Implantable sensing signal classification method and system based on nerve regulation and control processing

Abstract

The invention provides an implantable sensing signal classification method and system based on neural regulation and control processing, and relates to the field of data processing, and the method comprises the steps: obtaining a second sensing signal feature vector for guiding the first sensing signal feature vector to carry out interpolation for a first sensing signal feature vector to be subjected to interpolation processing; and determining a first characteristic value of the sampling point x according to the second sensing signal characteristic vector. Performing center gradient filtering on the first characteristic value through the characteristic values of the plurality of first adjacent sampling points to determine the signal high-frequency component of the sampling point x and the generality between the sampling point x and each first adjacent sampling point, obtaining a target filtering result, performing reasoning to obtain a first generality score between the sampling point x and each first adjacent sampling point, and obtaining a second generality score between the sampling point x and each first adjacent sampling point; and obtaining a self-adaptive interpolation matrix corresponding to the sampling point x based on each first generality score so as to improve the precision of the self-adaptive interpolation matrix and improve the integrity and accuracy of the interpolated signal, so that a more accurate signal classification result is obtained.

Description

Technical Field

[0001] This application relates to the field of data processing, and more specifically, to an implantable sensor signal classification method and system based on neural modulation processing. Background Technology

[0002] In modern medicine and biomonitoring, implantable sensors are widely used to acquire various physiological information from within the human body, such as heart rate, blood pressure, and neural electrical activity. The sensing signals generated by these implantable sensors contain a wealth of information. However, due to the complex physiological environment within the human body and the inherent characteristics of the sensors themselves, the acquired sensing signals often suffer from noise interference and poor signal integrity. For example, electromagnetic interference within the body and signal attenuation by tissues can affect the quality of the sensing signals. Furthermore, to accurately extract useful information from these sensing signals, they need to be classified. However, traditional signal classification methods have limitations when processing implantable sensor signals. Traditional signal classification methods often struggle to fully utilize the various characteristic information within the signals, especially the high-frequency components. During signal acquisition, various interference factors may weaken or lose high-frequency components, and traditional methods do not effectively recover and utilize these high-frequency components during processing, thus affecting the accuracy of signal classification. For example, some methods based on simple feature extraction and classifiers only focus on the low-frequency or overall features of the signal, ignoring the key physiological information contained in the high-frequency components, thus failing to accurately distinguish different types of physiological signals or identify abnormal states in the signal.

[0003] Furthermore, existing signal interpolation methods also suffer from low interpolation accuracy and poor adaptability to signal characteristics when processing implanted sensor signals. Inaccurate interpolation can lead to signal distortion, further affecting subsequent classification results. For example, traditional interpolation methods may use a fixed interpolation matrix or a simple linear interpolation approach, failing to consider the complex relationships between different sampling points in the signal and the inherent characteristic variations of the signal itself. This results in the interpolated signal not accurately reflecting the true state of the original signal, thus affecting the accuracy of classification operations based on the interpolated signal. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for classifying implantable sensor signals based on neuromodulation processing. This is achieved as follows:

[0005] In a first aspect, this application provides a method for classifying implantable sensor signals based on neural modulation processing, comprising: acquiring a first sensor signal feature vector to be interpolated, and acquiring a second sensor signal feature vector; the first sensor signal feature vector and the second sensor signal feature vector are obtained by extracting representational information at different scales for the same implantable sensor signal, and the feature values ​​in the second sensor signal feature vector contain high-frequency components of the signal; determining multiple first adjacent sampling points of sampling point x and the feature value of each first adjacent sampling point based on the first sensor signal feature vector, and determining a first feature value of sampling point x based on the second sensor signal feature vector; x is a positive integer not greater than M. M is the number of sampling points in the feature vector of the second sensing signal; based on the feature values ​​of the plurality of first adjacent sampling points, the first feature value of the sampling point x is subjected to central gradient filtering to obtain a target filtering result; based on the target filtering result, a first commonality score between the sampling point x and each first adjacent sampling point is inferred; and based on each first commonality score obtained by inference, an adaptive interpolation matrix corresponding to the sampling point x is generated; the first sensing signal feature vector is interpolated through the adaptive interpolation matrix corresponding to the sampling point x to obtain an interpolated sensing signal feature vector; based on the interpolated sensing signal feature vector, classification is performed to obtain an implantable sensing signal classification result.

[0006] In a second aspect, this application provides a computer system comprising: one or more processors; a memory; and one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, they implement the method described above.

[0007] The beneficial effects of this application are as follows: This application obtains a second sensor signal feature vector that guides the interpolation of the first sensor signal feature vector for a first sensor signal feature vector to be interpolated, and determines a first feature value of sampling point x based on the second sensor signal feature vector. Because the feature value in the second sensor signal feature vector contains high-frequency components of the signal, and the second sensor signal feature vector and the first sensor signal feature vector are obtained by extracting representational information at different scales for the same implanted sensor signal, determining the first feature value of sampling point x based on the second sensor signal feature vector ensures that the first feature value contains rich high-frequency components of the signal. Then, by using the feature values ​​of multiple first adjacent sampling points of sampling point x determined based on the feature vector of the first sensing signal, a central gradient filter is applied to the first feature values ​​of the rich high-frequency signal components contained in sampling point x. This is to determine the high-frequency signal components of sampling point x and the commonalities between sampling point x and each first adjacent sampling point, thereby obtaining a target filtering result with high-frequency signal components and commonality information. Based on the target filtering result, the first commonality score between sampling point x and each first adjacent sampling point is obtained more accurately. Based on each first commonality score, the adaptive interpolation matrix corresponding to sampling point x is obtained to increase the accuracy of the adaptive interpolation matrix. At the same time, the adaptive interpolation matrix contains high-frequency signal components, so that the feature vector of the first sensing signal is interpolated based on the accurate adaptive interpolation matrix with high-frequency signal components. This completes the feature interpolation, improves the signal details, and increases the integrity and accuracy of the interpolated signal, so as to obtain a more accurate signal classification result. Attached Figure Description

[0008] Figure 1 This is a flowchart of an implantable sensor signal classification method based on neuromodulation processing provided in an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of this application. Detailed Implementation

[0010] In this embodiment of the application, the execution entity of the implantable sensor signal classification method based on neuromodulation processing is a computer system, including but not limited to servers, personal computers, laptops, tablets, and smartphones. Figure 1 As shown, the method includes:

[0011] Step S100: Obtain the first sensor signal feature vector to be interpolated and the second sensor signal feature vector; the first sensor signal feature vector and the second sensor signal feature vector are obtained by extracting characterization information at different scales for the same implanted sensor signal, and the feature values ​​in the second sensor signal feature vector contain high-frequency components of the signal.

[0012] In step S100 of this embodiment, the first sensing signal feature vector is obtained by extracting specific representational information from the implanted sensing signal. This representational information extraction is an operation that converts the original implanted sensing signal into a vector form that better reflects the signal characteristics. For example, assuming the implanted sensing signal is a physiological signal within a biological body (such as an electrocardiogram signal), a computer system may process it using a specific algorithm, converting continuous electrical signal data into a vector. Each element in this vector represents a characteristic of the original signal in a certain aspect; this is the first sensing signal feature vector.

[0013] The second sensor signal feature vector is also obtained from the implanted sensor signal, but unlike the first sensor signal feature vector, it is obtained by extracting representational information from the same signal at different scales. Furthermore, the feature values ​​in the second sensor signal feature vector contain high-frequency components of the signal, i.e., detailed information. For example, when processing electrocardiogram (ECG) signals, the first sensor signal feature vector may reflect more of the overall trend and other macroscopic features of the ECG signal, while the second sensor signal feature vector can capture the rapidly changing parts of the ECG signal, such as subtle fluctuations and other high-frequency details.

[0014] For the technical means of obtaining the feature vector of the first sensing signal in a computer system, signal processing-based algorithms can be employed. For example, the time-domain implanted sensing signal can be converted into a frequency-domain signal using the Discrete Fourier Transform (DFT), and then information from specific frequency bands can be selected and combined to form the feature vector of the first sensing signal. The formula for DFT is: Where X(k) is the result in the frequency domain (which can be used to construct feature vectors), x(n) is the input signal in the time domain (the original implanted sensing signal), N is the number of points in the signal, k is the index in the frequency domain, and j is the imaginary unit.

[0015] Obtaining the feature vector of the second sensing signal also requires corresponding technical means. One possible approach is to utilize wavelet transform. Wavelet transform can decompose a signal into components of different scales, and the computer system can select those components corresponding to the high-frequency part to construct the feature vector of the second sensing signal. The basic formula for wavelet transform is: Where WT(a,b) is the result of wavelet transform, x(t) is the original signal (implanted sensor signal), ψ is the wavelet function, and a and b are the scale and translation parameters, respectively. By adjusting a and b, different features of the signal can be obtained, thereby constructing a second sensor signal feature vector containing high-frequency components.

[0016] Step S200: Determine multiple first adjacent sampling points of sampling point x and the feature value of each first adjacent sampling point based on the first sensor signal feature vector, and determine the first feature value of sampling point x based on the second sensor signal feature vector; x is a positive integer not greater than M, and M is the number of sampling points in the second sensor signal feature vector.

[0017] In step S200 of this embodiment, the computer system determines multiple first adjacent sampling points of sampling point x and the feature value of each first adjacent sampling point based on the first sensing signal feature vector. Here, the sampling points are discrete data points that constitute the sensing signal feature vector. For example, if the first sensing signal feature vector represents the bioelectrical signal characteristics over a certain time period, then these sampling points are like representative points selected at certain intervals on this time axis. Assume the first sensing signal feature vector is V1 = [v1, v2, v3, ..., v...]. n ], where v i For example, if x = 3, the computer system might determine x - 1 = 2 and x + 1 = 4 as the first adjacent sampling points and obtain their corresponding feature values ​​v2 and v4. One technique for determining adjacent sampling points is an index-based search method. Based on a pre-defined adjacency relationship (such as sequential adjacency), adjacent sampling points are located in the feature vector using their index values.

[0018] Simultaneously, the computer system determines the first eigenvalue of sampling point x based on the second sensor signal feature vector. Since the second sensor signal feature vector contains high-frequency components of the signal, this first eigenvalue reflects richer details. For example, in the case of bioelectrical signals, this first eigenvalue may reflect the contribution of high-frequency details such as minute electrical fluctuations at a specific moment to the overall signal characteristics. To obtain the eigenvalue of a specific sampling point from the second sensor signal feature vector, the computer system can directly locate it using an index, assuming the second sensor signal feature vector is V2 = [u1, u2, u3, ..., u...]. m Once x is determined, u is obtained directly. x This serves as the first feature value of the sampling point x.

[0019] In this process, the computer system processes two different feature vectors from the sensor signals. To enable them to work together in subsequent processing, some adjustments to the data format or feature representation may be necessary. For example, if the feature values ​​of the two vectors represent different physical meanings or have significantly different numerical ranges, the computer system may perform normalization. A common normalization technique is min-max normalization, expressed by the formula: Where x is the original feature value, x min and x max These are the minimum and maximum values ​​of the eigenvalue in the vector, respectively, x. newThis is the result after normalization. This maps feature values ​​of different ranges to the [0,1] interval, which facilitates subsequent calculations.

[0020] Step S300: Based on the feature values ​​of multiple first adjacent sampling points, perform central gradient filtering on the first feature value of sampling point x to obtain the target filtering result.

[0021] In step S300, based on the feature values ​​of the multiple first adjacent sampling points determined in step S200, center gradient filtering is performed on the first feature value of sampling point x to obtain the target filtering result. Center gradient filtering is a filtering method based on local signal features, which aims to optimize the feature value of a sampling point by analyzing the relationship between the sampling point and its neighboring sampling points.

[0022] Taking a simple bio-implanted sensor signal as an example, suppose this sensor signal records the fluctuations of a certain physiological indicator in an organism. The first sensor signal feature vector reflects the macroscopic changes of the physiological indicator over a longer period of time, while the second sensor signal feature vector contains more detailed high-frequency component information. After determining the first feature value of sampling point x and the feature values ​​of its multiple first adjacent sampling points in step S200, these feature values ​​all characterize the characteristic state of the physiological indicator at different times.

[0023] When a computer system performs central gradient filtering, it considers the gradient relationship between the eigenvalue of each first adjacent sampling point and the first eigenvalue of sampling point x. From a mathematical perspective, for a discrete sensor signal feature vector, if we assume the first eigenvalue of sampling point x is V... x The feature value of its first adjacent sampling point is V x-i and V x+i (where i represents the relative distance to sampling point x), center gradient filtering can represent gradient information by calculating the difference between adjacent sampling points and sampling point x. For example, a simple way to calculate the gradient difference between adjacent sampling points is ΔV. x-i =V x -V x-i and ΔV x+i =V x -V x+i .

[0024] In practical center gradient filtering operations, computer systems may employ a weighted summation technique to comprehensively consider the influence of each adjacent sampling point. Assume there are n first adjacent sampling points, with corresponding weights w. i (i = 1, 2, ..., n), then the result F_1 after one central gradient filter can be calculated using the following formula: The weight w here iThe weights can be set based on the signal characteristics and empirical values. For example, the weights of adjacent sampling points closer to sampling point x can be set larger to reflect their more critical influence on the feature values ​​of sampling point x. The computer system performs this central gradient filtering multiple times, assuming w times (w≥2). Each filtering yields a central gradient filter value; for example, F_2 is obtained after the second filtering, F3 after the third filtering, and so on. These filter values ​​each contain feature optimization information of sampling point x under different filtering processes.

[0025] Finally, the computer system generates the target filtering result using these w central gradient filter values. This target filtering result integrates information from multiple filters, more comprehensively reflecting the characteristic state of sampling point x after considering the influence of its neighboring sampling points. It contains information from the original first feature value and also incorporates the feature influence of neighboring sampling points, thus providing a more accurate and representative data foundation for subsequent steps (such as inferring the commonality score between the sampling point and its neighboring sampling points). The central gradient filtering operation in step S300 effectively utilizes the information from neighboring sampling points to optimize the feature value of sampling point x, improving the accuracy and effectiveness of the entire implanted sensor signal processing.

[0026] Step S400: Based on the target filtering result, infer the first commonality score between the sampling point x and each first adjacent sampling point; and based on each first commonality score obtained by inference, generate the adaptive interpolation matrix corresponding to the sampling point x.

[0027] After obtaining the target filtering result in step S300, two key operations are performed in step S400 based on this result. First, the computer system infers a first commonality score between the sampling point x and each of its first adjacent sampling points. This first commonality score quantifies the similarity or commonality between the sampling point x and its neighboring sampling points after the previous processing steps. For example, assuming the implanted sensor signal is a certain electrical signal within a biological body, after the previous processing, the feature values ​​of these sampling points have incorporated multiple pieces of information. If a sampling point x and one of its neighboring sampling points are very similar in terms of the trend and magnitude of their filtered feature values, then their first commonality score will be relatively high.

[0028] To achieve this reasoning, the computer system can employ distance-based techniques. Assume the target filtering result is a vector F = [f1, f2, ..., f...]. w For a sampling point x and its first neighboring sampling point y, a certain distance d(x,y) between them can be calculated, such as the Euclidean distance, as shown in the formula. Where f x,i and f y,iThese are the i-th elements of sampling points x and y in the target filtered result vector, respectively. Then, the first commonality score is calculated based on this distance; a simple way to do this is S. x,y = 1 / (1+d(x,y)), so when the distance d(x,y) is small, the commonality score S x,y A higher commonality score indicates a greater degree of commonality between the two. Secondly, the computer system generates an adaptive interpolation matrix corresponding to sampling point x based on the first commonality scores obtained through inference. This adaptive interpolation matrix is ​​constructed based on the commonality relationship between sampling point x and its neighboring sampling points; it will be used for subsequent interpolation operations on the feature vector of the first sensing signal. For example, continuing with the bioelectric signal example above, if the first commonality score between sampling point x and a certain neighboring sampling point is high, then the weight of this neighboring sampling point in the adaptive interpolation matrix will be relatively large.

[0029] One possible technique for constructing adaptive interpolation matrices in computer systems is based on matrix decomposition and reconstruction. Assume the first commonality score set is S = {S...} x,1 ,S x,2 ,…,S x,n}, where n is the number of the first adjacent sampling points. First, this set is constructed into an initial matrix M0, and then singular value decomposition (SVD) is performed on M0, with the formula M0 = UΣV T Let U, Σ, and V be the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix, respectively. U, Σ, and V are adjusted according to certain rules (such as the relationship between the magnitude of the singular values ​​and the corresponding first commonality score), and then reconstructed to obtain the adaptive interpolation matrix M.

[0030] Step S400 further processes the target filtering results through the computer system to obtain the quantified value of the common relationship between sampling point x and adjacent sampling points (first common score), and constructs an adaptive interpolation matrix. This lays a solid foundation for the subsequent accurate interpolation of the first sensor signal feature vector and the final classification of implanted sensor signals.

[0031] Step S500: Interpolate the first sensing signal feature vector using the adaptive interpolation matrix corresponding to the sampling point x to obtain the interpolated sensing signal feature vector.

[0032] In this step, the first sensing signal feature vector is interpolated using the adaptive interpolation matrix corresponding to the sampling point x generated in step S400, thereby obtaining the interpolated sensing signal feature vector. Interpolation is a technique for estimating new data points between discrete data points; in this context, it is used to refine the signal information represented by the first sensing signal feature vector.

[0033] Taking implanted sensor signals within a living organism as an example, suppose the feature vector of the first sensor signal represents the macroscopic characteristics of the bioelectrical signal over a certain period of time. Due to the potential for discreteness in the sampling process, the signal information may be incomplete. Interpolation is used to supplement information between these discrete sampling points, enabling the signal feature vector to more comprehensively and accurately reflect the characteristics of the original bioelectrical signal.

[0034] In computer systems performing interpolation operations, the adaptive interpolation matrix plays a crucial role. This matrix is ​​constructed based on the first commonality score between the sample point x and each of its first neighboring sample points, determined in previous steps. It contains information about the contribution of each neighboring sample point to the interpolation of sample point x. For example, if a certain element in the adaptive interpolation matrix has a large value, it indicates that the corresponding neighboring sample point has a significant influence on the interpolation of sample point x.

[0035] One interpolation technique is the weighted summation method. Assume the feature vector of the first sensing signal is V = [v1, v2, ..., v...]. n Let A be the adaptive interpolation matrix corresponding to the sampling point x, with dimensions d×d (where d is a parameter determined based on the specific situation, related to the number of adjacent sampling points, etc.). For the sampling point x to be interpolated, the computer system first determines a set of adjacent sampling points related to it (the selection and determination of these adjacent sampling points have been completed based on certain rules in previous steps). Let the eigenvector corresponding to this set of adjacent sampling points be denoted as . (m is the number of adjacent sampling points). Then, the interpolation result is calculated by weighted summation using an adaptive interpolation matrix. The specific formula can be expressed as: in It is the feature value after interpolation of the sampling point x, a i This refers to the weight element in the adaptive interpolation matrix A corresponding to the i-th adjacent sampling point. Through this interpolation operation, the computer system performs similar processing on each sampling point in the first sensor signal feature vector that requires interpolation, ultimately obtaining the interpolated sensor signal feature vector. Compared to the original first sensor signal feature vector, this interpolated feature vector contains more information obtained through interpolation estimation, and can more completely and meticulously reflect the characteristics of the original implanted sensor signal, providing a richer and more accurate data foundation for subsequent implanted sensor signal classification operations.

[0036] Step S600: Classify based on the interpolated sensor signal feature vector to obtain the classification result of the implanted sensor signal.

[0037] In step S600 of this embodiment, the classification operation determines the category to which the implanted sensing signal belongs based on the information contained in the interpolated sensing signal feature vector. For example, suppose the implanted sensing signal comes from a human physiological monitoring device, such as an electrocardiogram (ECG) signal monitoring heart activity or an electroencephalogram (EEG) signal monitoring brain neural activity. Different physiological states correspond to signals with different characteristics. For example, the ECG signal of a normal heartbeat has specific waveform characteristics, while the ECG signal waveform during arrhythmia will show abnormal changes. Similarly, EEG signals in different brain activity states also have different frequency, amplitude, and other characteristics.

[0038] To classify the feature vectors of interpolated sensing signals, computer systems can employ various techniques. A common approach is based on machine learning classification algorithms, such as Support Vector Machines (SVM). The basic principle of SVM is to find an optimal hyperplane to separate data points of different classes. For the feature vectors of interpolated sensing signals X = [x1, x2, ..., x...], ... n (where x) i (These are elements in the feature vector). Assume there are two types of implantable sensor signals: y = +1 (representing one type of physiological state) and y = -1 (representing another type of physiological state).

[0039] SVM determines the parameters of the hyperplane by solving an optimization problem. Its goal is to minimize... (where w is the normal vector of the hyperplane), and simultaneously satisfies the constraint y i (w·x i +b)≥1(for all training samples (x) i ,y i (where b is the bias term). During the training phase, the computer system uses interpolated sensor signal feature vectors of known categories as training data to determine the parameters w and b of the SVM.

[0040] During the classification phase, for a new interpolated sensor signal feature vector x, the computer system calculates the value of w·x+b. If this value is greater than 0, it is classified as class y=+1; if it is less than 0, it is classified as class y=-1.

[0041] Besides SVM, computer systems can also use neural networks for classification. A neural network consists of layers of multiple neurons, such as an input layer, hidden layers, and an output layer. The interpolated sensor signal feature vector serves as the input to the input layer, undergoes nonlinear transformations by neurons in the hidden layer, and finally yields the classification result at the output layer. For example, in a simple three-layer neural network, the input layer has n neurons (corresponding to the n elements of the interpolated sensor signal feature vector), the hidden layer has m neurons, and the output layer has c neurons (where c represents the number of categories).

[0042] The computation from the input layer to the hidden layer can be represented as h = f(W1x + b1), where W1 is the weight matrix from the input layer to the hidden layer, b1 is the bias vector, and f is the activation function (such as the ReLU function f(x) = max(0,x). The computation from the hidden layer to the output layer is y = g(W2h + b2), where W2 is the weight matrix from the hidden layer to the output layer, b2 is the bias vector, and g is the activation function of the output layer (such as the softmax function used for multi-class classification problems). The computer system determines the category to which the interpolated sensor signal feature vector belongs based on the results of the output layer.

[0043] Through step S600, the computer system can accurately classify implanted sensor signals using information from the interpolated sensor signal feature vector. This helps to identify different physiological states or signal patterns in fields such as medical diagnosis and biosignal monitoring.

[0044] As one implementation method, the first sensing signal feature vector is generated by extracting high-level characterization information from the target sensing signal, and the high-level characterization information extraction is the process of extracting high-level characterization information.

[0045] In step S100, obtaining the feature vector of the second sensing signal may include:

[0046] Step S110: Acquire the target sensing signal;

[0047] Step S120: Extract low-level representation information from the target sensing signal to obtain the second sensing signal feature vector; wherein, low-level representation information extraction is the process of extracting low-level representation information, which contains high-frequency components of the signal.

[0048] The target sensing signal is, for example, human physiological signals acquired from implanted sensors, such as electrocardiogram (ECG) or electroencephalogram (EEG). These signals contain rich physiological information and are raw, unprocessed electrical signal data. In step S120, the computer system extracts low-level representation information from the target sensing signal to obtain a second sensing signal feature vector. This step involves multiple sub-operations.

[0049] Specifically, the computer system extracts features from the target sensing signal through a third filtering unit to obtain the initial sensing signal feature vector. The third filtering unit is a signal processing unit designed to extract representative feature information from the original target sensing signal. Taking electroencephalogram (EEG) signals as an example, EEG signals contain multiple frequency components, each related to different brain activity states. The third filtering unit may employ bandpass filtering techniques. For instance, it might set a specific frequency range (e.g., 8-13Hz, which is associated with alpha waves, typically related to brain activity in a relaxed state), allowing only signals within this band to pass while suppressing signals from other frequency bands. The formula for bandpass filtering is: y(t) = x(t) * h(t), where x(t) is the input target sensing signal (EEG signal), h(t) is the impulse response of the bandpass filter, and y(t) is the filtered signal, which is part of the initial sensing signal feature vector. By filtering different frequency bands, the original EEG signal can be decomposed into multiple feature signals of different frequency bands, which, when combined, constitute the initial sensing signal feature vector. Next, the computer system performs group standardization on the initial sensor signal feature vector to obtain a group-standardized sensor signal feature vector. Group standardization is a data standardization technique aimed at making data from different groups have similar distribution characteristics, facilitating subsequent processing. For example, suppose the initial sensor signal feature vector contains data from multiple groups, each corresponding to a different feature dimension or signal frequency band. The group standardization operation calculates the mean μ and standard deviation σ of the data within each group, and then follows the formula... Each data point x within a group is standardized. This standardization results in a more regular feature vector within each group, with a more uniform numerical range and distribution, which is beneficial for subsequent nonlinear transformations. Then, the computer system performs a nonlinear transformation on the standardized sensor signal feature vectors to obtain the transformed sensor signal feature vectors. Nonlinear transformations enhance the ability of feature vectors to represent complex signal characteristics. For example, the Sigmoid function can be used for nonlinear transformation; the formula for the Sigmoid function is... Here, x represents an element in the standardized sensor signal feature vector. Through this function transformation, the originally linear data relationship is converted into a nonlinear relationship, better capturing the nonlinear characteristics of the signal. The feature vector after nonlinear transformation can more effectively represent these complex relationships. Finally, the computer system uses a fourth filtering unit to extract features from the transformed sensor signal feature vector to obtain the second sensor signal feature vector. Similar to the third filtering unit, the fourth filtering unit also extracts features from the signal, but it processes the transformed sensor signal feature vector after the preceding series of operations. For example, the fourth filtering unit might employ adaptive filtering technology, which automatically adjusts the filter parameters based on the dynamic characteristics of the signal. For the transformed EEG signal feature vector, adaptive filtering can dynamically adjust the filtering parameters based on the changes in the EEG signal at different time points, thereby extracting the part that best represents the signal characteristics and ultimately obtaining the second sensor signal feature vector.

[0050] Specifically, step S120 involves extracting low-level representation information from the target sensing signal to obtain a second sensing signal feature vector, including:

[0051] Step S121: Extract features from the target sensing signal using the third filtering unit to obtain the initial sensing signal feature vector;

[0052] Step S122: Perform group normalization on the initial sensor signal feature vector to obtain the group-normalized sensor signal feature vector;

[0053] Step S123: Perform a nonlinear transformation on the standardized sensor signal feature vector to obtain the transformed sensor signal feature vector;

[0054] Step S124: The transformed sensor signal feature vector is extracted by the fourth filtering unit to obtain the second sensor signal feature vector.

[0055] In step S121, the computer system extracts features from the target sensing signal using the third filtering unit to obtain an initial sensing signal feature vector. The third filtering unit is a module specifically designed to filter and extract key feature information from the original target sensing signal. In practical applications, the target sensing signal is assumed to be a weak physiological electrical signal within a living organism. This signal contains various frequency components, amplitude variations, and complex time-related information. For example, taking an electrocardiogram (ECG) signal as an example, it contains characteristic waveforms such as the P wave, QRS complex, and T wave, which correspond to different physiological processes of the heart.

[0056] To extract features from such complex target sensing signals, the computer system employs a third filtering unit. One feasible technique is wavelet transform filtering. Wavelet transform decomposes a signal into sub-bands of different scales and frequencies, thereby highlighting the signal's characteristics in different frequency bands. Its basic formula is: Where WT(a,b) represents the wavelet transform result, x(t) is the original target sensing signal (e.g., an electrocardiogram signal), ψ is the wavelet function, and a and b are the scale and translation parameters, respectively. By adjusting a and b, the computer system can analyze the signal at different scales and locations, extracting frequency and amplitude information that characterizes the waveform features of the electrocardiogram signal. Combining this information forms the initial sensing signal feature vector. Compared to the traditional Fourier transform, this filtering method can better handle non-stationary signals because it considers the time-frequency characteristics of the signal simultaneously.

[0057] Next, in step S122, the computer system performs group normalization on the initial sensor signal feature vector to obtain a group-normalized sensor signal feature vector. Group normalization is a data processing technique designed to make the data distribution in the initial sensor signal feature vector more regular, facilitating subsequent processing and analysis. Assume the initial sensor signal feature vector contains multiple feature groups, which may correspond to different types of signal features or signal information in different frequency bands. For example, when processing electrocardiogram (ECG) signals, one group may contain feature information related to the P wave, while another group may contain feature information related to the QRS complex, etc.

[0058] During group standardization, the computer system calculates the mean μ and standard deviation σ for each group separately. For each element x within a group, the standard deviation is calculated according to the formula... Standardization is performed. To illustrate with a simple numerical example, if a group of data is [1,3,5,7,9], with a mean μ = 5 and a standard deviation σ = 2.83 (rounded to two decimal places), then the standardized values ​​are [-1.41, -0.71, 0, 0.71, 1.41] (rounded to two decimal places). This group standardization operation makes the data from different groups more consistent in numerical range and distribution, preventing certain features from dominating or being ignored in subsequent processing due to excessively large or small values. This allows subsequent nonlinear transformations and other operations to treat each feature group more fairly.

[0059] In step S123, the computer system performs a nonlinear transformation on the standardized sensor signal feature vector to obtain the transformed sensor signal feature vector. The nonlinear transformation aims to enhance the feature vector's ability to represent complex signal characteristics, because many real-world sensor signals, especially bioelectrical signals, often exhibit nonlinear inherent characteristic relationships. For example, in electrocardiogram (ECG) signals, the relationship between different waveforms is not a simple linear combination, but rather exhibits a nonlinear relationship influenced by the complex physiological mechanisms of the heart.

[0060] One technique for nonlinear transformation is to use the hyperbolic tangent function (tanh function). The formula for the tanh function is: Here, x represents an element in the standardized feature vector of the sensor signal. As the value of x varies within different ranges, y exhibits a non-linear trend. By transforming each element in the standardized feature vector of the sensor signal using the tanh function, the originally approximately linear relationship is converted into a non-linear one, which can better capture the complex relationships between signal features. This non-linearly transformed feature vector can more accurately reflect the complex feature structure in the original target sensor signal, such as the complex interrelationships between the P wave, QRS complex, and T wave in an electrocardiogram (ECG) signal, as well as their variations under different physiological states.

[0061] Finally, in step S124, the computer system extracts features from the transformed sensor signal feature vector using the fourth filtering unit to obtain the second sensor signal feature vector. The fourth filtering unit is also a module for feature extraction, but it processes the transformed sensor signal feature vector after the preceding series of operations (group normalization and nonlinear transformation). For example, when processing the bioelectrical signal feature vector after nonlinear transformation, the fourth filtering unit might employ a filtering technique based on principal component analysis (PCA).

[0062] Assume the transformed sensor signal feature vector is X, and its covariance matrix is... (where n is the number of samples), the eigenvalue λ is obtained by solving the characteristic equation |C-λI|=0 (where λ is the eigenvalue and I is the identity matrix). i and the corresponding feature vector v i Then, the eigenvectors are sorted according to their eigenvalues, and the eigenvectors corresponding to the first k eigenvalues ​​are selected to construct a projection matrix P. The transformed sensor signal eigenvector X is then projected onto a low-dimensional space to obtain Y = P. T X. Y is the feature vector of the second sensing signal obtained after processing by the fourth filtering unit.

[0063] In one implementation, the sampling rate of the first sensing signal feature vector is less than the sampling rate of the second sensing signal feature vector, and the first sensing signal feature vector is located in a higher-level feature domain. Based on this, step S200, determining multiple first adjacent sampling points of sampling point x and the feature value of each first adjacent sampling point according to the first sensing signal feature vector, and determining the first feature value of sampling point x according to the second sensing signal feature vector, may include:

[0064] Step S210: Resample the first sensor signal feature vector to obtain the third sensor signal feature vector. The sampling rate of the third sensor signal feature vector is equal to the sampling rate of the second sensor signal feature vector.

[0065] Step S220: Map the second sensor signal feature vector to the higher-level feature domain where the first sensor signal feature vector is located to obtain the fourth sensor signal feature vector.

[0066] Step S230: In the high-level feature domain, perform matching correction on the feature vectors of the fourth and third sensing signals, so that two sampling points at the same position in the feature vectors of the fourth and third sensing signals match each other.

[0067] Step S240: After performing matching correction, determine the multiple first adjacent sampling points of sampling point x in the third sensing signal feature vector and the feature value of each first adjacent sampling point, and obtain the feature value of sampling point x in the fourth sensing signal feature vector as the first feature value of sampling point x.

[0068] In step S210, the computer system resamples the first sensor signal feature vector to obtain a third sensor signal feature vector, and the sampling rate of the third sensor signal feature vector is equal to the sampling rate of the second sensor signal feature vector. Signal resampling is an operation that adjusts the signal sampling rate. For example, suppose the first sensor signal feature vector represents the bioelectrical signal characteristics collected by an implantable sensor; the original sampling rate may be low and cannot work well with the second sensor signal feature vector in subsequent processing.

[0069] Computer systems can use linear interpolation for resampling. Assume the original first sensor signal feature vector is V1 = [v1, v2, v3, ..., v...]. n The original sampling rate is f1, and the target sampling rate (i.e., the sampling rate of the second sensor signal feature vector) is f2. If a sampling point i is to be resampled to a new sampling point j, the linear interpolation formula is: (Here, it is assumed that the interval between adjacent sampling points is 1). Each sampling point in the first sensor signal feature vector is processed in this way to obtain the third sensor signal feature vector with a sampling rate of f2. The purpose of this is to ensure that the two different sensor signal feature vectors maintain a consistent sampling rate, facilitating subsequent matching and processing operations.

[0070] In step S220, the computer system maps the second sensor signal feature vector to the higher-level feature domain where the first sensor signal feature vector resides, obtaining the fourth sensor signal feature vector. The higher-level feature domain represents a specific feature space where the first sensor signal feature vector already exists. For example, the first sensor signal feature vector may have undergone some higher-level feature extraction operations and resides in a feature domain that reflects the macroscopic features of the signal. The computer system can implement this mapping using a kernel function-based mapping method. For example, a Gaussian kernel function can be used. Where x is a sample point in the second sensing signal feature vector, y is a reference point related to the first sensing signal feature vector (which can be a point representing the features of a higher-level feature domain selected in some way), and σ is the parameter of the kernel function. By applying this kernel function to each sample point in the second sensing signal feature vector, it is mapped to a higher-level feature domain, thus obtaining the fourth sensing signal feature vector. This mapping method can establish a connection between different feature domains, allowing the second sensing signal feature vector to be processed in the same feature representation space as the first sensing signal feature vector.

[0071] In step S230, within the high-level feature domain, the computer system performs matching correction on the feature vectors of the fourth and third sensing signals, ensuring that two sampling points at the same position in the feature vectors of the fourth and third sensing signals are matched (i.e. aligned). This step is to ensure that the correspondence between the two vectors is accurate in subsequent processing.

[0072] For example, assuming the feature vector of the third sensor signal is [3,5,7,9] and the feature vector of the fourth sensor signal is [2,4,6,8], the computer system needs to determine that the sampling points at the same positions in these two vectors correspond to each other. For instance, the 3 in the feature vector of the third sensor signal corresponds to the 2 in the feature vector of the fourth sensor signal, and so on. In practice, more complex signal feature matching may be involved. For example, if the sampling points in the feature vector of the third sensor signal are based on the average feature values ​​of a certain time window, while the sampling points in the feature vector of the fourth sensor signal are based on feature values ​​of different frequency components, the computer system needs to find a reasonable matching method through adjustment and correction so that the sampling points in the two vectors can correspond accurately. This may require determining the matching rules based on the characteristics of the signal and some information obtained in previous processing steps.

[0073] In step S240, after matching correction, the computer system determines multiple first adjacent sampling points of sampling point x in the third sensor signal feature vector and the feature value of each first adjacent sampling point, and obtains the feature value of sampling point x in the fourth sensor signal feature vector as the first feature value of sampling point x. Here, adjacent sampling points refer to sampling points that are spatially close to sampling point x in the feature vector.

[0074] For example, assuming the feature vector of the third sensor signal after matching correction is [10, 12, 14, 16, 18], if the sampling point x = 3 (corresponding to feature value 14), the computer system may determine x-1 = 2 (feature value 12) and x+1 = 4 (feature value 16) as the first adjacent sampling points and obtain their feature values. For the feature vector of the fourth sensor signal, for example, [9, 11, 13, 15, 17], the computer system directly obtains the feature value 13 corresponding to the sampling point x = 3 as the first feature value of the sampling point x. This operation of determining adjacent sampling points and obtaining the first feature value is based on the previous matching correction results, providing basic data for subsequent operations such as central gradient filtering and inference commonality scoring, playing a crucial role in the entire implanted sensor signal classification process.

[0075] As one implementation, step S220, mapping the second sensing signal feature vector to the higher-level feature domain where the first sensing signal feature vector is located to obtain the fourth sensing signal feature vector, may include:

[0076] Step S221: Obtain the algorithm library for linear transformation smoothing, and obtain the third sensor signal feature vector obtained by resampling the first sensor signal feature vector;

[0077] Step S222: Generate a reference signal for linear transformation smoothing based on the feature vector of the second sensor signal, and generate the signal to be processed in linear transformation smoothing based on the feature vector of the third sensor signal;

[0078] Step S223: Using a linear transformation smoothing algorithm library, the reference signal is mapped to the higher-level feature domain where the signal to be processed is located to obtain the feature vector of the fourth sensing signal.

[0079] In step S221 of this embodiment, the computer system acquires a linear transform smoothing algorithm library and a third sensor signal feature vector obtained by resampling the first sensor signal feature vector. The linear transform smoothing algorithm library contains a series of algorithms, functions, and related parameters for implementing linear transform smoothing operations. These algorithms aim to smooth signals through linear relationships, thereby achieving mapping between different feature domains.

[0080] For example, assuming the implanted sensing signal is a bioelectrical signal, the first sensing signal feature vector, after processing, reflects certain macroscopic characteristics of the bioelectrical signal. In previous steps, this signal was resampled to obtain a third sensing signal feature vector. This third sensing signal feature vector may be a vector containing multiple elements, each element corresponding to a feature value of the bioelectrical signal after resampling at different times or frequency bands. For example, the third sensing signal feature vector can be represented as V2 = [v 21 ,v 22 ,v 23 ,…,v 2m ], where n is the dimension of the vector. A library of linear transformation smoothing algorithms may include various implementations, such as linear transformation algorithms optimized using the least squares method.

[0081] In step S222, the computer system generates a reference signal for linear transformation smoothing based on the feature vector of the second sensor signal, and generates the signal to be processed in linear transformation smoothing based on the feature vector of the third sensor signal. The reference signal acts as a reference standard in the mapping process; it is generated from the feature vector of the second sensor signal. Because the feature vector of the second sensor signal contains high-frequency components of the signal, its eigenvalues ​​can reflect more detailed signal information.

[0082] For example, suppose the feature vector of the second sensing signal is V2 = [v 21 ,v 22 ,v 23 ,…,v 2m The computer system may select elements from V2 or combine them in a certain way according to some rules to generate a reference signal R. For example, selecting an odd number of elements from V2 to form the reference signal R = [v 21 ,v 23 ,v 25 At the same time, the computer system directly uses the feature vector of the third sensor signal as the signal to be processed, T = V3 = [v 31 ,v 32 ,v 33 ,…,v 3n This method of generating the reference signal and the signal to be processed is based on the characteristics of the two vectors and the mapping requirements. The reference signal can guide the smoothing and mapping operations on the signal to be processed.

[0083] In step S223, the computer system uses a linear transform smoothing algorithm library to map the reference signal to the higher-level feature domain of the signal to be processed, thereby obtaining the feature vector of the fourth sensor signal. The algorithms in the linear transform smoothing algorithm library operate based on the relationship between the reference signal and the signal to be processed. For example, a guided filtering algorithm (a type of linear transform smoothing algorithm) is used, whose basic principle is based on the assumption of a local linear model, using the reference signal to guide the filtering and smoothing operations on the signal to be processed.

[0084] For example, the formula for guided filtering is: Where q i It is the output signal (i.e., the element in the feature vector of the mapped fourth sensing signal), I j It is the element in the feature vector of the signal to be processed (i.e., the element in the feature vector of the third sensor signal), w ij It is the weight, ω i It is a local window, a k and b k These are linear coefficients, calculated based on the reference signal. In this process, the computer system calculates the weights w within each local window based on the reference signal R and the signal to be processed T. ij Linear coefficient a k and b k Then, each element is calculated according to the formula to map the reference signal to the high-level feature domain where the signal to be processed is located, thus obtaining the fourth sensor signal feature vector V4. This fourth sensor signal feature vector can be matched and corrected with the first sensor signal feature vector in the same high-level feature domain, laying an important foundation for the entire implantable sensor signal classification process.

[0085] As one implementation method, after performing matching correction, the method may further include:

[0086] Step S250: Perform standardization operations on the feature vectors of the third and fourth sensing signals respectively to obtain the standardized feature vectors of the third and fourth sensing signals.

[0087] Step S260: Obtain the general transformation parameters between the feature vectors of the third and fourth sensing signals;

[0088] Step S270: Using the obtained transformation parameters, perform linear transformations on the standardized third sensor signal feature vector and the standardized fourth sensor signal feature vector, respectively.

[0089] Step S280: After performing the linear transformation, determine the multiple first adjacent sampling points of sampling point x in the third sensing signal feature vector and the feature value of each first adjacent sampling point, and obtain the feature value of sampling point x in the fourth sensing signal feature vector as the first feature value of sampling point x.

[0090] In step S250 of this embodiment, the computer system performs standardization operations on the feature vectors of the third and fourth sensor signals respectively to obtain standardized feature vectors of the third and fourth sensor signals. The standardization operation is to make different feature vectors numerically comparable and to facilitate subsequent mathematical operations and model fitting.

[0091] For example, suppose the feature vector of the third sensor signal is V3 = [v 31 ,v 32 ,v 33 ,…,v 3n ], calculate its mean Standard deviation The standardized third sensor signal feature vector

[0092] Similarly, for the fourth sensor signal feature vector V4 = [v 41 ,v 42 ,v 43 ,…,v 4m (m may equal n, depending on the matching correction), calculate the mean μ4 and standard deviation σ4 to obtain the standardized fourth sensor signal feature vector. This standardized operation, based on the statistical characteristics of the data, can eliminate the influence of different vectors due to factors such as differences in dimensions and data distribution.

[0093] In step S260, the computer system obtains a general transformation parameter between the third and fourth sensor signal feature vectors. This transformation parameter is a parameter that can establish a relationship and perform a transformation between two vectors. For example, in some cases, the transformation parameter can be obtained by calculating the covariance matrix of the two vectors.

[0094] Let the eigenvectors of the third sensor signal V3 and the eigenvectors of the fourth sensor signal V4 have a covariance matrix. (where k is the number of samples, and (These are the mean vectors of V3 and V4, respectively). The elements of this covariance matrix C can be used as part of the transformation parameters. Alternatively, in other cases, the transformation parameters can be determined based on the known physical relationship or model assumptions between the two vectors.

[0095] In step S270, the computer system performs linear transformations on the standardized third sensor signal feature vector and the standardized fourth sensor signal feature vector using the obtained transformation parameters. The linear transformation is a mathematical transformation of the vectors based on the previously obtained transformation parameters, aiming to adjust the numerical relationships of the vectors to better meet the requirements of subsequent processing.

[0096] Assuming the transformation parameter is matrix A, for the standardized eigenvector of the third sensor signal... Vector after linear transformation Similarly, for the standardized fourth sensor signal feature vector Vector after linear transformation Matrix multiplication here is a common linear transformation method. By multiplying the transformation parameter matrix A with the vector, the value of each element in the vector is changed. This change is based on the relationship established by the transformation parameter, which helps to further establish a connection between the two vectors or make them more suitable for subsequent operations.

[0097] In step S280, after performing the linear transformation, the computer system determines multiple first adjacent sampling points of sampling point x in the third sensor signal feature vector and the feature value of each first adjacent sampling point, and obtains the feature value of sampling point x in the fourth sensor signal feature vector as the first feature value of sampling point x. This step redetermines the relevant sampling points and feature values ​​based on the results of the previous series of processing.

[0098] For example, suppose the feature vector of the third sensing signal after linear transformation If the sampling point x = 3, the computer system may determine x - 1 = 2 and x + 1 = 4 as the first adjacent sampling points based on certain adjacency relationships, and obtain their feature values. and etc. For the fourth sensor signal feature vector The computer system acquires the eigenvalues ​​corresponding to the sampling point x = 3. This serves as the first feature value for sampling point x. This determination method is based on the vector state after standardization, obtaining transformation parameters, and linear transformation. It provides accurate basic data for subsequent operations such as central gradient filtering and inference commonality scoring, playing a crucial role in the entire implantable sensor signal classification process.

[0099] As an implementation manner, the size of the adaptive interpolation matrix is d*d, and the multiple first adjacent sampling points are z first adjacent sampling points, where d≥1 and z = d2. In step S240 or step S280, determining the multiple first adjacent sampling points of the sampling point x in the third sensing signal feature vector and the eigenvalue of each first adjacent sampling point may include:

[0100] Step S201: Obtain the displacement data x corresponding to the sampling point x, where the displacement data x includes z displacement values;

[0101] Step S202: In the third sensing signal feature vector, take the sampling point that matches the sampling point x as a reference point, and delimit a target range, where the target range includes z sampling points;

[0102] Step S203: Based on the g-th displacement value of the displacement data x, move the coordinate of the g-th sampling point in the target range to obtain a target coordinate, where g is a positive integer not greater than z;

[0103] Step S204: Take the sampling point at the target coordinate in the third sensing signal feature vector as the g-th first adjacent sampling point of the sampling point x in the third sensing signal feature vector, and obtain the eigenvalue of the g-th first adjacent sampling point in the third sensing signal feature vector.

[0104] In the embodiments of the present application, steps S201 - S204 in the specific implementation manners of step S240 or S280 are important operation processes for determining sampling point-related information, and the computer system undertakes the execution task therein.

[0105] In step S201, the computer system obtains the displacement data x corresponding to the sampling point x, where the displacement data x includes z displacement values. The displacement data is a data form used to describe the relative position change information of the sampling point in the feature vector. For example, when processing the sensing signal feature vector of a certain implantable bioelectric signal, assuming that this feature vector represents the eigenvalue of the bioelectric signal at different time points or different frequency bands. Each displacement value in the displacement data x corresponding to the sampling point x may be associated with a specific signal feature change.

[0106] A possible technical means to obtain displacement data is based on the local gradient information of the feature vector. Assuming the feature vector is V = [v1, v2, …, v n , for the sampling point x (1 < x < n), the computer system can calculate the gradient difference between it and the adjacent sampling points, such as Δv x-1 = v x - v x-1 and Δv x+1 = v x - v x+1, these gradient differences are combined to form a part of the displacement values in the displacement data x. If z displacement values are to be obtained, the gradient information of more adjacent sampling points around the sampling point x can be considered, and these gradient differences are combined according to a certain rule (such as the order from near to far) to form the displacement data x.

[0107] In step S202, in the third sensing signal feature vector, the computer system designates a target range with a sampling point that matches the sampling point x (such as the midpoint) as the reference point. This target range includes z sampling points. The designation of this target range is to determine a group of sampling points related to the sampling point x so that subsequent operations can be performed on these sampling points based on the displacement data.

[0108] For example, assume that the third sensing signal feature vector is V3 = [v 31 , v 32 , …, v 3m . If the sampling point x corresponds to v 3i (1 < i < m), the computer system takes v 3i as the reference point, and then determines the target range according to the value of z. If z = 5, it may select v 3(i-2) , v 3(i-1) , v 3i , v 3(i+1) , v 3(i+2) These 5 sampling points form the target range. This way of determining the target range is based on the position of the sampling point x and the preset value of z, ensuring that there is a clear group of sampling points for subsequent processing.

[0109] In step S203, based on the g-th displacement value of the displacement data x, the computer system moves the coordinates of the g-th sampling point in the target range to obtain the target coordinates, where g is a positive integer not greater than z. This step uses the displacement data to adjust the coordinates of the sampling points in the target range to reflect the relative position change relationship between the sampling points.

[0110] For example, assume that the coordinate of the g-th sampling point in the original feature vector in the target range is j, and the g-th displacement value in the displacement data x is δ. Then the moved target coordinate j ′=j+δ. Here, the coordinate movement is a linear adjustment based on the displacement value. This method integrates displacement data information into the coordinates of the sampling point, allowing the sampling point's position to be dynamically adjusted according to the displacement data. In step S204, the computer system takes the sampling point at the target coordinates in the third sensor signal feature vector as the g-th first adjacent sampling point of sampling point x in the third sensor signal feature vector, and obtains the feature value of the g-th first adjacent sampling point in the third sensor signal feature vector. This step determines the adjacent sampling points and their feature values ​​of sampling point x based on the preceding coordinate movement operation.

[0111] For example, if the target coordinates j' after movement were determined in the previous steps, and the sampling point corresponding to coordinate j' in the third sensor signal feature vector v3 is v... 3j′ Then the computer system will v 3j′ As the g-th first neighboring sampling point of sampling point x, its feature value v is obtained. 3j′ In this way, the computer system can determine multiple first adjacent sampling points and their feature values ​​of sampling point x based on displacement data and target range, providing the necessary data foundation for subsequent filtering, commonality scoring and other operations, which is of great significance in the entire implantable sensor signal classification process.

[0112] As one implementation method, the displacement data x corresponding to the sampling point x is the x-th displacement data in the target displacement data set. Based on this, before obtaining the displacement data x corresponding to the sampling point x, the method may further include:

[0113] Step S201a: Walk through each sampling point in the first sensing signal feature vector one by one, and determine the neighbors of the sampling point y that is currently being walked through in the first sensing signal feature vector; wherein, the neighbors of sampling point y include z sampling points, y is greater than or equal to 1, and less than or equal to the number of sampling points in the first sensing signal feature vector.

[0114] Step S201b: Based on the feature values ​​of each of the neighboring sampling points of sampling point y, perform z-fold central gradient filtering on the feature values ​​of sampling point y to obtain the displacement data of sampling point y; wherein, the displacement data of sampling point y includes z displacement values, and the a-th displacement value among the z displacement values ​​is obtained based on the a-th central gradient filtering, where a is a positive integer not greater than z.

[0115] Step S201c: When each sampling point in the first sensing signal feature vector has been traversed, an initial displacement data set is obtained; wherein, the initial displacement data set includes the displacement data of each sampling point in the first sensing signal feature vector;

[0116] Step S201d: Perform interpolation processing on the initial displacement data set to obtain the target displacement data set, which includes M displacement data.

[0117] In step S201a of this embodiment, the computer system iterates through each sampling point in the first sensing signal feature vector, determining the neighbors of the currently visited sampling point y in the first sensing signal feature vector. The neighbors of sampling point y include z sampling points, where y is greater than or equal to 1 and less than or equal to the number of sampling points in the first sensing signal feature vector. Here, "iterates" refers to the computer system sequentially visiting each sampling point in the first sensing signal feature vector in a certain order. Determining the adjacency relationship of sampling points is for subsequent processing based on the information of adjacent sampling points.

[0118] For example, suppose the first sensing signal feature vector V1 = [v1, v2, v3, ..., v n When y = 3, if z = 5, the computer system might determine v1, v2, v4, v5, and v6 (assuming a simple linear adjacency relationship; the actual situation may differ depending on the specific signal characteristics and algorithm) as adjacent sampling points of sampling point y = 3 (i.e., v3). The computer system can determine adjacent sampling points by setting an index range. For example, for sampling point y, the index range of its adjacent sampling points could be... And ensure that the index value is between 1 and n.

[0119] In step S201b, the computer system performs z-fold central gradient filtering on the feature values ​​of sampling point y based on the feature values ​​of each of its neighboring sampling points to obtain the displacement data of sampling point y. The displacement data of sampling point y includes z displacement values, and the a-th displacement value among these z values ​​is obtained based on the a-th central gradient filtering, where a is a positive integer not greater than z. Central gradient filtering is a filtering method for analyzing the relationship between a sampling point and its neighboring sampling points.

[0120] To illustrate with a simple numerical example, suppose the feature value of the sampling point y is v. y =10, and the feature values ​​of its adjacent sampling points are v respectively. y-1 =8, v y+1 =12 (here, for simplicity, we assume z=2). Center gradient filtering can represent gradient information by calculating the difference between adjacent sampling points and the sampling point y. For example, for the first filtering (a=1), calculate Δv... y-1 =v y -v y-1 =10-8=2, Δv y+1 =v y -v y+1=10-12=-2. Then calculate the displacement value based on these differences. A simple calculation method is to normalize the differences in some way, for example, let the displacement value be... (Here, d_1 is the first displacement value). Similarly, d_2 can be obtained for the second filtering (a=2). In practical applications, more complex calculation methods may be involved. For example, factors such as the weights of different adjacent sampling points may be considered. The weights can be set according to the distance between the adjacent sampling points and the sampling point y or the characteristics of the signal.

[0121] Then, in step S201c, when each sampling point in the first sensing signal feature vector has been traversed, the computer system obtains an initial displacement data set; wherein, the initial displacement data set includes the displacement data of each sampling point in the first sensing signal feature vector. This means that the computer system has performed the operations of steps S201a and S201b on each sampling point in the first sensing signal feature vector, thereby obtaining a set containing the displacement data of all sampling points.

[0122] For example, for the first sensing signal feature vector V1 = [v1, v2, v3, ..., v n After the preceding operations, the computer system obtained the data for each sampling point v. i Displacement data D corresponding to (i = 1, 2, ..., n) i =[d i1 ,d i2 ,…,d iz All these displacement data combined constitute the initial displacement data set D = {D1, D2, ..., D...} n This initial displacement data set reflects the relationship between each sampling point and its neighboring sampling points in the feature vector of the first sensing signal, and is an important basis for subsequent interpolation and other operations.

[0123] Finally, in step S201d, the computer system performs interpolation processing on the initial displacement data set to obtain the target displacement data set, which includes M displacement data points. The interpolation processing is used to supplement the original displacement data with additional data points through a specific algorithm to meet the requirements of subsequent steps regarding the quantity and distribution of displacement data.

[0124] For example, suppose the displacement data in the initial displacement dataset D is discrete and limited in number, making subsequent operations inconvenient. The computer system can use linear interpolation to process it. Suppose a segment of displacement data in the initial displacement dataset is D. section ={D i1 D i2 ,…,D ik}(k < M), it is necessary to interpolate it into M displacement data. Let x be the index of the original data point and y be the index of the target data point. The linear interpolation formula is (Here it is assumed that the interval between adjacent data points is 1). By performing such interpolation processing on each part of the initial displacement data set, the computer system finally obtains the target displacement data set, which provides a more comprehensive and more suitable data basis for subsequent steps (such as determining adjacent sampling points of the sampling point and their eigenvalue operations), and plays an important paving role in the entire implantable sensing signal classification process.

[0125] As an implementation, in step S300, according to the eigenvalues of multiple first adjacent sampling points, perform central gradient filtering on the first eigenvalue of the sampling point x to obtain the target filtering result, which may include:

[0126] Step S310: According to the eigenvalues of multiple first adjacent sampling points, perform w times of central gradient filtering on the first eigenvalue of the sampling point x to obtain w central gradient filtering values; where, one time of central gradient filtering is used to obtain one central gradient filtering value, w ≥ 2;

[0127] Step S320: Generate the target filtering result through w central gradient filtering values.

[0128] In step S310, the computer system performs w times of central gradient filtering on the first eigenvalue of the sampling point x according to the eigenvalues of multiple first adjacent sampling points to obtain w central gradient filtering values. Central gradient filtering is a method to optimize the eigenvalue of the sampling point by analyzing the gradient relationship between the sampling point and adjacent sampling points. For example, assume that the first eigenvalue of the sampling point x is V x , and it has n first adjacent sampling points, and the eigenvalues are V x-i and V x+i For one time of central gradient filtering, the computer system can reflect the gradient information by calculating the difference between the adjacent sampling point and the sampling point x. A simple calculation method is ΔV x-i = V x - V x-i and ΔV x+i = V x - V x+i . Then adjust V x according to these differences. Assume that the calculation formula for one time of central gradient filtering is where α iThe weighting coefficients are set based on the signal characteristics. The computer system performs w filtering operations in this manner, obtaining a central gradient filter value each time. Here, the eigenvalues ​​of each first adjacent sampling point and the first eigenvalue of sampling point x contain eigenelements under H components (i.e., feature channels), where H ≥ 1. In this case, the computer system needs to process each component separately during filtering. For example, the H components are divided into sets to obtain F component sets, each containing H / F components. During the e-th central gradient filtering operation, the target component set for filtering is first determined, and then each component in the target component set is iterated through. When a component is reached, it is taken as the current component. Based on the eigenelements of each first adjacent sampling point under the current component, the central gradient filter is applied to the eigenelements of sampling point x under the current component to obtain the filtering result under the current component. When all components in the target component set have been iterated through, the filtering results under each component are summed to obtain the e-th central gradient filter value.

[0129] In step S320, the computer system generates a target filtering result using w central gradient filter values. The target filtering result integrates the information obtained from w central gradient filters. For example, the computer system can form a vector F = [F1, F2, ..., F...] from these w central gradient filter values. w This vector F can then be used as the target filtering result. This target filtering result contains the feature optimization information of the sampling point x after multiple central gradient filterings. It considers the influence of different adjacent sampling points and integrates the results of multiple filterings, providing a more comprehensive and accurate data foundation for subsequent steps (such as inferring the commonality score between the sampling point and its adjacent sampling points).

[0130] In one implementation, the feature value of each first adjacent sampling point and the first feature value of sampling point x both contain feature elements under H components, where H ≥ 1; based on this, step S310, according to the feature values ​​of multiple first adjacent sampling points, performs w times of central gradient filtering on the first feature value of sampling point x to obtain w central gradient filtered values, including:

[0131] Step S311: Divide the H components into sets to obtain F component sets; where each component set includes H / F components, F≥1;

[0132] Step S312: From the F component sets, determine the target component set for performing the e-th central gradient filtering; where e is a positive integer not greater than w;

[0133] Step S313: Walk through each component in the target component set one by one, and take the component that is currently walked to as the current component.

[0134] Step S314: Based on the feature elements of each first adjacent sampling point in the current component, perform central gradient filtering on the feature elements of sampling point x in the current component to obtain the filtering result in the current component;

[0135] Step S315: When all components in the target component set have been traversed, the filtering results of each component in the target component set are summed to obtain the e-th central gradient filtering value.

[0136] In step S311, the computer system partitions the H components into sets to obtain F component sets. Here, the H components refer to the number of feature elements under the feature channel contained in the feature value of each first adjacent sampling point and the first feature value of sampling point x. Dividing them into component sets is for performing the central gradient filtering operation in batches and in an organized manner.

[0137] For example, assuming H = 10 and F = 2, then each component set contains H / F = 10 / 2 = 5 components. The computer system can sequentially divide these 10 components into two sets, such as set A = {1, 2, 3, 4, 5} and set B = {6, 7, 8, 9, 10}. This partitioning method allows for more detailed processing of different groups of components in subsequent filtering operations, possibly based on some intrinsic relationship between the components or to accommodate computational resource and algorithm complexity requirements.

[0138] Next is step S312, where the computer system determines the target component set for the e-th central gradient filtering from the F component sets. Here, e represents the nth central gradient filtering operation (e is a positive integer not greater than w), and each filtering operation requires determining a target component set for the operation.

[0139] For example, when e=1, if there are two component sets A and B according to the previous division, the computer system may determine the target component set as A according to some preset rule (such as alternating selection or a specific selection method based on signal characteristics). This rule can be set according to the specific application scenario and signal characteristics. For example, if some key features of the signal are more reflected in the first few components, the system may preferentially select the set containing these components as the target component set for early filtering operations.

[0140] Next is step S313, where the computer system iterates through each component in the target component set one by one, taking the currently visited component as the current component. This step is to process each component in the target component set individually, ensuring that each component is correctly processed in the filtering operation.

[0141] Taking the previously determined target component set A = {1,2,3,4,5} as an example, the computer system first processes component 1 as the current component, and then processes components 2, 3, 4, and 5 in sequence. When processing each current component, a specific filtering operation is performed based on the feature elements of that component.

[0142] The next step is S314, where the computer system performs central gradient filtering on the feature elements of each first adjacent sampling point in the current component, obtaining the filtering result for the current component. Central gradient filtering adjusts the feature values ​​of the sampling point by analyzing the differences between the sampling point and its neighboring sampling points in a specific component.

[0143] Suppose that the feature element of sampling point x in the current component is v x There are n first adjacent sampling points, and their feature elements under the current component are v respectively. x-i and v x+i One possible way to calculate the center gradient filter is to first calculate the difference between the adjacent sampling point and the sampling point x in the current component, such as Δv. x-i =v x -v x-i and Δv x+i =v x -v x+i Then, based on these differences and preset weighting coefficients, v is... x Adjustments are made. Assume the weighting coefficient is α. i Then the filtering result f under the current component can be expressed by the formula The weighting coefficient α is calculated here. i The weighting coefficient can be set based on the characteristics of the signal and empirical values. For example, if a certain adjacent sampling point has a greater impact on sampling point x, then its corresponding weighting coefficient can be set to be larger.

[0144] In step S315, when all components in the target component set have been processed, the computer system sums the filtering results for each component in the target component set to obtain the e-th central gradient filter value. This step integrates the results obtained from filtering each component individually in the target component set to obtain a central gradient filter value that comprehensively reflects the filtering effect on the target component set. For example, if the filtering results obtained after processing each component in the target component set A in step S314 are f1, f2, f3, f4, and f5, then the e-th central gradient filter value F e= f1 + f2 + f3 + f4 + f5. In this way, the computer system can comprehensively consider a portion of the H components (determined by the target component set) in each central gradient filtering. After processing and integrating these components one by one, a central gradient filter value reflecting the comprehensive feature optimization information of the sampling point x under this filtering is obtained, which lays the foundation for finally obtaining w central gradient filter values ​​and generating the target filtering result.

[0145] As one implementation, the size of the adaptive interpolation matrix is ​​d*d, and the multiple first adjacent sampling points are z first adjacent sampling points, where d≥1 and z=d2; based on this, step S314, according to the feature elements of each first adjacent sampling point in the current component, performs central gradient filtering on the feature elements of sampling point x in the current component to obtain the filtering result in the current component, may include:

[0146] Step S3141: Obtain the filter variable e for performing the e-th central gradient filtering. The filter variable e includes H / F filtering dimensions, and each filtering dimension contains z weights.

[0147] Step S3142: Determine the difference between the feature element of each first adjacent sampling point in the current component and the feature element of sampling point x in the current component, and obtain z element differences;

[0148] Step S3143: By using the z weights in the p-th filtering dimension included in the filtering variable e, the differences of the z elements are fused to obtain the filtering result under the current component; where the value of p is the sequential value of the current component in the target component set.

[0149] In step S3141, the computer system obtains the filter variable e for performing the e-th central gradient filtering. The filter variable e includes H / F filtering dimensions (i.e., channels), and each filtering dimension contains z weights. The filter variable e here is set to quantize and adjust the relationship between the sampling point and its neighboring sampling points according to different filtering dimensions and weights in the central gradient filtering operation.

[0150] For example, assuming H = 8 and F = 2, then H / F = 4, meaning the filter variable e contains 4 filtering dimensions. If z = 3, then there are 3 weights in each filtering dimension. These weights can be set according to the characteristics of the signal and the results of prior analysis. One technique for setting the weights is based on the results of Principal Component Analysis (PCA). Assuming that PCA analysis of historical data yields the weight relationships of each principal component's contribution to the signal characteristics in different dimensions, the computer system can set the weights in the filter variable e based on these weight relationships. For example, the 3 weights in the first filtering dimension might be [0.2, 0.3, 0.5], indicating that in this dimension, the quantization adjustment for different adjacent sampling point relationships has different emphases.

[0151] Next is step S3142, where the computer system determines the difference between the feature elements of each first adjacent sampling point in the current component and the feature elements of sampling point x in the current component, obtaining z element differences. This step calculates the degree of difference between sampling point x and its adjacent sampling points in the current component; these differences are the basis data for central gradient filtering.

[0152] For example, suppose the current component is a specific signal frequency characteristic component, and the characteristic element of the sampling point x under this component is v. x =10, with 3 first adjacent sampling points (z=3), and their feature elements in this component are v respectively. x-1 =8, v x+1 =12, v x+2 =9. Therefore, the calculated element differences are Δv. x-1 =v x -v x-1 =10-8=2, Δv x+1 =v x -v x+1 =10-12=-2, Δv x+2 =v x -v x+2 =10-9=1, and these three element differences will be used for subsequent fusion operations with the weights in the filter variable e.

[0153] Finally, in step S3143, the computer system fuses the differences of the z elements using the z weights in the p-th filtering dimension included in the filtering variable e, to obtain the filtering result for the current component. This step utilizes pre-set weights to comprehensively process the element differences, thereby obtaining the filtering result for the current component.

[0154] Continuing with the example above, assuming the three weights in the first filtering dimension (p=1) of the filtered variable e are [0.2, 0.3, 0.5], then the filtering result f under the current component can be expressed by the formula... Let's calculate: f = 0.2 × 2 + 0.3 × (-2) + 0.5 × 1 = 0.4 - 0.6 + 0.5 = 0.3. Here, w... p,i This represents the i-th weight Δv in the p-th filtering dimension of the filtered variable e. x±i This refers to the element difference calculated in the previous steps. Through this fusion operation, the computer system can quantify and integrate the differences between sampling point x and its neighboring sampling points in the current component based on different filtering dimensions and weights, thus obtaining the filtering result for the current component. This approach fully considers the impact of different filtering dimensions and weights on the filtering result, which helps to more accurately perform central gradient filtering on the feature values ​​of sampling point x, thereby providing a more accurate data foundation for subsequent operations in the entire implanted sensor signal classification process (such as generating target filtering results, inference commonality scoring, etc.).

[0155] As one implementation, the size of the adaptive interpolation matrix is ​​d*d, and the multiple first adjacent sampling points are z first adjacent sampling points, d≥1, z=d2; the target filtering result includes center gradient filter values ​​of w dimensions, w>z; based on this, in step S400, based on the target filtering result, the first commonality score between sampling point x and each first adjacent sampling point is inferred, which may include:

[0156] Step S410: Perform dimensionality reduction on the target filtering result using a multilayer perceptron to obtain a z-dimensional vector;

[0157] Step S420: Determine the g-th dimension vector in the z-dimensional vector as the first commonality score between the sampling point x and the g-th first adjacent sampling point; where g is a positive integer not greater than z.

[0158] In step S410, the computer system performs dimensionality reduction on the target filtering result using a multilayer perceptron to obtain a z-dimensional vector. A multilayer perceptron is an artificial neural network structure consisting of an input layer, hidden layers, and an output layer; it can perform complex nonlinear transformations on the input data. The target filtering result is obtained through central gradient filtering in the preceding steps and may contain a significant amount of dimensional information.

[0159] For example, suppose the target filtering result is an n-dimensional vector F = [f1, f2, ..., f n (n>z), the multilayer perceptron processes this vector through its internal neuron connections and weight parameters. A computer system can use a fully connected multilayer perceptron, receiving the target filtered result vector F at the input layer. The number of neurons in the hidden layer can be determined empirically or based on optimization algorithms; for example, cross-validation can be used to select the number of neurons in the hidden layer to achieve the best dimensionality reduction effect.

[0160] Assuming the hidden layer has m neurons, the connection weight matrix from the input layer to the hidden layer is W1, and the bias vector is b1, then the computation from the input layer to the hidden layer can be expressed as H = σ(W1F + b1), where σ is the activation function, such as the ReLU function σ(x) = max(0,x). The connection weight matrix from the hidden layer to the output layer is W2, and the bias vector is b2. The computation of the output layer is O = σ(W2H + b2), where O is the z-dimensional vector [o1, o2, ..., o] after dimension reduction. z ].

[0161] Next, in step S420, the computer system determines the g-th dimension of the z-dimensional vector as the first commonality score between the sampling point x and the g-th first neighboring sampling point. This first commonality score aims to quantify the similarity or degree of commonality between the sampling point x and its neighboring sampling points.

[0162] For example, suppose the z-dimensional vector O = [o1, o2, ..., o] obtained after step S410 z If z = 5, then when g = 3, o_3 is determined as the first commonality score between sampling point x and the third first adjacent sampling point. The magnitude of this commonality score reflects the degree of relationship between sampling point x and specific adjacent sampling points after the preceding filtering and multilayer perceptron processing. If the value of o_3 is large, it indicates that sampling point x and the third first adjacent sampling point have high commonality in features, which may mean that they have more similarities in signal features. For example, when the implanted sensing signal is a bioelectric signal, it may mean that they have high similarity in terms of amplitude, frequency, or phase of the electrical signal; conversely, if the value of o_3 is small, it indicates low commonality.

[0163] In one implementation, the multilayer perceptron includes a first filtering unit, which includes z first filtering matrices; step S410, performing dimensionality reduction on the target filtering result through the multilayer perceptron to obtain a z-dimensional vector, may include:

[0164] Step S411: Generate the execution data of the first filtering unit based on the target filtering result;

[0165] Step S412: Filter and smooth the execution data using each of the first filtering matrices in the first filtering unit to obtain a z-dimensional vector; wherein the g-th dimension vector in the z-dimensional vector is obtained by filtering and smoothing the execution data using the g-th first filtering matrix.

[0166] In step S411, the computer system generates execution data for the first filtering unit based on the target filtering result. The first filtering unit in the multilayer perceptron plays a crucial role in the dimensionality reduction process of the target filtering result, and generating appropriate execution data is its initial step.

[0167] The target filtering result is a vector containing multi-dimensional information. It is obtained through previous steps (such as center gradient filtering) and reflects the feature relationship between the sampling point x and its neighboring sampling points after a series of processing steps. For example, suppose the target filtering result is F = [f1, f2, ..., f n (n is the dimension of the target filtering result), each element in this vector contains information about the signal characteristics.

[0168] The computer system can generate the execution data for the first filtering unit in various ways. One possible technique is to normalize the target filtering result, ensuring its values ​​fall within a specific range to facilitate subsequent neural network calculations. For example, a min-max normalization method can be used, as shown in the formula... Where f new,i Let f_i be the i-th element after normalization, and f_i be the i-th element in the original target filtering result. min(F) and max(F) are the minimum and maximum values ​​in the target filtering result vector F, respectively. The normalized vector F... new =[f new,1 ,f new,2 ,…,f new,n This can be used as part of the execution data of the first filtering unit.

[0169] Furthermore, the computer system may also perform feature combination or add additional information as execution data to the target filtering result. For example, it may calculate the difference between adjacent elements in the target filtering result vector to obtain a new vector D = [d1, d2, ..., d...]. n-1 ], where d i =f i+1 -f i Then, this difference vector is concatenated with the normalized target filtering result vector to obtain the final execution data E = [f new,1 ,f new,2 ,…,f new,n ,d1,d2,…,d n-1 This approach can provide richer information to the first filtering unit, which helps to better perform dimensionality reduction operations.

[0170] In step S412, the computer system filters and smooths the execution data using the first filtering matrices in the first filtering unit to obtain a z-dimensional vector. The first filtering matrices in the first filtering unit are the key structures for implementing the filtering and smoothing operation. Each first filtering matrix performs a specific linear transformation on the execution data to achieve the filtering and smoothing effects, thereby reducing the dimension.

[0171] Assume the first filtering unit contains z first filtering matrices M1, M2, ..., M z For the executed data E, when it is filtered and smoothed by the first filtering matrix M1, V1 = M1E is calculated. Here, the multiplication is a matrix multiplication operation. For example, if E is an m×1 vector (m is the length of the executed data) and M1 is a k×m matrix (k is less than m), then V_1 is a k×1 vector.

[0172] In the same manner, the computer system sequentially passes through other first filter matrices M2, ..., M z Perform filtering and smoothing operations on the execution data to obtain V2,…,V z Finally, these results are combined to form a z-dimensional vector O = [o1, o2, ..., o2]. z ], where o1 is an element in V1 (for example, it could be the first element in V1 or an element obtained through some mapping), o2 is the corresponding element in V2, and so on. This z-dimensional vector is the result after dimensionality reduction. It is obtained by filtering and smoothing the execution data through the first filtering unit in the multilayer perceptron, which provides the basis for subsequently determining the commonality score between the sampling point x and its neighboring sampling points.

[0173] In one implementation, the multilayer perceptron further includes a second filtering unit and a nonlinear transformation unit; wherein, the second filtering unit includes S second filtering matrices, S≥1; based on this, step S411, generating execution data for the first filtering unit through the target filtering result, may include:

[0174] Step S4111: Filter and smooth the target filtering result using each of the second filtering matrices in the second filtering unit to obtain an S-dimensional vector; wherein the s-th dimension of the S-dimensional vector is obtained by filtering and smoothing the target filtering result using the s-th second filtering matrix.

[0175] Step S4112: Perform a nonlinear transformation on the S-dimensional vector using a nonlinear transformation unit to obtain the transformed S-dimensional vector;

[0176] Step S4113: Generate the execution data of the first filtering unit using the transformed S-dimensional vector.

[0177] In this embodiment of the application, steps S4111-S4113 in the specific implementation of step S411 are detailed operation procedures of the computer system generating the execution data of the first filtering unit through the target filtering result.

[0178] In step S4111, the computer system smooths the target filtering result using each of the second filtering matrices in the second filtering unit to obtain an S-dimensional vector. The second filtering matrices in the second filtering unit are tools used for preliminary processing of the target filtering result.

[0179] Assume the target filtering result is F = [f1, f2, ..., f n Here, n represents the dimension of the target filtering result. For example, if the second filtering unit contains three second filtering matrices B1, B2, and B3, when smoothing the target filtering result through B1, let B1 be an m1×n matrix (m_1 is less than n). According to the matrix multiplication rule, we can calculate V1 = B1F, which is an m1×1 vector. Similarly, we can obtain V2 = B2F and V3 = B3F through B2 and B3 respectively. Then, these results are combined to form an S-dimensional vector. For example, if Then the S-dimensional vector Here, S = m1 + m2 + m3. This filtering and smoothing operation can extract features and compress information in the target filtering result. Each second filtering matrix filters and integrates different information in the target filtering result according to its own parameter settings.

[0180] Next, in step S4112, the computer system performs a nonlinear transformation on the S-dimensional vector using a nonlinear transformation unit to obtain the transformed S-dimensional vector. The nonlinear transformation unit aims to introduce nonlinear relationships to better handle the complex information relationships in the target filtering results.

[0181] For example, the hyperbolic tangent function (tanh) can be used as the nonlinear transformation function. For each element g in the S-dimensional vector G... i After nonlinear transformation, the result is obtained The entire S-dimensional vector becomes after this transformation This nonlinear transformation can convert elements that may originally be linearly related into nonlinear relationships, thereby uncovering the complex feature relationships hidden in the target filtering results and making the data more suitable for subsequent processing steps.

[0182] In step S4113, the computer system performs a standardization operation on the transformed S-dimensional vector using a standardization unit to obtain the execution data for the first filtering unit. Standardization helps to give the data better numerical characteristics, facilitating subsequent calculations and processing.

[0183] For example, for a transformed S-dimensional vector G_{new}, the computer system calculates its mean. and standard deviation Then, according to the standardized formula For G newStandardize each element in the vector to obtain the standardized vector. This vector G std This refers to the execution data of the first filtering unit. Through this standardization operation, the mean of the data becomes 0 and the standard deviation becomes 1, avoiding the adverse effects caused by the data being too large or too small in subsequent calculations. This provides a suitable data foundation for the first filtering unit to perform filtering and smoothing operations on the execution data, thus playing an important connecting and preprocessing role in the entire implanted sensor signal classification process.

[0184] As one implementation, the multilayer perceptron also includes a normalization unit; based on this, step S4113, which generates execution data of the first filtering unit through the transformed S-dimensional vector, may include: performing a normalization operation on the transformed S-dimensional vector through the normalization unit to obtain the execution data of the first filtering unit.

[0185] Standardization aims to process data into a form with specific statistical properties to facilitate subsequent processing. Assume the transformed S-dimensional vector obtained after step S4112 is...

[0186] The computer system first calculates the mean μ of the vector, using the formula: For example, if S = 5, G new =[1,3,5,7,9], then Next, calculate the standard σ using the formula: For the above examples,

[0187] Then, the computer system follows the standardized formula. For G new Standardize each element in the dataset. After standardization

[0188] By analyzing G new Perform this standardization operation on all elements to obtain a standardized vector. This G std This refers to the execution data of the first filtering unit. Standardized operations make the data distribution more regular, avoiding the adverse effects of data of different magnitudes on subsequent filtering operations, and laying the foundation for the first filtering unit to accurately process the data.

[0189] As one implementation, the size of the adaptive interpolation matrix is ​​d*d, and the multiple first adjacent sampling points are z first adjacent sampling points, where d≥1; in step S400, based on the first commonality scores obtained through inference, the adaptive interpolation matrix corresponding to sampling point x is generated, which may include:

[0190] Step S430: Perform low-pass filtering on the feature vector of the second sensing signal to obtain the feature vector of the control sensing signal, and determine the z second adjacent sampling points of sampling point x in the feature vector of the control sensing signal;

[0191] Step S440: Obtain the second commonality score between sampling point x and each second adjacent sampling point;

[0192] Step S450: Combine the g-th second commonality score with the g-th first commonality score obtained through reasoning to obtain the g-th combined commonality score corresponding to the sampling point x; where g is a positive integer not greater than z;

[0193] Step S460: Generate the adaptive interpolation matrix corresponding to sampling point x by using the z merged common scores corresponding to sampling point x.

[0194] In step S430, the computer system performs low-pass filtering on the feature vector of the second sensing signal to obtain the feature vector of the control sensing signal, and determines the z second adjacent sampling points of sampling point x in the feature vector of the control sensing signal. Low-pass filtering is a signal processing technique designed to preserve low-frequency components in a signal while removing high-frequency noise or details.

[0195] For example, suppose the feature vector of the second sensing signal is V2 = [v 21 ,v 22 ,v 23 ,…,v 2n The computer system uses a Butterworth low-pass filter for filtering. The transfer function of the Butterworth low-pass filter is... (where s is the complex frequency variable, ω) c (Where N is the cutoff frequency, and N is the filter order). By transforming the discrete second sensor signal feature vector to the s-domain (e.g., using the z-transform), filtering it, and then transforming it back to the discrete domain, the control sensor signal feature vector is obtained.

[0196] To determine the z second neighboring sampling points of sampling point x in the feature vector of the reference sensing signal, assuming x = 5 and z = 3, the computer system can determine them based on a certain predefined adjacency relationship. This is the second adjacent sampling point of sampling point x. This adjacency relationship can be based on the relative position of the index, for example, selecting a certain number of sampling points before and after sampling point x as the center as adjacent sampling points.

[0197] In step S440, the computer system obtains a second commonality score between sampling point x and each of its second adjacent sampling points. This step is similar to obtaining the first commonality score, also aimed at quantifying the closeness of the relationship between a sampling point and its adjacent sampling points, but here it is based on the feature vector of the reference sensor signal. The computer system determines the second feature value of sampling point x and the feature value of each second adjacent sampling point based on the feature vector of the reference sensor signal. For example, the feature value of sampling point x in the feature vector of the reference sensor signal is... Its second adjacent sampling point The eigenvalues ​​are respectively

[0198] Then, based on the feature values ​​of these z second adjacent sampling points, a central gradient filter is applied to the second feature value of sampling point x to obtain the comparison filter result. Assuming a similar central gradient filter calculation method as before is used, the difference between the adjacent sampling point and sampling point x is first calculated, such as... Then, based on these differences and preset weighting coefficients (e.g., set according to signal characteristics and experience), Adjustments are made to obtain the control filtering result. Finally, based on the control filtering result, a second commonality score between sampling point x and each second adjacent sampling point is inferred. For example, a similar reasoning method to the first commonality score can be used, through a mapping function (which can be an empirical function trained on a large amount of data or a mathematical function defined according to signal characteristics) to convert the control filtering result into the second commonality score.

[0199] In step S450, the computer system merges the g-th second commonality score with the g-th first commonality score obtained through inference to obtain the g-th merged commonality score corresponding to sampling point x. For example, suppose the first commonality score obtained through inference is C. 1g The second commonality score is C. 2g A simple merging method is to use weighted summation, i.e., the g-th merged commonality score C. g =αC 1g +(1-α)C 2g α is a weighting coefficient set according to specific circumstances (0 < α < 1). This combined commonality score integrates the relationship information between sampling points and adjacent sampling points obtained from different perspectives (based on different sensor signal feature vectors and processing methods).

[0200] In step S460, the computer system generates an adaptive interpolation matrix corresponding to sampling point x using the z merged common scores. For example, assuming = 2 × 2 = 4, the adaptive interpolation matrix is: Here, C1, C2, C3, and C4 are the combined common scores corresponding to sampling point x. This adaptive interpolation matrix will be used for subsequent interpolation operations on the feature vector of the first sensor signal. It comprehensively considers the common scores obtained from multiple factors, which can more accurately reflect the relationship between sampling points and adjacent sampling points, thereby better preserving signal features during the interpolation process and improving the accuracy of implanted sensor signal classification.

[0201] As one implementation, step S440, obtaining the second commonality score between sampling point x and each second adjacent sampling point, may include:

[0202] Step S441: Determine the second feature value of sampling point x based on the feature vector of the second sensing signal, and determine the feature value of each second adjacent sampling point;

[0203] Step S442: Based on the feature values ​​of the z second adjacent sampling points, perform central gradient filtering on the second feature value of sampling point x to obtain the comparison filtering result;

[0204] Step S443: Based on the comparison filtering results, infer the second commonality score between the sampling point x and each second adjacent sampling point.

[0205] In this embodiment of the application, steps S441-S443 in the specific implementation of step S440 are key operation procedures for the computer system to obtain the second commonality score between sampling point x and each second adjacent sampling point.

[0206] The first step is S441, where the computer system determines the second feature value of sampling point x based on the second sensor signal feature vector, and also determines the feature value of each second adjacent sampling point. The second sensor signal feature vector contains rich signal information and is the basis for this operation.

[0207] For example, suppose the feature vector of the second sensing signal is V2 = [v 21 ,v 22 ,v 23 ,…,v 2n If sampling point x = 3 (here, the sampling point index is in the context of the second sensor signal feature vector), then the second eigenvalue of sampling point x is v. 23 To determine the second adjacent sampling point, assuming a certain adjacency rule (e.g., taking one sampling point to the left and one to the right of the sampling point x as the center), when x = 3, the second adjacent sampling points are x - 1 = 2 and x + 1 = 4. Then their eigenvalues ​​are vi and vj, respectively. 22 and v 24This method of determining feature values ​​directly obtains the corresponding values ​​from the feature vector of the second sensing signal based on the index of the sampling point, providing basic data elements for subsequent calculation of commonality scores.

[0208] In step S442, the computer system performs central gradient filtering on the second feature value of sampling point x based on the feature values ​​of the z second adjacent sampling points to obtain the comparison filtering result. Central gradient filtering is a filtering method based on the gradient relationship between the feature values ​​of adjacent sampling points and the feature value of the target sampling point.

[0209] Continuing with the previous example, let v be the second eigenvalue of sampling point x. 23 The feature value v of adjacent sampling points 22 and v 24 Calculate the gradient difference between adjacent sampling points and sampling point x. For example, for the left adjacent sampling point x-1 = 2, the gradient difference Δv 22-3 =v 23 -v 22 For the adjacent sampling point x+1 = 4 on the right, the gradient difference Δv 24-3 =v 24 -v 23 Then, a center gradient filtering operation is performed based on these gradient differences. A simple formula for center gradient filtering can be F = v 23 +β(Δv 22-3 +Δv 24-3 ), where β is a pre-set weighting coefficient based on signal characteristics. This F is the result of the contrast filtering, reflecting a filtered characteristic representation of sample point x after considering the influence of adjacent sample points. Through this filtering method, the intrinsic relationship between sample point x and its adjacent sample points in terms of feature values ​​can be uncovered. This relationship is an important basis for subsequent commonality scoring inference.

[0210] Finally, in step S443, based on the contrast filtering results, the computer system infers a second commonality score between the sample point x and each of its second adjacent sample points. This step converts the contrast filtering results into a score that can quantify the degree of commonality between the sample point and its adjacent sample points.

[0211] For example, a distance-metric-based inference method can be used. Assuming the comparison filtering result obtained after step S442 is F, for each second adjacent sampling point, a certain distance metric between it and F is calculated. Taking Euclidean distance as an example, for the left adjacent sampling point x-1 = 2, let its characteristic value after some transformation (e.g., linear transformation or a specific mapping based on signal characteristics) be... Then the distance from F Then, a second commonality score is calculated based on this distance value. One possible calculation method is... Here, C_{22} represents the second commonality score between sampling point x and its left-adjacent sampling point x-1=2. The same method can be used to calculate the second commonality score between sampling point x and other second-adjacent sampling points. This second commonality score quantifies the similarity or commonality between sampling point x and each second-adjacent sampling point, providing crucial data support for subsequent operations such as merging commonality scores and generating adaptive interpolation matrices in the entire implanted sensor signal classification method.

[0212] As one implementation, step S500, interpolating the feature vector of the first sensing signal using the adaptive interpolation matrix corresponding to the sampling point x, may include:

[0213] Step S510: Obtain the displacement data x corresponding to the sampling point x, where the displacement data x includes z displacement values;

[0214] Step S520: Project the sampling point x onto the feature vector of the first sensing signal to obtain the projection result corresponding to the sampling point x;

[0215] Step S530: Based on the projection results and the z displacement values ​​in the displacement data x, determine the z third adjacent sampling points of the sampling point x in the first sensing signal feature vector and the feature value of each third adjacent sampling point;

[0216] Step S540: Interpolate the feature values ​​of the z third adjacent sampling points using the adaptive interpolation matrix corresponding to the sampling point x.

[0217] In this embodiment of the application, steps S510-S540 in the specific implementation of step S500 are the detailed operation flow of the computer system interpolating the feature vector of the first sensing signal through the adaptive interpolation matrix corresponding to the sampling point x.

[0218] The first step is S510, where the computer system acquires the displacement data x corresponding to the sampling point x. The displacement data x includes z displacement values. The displacement data describes the relative position change of the sampling point in the feature vector, and it plays an important role in the interpolation process.

[0219] For example, suppose the first sensor signal feature vector represents the characteristics of an implanted bioelectrical signal at different time points, and this feature vector is V = [v1, v2, ..., v nFor a sampling point x, the displacement value in its displacement data x may be related to the fluctuation characteristics of the bioelectrical signal on the time axis. If z = 3, the displacement data x = [d1, d2, d3], and these displacement values ​​may be obtained by analyzing the changes in signal characteristics around the sampling point x. One way to obtain the displacement value is based on local gradient calculation, such as calculating the difference in eigenvalues ​​between the sampling point x and its neighboring sampling points, and then determining the displacement value according to certain rules (e.g., quantization based on the magnitude and direction of the difference). For example, let v x-1 and v x+1 d_1 is the feature value of the neighboring sampling points of sampling point x, which can be expressed by the formula... The calculations show that d2 and d3 can be obtained similarly based on different calculation rules or by considering more information from adjacent sampling points.

[0220] In step S520, the computer system projects the sampling point x onto the feature vector of the first sensing signal to obtain the projection result corresponding to the sampling point x. The projection operation is a way to establish a connection between the sampling point x and the feature vector of the first sensing signal, so that the specific location information of the sampling point x in the feature vector can be determined based on the projection result.

[0221] For example, continuing with the first sensing signal feature vector V = [v1, v2, ..., v] of the above bioelectrical signal, n For example, assuming an index-based projection method is used, if the index of sampling point x in a specific index space (which may be a relative index space obtained after the previous steps) is i, then when sampling point x is projected onto the feature vector of the first sensing signal, the projection result can be represented as the positional information corresponding to index i. For instance, if i = 5, the projection result may represent the relative positional relationship between sampling point x and v5 in the feature vector. This relationship can be a logical correspondence or a description that includes more information about the distribution of feature values ​​around v5.

[0222] In step S530, the computer system determines the z third adjacent sampling points of sampling point x in the first sensing signal feature vector, as well as the feature value of each third adjacent sampling point, based on the projection result and the z displacement values ​​in the displacement data x. This step uses the projection result and displacement data to accurately determine the adjacent sampling points and their feature values ​​related to sampling point x, providing accurate data for the interpolation operation.

[0223] For example, after determining the approximate position of sampling point x in the first sensor signal feature vector based on the projection results, this position is adjusted using the displacement value in the displacement data x to determine the third adjacent sampling point. Assuming the projection results indicate that sampling point x is associated with v5, and the displacement data x = [d1, d2, d3], if the displacement value d1 determines a one-unit shift to the left (this shift unit is determined based on the characteristics of the feature vector and the previously defined rules), then the first third adjacent sampling point might be v4; based on d2, a two-unit shift to the right results in the second third adjacent sampling point, v7; further adjustments based on d3 yield the third third adjacent sampling point (e.g., v8). Then, the feature values ​​of these third adjacent sampling points, i.e., the specific values ​​of v4, v7, and v8, are obtained.

[0224] In step S540, the computer system interpolates the feature values ​​of the z third adjacent sampling points using the adaptive interpolation matrix corresponding to the sampling point x. The adaptive interpolation matrix is ​​constructed based on the common relationship between the sampling point x and its adjacent sampling points obtained in the previous steps, and it determines the contribution of each adjacent sampling point to the interpolation result during the interpolation process.

[0225] For example, suppose the adaptive interpolation matrix is The feature values ​​of the three third adjacent sampling points are v4, v7, and v8, respectively. Interpolation can be performed using a weighted summation method, and the interpolated result... It can be done through formula We obtain the weight m here. ij (i = 1, 2, 3; j = 1, 2, 3) are elements in the adaptive interpolation matrix, determined based on information such as the commonality score between sampling point x and its neighboring sampling points. Through this interpolation operation, the computer system can interpolate the sampling points in the feature vector of the first sensing signal based on relevant information of sampling point x (such as displacement data, projection results, etc.) and the adaptive interpolation matrix, obtaining interpolated feature values. This improves the information in the feature vector of the first sensing signal, enhancing its ability to represent the original implanted sensing signal and providing a more accurate data foundation for subsequent implanted sensing signal classification. This interpolation operation comprehensively considers multiple factors, making the interpolation results more consistent with the actual characteristics of the signal, thus contributing to improving the accuracy and effectiveness of the entire signal classification method.

[0226] This application provides a computer system, such as... Figure 2As shown, the computer system 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one type, and the structure of this computer system 100 does not constitute a limitation on the embodiments of this application.

Claims

1. A method for classifying implantable sensor signals based on neural modulation processing, characterized in that, include: A first sensor signal feature vector to be interpolated is obtained, and a second sensor signal feature vector is obtained; the first sensor signal feature vector and the second sensor signal feature vector are obtained by extracting characterization information at different scales for the same implanted sensor signal, and the feature values ​​in the second sensor signal feature vector contain high-frequency components of the signal. Based on the first sensor signal feature vector, determine multiple first adjacent sampling points of sampling point x and the feature value of each first adjacent sampling point, and based on the second sensor signal feature vector, determine the first feature value of sampling point x; x is a positive integer not greater than M, where M is the number of sampling points in the feature vector of the second sensing signal; Based on the feature values ​​of the plurality of first adjacent sampling points, the first feature value of the sampling point x is subjected to central gradient filtering to obtain the target filtering result; Based on the target filtering result, the first common score between the sampling point x and each first adjacent sampling point is inferred; And based on the first commonality scores obtained through reasoning, an adaptive interpolation matrix corresponding to the sampling point x is generated; The first sensing signal feature vector is interpolated using the adaptive interpolation matrix corresponding to the sampling point x to obtain the interpolated sensing signal feature vector. Based on the interpolated sensor signal feature vector, classification results for implanted sensor signals are obtained.

2. The method according to claim 1, characterized in that, The sampling rate of the first sensing signal feature vector is less than the sampling rate of the second sensing signal feature vector, and the first sensing signal feature vector is located in a higher-level feature domain. The step of determining multiple first adjacent sampling points of sampling point x and the feature value of each first adjacent sampling point based on the first sensing signal feature vector, and determining the first feature value of sampling point x based on the second sensing signal feature vector, includes: The first sensing signal feature vector is resampled to obtain a third sensing signal feature vector, wherein the sampling rate of the third sensing signal feature vector is equal to the sampling rate of the second sensing signal feature vector. The second sensing signal feature vector is mapped to the higher-level feature domain where the first sensing signal feature vector is located to obtain the fourth sensing signal feature vector. In the high-level feature domain, the fourth sensor signal feature vector and the third sensor signal feature vector are matched and corrected so that two sampling points at the same position in the fourth sensor signal feature vector and the third sensor signal feature vector are matched with each other. After matching correction, multiple first adjacent sampling points of sampling point x in the third sensing signal feature vector and the feature value of each first adjacent sampling point are determined, and the feature value of sampling point x in the fourth sensing signal feature vector is obtained as the first feature value of sampling point x.

3. The method according to claim 2, characterized in that, The step of mapping the second sensing signal feature vector to the higher-level feature domain where the first sensing signal feature vector is located to obtain the fourth sensing signal feature vector includes: Obtain a library of algorithms for linear transformation smoothing, and obtain a third sensor signal feature vector obtained by resampling the first sensor signal feature vector. A reference signal for linear transformation smoothing is generated based on the feature vector of the second sensing signal, and a signal to be processed in the linear transformation smoothing is generated based on the feature vector of the third sensing signal. Using the linear transformation smoothing algorithm library, the reference signal is mapped to the high-level feature domain where the signal to be processed is located, and the fourth sensing signal feature vector is obtained. After performing matching correction, the method further includes: The third sensor signal feature vector and the fourth sensor signal feature vector are standardized respectively to obtain the standardized third sensor signal feature vector and the standardized fourth sensor signal feature vector. Obtain a common transformation parameter between the feature vector of the third sensing signal and the feature vector of the fourth sensing signal; Using the obtained transformation parameters, linear transformations are performed on the standardized third sensor signal feature vector and the standardized fourth sensor signal feature vector, respectively. After performing the linear transformation, the steps are as follows: determining multiple first adjacent sampling points of sampling point x in the third sensor signal feature vector and the feature value of each first adjacent sampling point, and obtaining the feature value of sampling point x in the fourth sensor signal feature vector as the first feature value of sampling point x.

4. The method according to claim 2 or 3, characterized in that, The size of the adaptive interpolation matrix is ​​d*d, and the plurality of first adjacent sampling points are z first adjacent sampling points, d≥1, z=d2. Determining the plurality of first adjacent sampling points of sampling point x in the feature vector of the third sensing signal and the feature value of each first adjacent sampling point includes: Obtain the displacement data x corresponding to the sampling point x, wherein the displacement data x includes z displacement values; In the third sensing signal feature vector, a target range is defined with the sampling points that match the sampling point x as reference points, and the target range includes z sampling points; Based on the g-th displacement value of the displacement data x, the coordinates of the g-th sampling point in the target range are moved to obtain the target coordinates, where g is a positive integer not greater than z; The sampling point located at the target coordinate in the feature vector of the third sensing signal is taken as the g-th first adjacent sampling point of the sampling point x in the feature vector of the third sensing signal, and the feature value of the g-th first adjacent sampling point is obtained in the feature vector of the third sensing signal. The displacement data x corresponding to the sampling point x is the x-th displacement data in the target displacement data set. Before obtaining the displacement data x corresponding to the sampling point x, the method further includes: For each sampling point in the first sensing signal feature vector, the neighbors of the sampling point y that is currently visited are determined in the first sensing signal feature vector; wherein, the neighbors of the sampling point y include z sampling points, y is greater than or equal to 1, and less than or equal to the number of sampling points in the first sensing signal feature vector. Based on the feature values ​​of each of the neighboring sampling points of the sampling point y, the feature values ​​of the sampling point y are subjected to z-time central gradient filtering to obtain the displacement data of the sampling point y; wherein, the displacement data of the sampling point y includes z displacement values, and the a-th displacement value among the z displacement values ​​is obtained based on the a-th central gradient filtering, where a is a positive integer not greater than z. When each sampling point in the first sensing signal feature vector has been traversed, an initial displacement data set is obtained; wherein, the initial displacement data set includes the displacement data of each sampling point in the first sensing signal feature vector; The initial displacement data set is interpolated to obtain the target displacement data set, which includes M displacement data points.

5. The method according to claim 1, characterized in that, The step of performing center gradient filtering on the first feature value of sampling point x based on the feature values ​​of the plurality of first adjacent sampling points to obtain the target filtering result includes: Based on the feature values ​​of the plurality of first adjacent sampling points, the first feature value of the sampling point x is subjected to w central gradient filtering times to obtain w central gradient filtered values; wherein, one central gradient filtering is used to obtain one central gradient filtered value, and w≥2; The target filtering result is generated using the w central gradient filtering values.

6. The method according to claim 5, characterized in that, Each first adjacent sampling point's feature value and the first feature value of sampling point x contain feature elements under H components, where H ≥ 1; the step of performing w times of central gradient filtering on the first feature value of sampling point x based on the feature values ​​of the multiple first adjacent sampling points to obtain w central gradient filtered values ​​includes: The H components are partitioned into sets to obtain F component sets; where each component set includes H / F components, and F≥1; From the F component sets, determine the target component set for performing the e-th central gradient filtering; where e is a positive integer not greater than w; Each component in the target component set is walked one by one, and the component that is currently walked to is taken as the current component. Based on the feature elements of each first adjacent sampling point in the current component, a central gradient filter is performed on the feature elements of the sampling point x in the current component to obtain the filtering result in the current component. When all components in the target component set have been traversed, the filtering results of each component in the target component set are summed to obtain the e-th central gradient filtering value.

7. The method according to claim 6, characterized in that, The size of the adaptive interpolation matrix is ​​d*d, and the plurality of first adjacent sampling points are z first adjacent sampling points, d≥1, z=d2; the step of performing center gradient filtering on the feature elements of the sampling point x in the current component based on the feature elements of each first adjacent sampling point in the current component to obtain the filtering result in the current component includes: Obtain the filter variable e used for the e-th central gradient filtering, wherein the filter variable e includes H / F filter dimensions, and each filter dimension contains z weights; Each of the first adjacent sampling points determines the difference between the feature element of the sampling point x under the current component and the feature element of the sampling point x under the current component, and obtains z element differences; By fusing the z-th weights in the p-th filtering dimension of the filtering variable e, the filtering result of the current component is obtained; where the value of p is the sequential value of the current component in the target component set.

8. The method according to claim 1, characterized in that, The size of the adaptive interpolation matrix is ​​d*d, and the plurality of first adjacent sampling points are z first adjacent sampling points, d≥1, z=d2; the target filtering result includes center gradient filter values ​​of w dimensions, w>z; based on the target filtering result, the first commonality score between the sampling point x and each first adjacent sampling point is inferred, including: The target filtering result is reduced in dimension by using a multilayer perceptron to obtain a z-dimensional vector. The g-th dimension vector in the z-dimensional vector is determined as the first commonality score between the sampling point x and the g-th first adjacent sampling point; where g is a positive integer not greater than z.

9. The method according to claim 8, characterized in that, The multilayer perceptron includes a first filtering unit, which includes z first filtering matrices; the step of performing dimensionality reduction on the target filtering result through the multilayer perceptron to obtain a z-dimensional vector includes: The execution data of the first filtering unit is generated based on the target filtering result; The execution data is filtered and smoothed using each of the first filtering matrices in the first filtering unit to obtain a z-dimensional vector; wherein the g-th dimension of the z-dimensional vector is obtained by filtering and smoothing the execution data using the g-th first filtering matrix. The multilayer perceptron further includes a second filtering unit and a nonlinear transformation unit; wherein, the second filtering unit includes S second filtering matrices, S≥1; the step of generating the execution data of the first filtering unit through the target filtering result includes: The target filtering result is filtered and smoothed by each of the second filtering matrices in the second filtering unit to obtain an S-dimensional vector; wherein the s-th dimension of the S-dimensional vector is obtained by filtering and smoothing the target filtering result by the s-th second filtering matrix. The S-dimensional vector is nonlinearly transformed by the nonlinear transformation unit to obtain the transformed S-dimensional vector. The execution data of the first filtering unit is generated by the transformed S-dimensional vector; The multilayer perceptron further includes a normalization unit; the step of generating the execution data of the first filtering unit through the transformed S-dimensional vector includes: The standardized S-dimensional vector is standardized by the standardized unit to obtain the execution data of the first filtering unit.

10. A computer system, characterized in that, include: One or more processors; Memory; One or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, they implement the method as described in any one of claims 1 to 9.

Images