Deep learning-based electroencephalogram regulation and control monitoring data feature mining method and computer system

Abstract

The invention provides an electroencephalogram regulation and control monitoring data feature mining method based on deep learning and a computer system, relates to the technical field of data processing, and aims to prune network configuration variables of an initially adjusted monitoring data feature mining network and secondarily adjust the pruned monitoring data feature mining network. And obtaining a monitoring data feature mining network in a convergence state. According to the method, the scale of the obtained monitoring data feature mining network is reduced, so that the situation that the monitoring data feature mining network is difficult to operate on the deployed electroencephalogram regulation and control device is prevented.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a method and a computer system for mining features of electroencephalogram regulation and monitoring data based on deep learning. Background Art

[0002] Electroencephalogram (EEG) signals contain rich information about brain activities and have great application potential in many fields such as medicine, neuroscience research, and brain-computer interface technology. In the medical field, EEG signals can be used for disease diagnosis, treatment monitoring, and assessment of brain function status. For example, neurological diseases such as epilepsy and Alzheimer's disease are often accompanied by characteristic changes in EEG signals, and analyzing EEG signals can assist doctors in early diagnosis and disease monitoring. In neuroscience research, EEG signals help to deeply understand complex mechanisms such as brain cognitive processes and neuroplasticity. In the aspect of brain-computer interface technology, EEG signals are the key medium for realizing communication between the brain and external devices, providing the possibility for paralyzed patients to control auxiliary devices, etc. With the application of deep learning technology in the processing of electroencephalogram regulation and monitoring data, the constructed neural network model often requires a large scale when dealing with complex electroencephalogram data feature mining tasks, including numerous neurons and connection weights. However, in actual application scenarios, especially on some portable or resource-constrained devices (such as wearable electroencephalogram monitoring devices), limited computing resources (such as memory and processing power) pose a huge challenge. A large-scale neural network may lead to slow running speed, high energy consumption, or even unable to run normally on these devices. Summary of the Invention

[0003] In view of this, the present invention provides a method and a computer system for mining features of electroencephalogram regulation and monitoring data based on deep learning. The technical solution of the present invention is realized as follows:

[0004] In a first aspect, an embodiment of the present invention provides a method for mining features of electroencephalogram regulation monitoring data based on deep learning. The method includes: obtaining a knowledge template library of the monitoring data feature mining network, where the knowledge template library includes one or more target knowledge templates and one or more non-target knowledge templates. The target knowledge template represents electroencephalogram regulation monitoring data with target features, and the non-target knowledge template represents electroencephalogram regulation monitoring data without the target features; calibrating the monitoring data feature mining network through the knowledge template library to obtain an initially calibrated monitoring data feature mining network; pruning network configuration variables of the initially calibrated monitoring data feature mining network to obtain a pruned monitoring data feature mining network, where the memory occupancy of the pruned monitoring data feature mining network is less than that of the initially calibrated monitoring data feature mining network; calibrating the pruned monitoring data feature mining network through the knowledge template library to obtain a monitoring data feature mining network in a converged state; extracting a data feature vector of the electroencephalogram regulation monitoring data to be mined based on a characterization information extraction component of the monitoring data feature mining network in the converged state, and determining whether the electroencephalogram regulation monitoring data to be mined has the target features based on the data feature vector of the electroencephalogram regulation monitoring data to be mined.

[0005] In a second aspect, the present invention provides a computer system, including a memory and a processor. The memory stores a computer program that can run on the processor, and when the processor executes the program, the steps in the above method are implemented.

[0006] Advantages of the present invention: The method and computer system for mining features of electroencephalogram regulation monitoring data based on deep learning provided by the present invention prune network configuration variables of the initially calibrated monitoring data feature mining network and perform secondary calibration on the pruned monitoring data feature mining network to obtain a monitoring data feature mining network in a converged state. The present invention reduces the scale of the obtained monitoring data feature mining network to prevent the monitoring data feature mining network from running difficultly on the deployed electroencephalogram regulation device. Description of the Drawings

[0007] Figure 1 It is a schematic flowchart of the implementation of a method for mining features of electroencephalogram regulation monitoring data based on deep learning provided by an embodiment of the present invention.

[0008] Figure 2 It is a schematic diagram of the hardware entity of a computer system provided by an embodiment of the present invention. Detailed Embodiments

[0009] To make the objectives, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be further elaborated in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be construed as limiting the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0010] An embodiment of the present invention provides a method for mining features of electroencephalogram regulation monitoring data based on deep learning, and this method can be executed by a processor of a computer system. Among them, the computer system may refer to devices with data processing capabilities such as servers, laptop computers, tablet computers, desktop computers, electroencephalogram regulation devices, etc.

[0011] Figure 1 It is a schematic flowchart of the implementation of a method for mining features of electroencephalogram regulation monitoring data based on deep learning provided by an embodiment of the present invention. As Figure 1 shown, this method includes the following steps:

[0012] Step S100: Obtain a knowledge template library of the monitoring data feature mining network. The knowledge template library includes one or more target knowledge templates and one or more non-target knowledge templates. The target knowledge template represents electroencephalogram regulation monitoring data with target features, and the non-target knowledge template represents electroencephalogram regulation monitoring data without target features.

[0013] The knowledge template library contains one or more targeted knowledge templates and one or more non-targeted knowledge templates. The targeted knowledge templates represent the electroencephalogram (EEG) regulation monitoring data with target features, while the non-targeted knowledge templates represent the EEG regulation monitoring data without such target features. To obtain this knowledge template library, an original template library needs to be constructed. Taking the detection of seizure features in EEG signals as an example, the construction of the original template library is sourced from a large amount of EEG monitoring data. The original targeted knowledge templates may be extracted from the EEG data collected during seizures from patients who have been diagnosed with epilepsy. These data have some obvious features, such as a sudden decrease in the energy of EEG signals in a specific frequency band (such as the α band of 8 - 12 Hz), and at the same time, there are spike-like energy bursts in the high-frequency band (such as the γ band of 30 - 80 Hz). These features can be represented by feature vectors. Suppose a feature vector is [α-band energy reduction amplitude, γ-band spike count, spike amplitude], where each element is a specific numerical range, such as [0.3 - 0.5 (representing the proportion of energy reduction), 3 - 5 (spike count), 0.8 - 1.2 (spike amplitude in microvolts)]. The original non-targeted knowledge templates are obtained from the EEG data of healthy people or patients without seizures. The feature vector of their EEG signals may be [α-band stable energy, low γ-band activity], and the corresponding numerical range may be [0.9 - 1.1 (α-band energy proportion), 0 - 1 (quantification value of low γ-band energy activity)].

[0014] After constructing the original template library, the computer system needs to process the target features to enrich the template library. Suppose the set of target features includes features such as the frequency stability, phase synchrony, and signal complexity of EEG signals. The computer system will perform combination operations on these target features. For example, combine the two target features of frequency stability and signal complexity into a data item sequence. For frequency stability, it can be represented by an array indicating the frequency fluctuations in different time periods, such as [0.1, 0.05, 0.08] (representing the amplitude of frequency fluctuations in three consecutive time periods), and the signal complexity can be represented by a specific numerical value, such as 0.6 (obtained based on a certain complexity calculation algorithm). The combined data item sequence contains the relevant data of these two target features.

[0015] Next, further operations are performed on this data item sequence. Suppose a part is extracted from this data item sequence as a subsequence, for example, only extract the subsequence of signal complexity. Then fuse this subsequence with the original non-targeted knowledge template. The original non-targeted knowledge template's EEG data originally did not contain this specific signal complexity-related information. After fusion, this non-targeted knowledge template is enhanced.

[0016] If the feature represented by this subsequence belongs to the target feature in the target feature set, for example, the feature of signal complexity is a member of the target feature set, then this enhanced template can be regarded as an enhanced target knowledge template. This means that this template now contains information related to the target feature and can be used as a target knowledge template in subsequent operations.

[0017] Conversely, if the feature represented by the subsequence does not belong to the target feature in the target feature set, then this enhanced template is still regarded as an enhanced non-target knowledge template.

[0018] Finally, the computer system constructs a knowledge template library for the monitoring data feature mining network based on the original template library and the newly generated enhanced target knowledge templates and enhanced non-target knowledge templates. The templates in this knowledge template library will provide an important basis for subsequent operations such as network calibration. For example, in a neural network model based on deep learning, the data features in these templates can be used as training data or reference data. During the training process of the neural network, the data in the target knowledge templates can be used as positive examples, and the data in the non-target knowledge templates can be used as negative examples to help the neural network learn how to distinguish electroencephalogram regulation monitoring data with target features from those without target features, so as to accurately mine the target features in the data.

[0019] Step S200: Calibrate the monitoring data feature mining network through the knowledge template library to obtain an initially calibrated monitoring data feature mining network.

[0020] In step S200, the knowledge template library obtained in step S100, which contains target knowledge templates and non-target knowledge templates, plays a key role in calibrating the monitoring data feature mining network.

[0021] Suppose the monitoring data feature mining network is a deep neural network model, and the purpose of this network is to mine specific features from electroencephalogram regulation monitoring data. For example, in electroencephalogram signal research, if the goal is to detect features related to a certain cognitive state (such as the focused state), the network needs to learn to distinguish the associations between different electroencephalogram signal patterns and the focused state.

[0022] The targeted knowledge templates in the knowledge template library represent EEG regulation monitoring data with target features. Taking the focused state as an example, the data features in the targeted knowledge templates may be manifested as the energy changes of EEG signals in specific frequency bands and the signal synchrony between different brain regions, etc. For example, in the frontal lobe brain region, the energy of the EEG signal in the α frequency band (8 - 12 Hz) may relatively decrease in the focused state, and at the same time, the signal synchrony between the frontal lobe and parietal lobe brain regions in the β frequency band (13 - 30 Hz) increases. These features can be represented by feature vectors, such as [relative value of α frequency band energy in the frontal lobe, signal synchrony index of β frequency band between the frontal lobe and parietal lobe], and the specific values may be [0.8 (indicating the energy ratio after relative decrease), 0.6 (indicating the quantified value of signal synchrony)].

[0023] The non - targeted knowledge templates represent EEG regulation monitoring data without target features. For focused state detection, the data in the non - targeted knowledge templates may come from EEG monitoring in non - focused states, such as EEG signals in the relaxed state or distracted state. These signals are different from the focused state in terms of frequency band energy distribution and brain region synchrony. For example, in the relaxed state, the energy of the α frequency band in the frontal lobe may be relatively high, and the signal synchrony of the β frequency band between brain regions is weak, and the corresponding feature vector may be [1.2 (relative value of α frequency band energy in the frontal lobe), 0.3 (signal synchrony index of β frequency band between the frontal lobe and parietal lobe)]. The computer system inputs the data in these targeted and non - targeted knowledge templates into the monitoring data feature mining network. In the deep neural network model, each layer of the network contains multiple neurons, and the neurons are connected by weights. When the data in the targeted knowledge templates is input, the network processes the data according to a preset activation function (such as the ReLU function). For example, for the input feature vector [0.8, 0.6] in the targeted knowledge template, the neuron will perform a dot - product operation on this vector and its own weight vector, and then obtain an output value through the ReLU function, and this output value is then passed to the next - layer neuron. The same operation is also performed on the data in the non - targeted knowledge templates.

[0024] By continuously inputting the data in the knowledge template library into the network, the weights in the network are adjusted according to the error between the input data and the expected output (the output corresponding to the target knowledge template has the target feature, and the output corresponding to no target knowledge template does not have the target feature). This adjustment process is based on the backpropagation algorithm. For example, if the network outputs an incorrect result for the data in the target knowledge template, it indicates that the weights of the network need to be adjusted. The backpropagation algorithm calculates the contribution degree of each weight to the error, and then adjusts the weights according to this contribution degree. After a large amount of data input and weight adjustment, the network gradually learns the data features in the target knowledge template and the non-target knowledge template, thereby realizing the calibration of the monitoring data feature mining network. After such a calibration process, the computer system obtains the monitoring data feature mining network that has been initially calibrated. This initially calibrated network already has a certain learning and discrimination ability for the features of the target and non-target electroencephalogram regulation monitoring data, laying a foundation for subsequent further optimization (such as network configuration variable pruning in step S300). Through this calibration method, the monitoring data feature mining network can better adapt to the feature mining task of electroencephalogram regulation monitoring data and improve the accuracy of the mining results.

[0025] Step S300: Prune the network configuration variables of the initially calibrated monitoring data feature mining network to obtain the pruned monitoring data feature mining network; wherein, the memory occupancy of the pruned monitoring data feature mining network is less than that of the initially calibrated monitoring data feature mining network.

[0026] In step S300, the network configuration variables of the initially calibrated monitoring data feature mining network are pruned to obtain the pruned monitoring data feature mining network, and the memory occupancy of the pruned network is less than that of the initially calibrated network.

[0027] In the embodiment of the present invention, the monitoring data feature mining network is a deep neural network structure, such as a convolutional neural network (CNN) or a recurrent neural network (RNN). This network is composed of multiple network components, and each network component contains numerous parameters, and these parameters are the network configuration variables.

[0028] Assume that a network component in the monitoring data feature mining network is a fully connected layer. The weight tensor in the fully connected layer can be represented as a two-dimensional array. For example, the weight tensor is an i*j matrix, where i = 5 and j = 4, and each element in the matrix is a weight parameter, such as numerical values like 0.1, -0.2, etc. This weight tensor determines the transformation method of the input data in this layer.

[0029] The computer system operates on this weight tensor. First, the weight tensor of the first network component is subjected to eigenvalue decomposition to obtain the product result of the first rotation tensor, the eigenvalue tensor, and the second rotation tensor. This process is similar to matrix eigenvalue decomposition. For example, the first rotation tensor is an i*i (where i = 5) matrix, the eigenvalue tensor is an i*j (i = 5, j = 4) matrix, and the second rotation tensor is a j*j (j = 4) matrix.

[0030] To understand this process, take a simple 2*2 matrix as an example. Assume a simple weight tensor After eigenvalue decomposition, we get A = QΛR, where Q is the first rotation tensor, Λ is the eigenvalue tensor, and R is the second rotation tensor.

[0031] In this eigenvalue tensor, there are multiple eigenvalues. The computer system then determines the target eigenvalues that meet the first requirement. For example, the eigenvalues in the eigenvalue tensor may be λ 1 = 0.5, λ 2 = 0.3, λ 3 = 0.1, λ 4 = 0.05 (here assume i*j = 4 eigenvalues). According to a certain determined rule, such as selecting the larger eigenvalues in the order of eigenvalue magnitude, assume the first p = 2 eigenvalues (λ 1 = 0.5 and λ 2 = 0.3) are selected as the target eigenvalues.

[0032] Then, in the first rotation tensor, the eigenvalue tensor, and the second rotation tensor, the configuration variables corresponding to the target eigenvalues are respectively extracted, so as to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor.

[0033] Continuing with the above example, if the first rotation tensor Since there are 2 target eigenvalues, the columns corresponding to these 2 target eigenvalues are extracted from Q. Assume they are the 1st column and the 3rd column, and the pruned first rotation tensor is obtained This pruned first rotation tensor contains j*p (here j = 4, p = 2) weight parameters.

[0034] For the eigenvalue tensor, originally After pruning, we get It contains p*p (p = 2) weight parameters.

[0035] For the second rotation tensor The rows corresponding to the target eigenvalues are extracted. Assume they are the 1st row and the 2nd row, and the pruned second rotation tensor is obtained It contains p*j (p = 2, j = 4) weight parameters.

[0036] Through such pruning operations on each tensor, the computer system successfully streamlines the configuration variables of the first network component. After performing similar operations on other network components in the monitoring data feature mining network, the pruned monitoring data feature mining network is finally obtained. Such pruning operations can effectively reduce redundant parameters in the network, lower the network complexity, thereby reducing the network's memory occupancy, while to a certain extent improving the network's operating efficiency, and without excessive loss of the network's ability to mine the features of electroencephalogram regulation monitoring data, providing a more optimized network structure for subsequent further tuning and other operations.

[0037] Step S400: Tune the pruned monitoring data feature mining network through the knowledge template library to obtain a monitoring data feature mining network in a converged state.

[0038] In step S400, the computer system tunes the pruned monitoring data feature mining network through the knowledge template library, thereby obtaining a monitoring data feature mining network in a converged state.

[0039] In the task of mining electroencephalogram regulation monitoring data features based on deep learning, after the monitoring data feature mining network has undergone previous steps (such as initial tuning and pruning operations), its structure and parameters have been adjusted to a certain extent. Although the pruned monitoring data feature mining network reduces memory occupancy, it may not yet reach the optimal state and still needs further tuning to reach a converged state.

[0040] Suppose the monitoring data feature mining network is a multi-layer perceptron (MLP) neural network for mining specific physiological or pathological features from electroencephalogram regulation monitoring data. For example, the goal is to mine features related to epileptic seizures from electroencephalogram signals, such as the energy change of electroencephalogram signals in a specific frequency band, the complexity of the signal, etc.

[0041] The knowledge template library contains target knowledge templates and non-target knowledge templates. Taking epilepsy detection as an example, the data in the target knowledge templates may come from electroencephalogram monitoring data during the seizures of patients already diagnosed with epilepsy. These data have specific feature vectors, assumed to be represented as an array, such as [the increase amplitude of the energy of electroencephalogram signals in a specific frequency band (such as 3 - 8 Hz) is 0.5, and the signal complexity reaches 0.8 (based on a certain complexity calculation method)]. The data in the non-target knowledge templates may come from electroencephalogram monitoring data of healthy people, and its feature vectors may be [the energy of electroencephalogram signals in the corresponding frequency band is stable, and the signal complexity is 0.2].

[0042] The computer system inputs the data in the knowledge template library into the pruned monitoring data feature mining network. For a multi-layer perceptron network, the input layer receives the feature vector data in the knowledge template library and then processes it through the hidden layer. The neurons in the hidden layer perform weighted summation on the input data and apply an activation function (such as the sigmoid function or ReLU function) for non-linear transformation.

[0043] For example, when the input is the data [0.5, 0.8] in the target knowledge template, assume that the first hidden layer has 3 neurons, and each neuron has a connection weight with the input layer. For the first neuron, the connection weight vector is assumed to be [0.1, 0.2], then the weighted summation is 0.1×0.5 + 0.2×0.8 = 0.21, and then through the sigmoid function to obtain the output value, where sigmoid(0.21) ≈ 0.55. This output value will continue to be passed to the next layer of neurons or directly used as the output (if it is the last hidden layer).

[0044] During the backpropagation process, the network adjusts the weights of the network according to the error between the output result and the expected result (the target knowledge template corresponds to having target features, and the non-target knowledge template corresponds to not having target features). If the output result indicates that the network has misjudged the data in the target knowledge template (i.e., judged as not having target features), then the weights need to be adjusted to reduce this error. Assume that the error between the output result and the expected result after one input of the target knowledge template data is E. Calculate the contribution degree of each weight to the error according to the error. For example, for the connection weights 0.1 and 0.2 of the first neuron mentioned above, calculate the amount Δw that they should be adjusted by a specific backpropagation algorithm (such as the gradient descent algorithm). 1 and Δw 2 . Then update the weights to Δw 1 = 0.1 + Δw 1 , Δw 2 = 0.2 + Δw 2 .

[0045] The computer system continuously repeats the input of the data in the knowledge template library into the pruned monitoring data feature mining network, performs forward propagation and backpropagation operations. As the number of iterations increases, the weights of the network are continuously adjusted. During this process, the error between the output result of the network and the expected result gradually decreases. When the error decreases to a certain extent, for example, the change in the error is less than a preset minimum value (such as 10 -6) It can be considered that the monitoring data feature mining network has reached a convergence state. At this time, the computer system has obtained the monitoring data feature mining network in the convergence state. This network in the convergence state already has a good ability to distinguish the target features and non-target features in the EEG regulation monitoring data, and can more accurately mine the features in the EEG regulation monitoring data, laying a solid foundation for the subsequent feature mining of the EEG regulation monitoring data based on this network.

[0046] Step S500: Based on the representation information extraction component of the monitoring data feature mining network in the convergence state, obtain the data feature vector of the EEG regulation monitoring data to be mined. Based on the data feature vector of the EEG regulation monitoring data to be mined, determine whether there are target features in the EEG regulation monitoring data to be mined.

[0047] In step S500, the computer system obtains the data feature vector of the EEG regulation monitoring data to be mined based on the representation information extraction component of the monitoring data feature mining network in the convergence state, and then determines whether there are target features in the EEG regulation monitoring data to be mined based on this data feature vector.

[0048] In the context of EEG regulation monitoring data feature mining based on deep learning, the monitoring data feature mining network in the convergence state has gone through a series of previous calibration processes (such as initial calibration through the knowledge template library, re-calibration after pruning, etc.) and has the ability to extract effective features from the EEG regulation monitoring data.

[0049] Suppose the monitoring data feature mining network in the convergence state is a deep neural network (such as a convolutional neural network), and its representation information extraction component can be regarded as the part of the network dedicated to extracting data feature representations. This part may consist of the last few layers of the network, which transform the input EEG regulation monitoring data into a data feature vector.

[0050] For example, the EEG regulation monitoring data can be multi-channel EEG signal data collected over a period of time. The data of each channel may contain information such as the EEG signal intensity in different frequency bands (such as α band, β band, γ band, etc.). Suppose the data of each channel is regarded as an array. For a simple example of three-channel EEG data, the data array of channel 1 is [α band intensity value is 0.3, β band intensity value is 0.5, γ band intensity value is 0.2], the data array of channel 2 is [α band intensity value is 0.4, β band intensity value is 0.3, γ band intensity value is 0.3], and the data array of channel 3 is [α band intensity value is 0.2, β band intensity value is 0.4, γ band intensity value is 0.4].

[0051] When these EEG regulation monitoring data are input into the monitoring data feature mining network in a convergent state, the representation information extraction component of the network processes these data. It may extract local features from data in different channels through convolutional layers, then compress the features through pooling layers, and finally combine these features into a data feature vector through fully connected layers. Suppose the obtained data feature vector is [0.1, 0.2, 0.3, 0.4], where each numerical value represents a quantization value related to different features after being processed by the network.

[0052] The computer system determines whether the EEG regulation monitoring data to be mined has target features based on this data feature vector. Suppose the target feature is the EEG signal feature related to a certain cognitive state (such as the focused state). The target feature set may contain multiple target data item sequences and the corresponding data item attribute sequences for each target data item sequence.

[0053] For example, one of the target data item sequences may be a sequence regarding the energy ratio of EEG signal frequency bands, such as [the energy ratio of the α band should be lower in the focused state, and the energy ratio of the β band should be higher in the focused state], and the corresponding data item attribute sequence may be [the threshold range of the energy ratio, the threshold range of the energy ratio], supposed to be [α band: 0 - 0.3, β band: 0.6 - 1].

[0054] The computer system matches the data feature vector [0.1, 0.2, 0.3, 0.4] with the target data item sequences in the target feature set. If the numerical values in the data feature vector conform to the conditions in the target data item sequence, for example, suppose the first numerical value 0.1 in the data feature vector represents the energy ratio of the α band, and 0.2 represents the energy ratio of the β band, and they respectively fall within the threshold ranges of the corresponding attribute sequences of the target data item sequence, then it can be determined that the EEG regulation monitoring data to be mined has the target feature (the feature related to the focused state).

[0055] Another example, if the target data item sequence is about the synchrony between different brain regions of the EEG signal, such as [the β - band synchrony between the frontal lobe and the parietal lobe is higher in the focused state], and the corresponding attribute sequence is [the threshold range of the synchrony index, such as 0.6 - 1], and a certain numerical value in the data feature vector represents that the β - band synchrony index between the frontal lobe and the parietal lobe is 0.8, which falls within the threshold range, this also indicates that the EEG regulation monitoring data to be mined has the target feature.

[0056] In this way, the computer system uses the data feature vector extracted by the monitoring data feature mining network in a convergent state to compare and match with the target feature set, thereby accurately judging whether there are target features in the EEG regulation monitoring data to be mined, which is of great significance for tasks such as feature analysis, disease diagnosis, or cognitive state assessment in EEG regulation monitoring.

[0057] As an implementation, in step S300, pruning the network configuration variables of the initially calibrated monitoring data feature mining network to obtain the pruned monitoring data feature mining network may include:

[0058] Step S310: For the first network component in the initially calibrated monitoring data feature mining network, perform eigenvalue splitting on the weight tensor of the first network component to obtain the product result of the first rotation tensor, eigenvalue tensor, and second rotation tensor;

[0059] Step S320: Determine the target eigenvalue that meets the first requirement among the eigenvalues of the eigenvalue tensor;

[0060] Step S330: In the first rotation tensor, eigenvalue tensor, and second rotation tensor, respectively extract the configuration variables corresponding to the target eigenvalue to obtain the pruned first rotation tensor, pruned eigenvalue tensor, and pruned second rotation tensor; wherein, the configuration variables of the first network component in the pruned monitoring data feature mining network are composed of the pruned first rotation tensor, pruned eigenvalue tensor, and pruned second rotation tensor.

[0061] A weight tensor is a multi-dimensional array used to represent the strength of the connection between neurons. In this scenario, for the first network component (such as a fully connected layer), the weight tensor is in the form of an i*j matrix (where i and j respectively represent the values related to the dimensions of the input and output). Taking a simple example, if the input layer has 4 neurons and the output layer has 3 neurons, then the weight tensor may be a 4*3 matrix. Each element (weight parameter) determines the signal transmission strength from the input neuron to the output neuron. For example, the element w ij in the weight tensor represents the connection weight from the i-th input neuron to the j-th output neuron. The weight tensor plays a key role in the forward propagation process of the neural network. When the input data passes through the first network component, each element of the input data is multiplied by the corresponding weight in the weight tensor, and then operations such as summation are performed, so as to convert the input data into a form that can be processed by the next layer. It determines how the network linearly transforms the input data and is an important basis for the neural network to learn and represent data features. Different weight values will result in different propagation paths and results of the input data in the network. By adjusting the weight parameters in the weight tensor, the network can learn different input-output mapping relationships, so as to realize the mining of the features of the electroencephalogram regulation monitoring data.

[0062] The First Rotation Tensor is an i*i matrix obtained by performing eigenvalue decomposition on the weight tensor (where i is related to one dimension of the weight tensor). For example, when the weight tensor is 4*3, if eigenvalue decomposition is performed, the First Rotation Tensor might be a 4*4 matrix. The values of its elements represent a rotation relationship in the feature space. The First Rotation Tensor plays a role in transforming the space of the original weight tensor within the overall framework of eigenvalue decomposition. It works together with the Eigenvalue Tensor and the Second Rotation Tensor to re - represent the structure of the weight tensor. In data processing, it helps decompose the complex linear transformation represented by the original weight tensor into a form that is easier to understand and operate on. For example, it can represent the complex relationships between different dimensions in the original weight tensor through rotation operations in a new feature space, providing a new perspective for subsequent determination of important eigenvalues and pruning operations.

[0063] The Eigenvalue Tensor is an i*j matrix (i and j are related to the dimensions of the weight tensor), and the elements on its diagonal are eigenvalues, and the off - diagonal elements may be 0 (in the case of ideal eigenvalue decomposition). For example, in the case where the weight tensor mentioned earlier is 4*3, the Eigenvalue Tensor might be a 4*3 matrix, and the eigenvalues on the diagonal represent the scaling factors in the corresponding eigenvector directions. The eigenvalues represent the degree of importance (in the form of scaling factors) in each eigen - direction after the rotation operation of the First Rotation Tensor. Larger eigenvalues mean that the changes in the data in the corresponding eigenvector direction have a greater influence in the linear transformation represented by the weight tensor. By analyzing the eigenvalues in the Eigenvalue Tensor, it is possible to determine which eigen - directions are more critical for the transformation represented by the weight tensor, providing a basis for determining the target eigenvalues for pruning operations.

[0064] The Second Rotation Tensor is a j*j matrix (j is related to one dimension of the weight tensor). For example, when the weight tensor is 4*3, the Second Rotation Tensor might be a 3*3 matrix. Similar to the First Rotation Tensor, it is also a representation of a rotation operation in the feature space, but from the perspective of another dimension. The Second Rotation Tensor works together with the First Rotation Tensor and the Eigenvalue Tensor to perform a complete eigenvalue decomposition representation of the weight tensor. It plays an important role in decomposing the weight tensor into a form that is easier to analyze and operate on. Specifically, together with the First Rotation Tensor, through multiplication operations with the Eigenvalue Tensor, it reconstructs the representation form of the weight tensor, enabling the understanding of the linear transformation represented by the weight tensor from different perspectives (i.e., by analyzing the eigenvalues and the corresponding rotation relationships), providing a basis for subsequent eigenvalue - based pruning operations.

[0065] The multiplication result of the first rotation tensor, the eigenvalue tensor, and the second rotation tensor refers to the result obtained by multiplying these three tensors according to specific multiplication rules. This result should be equivalent to the original weight tensor (theoretically, ignoring numerical errors in the calculation process). Mathematically, if the first rotation tensor is Q, the eigenvalue tensor is Λ, and the second rotation tensor is R, then the multiplication result is QΛR, and this result should be equal to the original weight tensor W, that is, W = QΛR. The role of this multiplication result is to verify the correctness of the eigenvalue splitting operation. If the multiplication result can be accurately restored to the original weight tensor, it indicates that the eigenvalue splitting operation is successful. At the same time, this multiplication result also provides a complete framework for subsequent pruning operations based on eigenvalues and related tensors. By decomposing the weight tensor into the multiplication form of these three tensors, it is more convenient to analyze and operate on the elements in each tensor (especially the part related to eigenvalues), so as to achieve pruning of network configuration variables and reduce network memory occupancy and improve network efficiency.

[0066] The computer system performs eigenvalue splitting on the weight tensor of the first network component in the monitoring data feature mining network after initial tuning, and obtains the multiplication result of the first rotation tensor, the eigenvalue tensor, and the second rotation tensor.

[0067] Under the neural network framework of deep learning, the monitoring data feature mining network is a complex structure, and the first network component therein can be, for example, a fully connected layer. The weight tensor in the fully connected layer is a key component, which determines the connection strength between neurons and the transformation mode of data when passing between layers.

[0068] Suppose the weight tensor of this fully connected layer is an i*j matrix (here i = 4, j = 3), and each element in the matrix is a weight parameter. Take a simple weight tensor example

[0069] The computer system performs eigenvalue splitting operation on this weight tensor W. This eigenvalue splitting is similar to the matrix decomposition operation in linear algebra, and the purpose is to decompose the weight tensor into a form that is easier to process and analyze. Through a specific algorithm (such as singular value decomposition or eigenvalue decomposition algorithm), the weight tensor W is decomposed into the first rotation tensor Q, the eigenvalue tensor Λ, and the second rotation tensor R, such that W = QΛR.

[0070] Suppose the first rotation tensor Q obtained through calculation is an i*i matrix (here i = 4), for example The eigenvalue tensor Λ is an i*j (i = 4, j = 3) matrix, for example The second rotation tensor R is a matrix of j*j (j = 3), for example In this way, the multiplication result of the first rotation tensor, the eigenvalue tensor, and the second rotation tensor corresponds to the original weight tensor W, providing a basis for subsequent pruning operations.

[0071] In the eigenvalues of the eigenvalue tensor of the computer system, the target eigenvalues that meet the first requirement are determined. Continuing with the above example, the eigenvalues in the eigenvalue tensor Λ are 2.0, 1.5, and 1.0 respectively (here, the case where the last row is all 0 is ignored because in actual eigenvalue decomposition, there may be some 0 eigenvalue cases that are relatively unimportant compared to the main eigenvalues).

[0072] To determine the target eigenvalues that meet the first requirement, a clear rule or standard is needed. This rule can be formulated based on various factors, such as the size of the eigenvalues, the degree of influence of the eigenvalues on the network output, etc. Suppose the first requirement here is to select larger eigenvalues because larger eigenvalues often represent the principal component directions that have a greater impact on data transformation.

[0073] In this example, in ascending order, 2.0 and 1.5 are relatively large, so the computer system determines 2.0 and 1.5 as the target eigenvalues. The purpose of this step is to screen out the eigenvalues in the eigenvalue tensor after the weight tensor decomposition that are significant for network performance or data representation, so that in subsequent pruning operations, the network configuration variables related to these important eigenvalues are retained, and the relatively unimportant parts are removed, thereby achieving the purpose of streamlining the network and reducing memory occupancy.

[0074] In step S330, the computer system extracts the configuration variables corresponding to the target eigenvalues from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor, and obtains the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor.

[0075] For the first rotation tensor Q, since there are 2 target eigenvalues (2.0 and 1.5), and the first rotation tensor Q is a 4*4 matrix, we need to extract the configuration variables corresponding to these 2 target eigenvalues from Q. Suppose the corresponding relationship here is according to the order of the eigenvalues in the eigenvalue tensor, corresponding to the columns in the first rotation tensor. Then, we extract the 1st and 2nd columns in the first rotation tensor Q, and the pruned first rotation tensor Q' obtained is a 4*2 matrix, for example

[0076] For the eigenvalue tensor Λ, originally a 4* matrix, because there are 2 target eigenvalues, the pruned eigenvalue tensor Λ′ only retains the parts related to these 2 target eigenvalues, and a 2*2 matrix is obtained, for example

[0077] For the second rotation tensor R, which is a 3× matrix, according to the correspondence with the target eigenvalues (assuming in the order of eigenvalues in the eigenvalue tensor, corresponding to the rows in the second rotation tensor), the first and second rows are extracted, and the pruned second rotation tensor R' obtained is a 2× matrix. For example

[0078] By such operations of extracting the configuration variables corresponding to the target eigenvalues in the first rotation tensor, the eigenvalue tensor, and the second rotation tensor respectively, the computer system obtains the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor. These pruned tensors constitute the configuration variables of the pruned first network component, thus realizing the pruning operation of the first network component. The pruning operation can also be performed on other network components in the monitoring data feature mining network in a similar way, and finally the entire pruned monitoring data feature mining network is obtained. This pruning operation can effectively reduce the redundant information in the network, reduce the complexity of the network, reduce the memory occupancy, and at the same time maintain the ability of the network to mine the features of the electroencephalogram regulation monitoring data to a certain extent, providing a more optimized network structure for the subsequent network tuning and data feature mining tasks.

[0079] As an implementation manner, in step S320, to determine the target eigenvalues that meet the first requirement among the eigenvalues of the eigenvalue tensor may include:

[0080] Step S321: Sort each eigenvalue of the eigenvalue tensor in a descending manner to obtain an eigenvalue vector;

[0081] Step S322: Determine the first p eigenvalues in the eigenvalue vector to obtain p target eigenvalues, where p≥2.

[0082] In step S320 of the computer system, the purpose is to determine the target eigenvalues that meet the first requirement among the eigenvalues of the eigenvalue tensor, and this process is realized through sub-steps S321 to S322.

[0083] In step S321, the computer system sorts each eigenvalue of the eigenvalue tensor in a descending manner to obtain an eigenvalue vector.

[0084] Under the neural network architecture of deep learning, assume that in the previous step S310, after eigenvalue splitting of the weight tensor of the first network component (for example, a specific hidden layer), an eigenvalue tensor is obtained. Assume that this eigenvalue tensor is an i×j matrix (for the convenience of example here, let i = 5 and j = 3), for example (The last two rows here are all 0 in the feature

[0085] The situations that may occur in value decomposition, the actual valid eigenvalues ​​are the first three).

[0086] The computer system will extract the non-zero eigenvalues ​​in this eigenvalue tensor, namely 3.0, 2.5, and 1.8. Then these eigenvalues ​​are sorted in descending order. The sorted result forms an eigenvalue vector. In this example, the eigenvalue vector is [3.0, 2.5, 1.8]. This eigenvalue vector represents the sequence of eigenvalues ​​extracted from the eigenvalue tensor according to importance (here the numerical value represents the importance, the larger the value, the more important it is). The significance of this sorting operation is that it provides an orderly basis for the subsequent determination of the target eigenvalue, so that the computer system can determine the eigenvalues ​​that have an important impact on the network structure and performance based on certain criteria (such as selecting the first few larger eigenvalues).

[0087] In step S322, the computer system determines the first p eigenvalues ​​in the eigenvalue vector and obtains p target eigenvalues, where p≥2.

[0088] Continuing with the above example, assume that p = 2 is set according to the network structure or experience. Then the computer system will select the first two eigenvalues, 3.0 and 2.5, from the obtained eigenvalue vector [3.0, 2.5, 1.8] as the target eigenvalues.

[0089] In the EEG control monitoring data feature mining network based on deep learning, the way to determine the target eigenvalue is based on the importance distribution of the eigenvalue in the network. Larger eigenvalues ​​often correspond to directions that have a greater impact on data transformation. For example, in the forward propagation process of the neural network, the weight tensor performs a linear transformation on the input data, and the eigenvalue plays a role similar to a scaling factor in this transformation. A larger eigenvalue means that in the direction of the corresponding eigenvector, the change in the input data will be amplified or reduced to a greater extent, thereby having a more significant impact on the output of the network.

[0090] To illustrate with a simple neural network example, assume that the input EEG control monitoring data is processed through a series of hidden layers and arrives at the first network component that is undergoing pruning. If the input data is regarded as a vector After the linear transformation of the weight tensor W (W here is represented by the tensor related to the previous eigenvalue split), the output is In this process, the eigenvalues ​​in the eigenvalue tensor affect the result of this linear transformation. For example, when the eigenvalue is large, the input vector The components in the direction of the corresponding eigenvectors will be changed to a greater extent, thus affecting the final output Have an important impact.

[0091] Therefore, by determining the first p eigenvalues ​​as target eigenvalues, the computer system can retain the network configuration variables related to these important eigenvalues ​​in the subsequent pruning operation, thereby reducing the network memory usage while retaining the network's ability to mine the features of EEG control monitoring data as much as possible. These target eigenvalues ​​will serve as the basis for extracting corresponding configuration variables from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor in subsequent steps (such as step S330) to achieve accurate pruning operations on network configuration variables.

[0092] As an implementation mode, the weight tensor of the first network component includes i*j weight parameters, the first rotation tensor includes i*i weight parameters, the eigenvalue tensor includes i*j weight parameters, the second rotation tensor includes j*j weight parameters, and the number of target eigenvalues ​​is p, i≥2; j≥2; p≥2; based on this, step S330, in the first rotation tensor, the eigenvalue tensor and the second rotation tensor, respectively extracts the configuration variables corresponding to the target eigenvalues, and obtains the pruned first rotation tensor, the pruned eigenvalue tensor and the pruned second rotation tensor, which may include:

[0093] Step S331: based on the target eigenvalue, obtain p columns of weight parameters corresponding to the target eigenvalue in the first rotation tensor to obtain a pruned first rotation tensor, where the pruned first rotation tensor includes j*p weight parameters;

[0094] Step S332: based on the target eigenvalue, obtaining a pruned eigenvalue tensor, wherein the pruned eigenvalue tensor includes p*p weight parameters;

[0095] Step S333: Based on the target eigenvalue, obtain p rows of weight parameters corresponding to the target eigenvalue in the second rotation tensor to obtain the pruned second rotation tensor, and the pruned second rotation tensor includes p*j weight parameters; each configuration variable in the eigenvalue tensor corresponds to a column of weight parameters in the first rotation tensor and a row of weight parameters in the second rotation tensor.

[0096] In step S330, the configuration variables corresponding to the target eigenvalues ​​are extracted from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor, respectively, to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor. This process is implemented through sub-steps S331 to S333.

[0097] In step S331, based on the target eigenvalue, the computer system obtains p column weight parameters corresponding to the target eigenvalue in the first rotation tensor, and obtains the pruned first rotation tensor. The pruned first rotation tensor includes j*p weight parameters. Assume that in the previous step, the first rotation tensor Q is an i*i matrix. Here, assume i = 5, then

[0098] In step S322, the number p of target eigenvalues is determined. Assume p = 3. And assume that the target eigenvalues are selected from the sorted eigenvalue vector. For example, they correspond to the 1st, 3rd, and 4th eigenvalues in the eigenvalue tensor (this is just for illustrative purposes to show the correspondence).

[0099] Since there is a certain correspondence between the position of each eigenvalue in the eigenvalue tensor and the columns in the first rotation tensor (this correspondence is determined by the algorithmic nature of eigenvalue splitting), the computer system extracts the columns in the first rotation tensor according to this correspondence. Here, it is to extract the 1st, 3rd, and 4th columns to obtain the pruned first rotation tensor Q'.

[0100] The pruned first rotation tensor Q' is an i*p (here i = 5, p = 3) matrix, that is This matrix contains j*p (here j = 5) weight parameters.

[0101] This pruning operation of the first rotation tensor helps reduce the complexity of the network. For example, in a deep neural network for electroencephalogram regulation and monitoring data classification, the first network component may be part of a fully connected layer, and the elements (weight parameters) of the first rotation tensor determine the way data is transformed in this component. By extracting the columns corresponding to the target eigenvalues, it is equivalent to removing those partial connections that have relatively little impact on data transformation. This is like removing some wire connections in a complex circuit network that have relatively little impact on the overall circuit function, thus simplifying the circuit structure. In a neural network, the simplified structure can reduce the computational amount and memory occupancy, while maintaining the network's ability to mine data features to a certain extent.

[0102] In step S332, based on the target eigenvalue, the computer system obtains the pruned eigenvalue tensor. The pruned eigenvalue tensor includes p*p weight parameters.

[0103] Assume that the previous eigenvalue tensor \(\Lambda\) is an i*j matrix. Here, assume i = 5, j = 3, for example where \(\lambda\) 1 、\(\lambda\) 2 、\(\lambda\) 3 are valid eigenvalues.

[0104] Since the number of target eigenvalues p = 3 is determined, and it is assumed that the target eigenvalues are λ 1 , λ 2 , λ 3 . The computer system constructs the pruned eigenvalue tensor Λ′ according to the target eigenvalues.

[0105] The pruned eigenvalue tensor Λ′ is a matrix of p*p (where p = 3), that is This matrix contains p*p weight parameters.

[0106] In the operation of the neural network, the eigenvalues in the eigenvalue tensor reflect the scaling ratios of the data in each eigen-direction during the linear transformation by the weight tensor. Larger eigenvalues mean that the changes in the data in the corresponding eigen-directions have a greater impact on the final result. By pruning the eigenvalue tensor and retaining the part related to the target eigenvalues, it is possible to reduce the network scale while capturing the eigen-directions that have a greater impact on the network performance. For example, in the process of feature mining of electroencephalogram regulation monitoring data, if the electroencephalogram signal is regarded as a data with multiple feature dimensions, the eigenvalues in the eigenvalue tensor determine the relative importance of different feature dimensions during the network propagation. The pruned eigenvalue tensor focuses on those important feature dimensions determined as target eigenvalues, which helps to improve the efficiency and accuracy of the network in processing electroencephalogram regulation monitoring data.

[0107] In step S333, the computer system obtains the p rows of weight parameters corresponding to the target eigenvalues in the second rotation tensor based on the target eigenvalues, and obtains the pruned second rotation tensor. The pruned second rotation tensor includes p*j weight parameters.

[0108] Assume that the second rotation tensor R is a matrix of j*j. Here, it is assumed that j = 3, then

[0109] Also based on the number of target eigenvalues p = 3 and the corresponding relationship between the target eigenvalues and the rows in the second rotation tensor (determined by the eigenvalue splitting algorithm), the computer system extracts the rows in the second rotation tensor. Here, it is to extract the 1st, 2nd, and 3rd rows to obtain the pruned second rotation tensor R'.

[0110] The pruned second rotation tensor R' is a matrix of p*j (where p = 3, j = 3), that is This matrix contains p*j weight parameters.

[0111] In a neural network, the second rotation tensor is similar to the first rotation tensor and participates in the transformation operation of the input data. By pruning the second rotation tensor and removing the rows related to non-target eigenvalues, the network structure can be simplified. For example, in a convolutional neural network for feature extraction of electroencephalogram regulation monitoring data, the network component where the second rotation tensor is located may perform transformations in the spatial or feature dimensions of the data. The pruned second rotation tensor reduces unnecessary calculations, enabling the network to more quickly focus on important feature information when processing electroencephalogram regulation monitoring data, improving the network's operating efficiency, and maintaining the network's ability to mine target features as much as possible while reducing memory occupancy.

[0112] Through steps S331 to S333, the computer system performs pruning operations on the first rotation tensor, the eigenvalue tensor, and the second rotation tensor respectively, obtaining the pruned tensors. These pruned tensors constitute the configuration variables of the pruned first network component, achieving the pruning of the network configuration variables of the first network component. For other network components in the monitoring data feature mining network, pruning operations can also be performed in a similar manner, ultimately obtaining the entire pruned monitoring data feature mining network, which helps reduce the network complexity and memory occupancy, while maintaining to a certain extent the network's ability to mine electroencephalogram regulation monitoring data features.

[0113] As an implementation manner, in step S100, obtaining the knowledge template library of the monitoring data feature mining network may include:

[0114] Step S110: Obtain the original template library of the monitoring data feature mining network, where the original template library includes one or more original targeted knowledge templates and one or more original non-targeted knowledge templates;

[0115] Step S120: Sample one or more target features from the set of target features to be mined, and combine the one or more target features to obtain a first data item sequence;

[0116] Step S130: Sample subsequences from the first data item sequence, and add the subsequences to the original non-targeted knowledge templates to obtain enhanced templates; where the subsequence is one data item or multiple consecutive data items in the first data item sequence;

[0117] Step S140: If the subsequence is a target feature in the set of target features, use the enhanced template as an enhanced targeted knowledge template;

[0118] Step S150: If the subsequence is not a target feature in the set of target features, use the enhanced template as an enhanced non-targeted knowledge template;

[0119] Step S160: Obtain the knowledge template library of the monitoring data feature mining network based on the original template library, the enhanced targeted knowledge templates, and the enhanced non-targeted knowledge templates.

[0120] In step S110, the computer system obtains the original template library of the monitoring data feature mining network, which includes one or more original targeted knowledge templates and one or more original non-targeted knowledge templates.

[0121] In the embodiments of the present invention, the construction of the original template library is the basis for the construction of the entire knowledge template library. For example, in the case of using electroencephalogram (EEG) signal analysis to detect epilepsy, the original targeted knowledge templates can be obtained from the EEG monitoring data of patients who have been diagnosed with epilepsy. These data have specific characteristic patterns, which may be manifested as abnormal EEG signals in specific frequency bands during seizures.

[0122] Taking a simple frequency feature as an example, the EEG signal is divided into different frequency bands, such as the α band (8 - 12 Hz), the β band (13 - 30 Hz), the γ band (30 - 80 Hz), etc. During an epileptic seizure, there may be a sudden increase in the energy of the γ band and a relative decrease in the energy of the α band. Then a feature vector in the original targeted knowledge template can be represented as an array: [the relative value of the α band energy is 0.3 (assuming normal is 1), the relative value of the γ band energy is 1.5 (assuming normal is 1)]. Here, the values are the energy ratios relative to the normal state. This feature vector represents the EEG signal characteristics related to epileptic seizures and belongs to a part of the original targeted knowledge template.

[0123] The original non-targeted knowledge templates can be obtained from the EEG monitoring data of healthy people or patients in a non-epileptic seizure state. For example, the EEG signals of healthy people are relatively stable in each frequency band, and their feature vectors may be [the relative value of the α band energy is 0.9, the relative value of the γ band energy is 1.1]. These feature vectors together constitute the original non-targeted knowledge template. The original template library contains multiple such original targeted knowledge templates and non-targeted knowledge templates, which are the initial representations of the characteristics of different types of EEG regulation monitoring data and provide the basic materials for constructing a richer knowledge template library in the future.

[0124] In step S120, the computer system samples one or more target features from the set of target features to be mined, and combines the one or more target features to obtain the first data item sequence.

[0125] Continuing with the example of using EEG signals for epilepsy detection, the set of target features to be mined may include the frequency features of the EEG signal, the peak-valley morphology features of the signal, the signal synchronization features between different brain regions, etc. The computer system samples from this set of target features.

[0126] Suppose frequency features and signal peak-valley morphological features are obtained by sampling. For frequency features, such as the energy ratios of different frequency bands mentioned above; for signal peak-valley morphological features, indicators such as the height of the peak, the depth of the valley, and the spacing between the peak and the valley can be used to describe. For example, the relative value of the peak height is 1.2 (assuming a certain standard height is 1), the relative value of the valley depth is 0.8, and the relative value of the peak-valley spacing is 1.1 (the relative values here are all relative to a certain standard scale).

[0127] Then these two target features are combined to obtain the first data item sequence. This first data item sequence can be represented as an array containing multiple elements, such as [relative value of α-band energy is 0.3, relative value of γ-band energy is 1.5, relative value of peak height is 1.2, relative value of valley depth is 0.8, relative value of peak-valley spacing is 1.1]. This sequence combines different target features together, providing a richer information source for the subsequent construction of the knowledge template library. This combination method can integrate the information of multiple target features, which helps to more comprehensively describe the feature patterns in the electroencephalogram regulation monitoring data. Especially when dealing with complex electroencephalogram signals, the combination of multiple features can improve the representation ability of target features (such as epilepsy seizure-related features).

[0128] In step S130, the computer system samples a subsequence from the first data item sequence and adds the subsequence to the original target-free knowledge template to obtain an enhanced template.

[0129] Taking the previously constructed first data item sequence [relative value of α-band energy is 0.3, relative value of γ-band energy is 1.5, relative value of peak height is 1.2, relative value of valley depth is 0.8, relative value of peak-valley spacing is 1.1] as an example, the computer system samples a subsequence from it. Suppose the sampled subsequence is [relative value of peak height is 1.2, relative value of valley depth is 0.8].

[0130] The original target-free knowledge template is obtained from the electroencephalogram monitoring data of healthy people or patients in a non-epileptic seizure state, such as the feature vector [relative value of α-band energy is 0.9, relative value of γ-band energy is 1.1] mentioned above. The computer system adds the sampled subsequence [relative value of peak height is 1.2, relative value of valley depth is 0.8] to this original target-free knowledge template.

[0131] The obtained enhanced template is [relative α - band energy value is 0.9, relative γ - band energy value is 1.1, relative peak height value is 1.2, relative trough depth value is 0.8]. This enhanced template incorporates the subsequence features sampled from the first data item sequence on the basis of the original template without target knowledge, thus enriching the content of the template. The purpose of this operation is to change the feature representation of the original template without target knowledge by introducing subsequences that may be related to the target features, enabling the template library to contain more diverse feature combinations, which helps improve the discrimination ability for different types of EEG regulation monitoring data. Especially when distinguishing data with target features (such as epilepsy - related features) from those without target features, this enhanced template can provide a more detailed basis for discrimination.

[0132] In step S140, if the subsequence is a target feature in the target feature set, the computer system takes the enhanced template as the enhanced template with target knowledge.

[0133] Taking the previous example, the subsequence is [relative peak height value is 1.2, relative trough depth value is 0.8]. Assuming that the peak - trough morphological features represented by this subsequence are target features in the target feature set (for epilepsy detection). Then the previously obtained enhanced template [relative α - band energy value is 0.9, relative γ - band energy value is 1.1, relative peak height value is 1.2, relative trough depth value is 0.8] will be taken as the enhanced template with target knowledge.

[0134] This means that this template now has information related to the target features and can participate in the optimization process of the monitoring data feature - mining network as a sample with target features in subsequent operations (such as network calibration, etc.). For example, in a neural - network - based EEG regulation monitoring data feature - mining model, the data in this enhanced template with target knowledge can be input into the network as positive examples to help the network learn the patterns related to the target features (epilepsy - related features), thereby improving the network's recognition ability for EEG regulation monitoring data with target features.

[0135] In step S150, if the subsequence is not a target feature in the target feature set, the computer system takes the enhanced template as the enhanced template without target knowledge.

[0136] Suppose in another case, the subsequence sampled from the first data item sequence is not a target feature in the target feature set. For example, the subsequence is a feature related to EEG signal noise (and the noise feature does not belong to the target feature set for epilepsy detection), then the enhanced template containing this subsequence is still regarded as the enhanced template without target knowledge.

[0137] Even though this enhanced template adds other features on the basis of the original non-target knowledge template, since the added subsequences are not target features, it is still used to represent EEG regulation monitoring data that does not have target features. In the neural network model, the data in this enhanced non-target knowledge template can be input into the network as counterexamples to help the network learn patterns related to non-target features, thereby improving the network's ability to distinguish EEG regulation monitoring data with and without target features.

[0138] In step S160, the computer system obtains the knowledge template library of the monitoring data feature mining network based on the original template library, the enhanced target knowledge template, and the enhanced non-target knowledge template.

[0139] The original template library contains target and non-target knowledge templates initially obtained from different sources (such as epilepsy patients and healthy people), and these templates provide a basic representation of EEG regulation monitoring data features. The enhanced target knowledge templates are obtained by adding subsequences related to target features on the basis of the original template library, which enriches the content of the knowledge templates with target features. The enhanced non-target knowledge templates are obtained by adding subsequences related to non-target features on the basis of the original non-target knowledge template, further refining the knowledge templates without target features.

[0140] The computer system integrates all the templates in the original template library, the enhanced target knowledge templates, and the enhanced non-target knowledge templates together to form the knowledge template library of the monitoring data feature mining network. This knowledge template library contains a rich and diverse representation of EEG regulation monitoring data features, including both the basic features obtained from the original data and the features enhanced by adding subsequences.

[0141] In a neural network model based on deep learning, the template data in this knowledge template library will be used as important inputs for operations such as network tuning. For example, when training a neural network for epilepsy detection in EEG signals, the target knowledge template data (including the original and enhanced ones) in the knowledge template library can be used as positive examples, and the non-target knowledge template data (including the original and enhanced ones) can be used as counterexamples. The network adjusts parameters such as weights in the network by continuously learning the feature patterns in these positive and counterexamples, thereby improving the ability to mine target features (epilepsy seizure-related features) in EEG regulation monitoring data. This knowledge template library provides a comprehensive feature information basis for the construction and optimization of the entire monitoring data feature mining network, ensuring that the network can accurately identify and mine target features in EEG regulation monitoring data.

[0142] As an implementation, the maximum loading size of the monitoring data feature mining network is M, where M≥1; based on this, after obtaining the knowledge template library of the monitoring data feature mining network, the method may further include:

[0143] Step S101: For the electroencephalogram (EEG) regulation monitoring data of the knowledge template to be loaded into the monitoring data feature mining network, if the size of the EEG regulation monitoring data of the knowledge template is greater than the maximum loading size M, filter out the non-target data items in the EEG regulation monitoring data of the knowledge template to obtain the simplified data of the knowledge template;

[0144] Step S102: If the size of the simplified data of the knowledge template is not greater than the maximum loading size M, use the simplified data of the knowledge template as the loading data of the monitoring data feature mining network;

[0145] Step S103: If the size of the simplified data of the knowledge template is greater than the maximum loading size M, then extract the EEG regulation monitoring data with a size not greater than the maximum loading size M from the simplified data of the knowledge template as the loading data of the monitoring data feature mining network.

[0146] In step S101, for the EEG regulation monitoring data of the knowledge template to be loaded into the monitoring data feature mining network, if the size of the EEG regulation monitoring data of the knowledge template is greater than the maximum loading size M, filter out the non-target data items in the EEG regulation monitoring data of the knowledge template to obtain the simplified data of the knowledge template.

[0147] In the task of EEG regulation monitoring data feature mining based on deep learning, assume that the maximum loading size M of the monitoring data feature mining network is 100 (here 100 is only for convenient example, representing a data volume limit). Taking EEG signal analysis as an example, the EEG regulation monitoring data in the knowledge template may contain data with multiple features, and these data can be represented by vectors or arrays.

[0148] For example, the EEG regulation monitoring data of a knowledge template contains a long feature vector, which is represented as [α-band energy value 1, α-band energy value 2, α-band energy value 3, β-band energy value 1, β-band energy value 2, signal peak height 1, signal peak height 2, signal trough depth 1, signal trough depth 2, signal synchronization index of different brain regions 1, signal synchronization index of different brain regions 2,...], and the total number of its data items is greater than 100.

[0149] Assume that the target features are the α-band energy feature and the signal peak height feature related to epilepsy detection. The computer system needs to determine which are non-target data items. In this example, β-band energy values, signal trough depths, and signal synchronization indexes of different brain regions, etc. may be regarded as non-target data items.

[0150] The computer system identifies these non-target data items through specific algorithms or rules. For example, it can be determined based on the definition of the target feature when the knowledge template library was previously constructed, and the correlation analysis between the feature and the target feature. Then, the system filters out these non-target data items and obtains simplified data, which may be simplified to [α band energy value 1, α band energy value 2, α band energy value 3, signal peak height 1, signal peak height 2].

[0151] The purpose of this filtering operation is to remove non-target data items that may not be important for target feature mining when the amount of EEG control monitoring data in the knowledge template is too large, thereby reducing the data size so that it can adapt to the maximum loading size limit of the monitoring data feature mining network, while retaining key information related to the target feature as much as possible, so that these data can still be effectively used in subsequent operations (such as network adjustment, etc.).

[0152] In step S102, if the size of the simplified data of the knowledge template is not greater than the maximum loading size M, the simplified data of the knowledge template is used as loading data for the monitoring data feature mining network.

[0153] Continuing with the above example, if the simplified data obtained after filtering in step S101, such as [α band energy value 1, α band energy value 2, α band energy value 3, signal peak height 1, signal peak height 2], the total number of data items is less than or equal to 100 (i.e., the maximum loading size M).

[0154] In the context of deep learning-based neural network models used for feature mining of EEG regulation monitoring data, this simplified data can be directly used as the loading data of the monitoring data feature mining network. For example, in a multi-layer perceptron (MLP) neural network, these simplified data can be used as input data for the input layer. The number of neurons in the input layer can be set according to the dimension of the simplified data, and each neuron receives the value of a data item. Then, the network can process these input data according to the pre-set weights and activation functions, and gradually perform feature mining and classification in the hidden layer and output layer (if it is a classification task, such as distinguishing EEG signals in epileptic seizure and non-seizure states). This method ensures that data related to the target features can be effectively input into the monitoring data feature mining network while meeting the network loading size limit, providing suitable data input for the normal operation of the network and accurate feature mining.

[0155] In step S103, if the size of the simplified data of the knowledge template is larger than the maximum loading size M, the computer system extracts EEG control monitoring data with a size not larger than the maximum loading size M from the simplified data of the knowledge template as loading data for the monitoring data feature mining network.

[0156] Suppose in another case, the simplified data after being filtered by step S101 is still large. For example, the simplified data is [alpha band energy value 1, alpha band energy value 2, alpha band energy value 3, alpha band energy value 4, alpha band energy value 5, signal peak height 1, signal peak height 2, signal peak height 3, signal peak height 4, signal peak height 5,...], and the total number of its data items is still greater than 100 (the maximum loading size M). The computer system needs to further extract some data as the loading data. There can be various extraction methods. For example, it can be extracted in a certain order (such as from front to back) or according to the importance of the data (if the importance of the data items can be pre-evaluated).

[0157] Suppose it is extracted in the order from front to back, and [alpha band energy value 1, alpha band energy value 2, alpha band energy value 3, alpha band energy value 4, alpha band energy value 5, signal peak height 1, signal peak height 2, signal peak height 3, signal peak height 4] is extracted. The size of this part of the extracted data is not greater than 100, then this part of the data can be used as the loading data for the monitoring data feature mining network. In the neural network model, this part of the extracted data is used as the input of the input layer, and the network will perform feature mining operations based on these input data. Although this is an incomplete use of the data (due to data size limitations), through a reasonable extraction method, it can still provide information related to the target features for the network to a certain extent, enabling the network to perform preliminary feature mining. Although there may be certain limitations due to the incompleteness of the data, in practical applications, this is an effective coping strategy under the data size limitations.

[0158] Through steps S101 - S103, the computer system can reasonably process the electroencephalogram regulation monitoring data in the knowledge template according to the maximum loading size M of the monitoring data feature mining network, ensure that appropriate data can be loaded into the network, thereby providing effective data support for subsequent network calibration, feature mining, etc. operations, and also adapting to the hardware resource limitations of the network (such as limitations in memory, etc., by limiting the loading data size to avoid problems such as memory overflow).

[0159] As an implementation manner, step S101, filtering out the non-target data items in the electroencephalogram regulation monitoring data of the knowledge template to obtain the simplified data of the knowledge template, may include:

[0160] Step S1011: Traverse each data item in the electroencephalogram regulation monitoring data of the knowledge template to obtain the identifier of the data item in the data dictionary; among them, different data items in the data dictionary correspond to different identifiers;

[0161] Step S1012: Filter out the data items in the EEG regulation monitoring data of the knowledge template whose identifiers do not belong to the target identifier range to obtain the simplified data of the knowledge template; wherein, the target identifier range is determined based on the identifiers of the data items of the target feature.

[0162] In step S101 of the computer system, when the size of the EEG regulation monitoring data of the knowledge template is greater than the maximum loading size M, it is necessary to filter out the non-target data items in the EEG regulation monitoring data of the knowledge template to obtain the simplified data of the knowledge template, and this process is implemented through sub-steps S1011 - S1012.

[0163] In step S1011, the computer system traverses each data item in the EEG regulation monitoring data of the knowledge template to obtain the identifier of the data item in the data dictionary. Among them, different data items in the data dictionary correspond to different identifiers.

[0164] In the context of EEG regulation monitoring data processing based on deep learning, the EEG regulation monitoring data of the knowledge template is a data set containing various information. For example, taking the analysis of EEG signals for the diagnosis of a certain brain disease as an example, the EEG regulation monitoring data in the knowledge template may include the energy values of EEG signals in different frequency bands (such as α band, β band, γ band, etc.), the peak and trough characteristics of EEG signals (such as peak height, trough depth, peak - to - peak distance, etc.), and the signal synchronization characteristics between different brain regions, etc.

[0165] Assume that the EEG regulation monitoring data is represented in the form of an array. A simple example may be [α - band energy value is 0.3, β - band energy value is 0.5, peak height is 1.2, trough depth is 0.8, signal synchronization index between frontal lobe and temporal lobe brain regions is 0.6].

[0166] The data dictionary is a reference for defining and encoding all possible data items. In this data dictionary, each data item has a unique identifier. For example, the identifier corresponding to the α - band energy value may be "001", the identifier corresponding to the β - band energy value is "002", the identifier corresponding to the peak height is "010", the identifier corresponding to the trough depth is "011", and the identifier corresponding to the signal synchronization index between frontal lobe and temporal lobe brain regions is "100".

[0167] The computer system traverses (i.e., wanders) each data item in the EEG regulation monitoring data in the knowledge template. Taking the array mentioned just now as an example, the system first processes the alpha band energy value of 0.3, and obtains its corresponding identifier "001" by looking up the data dictionary; then processes the beta band energy value of 0.5 and obtains its identifier "002"; then processes the peak height of 1.2 and obtains the identifier "010"; and so on until all data items are processed. The purpose of this wandering operation is to establish a connection between each data item and the data dictionary, so that the data items can be classified and operated on through the identifiers. In subsequent steps, these identifiers will become an important basis for judging whether a data item is a target data item. From the perspective of a machine learning model, when a neural network is used for feature mining of EEG regulation monitoring data, the data dictionary and identifiers help to standardize and classify the input data, enabling the network to better understand and process this data. For example, in the input layer, different data items can be assigned to different neurons or neuron groups for processing according to the identifiers, improving the processing efficiency and accuracy of the network for the data.

[0168] In step S1012, the computer system filters out the data items in the EEG regulation monitoring data of the knowledge template whose identifiers do not belong to the target identifier range, and obtains the simplified data of the knowledge template. Among them, the target identifier range is a range determined based on the identifiers of each data item having the target feature.

[0169] Continuing with the above example, assume that the target features are the two data items of the alpha band energy value and the peak height related to brain disease diagnosis. According to the previous steps (such as the process of determining the target identifier range), the identifier corresponding to the alpha band energy value is "001", and its target identifier range is assumed to be "001~005" (possibly considering factors such as small fluctuations in the alpha band energy value between different individuals or measurement errors, etc.); the identifier corresponding to the peak height is "010", and its target identifier range is assumed to be "010~015".

[0170] The computer system checks the identifiers of each data item obtained previously. For the β-band energy value, its identifier is "002", which does not belong to the target identifier range "001 - 005" of the α-band energy value, nor does it belong to the target identifier range "010 - 015" of the peak height. Therefore, this data item of the β-band energy value will be filtered out. For the trough depth, its identifier is "011", which also does not belong to the target identifier range and will be filtered out. For the frontal-temporal brain region signal synchronization index, its identifier is "100", which does not belong to any target identifier range and will also be filtered out. After filtering, the data composed of the remaining two data items, namely the α-band energy value and the peak height, is the simplified data of the knowledge template, that is, simplified to [α-band energy value is 0.3, peak height is 1.2].

[0171] The significance of this filtering operation is that when the size of the electroencephalogram regulation monitoring data in the knowledge template is too large, those data items that are not relevant to the target features are removed, thereby reducing the data scale. In practical applications, this helps to more effectively process the data related to the target features under limited resources (such as the maximum loading size limit of the network, computing resource limit, etc.). For example, when a neural network model based on deep learning is used for feature mining of electroencephalogram regulation monitoring data, if the input data volume is too large, it may lead to problems such as too long network training time and excessive memory occupation. Through this filtering operation, only the data items related to the target features are retained as the input, which can improve the training efficiency of the network, reduce unnecessary data interference at the same time, enable the network to focus more on mining the information related to the target features, and improve the accuracy of tasks such as brain disease diagnosis.

[0172] As an implementation manner, the method further includes a process for determining the target identifier range, including:

[0173] Step S101A: Obtain the identifiers of each data item with the target feature in the data dictionary;

[0174] Step S101B: Based on the identifiers and the extended range corresponding to each data item with the target feature, determine the identifier range corresponding to each data item with the target feature, where the extended range represents the numerical range of the identifier;

[0175] Step S101C: Based on the identifier ranges corresponding to each data item with the target feature, determine the target identifier range.

[0176] In step S101A, the computer system obtains the identifiers of each data item of the target feature in the data dictionary. In the embodiments of the present invention, the data dictionary is a reference for encoding and identifying various electroencephalogram regulation monitoring data features. Suppose we take electroencephalogram signal analysis for detecting specific brain diseases as an example. The electroencephalogram signal has multiple features, such as energy values in different frequency bands (alpha band, beta band, gamma band, etc.), peak-valley features of the signal (peak height, valley depth, peak spacing, etc.), and signal synchronization features between different brain regions. For example, the data dictionary may be in the form of a table, where each row corresponds to a data item, and each column contains different attributes of the data item, such as name, identifier, data type, value range, etc. For the data item of alpha band energy value, there may be a corresponding identifier in the data dictionary, suppose it is "001"; the identifier corresponding to the beta band energy value is "002"; the identifier corresponding to the peak height is "010", etc.

[0177] The computer system will traverse each data item in the target feature, and then look up and obtain their corresponding identifiers from the data dictionary. Suppose the target feature is the two data items of alpha band energy value and peak height related to a specific brain disease, then the computer system will obtain their identifiers "001" and "010" in the data dictionary. The purpose of this step is to provide basic elements for determining the target identifier range in the subsequent steps. By clarifying the identifiers of the target feature data items, these data related to the target feature can be accurately located and distinguished in the entire data system.

[0178] In step S101B, the computer system determines the identifier range corresponding to each data item of the target feature based on the identifiers and the extended range corresponding to each data item of the target feature, and the extended range represents the numerical range of the identifier.

[0179] Continuing with the above example, for the data item of alpha band energy value, its identifier in the data dictionary is "001". Suppose its extended range is determined based on the statistical analysis of a large amount of previous electroencephalogram monitoring data or based on specific medical knowledge. For example, in normal electroencephalogram signals, the fluctuation range of alpha band energy value is relatively stable, and the extended range of the corresponding identifier "001" may be defined as "001~005", and this range represents the possible change range of the identifier related to the alpha band energy value, which may be caused by minor differences between different individuals or measurement errors and other factors.

[0180] For the data item of peak height, its identifier is "010", and suppose its extended range is determined to be "010~015" according to the change law of peak height in different electroencephalogram states and factors such as measurement accuracy.

[0181] This way of determining the identifier range is based on an in-depth understanding of the target feature data items and the statistics and analysis of relevant data. For example, when a neural network model based on machine learning is used for feature mining of electroencephalogram (EEG) regulation monitoring data, these identifier ranges can help the network more accurately identify and process data related to the target features. If the network receives a data item with an identifier of "003", it can determine that this data item may be related to the alpha-band energy value according to the previously determined identifier range of the alpha-band energy value ("001 - 005"), and then perform corresponding processing.

[0182] In step S101C, the computer system determines the target identifier range based on the identifier ranges corresponding to each data item of the target feature.

[0183] In the previous steps, we have determined the identifier ranges corresponding to each target feature data item. For example, the identifier range corresponding to the alpha-band energy value is "001 - 005", and the identifier range corresponding to the peak height is "010 - 015". The computer system integrates these identifier ranges to determine the target identifier range.

[0184] In this example, the target identifier range may be the union of "001 - 005" and "010 - 015", that is, "001 - 005, 010 - 015". This target identifier range covers the identifier ranges of all data items related to the target features in the data dictionary.

[0185] This target identifier range is of great significance in subsequent operations. For example, when filtering out non-target data items in the EEG regulation monitoring data of the knowledge template in step S101, the computer system can determine whether to retain a data item by checking whether the identifier of the data item in the data dictionary is within the target identifier range. If the identifier of a data item is not within this target identifier range, for example, the identifier is "020", then this data item may be determined as a non-target data item and filtered out. This can ensure that only data related to the target features is retained during the processing of EEG regulation monitoring data, improving the efficiency and accuracy of data processing, providing a cleaner and more targeted data input for the neural network for feature mining of EEG regulation monitoring data based on deep learning, and helping to improve the network's ability to mine target features and distinguish different EEG states (such as normal and diseased states).

[0186] As an implementation, in step S400, after calibrating the pruned monitoring data feature mining network through the knowledge template library to obtain a converged monitoring data feature mining network, the method may further include: reducing the floating point of one or more weight parameters in the converged monitoring data feature mining network to obtain a converged monitoring data feature mining network with reduced floating point numbers.

[0187] When the computer system executes the method for mining features of electroencephalogram regulation monitoring data based on deep learning, after calibrating the pruned monitoring data feature mining network through the knowledge template library in step S400 to obtain a converged monitoring data feature mining network, it will also reduce the floating point of one or more weight parameters in the converged monitoring data feature mining network to obtain a converged monitoring data feature mining network with reduced floating point numbers.

[0188] In the context of using a neural network model based on deep learning for mining features of electroencephalogram regulation monitoring data, the monitoring data feature mining network is a complex structure, and the weight parameters therein determine the connection strength between neurons and the way data flows and transforms in the network. After the network is calibrated to reach a converged state, its weight parameters have certain values, which are usually represented in the form of floating point numbers.

[0189] For example, assume that the converged monitoring data feature mining network is a multi-layer perceptron (MLP), and the connection weight parameters between neurons in one hidden layer and the next layer are floating point numbers such as [0.123456, -0.234567, 0.345678]. The computer system performs an operation to reduce the floating point of these weight parameters.

[0190] This operation of reducing the floating point has various meanings and functions. First, from the perspective of data storage and computing resources, reducing the floating point of weight parameters can reduce the storage space requirement of data. For example, reducing the floating point number from double precision (64 bits) to single precision (32 bits). Taking a simple weight parameter 0.123456789 as an example, if stored in double precision, it requires 64 bits of storage space; while if reduced to single precision, it only requires 32 bits of storage space. In large-scale neural networks, especially those dealing with complex data such as electroencephalogram regulation monitoring, there are a large number of weight parameters. By reducing the number of floating point bits, the storage requirement of the entire network can be significantly reduced. This is very important for resource-constrained devices (such as some portable electroencephalogram monitoring devices), enabling the storage and operation of the network model within limited storage space.

[0191] Secondly, from the perspective of computational efficiency, lower precision floating point calculations are usually faster than high precision floating point calculations. In the forward and backward propagation of the neural network, a large number of multiplication and addition operations are required on the weight parameters. Take a simple neural network layer as an example. Assume that the input vector is [x1, x2, x3] and the weight matrix is where w ij is the weight parameter, and the output vector y is calculated as If the floating point of the weight parameter is reduced, for example from double precision to single precision, the computing unit (such as CPU or GPU) can complete the calculation faster when performing these multiplication and addition operations, thereby improving the operation speed of the entire network.

[0192] However, reducing the floating point of the weight parameter is not without cost. Due to the reduction of numerical precision, certain errors may be introduced. For example, in the process of feature mining of EEG signals, if the precision of the weight parameter is reduced too much, the network's ability to recognize some subtle changes in EEG signal features may be reduced. Suppose a weak feature change in the EEG signal that is related to a specific disease can be accurately identified in a network with high-precision weight parameters, but in a network with reduced floating point, due to the loss of precision of the weight parameter, this weak feature may be ignored or misjudged.

[0193] In order to balance the relationship between this loss of precision and resource saving, the computer system needs to determine the floating point reduction strategy according to the specific application scenarios and requirements. For example, if the main purpose of EEG control monitoring is to quickly classify obvious EEG signal features (such as distinguishing normal EEG signals from EEG signals during epileptic seizures, and the EEG signal features during epileptic seizures change more obviously), then the number of floating point bits can be appropriately reduced in exchange for faster computing speed and smaller storage requirements. However, if some subtle EEG signal features are to be accurately analyzed (such as studying the subtle differences in EEG signals under different cognitive states), it may be necessary to carefully reduce the number of floating point bits, or use some special quantization techniques to minimize the loss of precision.

[0194] In actual operation, computer systems can use a variety of methods to reduce the floating point of weight parameters. A common method is quantization operation. For example, the floating point number is mapped to a smaller numerical range according to certain rules and represented by an integer. Assuming that the value range of the weight parameter is [-1,1], this range can be divided into several intervals, such as [-1,-0.5), [-0.5,0), [0,0.5), [0.5,1], and then the floating point numbers in these intervals are represented by integers ~3, ~2, ~1, 0, 1, 2, 3, etc. In this way, the original floating point weight parameter is converted to integer representation, which greatly reduces the storage and calculation complexity of the data.

[0195] In summary, the computer system's operation of reducing the floating point of the weight parameters in the monitoring data feature mining network for the convergence state is a means of balancing resource utilization and model performance. Through a reasonable floating point reduction strategy, it is possible to improve the storage efficiency and computing efficiency of the network on the premise of meeting the requirements of the EEG regulation monitoring task, providing a more optimized solution for actual EEG regulation monitoring applications.

[0196] As an implementation, in step S500, based on the data feature vector of the EEG regulation monitoring data to be mined, determining whether the EEG regulation monitoring data to be mined has a target feature may include:

[0197] Step S510: Obtain a target feature set; the target feature set includes multiple target data item sequences, and a data item attribute sequence corresponding to each target data item sequence;

[0198] Step S520: Search for a matching target data item sequence that matches the data feature vector among the multiple target data item sequences, and extract the target data item attribute sequence corresponding to the matching target data item sequence;

[0199] Step S530: Identify the feature category that matches the data feature vector based on the matching target data item sequence and the target data item attribute sequence;

[0200] Step S540: Obtain the monitoring data attribute corresponding to the feature category as the mined monitoring data attribute;

[0201] Step S550: If the target data item sequence contains the mined monitoring data attribute, determine that the EEG regulation monitoring data to be mined has a target feature, and use the matching target data item sequence containing the mined monitoring data attribute as the target mined monitoring data item.

[0202] In step S500, the computer system determines whether the EEG regulation monitoring data to be mined has a target feature based on the data feature vector of the EEG regulation monitoring data to be mined, and this process is implemented through sub-steps S510 to S550.

[0203] In step S510, the computer system obtains a target feature set. The target feature set includes multiple target data item sequences, and a data item attribute sequence corresponding to each target data item sequence.

[0204] In the feature mining of EEG regulation monitoring data, taking the detection of features related to epileptic seizures in EEG signals as an example. The target feature set may contain various EEG signal feature patterns related to epileptic seizures.

[0205] Suppose one of the target data item sequences is a sequence regarding the energy change of EEG frequency bands. This sequence can be expressed as [relative value of α-band energy before seizure, relative value of α-band energy during seizure, relative value of α-band energy after seizure], where the relative value is relative to the energy value in the normal state. For example, in the normal state, the relative value of α-band energy is 1, it may become 0.9 before seizure, may decrease to 0.5 during seizure, and may recover to 0.8 after seizure. The corresponding data item attribute sequence for this target data item sequence may be [threshold range of relative energy value, threshold range of relative energy value, threshold range of relative energy value], such as [before seizure: 0.8 - 1.1, during seizure: 0 - 0.6, after seizure: 0.7 - 1]. These threshold ranges are obtained based on the statistical analysis of EEG signals of a large number of epilepsy patients and are used to define whether the relative value of α-band energy in different stages conforms to the characteristic pattern related to epileptic seizures. Suppose another target data item sequence is regarding the signal synchrony between different brain regions, such as [signal synchrony index between frontal lobe and temporal lobe before seizure, signal synchrony index between frontal lobe and temporal lobe during seizure, signal synchrony index between frontal lobe and temporal lobe after seizure]. The data item attribute sequence may be [threshold range of synchrony index, threshold range of synchrony index, threshold range of synchrony index], such as [before seizure: 0.6 - 0.9, during seizure: 0.2 - 0.5, after seizure: 0.6 - 0.9]. Here, the signal synchrony index can be a value obtained by calculating the correlation between EEG signals in the frontal lobe and temporal lobe brain regions and is used to measure the degree of signal synchrony between the two brain regions. These target data item sequences and their corresponding attribute sequences together constitute the target feature set, providing a comprehensive reference standard for subsequent judgment on whether the EEG regulation monitoring data to be mined has the target features.

[0206] In step S520, the computer system searches for the matching target data item sequence that matches the data feature vector among multiple target data item sequences, and extracts the corresponding target data item attribute sequence for the matching target data item sequence. Suppose in the previous step, the computer system obtains the data feature vector of the EEG regulation monitoring data to be mined as [0.85 (relative value of α-band energy before seizure), 0.4 (relative value of α-band energy during seizure), 0.75 (relative value of α-band energy after seizure), 0.7 (signal synchrony index between frontal lobe and temporal lobe before seizure), 0.3 (signal synchrony index between frontal lobe and temporal lobe during seizure), 0.7 (signal synchrony index between frontal lobe and temporal lobe after seizure)] through the representation information extraction component of the monitoring data feature mining network in the convergence state.

[0207] The computer system will compare this data feature vector with each target data item sequence in the target feature set. For the target data item sequence regarding the energy change of the EEG signal frequency band [relative value of α-band energy before seizure, relative value of α-band energy during seizure, relative value of α-band energy after seizure], the corresponding values in the data feature vector are within their corresponding threshold ranges (before seizure: 0.8 - 1.1, during seizure: 0 - 0.6, after seizure: 0.7 - 1).

[0208] For the target data item sequence regarding the signal synchrony between different brain regions [signal synchrony index between frontal lobe and temporal lobe before seizure, signal synchrony index between frontal lobe and temporal lobe during seizure, signal synchrony index between frontal lobe and temporal lobe after seizure], the corresponding values in the data feature vector are also within their corresponding threshold ranges (before seizure: 0.6 - 0.9, during seizure: 0.2 - 0.5, after seizure: 0.6 - 0.9).

[0209] Therefore, both of these two target data item sequences can be regarded as matching target data item sequences that match the data feature vector. Then, the computer system will extract the target data item attribute sequences corresponding to these two matching target data item sequences. For the matching target data item sequence of the frequency band energy change, the corresponding attribute sequence is [before seizure: 0.8 - 1.1, during seizure: 0 - 0.6, after seizure: 0.7 - 1]; for the matching target data item sequence of the signal synchrony, the corresponding attribute sequence is [before seizure: 0.6 - 0.9, during seizure: 0.2 - 0.5, after seizure: 0.6 - 0.9].

[0210] In step S530, the computer system identifies the feature category that matches the data feature vector based on the matching target data item sequence and the target data item attribute sequence.

[0211] Continuing with the above example, the computer system has determined two matching target data item sequences and their corresponding attribute sequences. For the matching target data item sequence of the frequency band energy change and its attribute sequence, it represents a type of EEG signal energy feature category related to epileptic seizures. The computer system can determine that this data feature vector conforms to the feature pattern related to epileptic seizures in terms of the frequency band energy change by comparing the values in the data feature vector with the threshold ranges in the attribute sequence.

[0212] For the matching target data item sequence of the signal synchrony and its attribute sequence, it represents a type of feature category of signal synchrony between brain regions related to epileptic seizures. Similarly, by comparing the values in the data feature vector with the threshold ranges in the attribute sequence, it can be determined that this data feature vector also conforms to the feature pattern related to epileptic seizures in terms of signal synchrony.

[0213] Therefore, based on these sequences of matching target data items and sequences of target data item attributes, the computer system identifies that the feature categories matching the data feature vectors are the frequency band energy change features related to epileptic seizures and the brain region signal synchronization features. This identification helps to associate the information in the data feature vectors with specific electroencephalogram (EEG) signal feature patterns, providing a more detailed basis for subsequent judgment on whether the EEG regulation monitoring data to be mined has the target features.

[0214] In step S540, the computer system obtains the monitoring data attributes corresponding to the feature categories as the attributes of the mined monitoring data.

[0215] For the feature category of frequency band energy change related to epileptic seizures identified previously, the corresponding monitoring data attribute may be the "epileptic seizure risk assessment index". This index can be a value comprehensively calculated based on the degree of frequency band energy change and its manifestations at different stages of the seizure. For example, if the relative value of the α-band energy is lower during the seizure and the energy change before and after the seizure is more obvious, this "epileptic seizure risk assessment index" may be higher.

[0216] For the feature category of brain region signal synchronization related to epileptic seizures, the corresponding monitoring data attribute may be the "brain functional connectivity stability index". This index reflects the impact of the change in signal synchronization between the frontal lobe and temporal lobe brain regions on the overall brain functional connectivity stability during epileptic seizures. For example, if the signal synchronization is lower during the seizure, this "brain functional connectivity stability index" may be lower, indicating that the brain functional connectivity is significantly affected during epileptic seizures.

[0217] These attributes of the mined monitoring data are key attributes further refined from the feature categories and related to EEG regulation monitoring. They can provide a more direct basis for judging whether the EEG regulation monitoring data to be mined has the target features.

[0218] In step S550, if the sequence of target data items contains the attributes of the mined monitoring data, the computer system determines that the EEG regulation monitoring data to be mined has the target features and uses the sequence of matching target data items containing the attributes of the mined monitoring data as the target mined monitoring data items.

[0219] Taking the "epileptic seizure risk assessment index" as an example of the attribute of the mined monitoring data, since it is related to the feature category of frequency band energy change related to epileptic seizures, and the sequence of target data items corresponding to this feature category contains content related to the relative value of the α-band energy at different stages of epileptic seizures, it can be determined that the EEG regulation monitoring data to be mined has the target features related to epileptic seizures. Moreover, the sequence of matching target data items containing the "epileptic seizure risk assessment index" (i.e., the sequence of target data items regarding the frequency band energy change of EEG signals) is used as the target mined monitoring data items.

[0220] Similarly, for the mining and monitoring data attribute of "brain functional connectivity stability index", since it is related to the category of signal synchronization characteristics in brain regions related to epileptic seizures, and the sequence of target data items corresponding to this characteristic category contains content related to signal synchronization indexes at different stages of frontal lobe and temporal lobe seizures, it can also be determined that the EEG regulation monitoring data to be mined has target characteristics related to epileptic seizures. The matching target data item sequence containing the mining and monitoring data attribute of "brain functional connectivity stability index" (i.e., the target data item sequence regarding signal synchronization between different brain regions) is also used as the target mining and monitoring data item. Through such a series of steps, the computer system can accurately judge whether it has target characteristics based on the data feature vector of the EEG regulation monitoring data to be mined. This judgment is of great significance for applications such as disease diagnosis and treatment effect evaluation in EEG regulation monitoring, and can provide a strong basis for relevant medical decisions.

[0221] Figure 2 FIG. is a schematic diagram of the hardware entity of a computer system provided by an embodiment of the present invention, as Figure 2 shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002. Among them, the memory 1002 stores a computer program that can run on the processor 1001, and when the processor 1001 executes the program, it implements the steps in the method of any one of the above embodiments.

Claims

1. A method for mining features of EEG control monitoring data based on deep learning, characterized in that: The method comprises: Acquire a knowledge template library of the monitoring data feature mining network, wherein the knowledge template library includes one or more targeted knowledge templates and one or more non-target knowledge templates, wherein the targeted knowledge template represents the EEG regulation monitoring data with target features, and the non-target knowledge template represents the EEG regulation monitoring data without the target features; The monitoring data feature mining network is adjusted by using the knowledge template library to obtain an initially adjusted monitoring data feature mining network; Pruning the network configuration variables of the initially calibrated monitoring data feature mining network to obtain a pruned monitoring data feature mining network; wherein the memory usage of the pruned monitoring data feature mining network is less than the memory usage of the initially calibrated monitoring data feature mining network; The pruned monitoring data feature mining network is adjusted through the knowledge template library to obtain a monitoring data feature mining network in a converged state; The characterization information extraction component of the monitoring data feature mining network based on the convergence state obtains the data feature vector of the EEG regulation monitoring data to be mined, and determines whether the EEG regulation monitoring data to be mined has the target feature based on the data feature vector of the EEG regulation monitoring data to be mined.

2. The method according to claim 1, characterized in that The pruning of the network configuration variables of the initially adjusted monitoring data feature mining network to obtain the pruned monitoring data feature mining network includes: For the first network component in the initially calibrated monitoring data feature mining network, perform eigenvalue splitting on the weight tensor of the first network component to obtain a multiplication result of the first rotation tensor, the eigenvalue tensor and the second rotation tensor; Determine, among the eigenvalues ​​of the eigenvalue tensor, a target eigenvalue that meets the first requirement; Extracting configuration variables corresponding to the target eigenvalues ​​from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor, respectively, to obtain a pruned first rotation tensor, a pruned eigenvalue tensor, and a pruned second rotation tensor; Among them, the configuration variables of the first network component in the pruned monitoring data feature mining network are composed of the pruned first rotation tensor, the pruned eigenvalue tensor and the pruned second rotation tensor.

3. The method according to claim 2, characterized in that The step of determining, among the eigenvalues ​​of the eigenvalue tensor, a target eigenvalue that meets the first requirement comprises: Sort the eigenvalues ​​of the eigenvalue tensor in descending order to obtain an eigenvalue vector; Determine the first p eigenvalues ​​in the eigenvalue vector, and obtain p target eigenvalues, where p≥2.

4. The method according to claim 2, characterized in that: The weight tensor of the first network component includes i*j weight parameters, the first rotation tensor includes i*i weight parameters, the eigenvalue tensor includes i*j weight parameters, the second rotation tensor includes j*j weight parameters, and the number of target eigenvalues ​​is p, i≥2; j≥2; p≥2; The method further comprises: extracting configuration variables corresponding to the target eigenvalues ​​from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor respectively to obtain a pruned first rotation tensor, a pruned eigenvalue tensor, and a pruned second rotation tensor, including: Based on the target eigenvalue, obtaining p columns of weight parameters corresponding to the target eigenvalue in the first rotation tensor to obtain the pruned first rotation tensor, wherein the pruned first rotation tensor includes j*p weight parameters; Based on the target eigenvalue, obtaining the pruned eigenvalue tensor, wherein the pruned eigenvalue tensor includes p*p weight parameters; Based on the target eigenvalue, p rows of weight parameters corresponding to the target eigenvalue are obtained in the second rotation tensor to obtain the pruned second rotation tensor, wherein the pruned second rotation tensor includes p*j weight parameters; each configuration variable in the eigenvalue tensor corresponds to a column of weight parameters in the first rotation tensor and a row of weight parameters in the second rotation tensor.

5. The method according to claim 1, characterized in that The step of obtaining a knowledge template library of the monitoring data feature mining network includes: Acquire an original template library of the monitoring data feature mining network, wherein the original template library includes one or more original target knowledge templates and one or more original non-target knowledge templates; Sampling one or more target features from the target feature set to be mined, and combining the one or more target features to obtain a first data item sequence; Sampling a subsequence in the first data item sequence, and adding the subsequence to the original target-free knowledge template to obtain an enhanced template; wherein the subsequence is one data item or multiple uninterrupted data items in the first data item sequence; If the subsequence is a target feature in the target feature set, using the enhanced template as an enhanced target knowledge template; If the subsequence is not a target feature in the target feature set, using the enhanced template as an enhanced non-target knowledge template; The knowledge template library of the monitoring data feature mining network is obtained based on the original template library, the enhanced targeted knowledge template and the enhanced non-target knowledge template.

6. The method according to claim 1, characterized in that The maximum loading size of the monitoring data feature mining network is M, M≥1; After acquiring the knowledge template library of the monitoring data feature mining network, the method further includes: For the EEG control monitoring data of the knowledge template to be loaded into the monitoring data feature mining network, if the size of the EEG control monitoring data of the knowledge template is greater than the maximum loading size M, non-target data items in the EEG control monitoring data of the knowledge template are filtered out to obtain simplified data of the knowledge template; If the size of the simplified data of the knowledge template is not greater than the maximum loading size M, the simplified data of the knowledge template is used as loading data of the monitoring data feature mining network; If the size of the simplified data of the knowledge template is larger than the maximum loading size M, then from the simplified data of the knowledge template, EEG control monitoring data with a size not larger than the maximum loading size M is extracted as loading data for the monitoring data feature mining network.

7. The method according to claim 6, characterized in that The filtering out non-target data items in the EEG control monitoring data of the knowledge template to obtain simplified data of the knowledge template includes: Walk through each data item in the EEG control monitoring data of the knowledge template to obtain an identifier of the data item in a data dictionary; wherein different data items in the data dictionary correspond to different identifiers; In the EEG control monitoring data of the knowledge template, the data items whose identifiers do not belong to the target identifier range are filtered out to obtain simplified data of the knowledge template; wherein the target identifier range is a range determined based on the identifiers of each data item possessed by the target feature.

8. The method according to claim 7, characterized in that The method further comprises: Obtaining identifiers of various data items of the target feature in a data dictionary; Determine the range of identifiers corresponding to the data items of the target feature based on the identifiers and the extended ranges corresponding to the data items of the target feature, wherein the extended range represents the numerical range of the identifiers; The target identifier range is determined based on the identifier ranges corresponding to the respective data items of the target feature.

9. The method according to claim 1, characterized in that: After the pruned monitoring data feature mining network is adjusted by the knowledge template library to obtain a converged monitoring data feature mining network, the method further includes: The floating point of one or more weight parameters in the convergence state monitoring data feature mining network is reduced to obtain the convergence state monitoring data feature mining network after the floating point number is reduced.

10. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 9 are implemented.

Images