Bio-electricity signal classification method, device, equipment and medium

Abstract

The invention discloses a bio-electricity signal classification method, device and equipment and a medium, and belongs to the technical field of signal processing. The method comprises the following steps: acquiring a target signal; the target signal comprises a bio-electricity signal of each channel; according to the target signal of each time window, constructing a physiological function connection diagram corresponding to each time window; inputting the physiological function connection diagram corresponding to each time window into a diagram convolutional network to obtain a diagram embedding sequence corresponding to each time window; and sequentially inputting the graph embedding sequence of each time window into the extended long-short-term memory network to obtain a classification result of the target signal. According to the scheme, on the basis that the time memory ability of the extended long-short-term memory network is improved through the matrix memory unit, the long-short-term memory network is combined with the graph convolutional network, feature extraction of the time dimension and the space dimension is considered, and the recognition accuracy of the bioelectricity signals is improved.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a method, apparatus, device, and medium for classifying bioelectrical signals. Background Technology

[0002] Bioelectric signals are the electrical potential changes generated by cells or tissues within a living organism, and are ubiquitous in the physiological activities of all living organisms. Current technologies show great promise for the application of bioelectric signals. For example, acquiring the electroencephalogram (EEG) signals of a target human body can enable real-time monitoring of the electrical activity in the cerebral cortex, thereby facilitating pathological identification; or, brain-computer interfaces can be used to read EEG signals to control corresponding electronic devices.

[0003] In current mainstream EEG recognition algorithms, since EEG is an electrical signal with certain spatial characteristics, convolutional neural networks are commonly used to extract features from EEG to achieve EEG signal recognition. On the other hand, EEG also has temporal characteristics to some extent, so recurrent neural networks or long short-term memory networks (LSTM) can also be used to recognize EEG signals.

[0004] However, neither convolutional neural networks nor long short-term memory networks can simultaneously extract the spatial and temporal features of bioelectrical signals, resulting in low accuracy in bioelectrical signal recognition. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, equipment, and medium for classifying bioelectric signals to address the above problems, thereby improving the accuracy of bioelectric signal recognition.

[0006] This application provides a method for classifying bioelectrical signals, the method comprising: Acquire the target signal; the target signal includes bioelectrical signals from each channel; Based on the target signal for each time window, a physiological function connectivity diagram is constructed for each time window; the physiological function connectivity diagram includes nodes and edges connecting the nodes; each node represents a channel; the edges connecting the nodes are used to represent the correlation between the bioelectrical signals of the channels; The physiological function connection graph corresponding to each time window is input into the graph convolutional network to obtain the graph embedding sequence corresponding to each time window. The graph embedding sequence of each time window is sequentially input into an extended long short-term memory network to obtain the classification result of the target signal; The extended long short-term memory network includes a matrix memory unit; the matrix memory unit is used to update the memory matrix of the previous time step according to the output of the graph convolutional network at the current time step to obtain the memory matrix at the current time step; the memory matrix is ​​used to control the weights of the gating mechanism in the extended long short-term memory network.

[0007] In one possible implementation, constructing a corresponding physiological function connectivity map based on the target signal for each time window includes: Establish the node corresponding to each channel in the target signal; Within the current time window, obtain the fusion correlation coefficient between the target channel and other channels; the target channel is any one of the channels in the target signal; the fusion correlation coefficient is used to indicate the correlation between the bioelectrical signals of the channels; Based on the fusion correlation coefficient between the target channel and other channels, a connected channel is selected from the other channels, and an edge connecting the target channel and the connected channel is generated.

[0008] In one possible implementation, obtaining the fusion correlation coefficient between the target channel and other channels within the current time window includes: Obtain the linear correlation coefficient, rank correlation coefficient, and Euclidean distance between the target electrical signal and the candidate electrical signals; the target electrical signal is the bioelectrical signal of the target channel within the current time window; the candidate electrical signal is the bioelectrical signal of the candidate channel within the current time window; the candidate channel is any one of the other channels; The fusion correlation coefficient between the target channel and the candidate channel is obtained by weighted summation based on the linear correlation coefficient, rank correlation coefficient and Euclidean distance between the target electrical signal and the candidate electrical signal.

[0009] In one possible implementation, the method further includes: During the training process of the graph convolutional network, the weights of the linear correlation coefficient, the rank correlation coefficient, and the Euclidean distance are updated using the backpropagation algorithm.

[0010] In one possible implementation, each graph embedding sequence includes several node vectors; each vector corresponds to one node. The step of sequentially inputting the graph embedding sequence of each time window into an extended long short-term memory network to obtain the classification result of the target signal includes: For each time window's graph embedding sequence, the node vectors corresponding to each node in the graph embedding sequence are fused to obtain a graph-level vector sequence; the graph-level vector sequence includes the graph-level vectors for each time window; and the graph-level vectors are sorted according to the time window. The graph-level vectors are sequentially input into the extended long short-term memory network according to the time window order to obtain the classification result of the target signal.

[0011] In one possible implementation, the method further includes: Project the vector sequence at each time step onto the vector space of the target dimension to obtain the projected vector; Calculate the outer product of the projection vectors, and update the memory matrix of the previous time step based on the outer product to obtain the memory matrix of the current time step.

[0012] In one possible implementation, the extended long short-term memory network includes multiple scalar long short-term memory units; each scalar long short-term memory unit has a different gating bias parameter. The step of sequentially inputting the graph-level vectors into the extended long short-term memory network according to the time window order to obtain the classification result of the target signal includes: Each graph-level vector is simultaneously input into the multiple scalar long short-term memory units in the order of the time window. Obtain the final vector composed of the scalars output by the multiple scalar long short-term memory units; The classification result of the target signal is determined based on the final vector.

[0013] This application also provides a bioelectrical signal classification device, the device comprising: A signal acquisition module is used to acquire a target signal; the target signal includes bioelectrical signals from various channels. The graph construction module is used to construct a corresponding physiological function connectivity graph based on the target signal of each time window. The physiological function connectivity graph includes nodes and edges connecting the nodes. Each node represents a channel. The edges connecting the nodes are used to represent the correlation between the bioelectrical signals of the channels. The graph convolution module is used to input the physiological function corresponding to each time window into the graph convolution network to obtain the graph embedding sequence corresponding to each time window. The signal classification module is used to sequentially input the graph embedding sequence of each time window into the extended long short-term memory network to obtain the classification result of the target signal; The extended long short-term memory network includes a matrix memory unit; the matrix memory unit is used to update the memory matrix of the previous time step according to the output of the graph convolutional network at the current time step to obtain the memory matrix at the current time step; the memory matrix is ​​used to control the weights of the gating mechanism in the extended long short-term memory network.

[0014] In another aspect, this application also provides an electronic device, including: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the above-mentioned bioelectric signal classification method by executing the computer instructions.

[0015] This application also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the above-described bioelectric signal classification method.

[0016] Compared with the prior art, the technical solution provided in this application has the following advantages: When this application acquires a target signal, which includes bioelectrical signals from various channels, it first needs to divide the time window and construct a physiological function connectivity graph within each time window. Nodes in the physiological function connectivity graph correspond to various channels, and the edges connecting the nodes represent the correlation between the bioelectrical signals of the channels. Therefore, the aforementioned physiological function connectivity graph has fused the bioelectrical signals of various channels in the target signal into a graph representation. This physiological function connectivity graph can then be input into a graph convolutional network for spatial feature extraction, yielding a graph embedding sequence corresponding to each time window. The graph embedding sequences can then be sequentially input into an extended long short-term memory (LSTM) network for processing according to the time window order. The LSTM network also includes a matrix memory unit, which continuously maintains a memory matrix based on the graph embedding sequences output by the graph convolutional network, thereby controlling the gating mechanism in the LSTM network. After the LSTM network processes the graph embedding sequences corresponding to the time windows sequentially, the final classification result of the target signal can be output. The above scheme improves the temporal memory capability of the extended long short-term memory network by using matrix memory units, and combines the long short-term memory network with the graph convolutional network to take into account the feature extraction of both time and space dimensions, thereby improving the accuracy of bioelectrical signal recognition. Attached Figure Description

[0017] Figure 1 A flowchart of a method for classifying bioelectrical signals according to an embodiment of the present invention is shown.

[0018] Figure 2 shows a signal processing flow logic diagram according to an embodiment of this application.

[0019] Figure 3 A flowchart illustrating a bioelectrical signal classification method according to an embodiment of the present invention is shown.

[0020] Figure 4 This illustration shows a logical diagram of constructing a dynamic physiological function connectivity diagram according to an embodiment of this application.

[0021] Figure 5 This is a schematic diagram of the structure of a bioelectric signal classification device provided in an embodiment of this application.

[0022] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an optional embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Bioelectric signals are the electrical potential changes generated by cells or tissues within a living organism, and they are ubiquitous in the physiological activities of all living things. Bioelectric signals have become one of the core technologies in many fields, including medicine, neuroscience, smart wearable devices, and brain-computer interfaces.

[0025] For example, in the medical field, epilepsy is a neurological disease caused by abnormal discharge of neurons in the brain. Electroencephalography (EEG), as a core technology for monitoring brain activity, collects weak electrical signals generated by the synchronous discharge of neuronal groups by placing an electrode array on the scalp. With its advantages such as non-invasiveness, millisecond-level time resolution, and low detection cost, it plays a key role in the diagnosis and seizure monitoring of epilepsy.

[0026] For example, in the field of brain-computer interfaces, brain-computer interfaces can acquire electroencephalograms (EEGs). By analyzing the EEGs, computer devices can achieve functions such as assisted communication (e.g., users can directly input text and generate images through brain signals) and motion control (e.g., users can control robotic arms or wheelchairs directly with their brains).

[0027] Traditional bioelectrical signal analysis methods mainly rely on manually designed features (such as time-domain statistics, frequency-domain power spectra, wavelet coefficients, entropy values, etc.) combined with shallow classifiers such as support vector machines (SVM) and linear discriminant analysis (LDA) for pattern recognition. These methods are highly dependent on prior knowledge, have difficulty adapting to individual differences and dynamic environmental changes, and are prone to the "curse of dimensionality" when processing high-dimensional multi-channel data, resulting in limited generalization ability.

[0028] In recent years, deep learning technology has been widely applied to bioelectrical signal processing. Convolutional Neural Networks (CNNs), due to their powerful local feature extraction capabilities, have achieved good results in tasks such as EEG emotion recognition and EMG gesture classification; Recurrent Neural Networks (RNNs) and their variants (such as LSTM and GRU) are adept at modeling long-term dependencies in time series and are suitable for time-series discrimination tasks such as arrhythmia detection and sleep staging. However, most of these models treat multi-channel signals as independent time series or simply splice them together, ignoring the spatial topology and functional coordination mechanisms between channels, thus limiting their ability to model the dynamics of the overall physiological network.

[0029] To address the aforementioned issues, this application also provides a method for classifying bioelectrical signals. Figure 1 A flowchart of a bioelectrical signal classification method according to an embodiment of the present invention is shown, as follows: Figure 1 As shown, the method flow includes: Step 101: Obtain the target signal.

[0030] The target signal includes bioelectrical signals from various channels.

[0031] In the embodiments of this application, the target signal may be at least one bioelectrical signal selected from electroencephalography (EEG), electromyography (EMG), and electrocardiography (ECG).

[0032] The target signal can be obtained by collecting data from the human body using a standard electrode array. Taking EEG as an example, a standard electrode array corresponding to the EEG signal can be used to collect 21 channels of EEG signal at a frequency of 256 Hz.

[0033] In this embodiment of the application, after the target signal is sampled, the target signal can be divided into various time windows according to a specified duration for subsequent timing processing.

[0034] Furthermore, when dividing the target signal according to duration, a sliding event window can be used for division. For example, the original signal can be divided into sliding time windows with a length of 1 second and an overlap rate of 50%, forming target signals {X1, X2, ..., X...} ordered by time. T}, where each X t ∈R N×L N is the number of channels, and L is the number of sampling points per window. Wherein, for any target signal X within a time window... t Each of these can be viewed as an N*L matrix, where each element represents a digital electrical signal sampled at a specific sequence and time.

[0035] Step 102: Based on the target signal of each time window, construct the physiological function connectivity diagram corresponding to each time window.

[0036] The physiological function connectivity diagram includes nodes and edges connecting the nodes; each node represents a channel; and the edges connecting the nodes represent the correlation between the bioelectrical signals of the channels.

[0037] Since the bioelectric signals of multiple channels in the target signal are usually collected from the human body through a standard electrode array, and the bioelectric signals of each channel are usually obtained by deploying different electrodes of the standard electrode array on the human body, the bioelectric signals of different channels have a certain degree of similarity (collected from the same human body part) and a certain degree of difference (different electrode deployment positions).

[0038] Therefore, in order to fully extract the spatial distribution of bioelectric signals, the embodiments of this application need to construct a physiological function connection diagram between various bioelectric data in order to fully extract the spatial relationship between the bioelectric data of each channel.

[0039] Specifically, in this embodiment, the physiological function connectivity graph can be an undirected weighted graph. Each node in the undirected weighted graph represents a channel in the target signal; then, the correlation between the signals of each channel is mapped to the edges between nodes, thereby mapping the bioelectrical signals of multiple channels in the same time window into a physiological function connectivity graph, so that the spatial features of the bioelectrical signals of multiple channels in the same time window can be extracted by a graph convolutional network.

[0040] Step 103: Input the physiological function connection graph corresponding to each time window into the graph convolutional network to obtain the graph embedding sequence corresponding to each time window.

[0041] Since the physiological function connectivity graph is an undirected weighted graph, and graph convolutional networks cannot directly process undirected weighted graphs, the embodiments of this application first need to convert the input physiological function connectivity graph into a low-dimensional vector, and then process it through a graph convolutional network to obtain the corresponding graph embedding vector. Since each time window has a physiological function connectivity graph, by processing the physiological function connectivity graph of each time window through a graph convolutional network, we can obtain the graph embedding vector corresponding to each time window, thus forming a graph embedding sequence ordered by time window.

[0042] In the process of feature extraction, graph convolutional networks extract the feature information of each node and its neighbors. Through the adjacency relationship of the graph, each node can perceive the state of its neighbors. Therefore, after processing by graph convolutional networks, not only can important features in the current node (i.e., a single channel) be extracted, but also the features between the current node and its neighbors (i.e., the current channel and highly correlated channels) can be fused to merge similar features in space, thus realizing the extraction of features of bioelectrical signals in the spatial dimension.

[0043] Step 104: Input the graph embedding sequence of each time window into the extended long short-term memory network in sequence to obtain the classification result of the target signal.

[0044] The extended long short-term memory network includes a matrix memory unit; the matrix memory unit is used to update the memory matrix of the previous time step according to the output of the graph convolutional network at the current time step to obtain the memory matrix at the current time step; the memory matrix is ​​used to control the weights of the gating mechanism in the extended long short-term memory network.

[0045] Since bioelectric signals also exhibit temporal correlation, simply using graph convolutional networks to extract spatial features from the target signal is insufficient to analyze the overall characteristics of the bioelectric signal. In this embodiment, after extracting spatial features through a graph convolutional network for each time window to obtain a graph embedding sequence, the graph embedding sequence is then input into an extended long short-term memory network to extract features in the temporal dimension, ultimately obtaining the classification result of the target signal. At this point, the classification result takes into account features in both the temporal and spatial dimensions, resulting in high classification accuracy.

[0046] Specifically, in commonly used Long Short-Term Memory (LSTM) networks, temporal features have been extracted on a time scale through mechanisms such as forget gates, input gates, and memory cells. However, traditional memory cells are prone to gradient vanishing and memory saturation problems in scenarios with long time sequences and large amounts of data. They are also prone to forgetting signals that appear briefly or have weak signal strength, resulting in low signal classification accuracy.

[0047] Therefore, in this embodiment, the Long Short-Term Memory (LSTM) network is extended by adding a matrix memory unit to memorize bioelectrical signals from a higher dimension. This increases the memory capacity and allows the higher-dimensional matrix to better store the spatial features extracted from the bioelectrical signals. As a result, the extended LSM network can extract the temporal features of bioelectrical signals in each time window based on the spatial features extracted by the graph convolutional network, thereby identifying the target signal and improving the accuracy of bioelectrical signal classification.

[0048] Please refer to Figure 2 This illustrates a signal processing flowchart according to an embodiment of this application. Figure 2 As shown. In this embodiment of the application, in order to process bioelectric signals, a bioelectric signal processing model can be pre-trained to identify subsequently acquired signals. For example, the bioelectric signal processing model may include a dynamic brain function map construction module, a GCN feature extraction module, an XLSTM time series modeling module, and a classifier and early warning output module.

[0049] Taking electroencephalogram (EEG) as an example of bioelectrical signals, after acquiring EEG signals from N channels, these signals can be input into a dynamic brain function mapping model to construct a physiological function connectivity map, as shown in step 102. After constructing the physiological function connectivity map, it can be input into a graph convolutional network (GCN) feature extraction module. Feature extraction is performed using the graph convolutional network within the GCN module to obtain graph embedding sequences. These sequences are then input into an XLSTM temporal modeling module, where temporal processing is performed using the extended long short-term memory (XLSTM) network to obtain output features. Finally, the output features are passed through a classifier to determine the category of the bioelectrical signal. If a warning is required for this category, a corresponding warning signal can be output through the warning output module (e.g., issuing a warning signal when abnormal brain discharge is detected).

[0050] In summary, when this application obtains the target signal, which includes bioelectrical signals from various channels, it first needs to divide the time window and construct a physiological function connectivity graph within each time window. Nodes in the physiological function connectivity graph correspond to each channel, and the edges connecting the nodes represent the correlation between the bioelectrical signals of the channels. Therefore, the aforementioned physiological function connectivity graph has fused the bioelectrical signals of each channel in the target signal into a graph representation. This physiological function connectivity graph can then be input into a graph convolutional network for spatial feature extraction, yielding a graph embedding sequence corresponding to each time window. The graph embedding sequences can then be sequentially input into an extended long short-term memory (LSTM) network for processing according to the time window order. The LSTM network also includes a matrix memory unit, which continuously maintains a memory matrix based on the graph embedding sequences output by the graph convolutional network, thereby controlling the gating mechanism in the LSTM network. After the LSTM network processes the graph embedding sequences corresponding to the time windows sequentially, the final classification result of the target signal can be output. The above scheme improves the temporal memory capability of the extended long short-term memory network by using matrix memory units, and combines the long short-term memory network with the graph convolutional network to take into account the feature extraction of both time and space dimensions, thereby improving the accuracy of bioelectrical signal recognition.

[0051] Figure 3 shows a flowchart of a bioelectrical signal classification method according to an embodiment of the present invention, as follows: Figure 2 As shown, the method includes: Step 301: Obtain the target signal.

[0052] The target signal includes bioelectrical signals from each channel. Step 301 is similar to step 101 described above and will not be repeated here.

[0053] Step 302: Establish the node corresponding to each channel in the target signal.

[0054] To construct an undirected weighted graph as a physiological function connectivity graph, it is first necessary to determine the basic nodes required for its construction. In this embodiment, the target signal includes various channels, so a unique corresponding node can be established for each channel, and then the correlation between channels is used to determine whether connections are needed between the nodes.

[0055] Step 303: Within the current time window, obtain the fusion correlation coefficient between the target channel and other channels.

[0056] The target channel is any one of the channels in the target signal; the fusion correlation coefficient is used to indicate the correlation between the bioelectrical signals of the channels.

[0057] In the embodiments of this application, for any target channel among the channels of the target signal, if it is necessary to determine which channels are connected to the target channel, it is first necessary to determine which channels are related to the target channel. Therefore, a parameter is needed to define the correlation between each channel and the target channel, which is the fusion correlation coefficient proposed in the embodiments of this application.

[0058] Figure 4 This illustration shows a logical diagram of a method for constructing a dynamic physiological function connectivity graph according to an embodiment of this application. Figure 4 As shown, embodiments of this application can obtain the linear correlation coefficient, rank correlation coefficient, and Euclidean distance between the target electrical signal and the candidate electrical signal; the target electrical signal is the bioelectrical signal of the target channel within the current time window; the candidate electrical signal is the bioelectrical signal of the candidate channel within the current time window; the candidate channel is any one of the other channels; the fusion correlation coefficient between the target channel and the candidate channel is obtained by weighted summation based on the linear correlation coefficient, rank correlation coefficient, and Euclidean distance between the target electrical signal and the candidate electrical signal.

[0059] In simple terms, in this embodiment of the application, in order to quantify the bioelectrical signal x between any two nodes (that is, between two channels i and j) i x jThe degree of similarity or association between (the two entities) can be judged by a combination of Pearson linear correlation coefficient, Spearman rank correlation coefficient, and Euclidean distance (specifically, normalized Euclidean distance).

[0060] The Pearson correlation coefficient measures the strength and direction of the linear relationship between signals. For example, if the bioelectrical signal in one channel increases, the bioelectrical signal in another channel also increases, indicating a positive linear correlation; conversely, if the bioelectrical signal in the other channel decreases, the correlation is negative.

[0061] Spearman's rank correlation coefficient is used to measure the strength of the monotonic relationship between signals, and is used to capture nonlinear but consistent changes in bioelectrical signals between two channels.

[0062] Euclidean distance, on the other hand, directly calculates the absolute difference between the bioelectrical signals of two channels in numerical space.

[0063] By calculating the Pearson linear correlation coefficient, Spearman rank correlation coefficient, and Euclidean distance between the bioelectrical signals of two channels, and then weighting and summing the three, the fusion correlation coefficient between the two channels can be obtained. This fusion correlation coefficient can, to a certain extent, guarantee the similarity and correlation between the two channels.

[0064] In this embodiment of the application, during the training process of the graph convolutional network, the weights of the linear correlation coefficient, the rank correlation coefficient, and the Euclidean distance can be updated using the backpropagation algorithm. This allows the algorithm with updated weights to focus on the parameters among the linear correlation coefficient, rank correlation coefficient, and Euclidean distance that contribute more to spatial feature extraction, thereby gradually improving the effect of spatial feature extraction.

[0065] Step 304: Based on the fusion correlation coefficient between the target channel and other channels, select connected channels from the other channels and generate edges connecting the target channel and the connected channels to generate a physiological function connection diagram.

[0066] After obtaining the fusion correlation coefficient between any two channels (for ease of understanding, let's take any target channel as an example), theoretically, the target channel should have a certain correlation with other channels, that is, the fusion correlation coefficient should not be zero. However, in order to reduce noise in the feature extraction process, when generating an undirected weighted graph, only the edges between channels with larger fusion correlation coefficients should be retained.

[0067] Optionally, in this embodiment, the fusion correlation coefficients of the target channel and other channels can be sorted, and the first N channels can be selected in order to retain the edges between the node corresponding to the target channel and the nodes corresponding to the first N channels; where N is greater than or equal to 1 and N is an integer. For example, when N equals 5, for the node corresponding to the target channel, the nodes of the 5 channels with the largest fusion correlation coefficient with the target channel can be found, and these 5 nodes can be connected to the node corresponding to the target channel.

[0068] Optionally, in this embodiment, the fusion correlation coefficient between the target channel and other channels can be compared with a preset coefficient threshold. If the fusion correlation coefficient is higher than the coefficient threshold, the connection relationship between the nodes of the target channel and the nodes of that channel is retained.

[0069] In this embodiment, each channel is treated as a target channel and the above logic is executed, which can construct a sparse but physiological function connectivity diagram that can reflect the overall characteristics of bioelectric signals.

[0070] Step 305: Input the physiological function connection graph corresponding to each time window into the graph convolutional network to obtain the graph embedding sequence corresponding to each time window.

[0071] After obtaining the physiological function connectivity map, it can be input into a graph convolutional network to extract features in the spatial dimension.

[0072] Specifically, the processing flow of the graph convolutional network in this application embodiment can satisfy the following formula:

[0073]

[0074] in, = , represents the adjacency matrix with added self-loops; It is a degree matrix; and These are the learnable weight matrices in a two-layer graph convolutional network; The output of the graph convolutional network is denoted as , where is the node-level embedding vector. Where t represents the time window, and each time window can generate a corresponding embedding vector according to the above formula. The embedding vectors of each time window can be sorted according to the time window to form a graph embedding sequence.

[0075] Step 306: For the graph embedding sequence of each time window, fuse the node vectors corresponding to each node in the graph embedding sequence to obtain a graph-level vector sequence.

[0076] The graph-level vector sequence includes graph-level vectors for each time window; and the graph-level vectors are sorted according to the time window.

[0077] Specifically, after obtaining the graph convolution vector output by the graph convolutional network... After that, it is also necessary to... Global average pooling or attention mechanism aggregation is performed to obtain graph-level embedding vectors. :

[0078] Where t represents the time window, each time window can generate a corresponding graph-level vector according to the above formula, and the graph-level vectors of each time window can be sorted according to the time window to form a graph-level vector sequence. .

[0079] Step 307: Input the graph-level vectors into the extended long short-term memory network in the order of the time window to obtain the classification result of the target signal.

[0080] The extended long short-term memory network in this application embodiment can add two modules, matrix memory unit and multi-head scalar LSTM, to the traditional long short-term memory network.

[0081] Optionally, for a matrix memory unit, the following steps can be performed in each iterative operation: project the vector sequence of each time step onto the vector space of the target dimension to obtain the projected vector; calculate the outer product of the projected vectors, and update the memory matrix of the previous time step according to the outer product to obtain the memory matrix of the current time step.

[0082] Specifically, let the current input be... The hidden state at the previous moment was Matrix memory is For example, r can be 16.

[0083] Then the current input can be changed in the following way. Projecting onto a lower-dimensional space yields a lower-dimensional vector. : = * ,in, ;in The projection vector is used to... Dimensional reduction.

[0084] After obtaining the low-dimensional vector Then, the memory matrix in the matrix memory unit can be updated based on the low-dimensional vector (an exponential moving average method can be used), for example: ,in =0.9.

[0085] During each iteration of the LSTM operation, the matrix memory unit can utilize Mt to control the gating mechanism or attention weights: For example, matrix memory units can be extracted. The main feature vectors are used to weight the hidden states at different time steps; For example, a matrix memory unit can also store a matrix The F-norm is used as an additional input to control the weight of the forget gate, that is, to control the degree to which the forget gate is opened and closed.

[0086] The aforementioned matrix memory unit allows for explicit modeling of the covariance structure evolution of cross-channel signals during bioelectric signal processing. This is suitable for capturing long-term physiological dynamics such as brain functional reorganization, muscle coordination changes, and heart rate variability trends, thereby improving the accuracy of bioelectric signal recognition.

[0087] Specifically, the extended long short-term memory network includes multiple scalar long short-term memory units, thereby forming a multi-head scalar LSTM; each scalar long short-term memory unit has different gating bias parameters; The computer device can simultaneously input the various graph-level vectors into the multiple scalar long short-term memory units in the order of the time window; and obtain the final vector composed of the scalars finally output by the multiple scalar long short-term memory units.

[0088] Finally, the computer device can determine the classification result of the target signal based on the final vector.

[0089] For example, in the embodiments of this application, K=4 independent scalar LSTM units can be constructed, with each unit maintaining only one scalar memory. However, they have different gating bias parameters, making them sensitive to rapid changes (such as spikes and R waves); medium-speed rhythms (such as alpha waves and muscle tremors); slow drifts (such as fatigue accumulation and autonomic regulation); and very slow trends (such as state transitions and sleep stage shifts).

[0090] At this point, the four independent scalar LSTM units can each output different values. , , , Then, the outputs of each scalar LSTM unit are fused through a fully connected layer to finally output the final hidden state of the entire sequence. .Will Input fully connected classifier: y=Softmax( + ) Output y∈ This represents the probability distribution of physiological state C (e.g., rest, exercise, fatigue, abnormal events, etc.), and then a threshold (e.g., P>0.8) is set to trigger an early warning or control command, which is then sent to the terminal device via Bluetooth / Wi-Fi.

[0091] In summary, when this application obtains the target signal, which includes bioelectrical signals from various channels, it first needs to divide the time window and construct a physiological function connectivity graph within each time window. Nodes in the physiological function connectivity graph correspond to each channel, and the edges connecting the nodes represent the correlation between the bioelectrical signals of the channels. Therefore, the aforementioned physiological function connectivity graph has fused the bioelectrical signals of each channel in the target signal into a graph representation. This physiological function connectivity graph can then be input into a graph convolutional network for spatial feature extraction, yielding a graph embedding sequence corresponding to each time window. The graph embedding sequences can then be sequentially input into an extended long short-term memory (LSTM) network for processing according to the time window order. The LSTM network also includes a matrix memory unit, which continuously maintains a memory matrix based on the graph embedding sequences output by the graph convolutional network, thereby controlling the gating mechanism in the LSTM network. After the LSTM network processes the graph embedding sequences corresponding to the time windows sequentially, the final classification result of the target signal can be output. The above scheme improves the temporal memory capability of the extended long short-term memory network by using matrix memory units, and combines the long short-term memory network with the graph convolutional network to take into account the feature extraction of both time and space dimensions, thereby improving the accuracy of bioelectrical signal recognition.

[0092] Furthermore, in the bioelectric signal processing of this application embodiment, by calculating and constructing a dynamic functional connectivity graph through fusion correlation coefficients, it is possible to effectively identify "hub channels" that maintain high synergy with multiple channels. These channels often correspond to core physiological functional areas (such as the motor cortex, cardiac pacing area, or major muscle activation sources). Combining the attention mechanism of graph neural networks (GNNs) can explain why certain channels are assigned higher weights by the model, providing a visual basis for physiological mechanism analysis.

[0093] Furthermore, the aforementioned technical solution of fusing correlation coefficients avoids the possibility of false connections caused by noise in a single indicator. After fusion, the consistency of multiple indicators is used to reduce the false judgment rate. Moreover, by fusing correlation coefficients, key connections can be preserved. Even if a certain indicator is not significant, other indicators may still support the existence of the connection, thus avoiding the omission of important functional pathways. In addition, the fusion correlation coefficient can also enhance the robustness of the graph and improve the generalization ability of the GNN model.

[0094] Furthermore, Pearson calculus may underestimate correlation due to nonlinear relationships; Spearman calculus ignores the absolute change in signal amplitude; and Euclidean distance is greatly affected by signal amplitude. Therefore, combining these three methods can compensate for each other's shortcomings in specific situations, improving the model's adaptability. The weighted fusion strategy can flexibly adjust the contribution of each indicator according to task requirements. By weighted fusion of these three factors, multiple correlation patterns such as linearity, nonlinearity, monotonicity, and morphological similarity can be captured simultaneously, avoiding the limitations of a single indicator and enabling the constructed functional network to more comprehensively and realistically reflect electrophysiological activities.

[0095] Furthermore, existing models mostly focus on acute changes occurring in the immediate future (such as R-wave detection and action potential triggering), making it difficult to identify gradual precursor signals from seconds to minutes in advance. This invention introduces matrix memory units from Extended Long Short-Term Memory (XLSTM) networks, which can accumulate the covariance structure evolution information of cross-channel signals online, forming structured long-term memory. Compared to the traditional LSTM method that relies solely on vector memory, the memory mechanism of this invention is more suitable for the long-term dynamic modeling needs of high-dimensional, multivariate physiological systems.

[0096] Furthermore, bioelectrical signal patterns vary significantly among individuals, including signal amplitude, frequency distribution, channel layout, and functional organization. Traditional models often experience performance degradation on new users due to overfitting to training data. This invention sets the weighting coefficients α, β, and γ in the fusion correlation coefficient as learnable parameters, allowing the model to automatically adjust its feature combination strategy during training or fine-tuning. This invention also introduces lightweight fine-tuning mechanisms (such as LoRA and channel attention retraining), requiring only a small amount of new user data for adaptation. Moreover, the matrix memory unit naturally possesses a certain standardization characteristic in modeling the covariance structure, reducing the impact of individual amplitude differences. In summary, the extended long short-term memory network of this invention can further improve the individual adaptability of bioelectrical signal recognition. Finally, this invention can also display dynamic heatmaps of coordinated changes in physiological networks. This visualization output can be used to show the functional reorganization process of the brain, heart, or muscle systems before task execution, state transitions, or abnormal events, thereby providing doctors, patients, or engineers with an intuitive view of physiological state change trends and helping them determine the reasonableness of the results.

[0097] This application also provides a bioelectrical signal classification device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0098] This application provides a bioelectrical signal classification device. Figure 5 This is a schematic diagram of a bioelectrical signal classification device provided in an embodiment of this application. The device includes: The signal acquisition module 501 is used to acquire the target signal; the target signal includes bioelectric signals from each channel. The graph construction module 502 is used to construct a corresponding physiological function connectivity graph based on the target signal of each time window; the physiological function connectivity graph includes each node and the edges connecting the nodes; each node represents each channel; the edges connecting the nodes are used to represent the correlation between the bioelectrical signals of the channels; Graph convolution module 503 is used to input the physiological function connection graph corresponding to each time window into the graph convolution network to obtain the graph embedding sequence corresponding to each time window. The signal classification module 504 is used to sequentially input the graph embedding sequence of each time window into the extended long short-term memory network to obtain the classification result of the target signal; The extended long short-term memory network includes a matrix memory unit; the matrix memory unit is used to update the memory matrix of the previous time step according to the output of the graph convolutional network at the current time step to obtain the memory matrix at the current time step; the memory matrix is ​​used to control the weights of the gating mechanism in the extended long short-term memory network.

[0099] In summary, when this application obtains the target signal, which includes bioelectrical signals from various channels, it first needs to divide the time window and construct a physiological function connectivity graph within each time window. Nodes in the physiological function connectivity graph correspond to each channel, and the edges connecting the nodes represent the correlation between the bioelectrical signals of the channels. Therefore, the aforementioned physiological function connectivity graph has fused the bioelectrical signals of each channel in the target signal into a graph representation. This physiological function connectivity graph can then be input into a graph convolutional network for spatial feature extraction, yielding a graph embedding sequence corresponding to each time window. The graph embedding sequences can then be sequentially input into an extended long short-term memory (LSTM) network for processing according to the time window order. The LSTM network also includes a matrix memory unit, which continuously maintains a memory matrix based on the graph embedding sequences output by the graph convolutional network, thereby controlling the gating mechanism in the LSTM network. After the LSTM network processes the graph embedding sequences corresponding to the time windows sequentially, the final classification result of the target signal can be output. The above scheme improves the temporal memory capability of the extended long short-term memory network by using matrix memory units, and combines the long short-term memory network with the graph convolutional network to take into account the feature extraction of both time and space dimensions, thereby improving the accuracy of bioelectrical signal recognition.

[0100] Please see Figure 6 , Figure 6 This is a schematic diagram of an electronic device provided in an optional embodiment of the present invention. This electronic device can be a computer device used to execute the aforementioned bioelectrical signal classification method. Figure 6 As shown, the electronic device includes one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces).

[0101] The processor 10 may further include a hardware chip. This hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0102] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0103] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the use of the electronic device based on the display of a mini-program landing page. Furthermore, the memory 20 may include high-speed random access memory (RAM), and may also include non-transient memory, such as at least one disk storage device, flash memory device, or other non-transient solid-state storage device. The memory 20 may include volatile memory, such as RAM; the memory may also include non-volatile memory, such as flash memory, hard disk, or solid-state drive; the memory 20 may also include combinations of the above types of memory.

[0104] The electronic device also includes a communication interface 30 for communicating with other devices or communication networks.

[0105] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method of classifying bioelectric signals, characterized by, The method comprises: acquiring a target signal; the target signal comprising bioelectric signals of each channel; constructing a physiological functional connection graph corresponding to each time window according to the target signal of each time window; the physiological functional connection graph comprising nodes and edges connecting the nodes; each node representing a channel; and the edges connecting the nodes representing the correlation between the bioelectric signals of the channels; inputting the physiological functional connection graph corresponding to each time window into a graph convolution network to obtain a graph embedding sequence corresponding to each time window; inputting the graph embedding sequence of each time window into an extended long short-term memory network in sequence to obtain a classification result of the target signal; wherein the extended long short-term memory network comprises a matrix memory unit; the matrix memory unit is used to update a memory matrix at a previous time according to the output of the graph convolution network at the current time to obtain a memory matrix at the current time; and the memory matrix is used to control the weight of a gating mechanism in the extended long short-term memory network.

2. The method of claim 1, wherein, The method comprises: establishing a node corresponding to each channel in the target signal; acquiring a fusion correlation coefficient between a target channel and other channels in a current time window; the target channel being any one of the channels in the target signal; and the fusion correlation coefficient indicating the correlation between the bioelectric signals of the channels; selecting a connected channel from the other channels according to the fusion correlation coefficient between the target channel and the other channels, and generating an edge connecting the target channel and the connected channel.

3. The method of claim 2, wherein, The method comprises: acquiring a linear correlation coefficient, a rank correlation coefficient and a Euclidean distance between a target electrical signal and a candidate electrical signal; the target electrical signal being the bioelectric signal of the target channel in the current time window; the candidate electrical signal being the bioelectric signal of the candidate channel in the current time window; and the candidate channel being any one of the other channels; performing weighted summation on the linear correlation coefficient, the rank correlation coefficient and the Euclidean distance between the target electrical signal and the candidate electrical signal to obtain the fusion correlation coefficient between the target channel and the candidate channel.

4. The method of claim 3, wherein, The method further comprises: updating the weight of the linear correlation coefficient, the weight of the rank correlation coefficient and the weight of the Euclidean distance by a back propagation algorithm during the training process of the graph convolution network.

5. The method according to any one of claims 1 to 4, characterized in that, Each graph embedding sequence comprises a plurality of node vectors; each vector corresponding to a node; The method comprises: for each graph embedding sequence of a time window, fusing the node vectors corresponding to the nodes in the graph embedding sequence to obtain a graph-level vector sequence; the graph-level vector sequence comprising a graph-level vector of each time window; and the graph-level vectors being sorted according to the time windows. The respective graph-level vectors are sequentially input into the extended long short-term memory network in order of time windows to obtain a classification result of the target signal.

6. The method of claim 5, wherein, The method further comprises: projecting the vector sequence of each time step into a vector space of a target dimension to obtain a projected vector; calculating an outer product of the projected vector, and updating a memory matrix of a previous time according to the outer product to obtain a memory matrix of a current time.

7. The method of claim 5, wherein, The extended long short-term memory network comprises a plurality of scalar long short-term memory units; each scalar long short-term memory unit has different gate bias parameters; The step of sequentially inputting the respective graph-level vectors into the extended long short-term memory network in order of time windows to obtain a classification result of the target signal comprises: simultaneously inputting the respective graph-level vectors into the plurality of scalar long short-term memory units in order of time windows; obtaining a final vector composed of scalar outputs of the plurality of scalar long short-term memory units; determining the classification result of the target signal according to the final vector.

8. A biological electrical signal classification device, characterized by, The device comprises: a signal acquisition module configured to acquire a target signal; the target signal comprises bioelectric signals of respective channels; a graph construction module configured to construct a corresponding physiological functional connection graph according to the target signal of each time window; the physiological functional connection graph comprises nodes and edges connecting the nodes; the nodes represent the respective channels; and the edges represent correlations between the bioelectric signals of the channels; a graph convolution module configured to input the physiological functional connection graph corresponding to each time window into a graph convolution network to obtain a graph embedding sequence corresponding to each time window; a signal classification module configured to sequentially input the graph embedding sequence of each time window into an extended long short-term memory network to obtain a classification result of the target signal; The extended long short-term memory network comprises a matrix memory unit; the matrix memory unit is configured to update a memory matrix of a previous time according to an output of a graph convolution network of a current time to obtain a memory matrix of the current time; and the memory matrix is configured to control weights of a gating mechanism in the extended long short-term memory network.

9. An electronic device, comprising: The device comprises: a memory and a processor in communication connection with each other; the memory stores computer instructions; and the processor executes the computer instructions to perform the bioelectric signal classification method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions; and the computer instructions are configured to cause a computer to perform the bioelectric signal classification method according to any one of claims 1 to 7.

Images