU activation function is: Z l [t,c]=RELU(Y l [t,c]); where Z l [t,c] is the convolution value at layer l, time step t, channel c after the activation function is applied, x is the input data, The shape remains unchanged.

[0080] Regularization technology uses the Dropout method to discard some neurons with corresponding probability to prevent the model from relying on specific neurons and overfitting. During each forward propagation, Dropout randomly "discards" some neurons and temporarily sets the output of these neurons to 0. The "discarded" neurons do not participate in the calculation in the current forward propagation and back propagation, but their weights will still be retained. During the next forward propagation, Dropout will randomly discard different neurons in the same way.

[0081] Regularization is performed through the Dropout method, discarding some neurons with corresponding probabilities, where the Dropout method is: in, The shape remains unchanged, but the values ​​of some neurons will become 0, and p is the corresponding probability.

[0082] Finally, after the L layer, a fully connected layer is used to obtain the global features of the biological time series signal extracted by the TCN convolution module. TCN =F L ·W o Output the global features of biological time series signals, where is the global feature of biological timing signals extracted by temporal convolutional network, d TCN is the number of output channels of TCN, is the convolution value after L layers of dilated causal convolution, is the weight matrix.

[0083] In addition, before the fully connected layer, in order to prevent gradient vanishing and information loss, a residual connection can be added to each convolutional layer. If the input dimension is the same as the output dimension, the residual connection is: F l [t,c]=Z l [t,c]+F l-1 [t,c], where F l [t,c] is the convolution value of the lth layer, time step t, and channel c after the dilated causal convolution, Z l [t,c] is the convolution value of channel c at layer l, time step t.

[0084] If the input dimension is different from the output dimension, use one-dimensional convolution to adjust the number of channels, F l [t,c]=Z l [t,c]+F conv [t,c],F conv [t,c]=W 1×1 ·F l-1 ,in, is the weight of the one-dimensional convolution, is the output of one-dimensional convolution.

[0085] like Figure 2 As shown in the figure, the Shapelets method can capture local and category-specific features from biological signals, such as subtle changes in heart rate in medical time series data. The Shapelets module consists of two parts: subsequence discovery and subsequence filter. The biological signal data is found through subsequence discovery to obtain a subsequence pool, and the subsequence filter is used to obtain the best fitting subsequence of a specific biological signal data, and finally the morphological features extracted by the Shapelets module are obtained.

[0086] Specifically, subsequence discovery is a technique for extracting local morphological features from time series, which can be used for biological signal feature extraction. In biological signal processing, physical representation information is usually reflected in the morphology of time series waveforms. The Shapelets method can discover and filter out these key morphological features, thereby helping the model to better classify.

[0087] The Offline Shapelet Discovery (OSD) method based on multivariate time series is used for subsequence discovery. The OSD method uses Perceptually Important Points (PIPs) to compress time series data by selecting points that are very similar to the original data based on the reconstruction distance.

[0088] like Figure 4As shown in the figure, important perception points (PIPs for short) are points in time series data or curves that are identified as being important to the data features. In time series analysis, important perception points (PIPs) are often used to identify key turning points in the curve, which can effectively represent the overall trend, simplify the data, improve the efficiency of feature extraction, and maintain the key information of the data. By extracting a limited number of important perception points (PIPs), data compression can be performed while retaining data features and reducing computational complexity. This makes the overall analysis more concise without losing important information.

[0089] Subsequence discovery includes two main stages: subsequence extraction and subsequence selection. In the subsequence extraction stage, the turning points of biological timing signals are obtained through the OSD method, and multiple rounds of important perception points are selected among the turning points.

[0090] After each selection, a first subsequence is generated based on the important perception points selected in the current round and the important perception points adjacent to the important perception points selected in the current round and selected in the previous round, and the first subsequence is used as the Shapelet candidate feature.

[0091] by Figure 5 For example, among the turning points, the first turning point and the last turning point are taken as important perception points and added to the important perception point set (referred to as PIPs set).

[0092] Among the remaining turning points, multiple rounds of selection of important perception points are performed. In each round of selection, the turning point with the largest reconstruction distance is selected as the important perception point selected in the current round, until the number of selected important perception points reaches a preset upper limit.

[0093] Among them, the reconstruction distance (PD) refers to the vertical distance between the target turning point and the straight line reconstructed between the two nearest selected important perception points, and the maximum reconstruction distance refers to the maximum value of this distance.

[0094] In each round, whenever a new important perception point (referred to as New PIP) is added, three consecutive important perception points are used to extract Shapelet candidate features. The three consecutive important perception points need to include the important perception points selected in the current round and are continuous, thereby generating at least one first subsequence. Figure 5 and Figure 6 For example, when the new important perception points added in the current round are used as the first important perception point, the second important perception point and the third important perception point of the first subsequence, at most three new first subsequences can be added to the PIPs set as three new Shapelet candidate features.

[0095] Here, the number of important perception points npip in the PIPs set can be set to: npip = 0.2 × M, where M represents the length of the time series. This method can greatly reduce the number of candidate objects extracted by the traditional shapelets method, thereby improving the algorithm performance.

[0096] After subsequence extraction, multiple first subsequences are obtained. At this time, subsequence selection is performed. For each label category corresponding to the biological time series signal, a corresponding number of first subsequences are selected for each label category. The label category is labeled for each biological time series signal during training and can be identified by the trained feature extraction model. Generally speaking, the number of first subsequences corresponding to each label category is the same.

[0097] During training, for the training data set (X i ,Y i ) in the label category Y i (When the model is actually used, it corresponds to the first subsequence of each label category that needs to be identified) (Shapelet candidate features) 1 , based on each first subsequence Sub 1 The perceptual subsequence distance (PSD) between each biological time series signal contained in the label category is determined to determine each first subsequence Sub 1 Several second subsequences Sub with the highest information gain relative to the label category 2 , and the second subsequence Sub 2 Add to the morphological subsequence pool SUB corresponding to the label category.

[0098] Specifically, through as well as To achieve selection. Among them, PSD(X i ,Sub 1 ) is the data set X i With the first subsequence Sub 1 The perceptual subsequence distance between them, M is the time series X i The length, l ′ The first subsequence Sub 1 The length of CID is the complexity invariant distance: Let Q = {q1,q2,…,q N} and C={c1,c2,…,c N} are any two time series of equal length, ED(Q,C) is the Euclidean distance between Q and C, Then use the PSD distance to find the optimal information gain and select the first g candidate subsequences with the highest information gain as the second subsequence Sub 2And store it in the morphological subsequence pool SUB to obtain the morphological subsequence pool SUB.

[0099] In this way, after the subsequence selection is completed, the subsequence discovery is also completed. At this time, the subsequence filter is executed. For the morphological subsequence pool, the best fitting subsequence corresponding to each second subsequence is selected through the filter, and the morphological features of the biological time series signal are generated through the best fitting subsequence.

[0100] Specifically, given a morphological subsequence pool SUB, a biological signal temporal waveform sequence X and its label Y, a subsequence filter is used to select the best fitting subsequence.

[0101] For each second subsequence sub in the morphological subsequence pool SUB 2 , calculate the best fitting subsequence I with the biological time series signal X i The search for the best fitting subsequence is restricted to the neighborhood within the hyperparameter window size w. calculate The complexity invariant distance CID between each subsequence of the biological time series signal X is calculated, and the subsequence of the biological time series signal X with the shortest complexity invariant distance CID is selected as the distance between the biological time series signal X and the subsequence pool. The best fitting subsequence I u .

[0102] Among them, for each second subsequence sub in the morphological subsequence pool 2 When calculating the corresponding complexity invariant distance CID, an upper limit on the number of calculations is preset, and the upper limit is positively linearly related to the hyperparameter window size. For example, the distances of at most 2w+1 subsequences in X are calculated.

[0103] Based on the linear projection difference of the best fitting subsequence and the morphological subsequence pool in the feature space, the difference feature corresponding to the best fitting subsequence is obtained, as shown in the formula: I u =X[index:index+l ′ ]. Here, argmin represents the subscript that obtains the minimum value in its domain.

[0104] The feature U obtained by the shapeles method using difference feature representation i , used to highlight the substantive differences between classes: in, P I , P Sub It is in space The linear projection of is the best fitting subsequence, where l ′ For Ui The length, d Shapelet is the difference feature U i The number of

[0105] The difference feature U of the best fitting subsequence i As a column vector, all column vectors are merged to generate the morphological feature F of the biological time series signal Shapelet .in,

[0106] S103: Input the global features and the morphological features into a fusion encoder for fusion, and output a feature signal.

[0107] Specifically, the global features and morphological features are input into a fusion encoder (herein referred to as TCNSFormer fusion encoder), and the global features and morphological features are fused through one-dimensional convolution fusion to obtain fusion features. The fusion features are input into the encoder, and the feature signal is output.

[0108] The TCNSFormer fusion encoder combines the advantages of TCN and Shapelets feature extraction, processes global features and morphological features through one-dimensional convolution fusion, and then inputs them into the Transformer Encoder. It consists of two parts: one-dimensional convolution fusion and encoder.

[0109] For the one-dimensional convolution fusion part, after extracting the global features and morphological characteristics After that, one-dimensional convolution is used to transform F TCN and F Shapelet Perform feature extraction and fusion, and treat the two time series data of different lengths as two channels. Convolution fusion is suitable for time series data, which can effectively extract local spatiotemporal features, identify patterns or structures in the spatial dimension, and automatically learn the interaction and correlation between different modalities and different source data through convolution kernels, and extract interactive information at different time steps through convolution.

[0110] Specifically, the global feature F TCN 、Morphological characteristicsF Shapelet The sequences were aligned in length and the morphological features F Shapelet Fill in the shorter sequence (such as F Shapelet ) Fill with 0 and expand from length l′ to M: F′ Shapelet =pad(F Shapelet ,Ml′). Thus, the morphological features after filling are obtained

[0111] Then, the global characteristics and morphological characteristics of the climate are spliced ​​together, and the columns are spliced ​​together in the feature dimension to obtain the fused feature data Z: in, Then use one-dimensional convolution to perform convolution operation on the concatenated features to extract the corresponding interactive features: O = Conv1D (Z, kernel size , filters), where kernel size is the size of the convolution kernel, filters is the number of output channels, T depends on the stride and padding of the convolution, and c is the number of output channels.

[0112] Finally, the interaction features are processed by the activation function ReLU to obtain the fusion feature F combined =ReLU(O), where

[0113] For the encoder part, TCN itself contains timing information, and the fused feature F combined It is further processed as the input of the encoder, and the temporal correlation of different time steps is further learned through the self-attention mechanism, and the fused features are processed to output the feature signal.

[0114] The formula of the self-attention mechanism is: Among them, Q, K, and V are query, key, and value matrices respectively, and d k is the dimension of the key vector.

[0115] The global time series features are extracted through the TCN convolution module, the input biological signal time series waveform is processed, the long-term dependence features of the entire time series are extracted, the local morphological features are extracted through the Shapelets module, and the local morphological features are extracted from the same signal. The global features and local morphological features extracted by the TCN convolution module and the Shapelets module are introduced into the fusion encoder for fusion to obtain the feature signal. In the task of dynamically combining the global and local, the present application can effectively capture and fully utilize the global and local features in the biological signal for data analysis, and at the same time solves the problem that local morphological features are difficult to combine organically with global features, thereby improving the model's overall understanding and classification performance of time series data, and ensuring the classification accuracy and robustness of the model.

[0116] like Figure 6 As shown, the embodiment of the present application also provides a waveform morphology feature extraction device for biological time series signals, including:

[0117] at least one processor; and,

[0118] a memory communicatively connected to the at least one processor; wherein,

[0119] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the waveform morphology feature extraction method of the biological timing signal as described in any of the above embodiments.

[0120] An embodiment of the present application also provides a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured to be: a method for extracting waveform morphological features of biological timing signals as described in any of the above embodiments.

[0121] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0122] The devices and media provided in the embodiments of the present application correspond one-to-one to the methods. Therefore, the devices and media also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0123] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0124] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0125] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0126] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0127] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0128] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0129] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0130] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0131] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for extracting waveform morphological features of biological time series signals, characterized in that: include: Inputting biological time series signals into a pre-trained feature extraction model; The feature extraction model includes a TCN convolution module and a Shapelets module; Extracting the global features of the biological time series signal through the TCN convolution module, and extracting the morphological features of the biological time series signal through the Shapelets module; The global features and the morphological features are input into a fusion encoder for fusion, and a feature signal is output.

2. The method according to claim 1, characterized in that The global features of the biological time series signal are extracted by the TCN convolution module, specifically including: Determine that the architecture of the TCN convolution module includes multiple convolution layers, the multiple convolution layers include an input layer, multiple hidden layers and an output layer; wherein the dilation factors corresponding to the input layer and the multiple hidden layers are increased in sequence according to the data forward processing order; Performing dilated causal convolution, weight normalization, activation function, and regularization in sequence through the input layer and at least one of the multiple hidden layers; The fully connected layer is used as the output layer to output the global features of the biological time series signal.

3. The method according to claim 2, characterized in that The method further comprises: Adding a residual connection in each of the convolutional layers; Among them, if the input dimension is the same as the output dimension, the residual connection is: F l [t, c] = Z l [t,c]+F l-1 [t, c], where F l [t, c] is the convolution value of the lth layer, time step t, and channel c after the dilated causal convolution, Z l [t, c] is the convolution value of channel c at layer l, time step t; If the input dimension is different from the output dimension, F l [t, c] = Z l [t,c]+F conv [t,c],F conv [t, c] = W 1×1 ·F l-1 ,in, is the weight of the one-dimensional convolution, is the output of one-dimensional convolution.

4. The method according to claim 1, characterized in that: The morphological features of the biological time series signal are extracted by the Shapelets module, specifically including: Based on the multivariate time series, obtaining the turning point of the biological timing signal; Performing multiple rounds of selection of important perception points at the turning point, and after each selection, generating a first subsequence based on the important perception points selected in the current round and the important perception points adjacent to the important perception points selected in the current round and selected in the previous round; For each label category corresponding to the biological time series signal, based on the perceptual subsequence distance between each first subsequence and each biological time series signal contained in the label category, determine a plurality of second subsequences having the highest information gain of each first subsequence relative to the label category, and add the second subsequences to the morphological subsequence pool corresponding to the label category; For the morphological subsequence pool, the best fitting subsequence corresponding to each second subsequence is selected through a filter, and the morphological features of the biological time series signal are generated through the best fitting subsequence.

5. The method according to claim 4, characterized in that Generate a first subsequence, specifically including: Among the turning points, the first turning point and the last turning point are taken as important perception points; Among the remaining turning points, multiple rounds of important perception points are selected until the number of selected important perception points reaches a preset upper limit; wherein, in each round of selection, the turning point with the largest reconstruction distance is selected as the important perception point selected in the current round; the reconstruction distance refers to the vertical distance between the target turning point and the straight line reconstructed between the two nearest selected important perception points; In each round, at least one first subsequence is generated according to three consecutive important perception points including the important perception points selected in the current round.

6. The method according to claim 4, characterized in that Generating the morphological features of the biological time series signal specifically includes: Restrict the search for the best-fitting subsequence to a neighborhood within the hyperparameter window size; For each second subsequence in the morphological subsequence pool, calculating the complexity invariant distance between it and the subsequence of the biological time series signal, and selecting the subsequence of the biological time series signal with the shortest complexity invariant distance as the best fitting subsequence of the biological time series signal; Obtaining a difference feature corresponding to the best fitting subsequence based on a linear projection of the best fitting subsequence and the morphological subsequence pool in a feature space; The differential features of all the best fitting subsequences are merged to generate the morphological features of the biological time series signal.

7. The method according to claim 6, characterized in that For each second subsequence in the morphological subsequence pool, when calculating the corresponding complexity invariant distance, an upper limit of the number of calculations is preset, and the upper limit of the calculations is positively linearly related to the hyperparameter window size; The difference features of all the best fitting subsequences are merged to generate the morphological features of the biological time series signal, specifically including: The difference features of the best fitting subsequence are used as column vectors, and all column vectors are merged to generate the morphological features of the biological time series signal.

8. The method according to claim 1, characterized in that Inputting the global features and the morphological features into a fusion encoder for fusion, and outputting feature signals, specifically includes: Inputting the global features and the morphological features into a fusion encoder, fusing the global features and the morphological features through one-dimensional convolution fusion to obtain a fusion feature; The fused features are input into an encoder, and a feature signal is output.

9. The method according to claim 8, characterized in that The global features and the morphological features are fused by one-dimensional convolution fusion to obtain fused features, which specifically include: Aligning the global features and the morphological features, and filling the morphological features; The global features and the morphological features are concatenated, and a one-dimensional convolution is used to perform a convolution operation on the concatenated features to extract corresponding interactive features; The interactive features are processed by an activation function to obtain fusion features; The fusion feature is input into an encoder, and a feature signal is output, specifically including: The fused features are processed through a self-attention mechanism to output feature signals.

10. The method according to claim 1, characterized in that Input biological time series signals into a pre-trained feature extraction model, including: Acquire a data set consisting of multiple biological time-series signals; The biological time series signals in the data set are input into a pre-trained feature extraction model.

11. The method according to claim 1, characterized in that: The biological timing signal includes at least one of a current value obtained based on continuous blood glucose monitoring and a glucose value obtained based on the current value.

12. A device for extracting waveform morphological features of biological time series signals, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the waveform morphology feature extraction method of the biological timing signal as described in any one of claims 1 to 11.

13. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions are configured as: a method for extracting waveform morphological features of biological timing signals as described in any one of claims 1 to 11.

Images