A nonlinear feature extraction method for medical internet of things data

Abstract

The application discloses a kind of nonlinear feature extraction methods of medical internet of things data, obtain medical internet of things data and carry out mean removal operation and obtain original data and are divided into training set, verification set and test set;Build suitable for medical internet of things data anomaly detection model, utilize sparse, binary random projection matrix original data is embedded from original feature space to high-dimensional transition space, the dimension of transition space is higher than the dimension of original feature space;Using PCA extracts principal component information and projects data from high-dimensional transition space to low-dimensional feature space to obtain nonlinear projection vector;Using training set and verification set, the best hyperparameter group is obtained by bayesian hyperparameter optimization to anomaly detection model;According to the best hyperparameter group and test set, the performance of the anomaly detection model is calculated;The medical internet of things data to be detected is input into the best final detection model to obtain the corresponding nonlinear feature extraction result.The application realizes very good nonlinear feature processing effect.

Description

Technical Field

[0001] This invention relates to the field of data mining in information processing technology, and more particularly to a method for nonlinear feature extraction of medical Internet of Things (IoT) data. Background Technology

[0002] The Internet of Things in Healthcare (IoMT) refers to the application of Internet of Things (IoT) technology in the medical field. By connecting medical devices, sensors, and mobile devices to the internet, it enables the collection, transmission, and analysis of medical information. The widespread application of IoMT in healthcare has brought about significant changes to the industry. Through IoMT technology, doctors can more easily monitor patients' health and take timely measures, greatly improving the quality and efficiency of medical services. At the same time, IoMT technology can also provide patients with more convenient, efficient, and personalized medical services, such as avoiding frequent outpatient visits through remote monitoring and diagnosis, thereby reducing the burden on patients.

[0003] IoMT anomaly detection refers to the process of quickly and accurately identifying and classifying abnormal situations by modeling and analyzing medical data generated by IoMT devices and using technologies such as machine learning and artificial intelligence. The main purpose of IoMT anomaly detection is to help healthcare professionals better monitor patients' health status, promptly identify potential disease risks, and improve the quality and efficiency of medical services. For example, by monitoring patients' physiological parameters, body temperature, heart rate, and respiration, potential disease risks can be quickly detected, further strengthening prevention and treatment measures. However, IoMT data typically exhibits complex characteristics such as large volume, high dimensionality, and nonlinearity. How to effectively process and analyze IoMT data and extract valuable information has become a major challenge for IoMT anomaly detection. Dimensionality reduction through feature extraction (FE) is one of the main methods to solve the above problems. FE can reduce the dimensionality of data without losing a large amount of information in the original feature space, making it very suitable as an effective means of processing high-dimensional data. Feature extraction is generally divided into two types: linear feature extraction and nonlinear feature extraction. Linear feature extraction is based on the assumption that the data has a linear structure, but it has strong limitations and the results are not ideal. Nonlinear feature extraction does not rely on linear assumptions and performs well in processing nonlinear structural data, making it one of the hot research directions in current data mining. Kernel methods and manifold learning, as the main means of extracting nonlinear features, have yielded many important research results, such as kernel-based KPCA and KICA, and manifold-based LLE, ISOMAP, LE, UMAP, TriMap, PaCMAP, and hNNE. Choosing a suitable kernel is not a simple task, as it often affects the success of the algorithm. Compared with kernel methods, manifold learning can better capture the inherent geometric structure of the original high-dimensional data, but it also suffers from computational cost and is susceptible to anomalous noise. When high-dimensional data is embedded into a low-dimensional feature space, the data topology structure undergoes significant distortion and deformation. Sanjoy et al. from the University of California proposed a Local Sensitive Hash (FLSH) algorithm based on the olfactory neural network of fruit flies. By dissecting the olfactory nervous system of fruit flies, they divided its odor information processing process into three stages: feedforward connection stage, dimensional projection stage, and winner-takes-all stage, achieving hash encoding of odor signals. The nonlinear mapping process of the fruit fly nervous system is particularly effective. Summary of the Invention

[0004] The purpose of this invention is to provide a method for nonlinear feature extraction of medical Internet of Things (IoT) data.

[0005] The technical solution adopted in this invention is:

[0006] A method for nonlinear feature extraction from medical IoT data includes the following steps:

[0007] Step 1: Acquire medical IoT data and perform mean removal to obtain raw data so that the data features of the raw data are better clustered around zero, thereby improving the execution efficiency of model building; at the same time, divide the raw data into training set X, validation set V and test set T.

[0008] Step 2: Construct an anomaly detection model suitable for medical IoT data. The specific execution steps of the anomaly detection model are as follows:

[0009] Step 2-1, using a sparse, binary random projection matrix M m×n Data with n samples are transferred from the original feature space X. d×n Embedded into a high-dimensional transition space Y m×n The dimension m of the transition space is higher than the dimension d of the original feature space;

[0010] Step 2-2: Use PCA to extract principal component information and transition the data from the high-dimensional transition space Y. m×n Projected onto the low-dimensional feature space Z k×n The nonlinear projection vector z is obtained, achieving linear dimensionality reduction;

[0011] Step 3: Use the training set X and the validation set V to perform Bayesian hyperparameter optimization on the anomaly detection model to obtain the optimal hyperparameter set X. best hyperparameter X best This includes the optimal dimension expansion m and the PCA dimension reduction k;

[0012] Step 4, based on the optimal hyperparameter set X best The performance of the anomaly detection model was calculated using the test set T;

[0013] Step 5: Select the best-performing anomaly detection model as the final detection model, and input the medical IoT data to be detected into the final detection model to obtain the corresponding nonlinear feature extraction results.

[0014] Furthermore, the specific steps of step 2-1 are as follows:

[0015] Step 2-1-1, construct the random projection matrix M m×n Random projection matrix M m×n In each row of elements, 6 positions are randomly selected and set to 1, while the rest are set to 0, thus satisfying the constraint:

[0016]

[0017] Step 2-1-2, let x = (x1, x2, ..., x... d ) T , x∈X d×nx represents a data object in the original feature space; y = (y1, y2, ..., y3) m ) T y∈Y m×n y represents the new data object generated after spatial transformation;

[0018] Step 2-1-3, the process of dimensional projection, is as follows:

[0019] y = Mx (2)

[0020] Where M represents the random projection matrix M m×n .

[0021] Furthermore, the specific steps of step 2-2 are as follows:

[0022] Step 2-2-1, let y = (y1, y2, ..., y m ) T y∈Y m×n y represents the data object in the dimensional transition space; z = (z1, z2, ..., z...) k ) T , z∈Z k×n z represents a new data object in the low-dimensional feature space, k is the dimension of the low-dimensional feature space Z and k < d < m;

[0023] Step 2-2-2, the process of PCA retaining principal components is as follows:

[0024] z = PCA(y) (3).

[0025] Furthermore, the specific steps for Bayesian hyperparameter optimization of the anomaly detection model are as follows:

[0026] S1, randomly initialize the hyperparameter set of the anomaly detection model: the dimension m of the up-dimensional projection and the dimension k of the PCA down-dimensionality reduction, and construct H from the initialized point set;

[0027] S2, construct the surrogate model p based on the probability distribution of H;

[0028] S3, combine the acquisition function of the surrogate model p to find a new set of hyperparameters x;

[0029] S4. Evaluate the objective function f based on the validation set V and the new hyperparameter set;

[0030] S5, add {x, f(x)} to H and update the knowledge base;

[0031] S6. Repeat steps S2-S5 until the maximum number of iterations is reached, and find the optimal hyperparameter set X. best = (m*, k*).

[0032] This invention employs the above technical solution, extending the original fruit fly olfactory neural network as the basis for the locality-sensitive hashing retrieval mechanism to the foundation of a high-dimensional space nonlinear characteristic extraction mechanism in the field of anomaly detection. The FDEPCA algorithm mainly consists of three steps: first, mean centering is used to normalize the data; then, a sparse binary random projection matrix is ​​used to quickly and with low distortion embed the original data into a higher-dimensional transition space to extract nonlinear structure data; finally, PCA is used for linear dimensionality reduction to extract key feature information. This invention overcomes the problems of severe distortion of the inherent geometric structure and high computational load in current nonlinear dimensionality reduction methods when processing large-volume, high-dimensional anomaly data. Applying the FDEPCA algorithm to an anomaly detection model achieves excellent nonlinear feature processing results. The dimensionality-up projection method of this invention has good distance-preserving performance, and the more significant the dimensionality expansion, the better the maximum regression distance-preserving performance. Analysis proves the projection time performance of the FDEPCA algorithm. This invention demonstrates excellent performance in terms of distance preservation and time complexity. Attached Figure Description

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0034] Figure 1 This is a schematic diagram of a nonlinear feature extraction model based on a fruit fly olfactory neural network.

[0035] Figure 2 This is a schematic diagram of the application framework of the present invention for anomaly detection;

[0036] Figure 3 This is a flowchart illustrating the specific application of the present invention in anomaly detection;

[0037] Figure 4 This is a schematic diagram comparing the impact of dimensionality expansion on anomaly detection accuracy under different datasets.

[0038] Figure 5 This diagram illustrates the comparison of runtime for different nonlinear feature extraction algorithms across different datasets. Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0040] This invention constructs a novel fruit fly olfactory neural network model that maps high-dimensional external input signals into new low-dimensional neural patterns, enabling fast and accurate nonlinear feature extraction. When the proposed method is applied to an anomaly detection model, it can be applied to different anomaly detection tasks in IoMT.

[0041] like Figures 1 to 4 As shown in one example, this invention discloses a nonlinear feature extraction method for medical Internet of Things (IoT) data, which includes the following steps:

[0042] Step 1: Acquire medical IoT data and perform mean removal to obtain raw data; this will allow the data features of the raw data to better cluster around the zero value, thereby improving the execution efficiency of model building.

[0043] Step 2, using a sparse, binary random projection matrix M m×n Data with n samples are transferred from the original feature space X. d×n Embedded into a high-dimensional transition space Y m×n The dimension m of the transition space is higher than the dimension d of the original feature space;

[0044] Step 3: Use PCA to extract principal component information and transition the data from the high-dimensional transition space Y. m×n Projected onto the low-dimensional feature space Z k×n The nonlinear projection vector z is obtained, achieving linear dimensionality reduction.

[0045] Furthermore, the specific steps of step 2 are as follows:

[0046] Step 2-1, construct the random projection matrix M m×n Random projection matrix M m×n In each row of elements, 6 positions are randomly selected and set to 1, while the rest are set to 0, thus satisfying the constraint:

[0047]

[0048] Step 2-2, let x = (x1, x2, ..., x...) d ) T , x∈X d×n x represents a data object in the original feature space; y = (y1, y2, ..., y3) m ) T y∈Y m×n y represents the new data object generated after spatial transformation;

[0049] Steps 2-3, the process of dimensional projection, are as follows:

[0050] y = Mx (2)

[0051] Where M represents the random projection matrix M m×n .

[0052] Furthermore, the specific steps of step 3 are as follows:

[0053] Step 3-1, let y = (y1, y2, ..., y m ) Ty∈Y m×n y represents the data object in the dimensional transition space; z = (z1, z2, ..., z...) k ) T , z∈Z k×n , represents a new data object in a low-dimensional feature space, and k < d < m;

[0054] Step 3-2, the process of PCA retaining principal components is as follows:

[0055] z = PCA(y) (3).

[0056] The specific principles of this invention will be explained in detail below:

[0057] The data in intelligent medical IoT systems exhibits complex characteristics such as large volume, high dimensionality, nonlinearity, and diversity, severely impacting the efficiency and performance of anomaly detection tasks. While nonlinear feature extraction, as an effective dimensionality reduction method, has yielded significant research results, it still falls short of meeting the feature processing requirements of current anomaly detection tasks. Researchers at the University of California, including Sanjoy Dasgupta, dissected the olfactory nervous system of fruit flies to study its odor signal processing. Their results show that the fruit fly's olfactory nervous system exhibits different neural patterns for different types of odor molecules, essentially assigning a unique feature code to each odor category. The odor molecule processing process can be divided into three stages: a feedforward connection stage, a dimensionality-up projection stage, and a winner-takes-all stage.

[0058] This invention simulates the olfactory neural system of fruit flies to perceive odors, constructing a novel fruit fly olfactory neural network model, FDEPCA, which maps high-dimensional external input signals into new low-dimensional neural patterns, achieving feature encoding of odor signals, such as... Figure 1 As shown. FDEPCA is essentially a nonlinear feature extraction model consisting of three steps: The first step is mean centering, where the input odor signal is transmitted from odor receptor neurons (ORNs) in the fruit fly's olfactory organs to the connected projection neurons (PNs). This process shifts the center of the spatial mean distribution of the odor signal to the origin, removing the influence of odor signal concentration on the neuron firing frequency. The second step is dimensionality-up projection, where the neural signal is transmitted from 50 projection neurons (PNs) to 2000 Kenyon cells (KCs), with each KC establishing connections with only 6 PNs. The third step is PCA with principal component retention, extracting the most significant principal components of the KCs as the feature encoding of the final input odor signal.

[0059] (1) Mean centering

[0060] Performing mean removal on the input data to achieve data normalization is a common strategy in the data preprocessing stage of model training.

[0061] (2) Upward random projection

[0062] Using a sparse, binary random projection matrix M m×d Data from the original feature space X d×n Embedded into a higher-dimensional transition space Y m×n The dimension *m* of the transition space is higher than the dimension *d* of the original feature space. The transformation matrix *M* is similar to the connection method of Drosophila neurons undergoing dimensionality upgrades; in each row, six elements are randomly selected and set to 1, while the rest are set to 0, thus satisfying the constraint:

[0063]

[0064] Let x = (x1, x2, ..., xn) d ) T , x∈X d×n y represents a data object in the original feature space. y = (y1, y2, ..., y...) m ) T y∈Y m×n This represents a new data object generated after spatial transformation. The process of dimensionality upscaling projection is as follows:

[0065] y = Mx (2)

[0066] (3) PCA retains principal components

[0067] PCA is used to extract principal component information, and the data is transitioned from the high-dimensional transition space Y. m×n Projected onto the low-dimensional feature space Z k×n This achieves linear dimensionality reduction. Let y = (y1, y2, ..., y... m ) T y∈Y m×n This represents a data object in the dimensional transition space. z = (z1, z2, ..., z...) k ) T , z∈Z k×n , representing a new data object in a low-dimensional feature space. k < d < m, the process of PCA preserving principal components is:

[0068] z = PCA(y) (3)

[0069] Furthermore, the FDEPCA nonlinear feature extraction method of this invention is a mathematical abstraction of the Drosophila neural process. The FDEPCA algorithm consists of three parts: mean centering, dimensionality-upgrading random projection, and PCA principal component extraction. The first step, mean centering, aims to better cluster data features near zero, improving the efficiency of model construction. The second step, dimensionality-upgrading random projection, is based on the JL theorem and uses a sparse binary transformation matrix for spatial transformation, embedding the original nonlinear data structure into a higher-dimensional transition space to achieve linear separability. This process exhibits excellent performance in terms of distance preservation and time complexity, enabling fast and low-distortion projection mapping. Finally, PCA is used for linear dimensionality reduction, extracting the most significant features as the final low-dimensional representation, facilitating data processing and storage.

[0070] Specifically, anomaly detection in cloud computing and IoT applications is a hot topic. Cloud computing and IoT environments often present large-scale, high-dimensional, and complex data scenarios. Nonlinear feature extraction can effectively capture hidden features in complex data to discover more valuable anomaly patterns and accurately model complex data structures, thereby improving anomaly detection performance. Traditional nonlinear feature extraction methods are completely different from the work of this invention. FDEPCA simulates the olfactory nerve perception process of fruit flies to achieve fast and low-distortion nonlinear feature projection transformation, extracting more valuable feature information. This overcomes the problems of structural information distortion and computational time consumption that current nonlinear feature extraction methods easily encounter when processing complex pattern data in anomaly detection.

[0071] This invention combines the fruit fly olfactory neural network FDEPCA with anomaly detection to form an anomaly detection application model based on FDEPCA nonlinear feature extraction. The performance of the model is evaluated by applying the dataset processed by the FDEPCA algorithm to the anomaly detector. The specific steps are as follows:

[0072] (1) Data partitioning. The original dataset X, y is partitioned into training sets X_train, y_train and test sets X_test, y_test.

[0073] (2) Nonlinear feature processing. FDEPCA is used to extract nonlinear features from the training set and the test set respectively, generating a new low-dimensional training set Z_train and a test set Z_test.

[0074] (3) Anomaly detection model training. Select a specific type of anomaly detection classifier (e.g., IFOrest) and train the model on the training set Z_train.

[0075] (4) Anomaly detection model prediction. Predict the classification results of the model on the test set Z_test.

[0076] (5) Anomaly detection model evaluation. The classification performance of the model was evaluated using ROC and AUC.

[0077] Furthermore, to enable the FDEPCA algorithm to adaptively extract appropriate feature dimensions in different anomaly detection tasks, this invention employs the SMBO method of Bayesian optimization to adaptively extract the dimensionality increase m and dimensionality decrease k of the FDEPCA algorithm. Let X and V be the training set and validation set, respectively; M x It is an FDEPCA anomaly detection model constructed using X and a set of hyperparameters x, where the hyperparameters x include the dimensionality increase m of the FDEPCA algorithm and the dimensionality reduction k of PCA; f(M x ) is the objective function, which evaluates the anomaly detection accuracy of the model after FDEPCA feature extraction on set V. The feature optimization task of the FDEPCA anomaly detection model can be described as an optimization problem:

[0078]

[0079] The specific steps for Bayesian hyperparameter optimization of the anomaly detection model are as follows:

[0080] S1, randomly initialize the hyperparameter set of the anomaly detection model: the dimension m of the up-dimensional projection and the dimension k of the PCA down-dimensionality reduction, and construct H from the initialized point set;

[0081] S2, construct the surrogate model p based on the probability distribution of H;

[0082] S3, combine the acquisition function of the surrogate model p to find a new set of hyperparameters x;

[0083] S4. Evaluate the objective function f based on the validation set V and the new hyperparameter set;

[0084] S5, add {x, f(x)} to H and update the knowledge base;

[0085] S6. Repeat steps S2-S5 until the maximum number of iterations is reached, and find the optimal hyperparameter set X. best = (m*, k*).

[0086] In summary, the olfactory nervous system of fruit flies generates different neural patterns in response to different external stimuli. To adapt to complex and ever-changing environments, organisms must be able to recognize different external stimuli, judge their quality, and respond quickly, while also implicitly responding to similar odors. Therefore, the olfactory neural network of fruit flies places high demands on the projection connection method of neurons. First, the dimensionality-upgrading connection process needs to have good similarity-preserving performance, which requires the transformation matrix of the dimensionality-upgrading projection stage of the FDEPCA algorithm to maximize the preservation of the relative distances of the projected data objects. Second, the fruit fly nervous system's ability to respond quickly to external stimuli requires the FDEPCA algorithm to minimize the time complexity of feature extraction. A detailed proof and analysis follow.

[0087] FDEPCA Upward Projection Analysis: The FDEPCA upward projection method simulates the connection between projection neurons and Kenyon cells in the olfactory nervous system of fruit flies. Each Kenyon cell establishes random connections with an average of only 6 projection neurons. This connection pattern exhibits sparsity and binaryity. Mathematically, this is represented by using sparse binary random projection to transform the original data into a higher-dimensional transition space.

[0088] Distance Preservation Analysis: The dimensionality-upgrading projection method of FDEPCA is based on the Johnson-Lindenstrauss (JL) theorem. After projection using the transformation matrix, a loss of similarity between data objects within the dataset is inevitable. The JL theorem ensures that the relative distances between data objects change as little as possible after the projection transformation. To analyze the distance-preserving performance of the dimensionality-upgrading projection in the FDEPCA algorithm, the distance-preserving definition of the transformation matrix is ​​given below.

[0089] Definition: (Mean Preservation) For a dataset X, where any subset is independent of the other subsets, the new dataset Y is obtained by projecting the dataset X through a transformation matrix V. For any element x in the original dataset, the ratio of the intersection of the S points surrounding x in the original dataset and the S points surrounding x in the projected dataset Y to S is defined as the mean of the mean preservation property of the transformation matrix V. If the S points before and after the projection completely coincide, then the dataset has undergone an isometric transformation.

[0090] Theorem: (Theorem of Preservation of Distance for FDEPCA Upward Projection) Let x, x′∈R d These are two data objects in the original sample space, y, y′∈R. m It is a sparse binary random matrix M∈R m×d The relative distances between the data objects before and after projection are as follows: The new data objects generated after projection have the following relationship:

[0091] E[||yy′|| 2|=mp((1-p)||xx′|| 2 (4)

[0092] Where E represents the expectation, and p = 6 / d. When m is sufficiently large, the time complexity of random projection is O(d), and the variance ||y|| 2 The data objects are tightly clustered around the expected value, meaning that the projected data objects maintain their neighborhood structure.

[0093] (1-∈)E||y|| 2 ≤||y|| 2 ≤(1+∈)E||y|| 2 (5)

[0094] Where 0 < ∈ < 1, it represents the fluctuation value near the expected value.

[0095] The proof of the theorem is as follows:

[0096] Let y = (y1, y2, ..., y3) m y = Mx, 1≤j≤m, assuming x is very sparse, and the only non-zero coordinate of x is x1, then y j =M j x has the following distribution:

[0097]

[0098] right have:

[0099]

[0100] In fruit flies, since x and x′ have approximately the same overall activation level, p(1·(xx′)) 2 ≈0, thus proving:

[0101] E(||yy′|| 2 )=mp((1-p)||xx′|| 2 (8)

[0102] This process is based on the JL theorem. When m is large enough, low-distortion embedding can be achieved within the error range, and the relative distance between the new data objects after projection is preserved to the maximum extent.

[0103] Maximum regression distance preservation analysis: JL's theorem provides a quantitative analysis of information loss, proving the relationship between the loss boundary of projection and the projection dimension. Below, we introduce JL's theorem for FDEPCA upscaling projection.

[0104] Theorem: (JL Theorem for FDEPCA Upgraded Projection) For and Let n be a positive integer, and m be a positive integer. Then the relationship between the loss parameter ε and the projected dimension m is as follows:

[0105]

[0106] To further analyze the distance-preserving performance of dimensional expansion projection in the FDEPCA algorithm, the concept of maximum regression distance preservation is introduced below to compare the distance-preserving performance of different dimensional expansions.

[0107] Definition: (Maximum Regression Preservation of Distance) Assume the original dataset has size n and dimension d. Expand the dimensions of the dataset to m1 and m2, where m1 < m2. The distance loss parameter for m1 is ε1, and the maximum regression preservation of distance is MRD1 = 1 - ε1. The distance loss parameter for m2 is ε2, and the maximum regression preservation of distance is MRD2 = 1 - ε2. Then the difference between the maximum regression preservation of distance between the two datasets is defined as: ΔMRD = MRD1 - MRD2.

[0108] Theorem: (Difference Theorem for Maximum Regression Preservation) Suppose the original dataset is X, and the dataset is expanded in different degrees, resulting in Y and Z datasets after projection. The more obvious the dimensional expansion, the higher the maximum regression preservation property can be obtained.

[0109] The proof of the theorem is as follows:

[0110]

[0111] Since m1 < m2, ΔMRD > 0, meaning MRD2 > MRD1. This indicates that the greater the dimensionality expansion, the better the maximum regression distance-preserving performance. Therefore, feature amplification of the data can better preserve the similarity between data objects, reduce the loss of data structure information caused by the projection process, and improve the detection accuracy of the model.

[0112] Time Complexity Analysis of FDEPCA: The olfactory nervous system of fruit flies can quickly perceive different odor information and generate different neural patterns. Therefore, the FDEPCA algorithm has efficient projection time performance. The FDEPCA algorithm mainly consists of three steps: the first step is to preprocess the data by mean centering; the second step is to perform dimensionality-upgrading projection on the data using a sparse binary random projection matrix, expanding the dimension to m dimensions; and the third step is to use PCA to extract the k most significant features of the transformed data. Therefore, the time complexity of the FDEPCA algorithm can be mainly divided into three parts:

[0113] The first step is to perform mean normalization on the original data, which requires subtracting the mean from each feature of the data. The time complexity is O(nd), where n represents the number of samples in the original data and d represents the dimension of the original data.

[0114] The second step is to perform dimensionality-upgrading random projection, generating a sparse binary random projection matrix M. m×d The required time complexity is O(md), and the time complexity required to project the normalized data to the next dimension is O(nmd), where n is the number of samples, m is the dimension after the projection, and d is the dimension of the original data.

[0115] The third step is PCA to extract principal components, retaining the k most significant features as new feature components. The time complexity of PCA depends on the computation of the covariance matrix, O(nm). 2 ) and perform eigenvalue decomposition O(m 3 Therefore, the total time complexity of PCA is O(nm). 2 +m 3 ).

[0116] Therefore, the time complexity of the FDEPCA algorithm proposed in this invention is: O(nd+md+nmd+nm). 2 +m 3 ).

[0117] Experiment and Results Analysis

[0118] Datasets and Experimental Environment: To evaluate the performance of the anomaly detection method of this invention, eight real-world datasets from the UCI dataset with different practical application scenarios were selected, as shown in Table 1. The specific information of these eight datasets is described in detail below:

[0119] (1) Ionosphere dataset: The original dataset contains statistical information about the ionosphere, including 34 features and 2 categories. In this experiment, the original dataset was preprocessed, in which features with all attribute values ​​of 0 were discarded, and the bad class was defined as the anomalous class and the good class as the normal class. The processed dataset contains a total of 351 data objects, each containing 33 dimensions, of which 126 data objects are considered anomalous data.

[0120] (2) Satellite dataset: The original dataset contains spectral information from satellite images, including 36 features and multiple categories. This experiment preprocessed the original dataset, merging the three categories with the fewest entries (categories 2, 4, and 5) into anomaly categories, and merging the remaining categories into normal categories. The processed dataset contains a total of 6435 data objects, each with 36 dimensions, of which 2036 data objects are considered anomalous.

[0121] (3) Optdigits dataset: Its original dataset records the optical information of handwritten digits, containing 64 features and 10 categories. In this experiment, the original dataset was preprocessed, with the digit 0 defined as an anomaly class and the digits 1-9 defined as normal classes. The processed dataset contains a total of 5216 data objects, each containing 64 dimensions, of which 150 data objects are considered anomalous data.

[0122] (4) MNIST dataset: Its original dataset contains image information of handwritten digits, including 784 features and 10 categories. In this experiment, the original dataset was preprocessed. The digit 0 was defined as the normal class, and the digits 1-9 were defined as the abnormal class. 100 features were randomly selected from the 784 features as new dimensions. The processed dataset contains a total of 7603 data objects, each containing 100 dimensions, of which 700 data objects are considered abnormal data.

[0123] (5) Musk Dataset: The original dataset contains information about musk molecules, including 166 features and multiple categories. This experiment preprocessed the original dataset, defining non-musk classes j146, j147, and 252 as normal classes, and musk classes 213 and 211 as anomalous classes, discarding the other categories. The processed dataset contains a total of 3062 data objects, each with 166 dimensions, of which 97 data objects were identified as anomalous data.

[0124] (6) Arrhythmia dataset: The original dataset contains information on cardiac arrhythmias, including 279 features and 16 categories. This experiment preprocessed the original dataset, discarding 5 useless features and classifying the 8 categories with the fewest entries (categories 3, 4, 5, 7, 8, 9, 14, and 15) as anomalous data, while the remaining categories were considered normal. The processed dataset contained a total of 274 data objects, each with 274 dimensions, of which 66 data objects were identified as anomalous data.

[0125] (7) Speech dataset: Its original dataset records speech information in the real world, containing 400 features and 8 categories. The American accent accounts for the largest proportion and is regarded as the normal class, while the other 7 categories are regarded as anomalous classes. The preprocessed dataset contains a total of 3686 data objects, each containing 400 dimensions, of which 61 data objects are regarded as anomalous data.

[0126] Table 1: Dataset Information

[0127]

[0128]

[0129] The experimental environment for this invention is a Windows 10 system with an i5-6500 processor, 8GB of memory, Jupyter Notebook, and PyCharm compiler.

[0130] Evaluation metrics: The anomaly detection and evaluation methods used in this invention are ROC curve (Receiver Operator Characteristic Curve, ROC) and AUC (Area Under Curve, AUC).

[0131] As shown in Table 2, the classification results were divided into two categories: outliers in the dataset were labeled as positive (P), and normal points were labeled as negative (N). The anomaly detection results for the sample data were categorized into four cases: actual outliers correctly predicted as outliers (TP), actual normal points incorrectly predicted as outliers (FP), actual normal points correctly predicted as normal points (TN), and actual outliers incorrectly predicted as normal points (FN).

[0132] Table 2: Confusion Matrix

[0133] Predicting the real Anomalies Normal Anomalies TP FN Normal FP TN

[0134] The ROC curve is a graphical representation of the relationship between the false positive rate and the true positive rate of a classifier. The horizontal and vertical axes of the curve represent the false positive rate (FPR) and the true positive rate (TPR), respectively. The calculation formula is as follows:

[0135] TPR = TP / (TP + FN)

[0136] FPR = FP / (FP + TN)

[0137] In this context, TPR represents the proportion of samples with positive actual labels that are correctly predicted, while FPR represents the proportion of samples with negative actual labels that are incorrectly predicted. Generating the ROC curve requires calculating the true positive rate and false positive rate of the samples using pre-set thresholds. As the thresholds change, corresponding coordinate points are obtained, and connecting these points forms a two-dimensional curve. The area under the curve is defined as AUC; a higher AUC value indicates higher accuracy and better performance of the anomaly detection method.

[0138] Isolation Forest (IForest) is a widely adopted anomaly detection algorithm that has received extensive attention due to its simplicity and efficiency. It constructs a forest consisting of multiple isolated trees. Each isolated tree recursively partitions the feature space by randomly selecting features and split values. Outliers are considered to be those data points that require fewer splitting times during the partitioning process. Compared with other anomaly detection methods, IForest has demonstrated excellent performance and scalability in various applications, especially suitable for handling large-scale and high-dimensional data. In addition, IForest has a low computational cost and requires little parameter adjustment, making it a practical choice in practical applications. Therefore, in the first two experiments, we used it as a fixed anomaly detection classifier to test the anomaly detection classification performance of the feature extraction algorithm applied to IForest.

[0139] Experiment 1: The purpose of this experiment is to compare the distance-preserving performance of different dimensional expansions of the FDEPCA algorithm. Assume that the dimension of the original dataset is d. The FDEPCA algorithm uses a sparse binary projection matrix to expand the dimension of the data to m, and then performs PCA linear dimensionality reduction to k, where k < d < m. To reduce the model computational complexity, in this experiment, k is fixed at a relatively small value of 10, and the FDEPCA algorithm is expanded in different degrees, and the AUC value of its action on IForest is calculated.

[0140] As Figure 4 shown, the impact of dimensional expansion on the anomaly detection accuracy. The abscissa represents the multiple of dimensional expansion, and the actually expanded dimension is expressed as: m = multiple of dimensional expansion × d. The ordinate represents the AUC value of the FDEPCA algorithm acting on IForest.

[0141] The experimental results show that: the greater the dimensional expansion, the higher the anomaly detection accuracy of the FDEPCA algorithm. It can be inferred that dimensional expansion can effectively preserve the intrinsic geometric structure of the data. As the degree of data feature expansion increases, normal data and abnormal data can be better separated, and the anomaly detection accuracy of the model is also higher. Therefore, dimensional expansion has very good distance-preserving performance, and the more obvious the dimensional expansion, the better the distance-preserving effect.

[0142] Experiment 2: The purpose of this experiment is to verify that the FDEPCA algorithm can achieve fast and low-distortion non-linear feature extraction. Compare it with the current mainstream non-linear feature extraction algorithms KPCA, LLE, ISOMAP, LE, UMAP, TriMap, PaCMAP, and hNNE, calculate the anomaly detection accuracy of their action on IForest (as shown in Table 3, the ranking is shown in parentheses (the lower the better)), and at the same time, compare the running time of non-linear feature extraction algorithms under different feature dimensions.

[0143] Table 3: Comparison of AUC values ​​of different nonlinear feature extraction algorithms

[0144] Dataset KPCA LLE ISOMAP LE UMAP TriMAP PaCMAP hNNE FDEPCA Ionosphere 0.9311(2) 0.7246(5) 0.8869(3) 0.7997(4) 0.4762(7) 0.4723(8) 0.5677(6) 0.3566(9) 0.9431(1) Satellite 0.6684(8) 0.8591(1) 0.7556(5) 0.7835(2) 0.6448(9) 0.7423(5) 0.6868(7) 0.6990(6) 0.7607(3) Mnist 0.8513(5) 0.7066(8) 0.8571(4) 0.8209(7) 0.8230(6) 0.9110(1) 0.8966(2) 0.5328(9) 0.8678(3) Musk 0.9993(6) 0.9455(7) 0.6549(8) 1(1) 1(1) 1(1) 1(1) 0.5994(9) 1(1) Arrhythmia 0.8261(2) 0.7388(8) 0.8174(3) 0.7979(6) 0.7815(7) 0.8102(4) 0.8073(5) 0.6254(9) 0.8487(1) Speech 0.4913(7) 0.5722(4) 0.4862(8) 0.5325(5) 0.6116(2) 0.4932(6) 0.4602(9) 0.6293(1) 0.5755(3) Avg 0.7946(2) 0.7578(4) 0.7430(5) 0.7891(3) 0.7229(8) 0.7382(6) 0.7364(7) 0.5738(9) 0.8074(1)

[0145] like Figure 5 The image shows a comparison of the running times of different nonlinear feature extraction algorithms. The horizontal axis represents the dimensionality of the dataset after dimensionality reduction, and the vertical axis represents the running time of the model, in seconds.

[0146] Experimental results show that, overall, the FDEPCA algorithm has better detection performance and a significant time efficiency advantage than current mainstream nonlinear feature extraction algorithms. Observations reveal that for ultra-large datasets (such as Satellite, MNIST, Musk, and Speech), the time efficiency advantage of the FDEPCA algorithm is even more pronounced. These phenomena indicate that the FDEPCA algorithm used in this invention exhibits excellent performance in terms of distance preservation and time complexity, effectively preserving the inherent geometric structure characteristics of the original data and achieving fast, low-distortion feature embedding. Therefore, the FDEPCA algorithm has significant feature processing advantages in handling anomaly detection problems with large-scale, high-dimensional nonlinear data.

[0147] Experiment 3: The purpose of this experiment is to demonstrate the applicability of the FDEPCA algorithm to different anomaly detection tasks. The FDEPCA algorithm is applied to different anomaly detectors based on nearest neighbors, probability, and ensemble methods, and the AUC values ​​are compared with the original anomaly detection algorithm. Bayesian hyperparameter optimization is used in this process to automatically determine the required feature dimension. Assuming the original data dimension is d, m represents the dimension of the fruit fly dimensionality increase, k represents the dimension of the PCA dimensionality reduction, m = ratio × d, the value range of k is [2, min(d-1, 120)], and the value range of ratio is [2, 20]. After 50 iterations, the anomaly detector uses the PyOD library with default parameters. Among them, the representative anomaly detection algorithms based on nearest neighbors are: FB, Iforest, LODA, INNE, and SUOD; the representative anomaly detection algorithms based on probability are: LOF, KNN, COF, CBLOF, and HBOS; and the representative anomaly detection algorithms based on ensembles are: ABOD, SOS, Sampling, COPOD, and ECOD. The experimental results are shown in Tables 4, 5, and 6.

[0148] Table 4: Comparison of AUC values ​​between the nearest neighbor anomaly detection algorithm based on FDEPCA and the original algorithm.

[0149] Dataset LOF FDEPCA-LOF KNN FDEPCA-KNN COF FDEPCA-COF CBLOF FDEPCA-CBLOF HBOS FDEPCA-HBOS Ionosphere 0.8743 0.8121 0.9287 0.9175 0.9142 0.8409 0.8858 0.9082 0.6473 0.9550 Satellite 0.5512 0.5726 0.6835 0.7280 0.5640 0.5953 0.7466 0.7537 0.7651 0.6738 Optdigits 0.5758 0.6042 0.4183 0.7269 0.3770 0.6273 0.9643 0.9663 0.8710 0.8948 <![CDATA[Mni s t]]> 0.6668 0.6839 0.8055 0.8069 0.7636 0.7817 0.7994 0.8298 0.6882 0.8461 Musk 0.4407 0.4458 0.4258 0.9687 0.5262 0.9202 0.8308 1 0.9989 1 Arrhythmia 0.7951 0.7999 0.8046 0.8211 0.7742 0.8372 0.7754 0.8389 0.8014 0.8282 Speech 0.5002 0.5644 0.4851 0.5807 0.4870 0.5141 0.4754 0.5622 0.4798 0.5467

[0150] Table 5: Comparison of AUC values ​​between the FDEPCA-based probabilistic anomaly detection algorithm and the original algorithm.

[0151]

[0152]

[0153] Table 6: Comparison of AUC values ​​between the FDEPCA-based integrated anomaly detection algorithm and the original algorithm.

[0154]

[0155] Experimental results show that most models perform better in anomaly detection than the original models after FDEPCA feature extraction. However, observations reveal that the anomaly detection accuracy of models based on nearest neighbor (LOF), KNN, and COF) algorithms on the Ionosphere dataset actually decreases after feature extraction. This indicates that the FDEPCA algorithm is more suitable for high-dimensional datasets and cannot extract effective key information from low-dimensional data. Analysis shows that FDEPCA can reduce the dimensionality of high-dimensional data while minimizing the loss of its inherent structural information. It effectively alleviates the "curse of dimensionality," quickly and accurately extracts nonlinear features from high-dimensional data, removes the influence of noisy data, and reduces the complexity of the learning model. Therefore, the FDEPCA algorithm has excellent applicability to various anomaly detection tasks involving high-dimensional data.

[0156] This invention conducts extensive comparative experiments on real datasets, using AUC to evaluate the anomaly detection performance of different nonlinear feature extraction methods. The FDEPCA algorithm is compared with current mainstream nonlinear feature extraction methods, and experimental results show that the proposed FDEPCA algorithm has significant advantages in detection performance and operational efficiency. The FDEPCA algorithm is extended to different anomaly detection techniques, applied to nearest-neighbor based, probability-based, and ensemble-based anomaly detection models respectively. Bayesian hyperparameter optimization is used to automatically extract the dimensionality m and dimensionality k of the FDEPCA algorithm. Experimental results show that the proposed FDEPCA algorithm has excellent adaptability in anomaly detection tasks.

[0157] This invention employs the above technical solution, extending the original fruit fly olfactory neural network as the basis for the locality-sensitive hashing retrieval mechanism to the foundation of a high-dimensional space nonlinear characteristic extraction mechanism in the field of anomaly detection. The FDEPCA algorithm mainly consists of three steps: first, mean centering is used to normalize the data; then, a sparse binary random projection matrix is ​​used to quickly and with low distortion embed the original data into a higher-dimensional transition space to extract nonlinear structure data; finally, PCA is used for linear dimensionality reduction to extract key feature information. This invention overcomes the problems of severe distortion of the inherent geometric structure and high computational load in current nonlinear dimensionality reduction methods when processing large-volume, high-dimensional anomaly data. Applying the FDEPCA algorithm to an anomaly detection model achieves excellent nonlinear feature processing results. The dimensionality-up projection method of this invention has good distance-preserving performance, and the more significant the dimensionality expansion, the better the maximum regression distance-preserving performance. Analysis proves the projection time performance of the FDEPCA algorithm. This invention demonstrates excellent performance in terms of distance preservation and time complexity.

[0158] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0159] References:

[0160] 【2017-S-CF-S】Dasgupta S, Stevens CF, Navlakha S.Aneural algorithm for a fundamental computing problem[J].Science,2017,358(6364):793-796.

[0161] 【2014-KK】Khalid S, Khalil T, Nasreen S.Asurvey of feature selection and feature extraction techniques in machine learning[C] / / 2014science and information conference.IEEE,2014:372-378.

[0162] 【2005-A-V】Lazarevic A,Kumar V.Feature bagging for outlier detection[C] / / Proceedings of the eleventh ACM SIGKDD international conference onKnowledge discovery in data mining.2005:157-166.

[0163] 【2008-FT-KM-ZH】Liu F T,Ting K M,Zhou Z H.Isolation forest[C] / / 2008eighth ieee international conference on data mining.IEEE,2008:413-422.

[0164] 【2016-T】 T.Loda:Lightweight on-line detector of anomalies[J].Machine Learning,2016,102(2):275-304.

[0165] 【2018-TR-KM-D】Bandaragoda T R,Ting K M,Albrecht D,et al.Isolation-based anomaly detection using nearest-neighbor ensembles[J].ComputationalIntelligence,2018,34(4):968-998.

[0166] 【2019-Y-Z-MK】Zhao,Y.,Nasrullah,Z.,Hryniewicki,M.K.and Li,Z.,2019,May.LSCP:Locally selective combination in parallel outlier ensembles.InProceedings of the 2019SIAM International Conference on Data Mining(SDM),pp.585-593.Society for Industrial and Applied Mathematics.

[0167] 【2021-Y-X-C】Zhao Y,Hu X,Cheng C,et al.SUOD:Accelerating large-scaleunsupervised heterogeneous outlier detection[J].Proceedings of MachineLearning and Systems,2021,3:463-478.

[0168] 【2000-MM-HP-RT】Breunig M M,Kriegel H P,Ng R T,et al.LOF:identifyingdensity-based local outliers[C] / / Proceedings of the 2000ACM SIGMODinternational conference on Management of data.2000:93-104.

[0169] 【2000-S-R-K】Ramaswamy S,Rastogi R,Shim K.Efficient algorithms formining outliers from large data sets[C] / / Proceedings of the 2000ACM SIGMODinternational conference on Management of data.2000:427-438.

[0170] 【2002-J-Z-AWC】Tang J,Chen Z,Fu AW C,et al.Enhancing effectiveness ofoutlier detections for low density patterns[C] / / Pacific-Asia conference onknowledge discovery and data mining.Springer,Berlin,Heidelberg,2002:535-548.

[0171] 【2003-Z-X-S】He Z,Xu X,Deng S.Discovering cluster-based local outliers[J].Pattern recognition letters,2003,24(9-10):1641-1650.

[0172] 【2009-HP-P-E】Kriegel H P, P,Schubert E,et al.Outlier detection inaxis-parallel subspaces of high dimensional data[C] / / Pacific-asia conferenceon knowledge discovery and data mining.Springer,Berlin,Heidelberg,2009:831-838.

[0173] 【2012-M-A】Goldstein M,Dengel A.Histogram-based outlier score(hbos):Afast unsupervised anomaly detectionalgorithm[J].KI-2012:poster and demotrack,2012,9.

[0174] 【2008-HP-M-A】Kriegel H P,Schubert M,Zimek A.Angle-based outlierdetection in high-dimensional data[C] / / Proceedingsof the 14th ACM SIGKDDinternational conference on Knowledge discovery and data mining.2008:444-452.

[0175] 【2011-JH-EO-JH】Janssens J H,Postma E O,van den Herik J H.Maritimeanomaly detection using stochastic outlierselection[C] / / MAD 2011 WorkshopProceedings.2011:121.

[0176] 【2013-M-K】Sugiyama M,Borgwardt K.Rapid distance-based outlierdetection via sampling[J].Advances in neuralinformation processing systems,2013,26.

[0177] 【2020-Z-Y-N】Li Z,Zhao Y,Botta N,et al.COPOD:copula-based outlierdetection[C] / / 2020 IEEE International Conferenceon Data Mining(ICDM).IEEE,2020:1118-1123.

[0178] 【2022-Z-Y-X】Li Z,Zhao Y,Hu X,et al.Ecod:Unsupervised outlierdetection using empirical cumulative distributionfunctions[J].IEEETransactions on Knowledge and Data Engineering,2022.

[0179] 【2015-D-A-S】Storcheus D,Rostamizadeh A,Kumar S.A survey of modernquestions and challenges in featureextraction[C] / / Feature Extraction:ModernQuestions and Challenges.PMLR,2015:1-18.

[0180] 【2006-YW-CJ】Feature extraction:foundations and applications[M].Springer,2008.

[0181] 【2003-IM-Andr】Guyon I,Elisseeff A.An introduction to variable andfeature selection[J].Journal of machine learningresearch,2003,3(Mar):1157-1182.

[0182] 【1998-B-A-KR】 B,Smola A,Müller K R.Nonlinear component analysisas a kernel eigenvalue problem[J].Neuralcomputation,1998,10(5):1299-1319.

[0183] 【2009-R】Jenssen R.Kernel entropy component analysis[J].IEEEtransactions on pattern analysis and machine intelligence,2009,32(5):847-860.

[0184] 【2012-S】Zafeiriou S.Subspace learning in krein spaces:Complete kernelfisher discriminant analysis with indefinitekernels[C] / / European Conferenceon Computer Vision.Springer,Berlin,Heidelberg,2012:488-501.

[0185] 【2000-JB-V-JC】Tenenbaum J B,Silva V,Langford J C.A global geometricframework for nonlinear dimensionalityreduction[J].science,2000,290(5500):2319-2323.

[0186] 【2008-MAA-TF】Carroll J D,Arabie P.Multidimensional scaling[J].Measurement,judgment and decision making,1998:179-250.

[0187] 【2000-ST-LK】Roweis S T,Saul L K.Nonlinear dimensionality reduction bylocally linear embedding[J].science,2000,290(5500):2323-2326.

[0188] 【2008-X-S】Li X,Shu L.Kernel based nonlinear dimensionality reductionfor microarray gene expression data analysis[J].Expert Systems withApplications,2009,36(4):7644-7650.

[0189] 【2000-ST-LK】Roweis S T,Saul L K.Nonlinear dimensionality reduction bylocally linear embedding[J].science,2000,290(5500):2323-2326.

[0190] 【2000-JB-V-JC】Tenenbaum J B,Silva V,Langford J C.A global geometricframework for nonlinear dimensionalityreduction[J].science,2000,290(5500):2319-2323.

[0191] 【2003-M-P】Belkin M,Niyogi P.Laplacian eigenmaps for dimensionalityreduction and data representation[J].Neuralcomputation,2003,15(6):1373-1396.

[0192] 【2003-X-P】He X,Niyogi P.Locality preserving projections[J].Advancesin neural information processing systems,2003,16.

[0193] 【2022-D-X-Q】Wei D,Shen X,Sun Q,et al.Neighborhood preservingembedding on Grassmann manifold for image-setanalysis[J].Pattern Recognition,2022,122:108335.

[0194] 【2009-GR-M】Rosman G,Bronstein M M,Bronstein AM,et al.Nonlineardimensionality reduction by topologicallyconstrained isometric embedding[J].International Journal of Computer Vision,2010,89(1):56-68.

[0195] 【2002-AM-AC】Martinez AM,Kak AC.Pca versus lda[J].IEEE transactions onpattern analysis and machine intelligence,2001,23(2):228-233.

[0196] 【2018-S-A-M】Smallman L,Artemiou A,Morgan J.Sparse generalisedprincipal component analysis[J].Pattern Recognition,2018,83:443-455.

[0197] 【2020-F-A】Dornaika F,Khoder A.Linear embedding by joint robustdiscriminant analysis and inter-class sparsity[J].NeuralNetworks,2020,127:141-159.

[0198] 【2020-HL-C】Li H,Zhang L,Huang B,et al.Cost-sensitive dual-bidirectional linear discriminant analysis[J].InformationSciences,2020,510:283-303.

[0199] 【2022-R-W-D】Zhou R,Gao W,Ding D,et al.Supervised dimensionalityreduction technology of generalized discriminantcomponent analysis and itskernelization forms[J].Pattern Recognition,2022,124:108450.

[0200] 【2006-Z-W】Zhang Z,Wang J.MLLE:Modified locally linear embedding usingmultiple weights[J].Advances in neuralinformation processing systems,2006,19.

[0201] 【2003-DL-C】Donoho D L,Grimes C.Hessian eigenmaps:Locally linearembedding techniques for high-dimensional data[J].Proceedings of the NationalAcademy of Sciences,2003,100(10):5591-5596.

[0202] 【2005-Z-H】Zhang Z,Zha H.Principal manifolds and nonlineardimensionality reduction via tangent space alignment[J].SIAM journal onscientific computing,2004,26(1):313-338.

[0203] 【2022-S-M-C】Sarfraz S,Koulakis M,Seibold C,et al.Hierarchical NearestNeighbor Graph Embedding for EfficientDimensionality Reduction[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and PatternRecognition.2022:336-345.

[0204] 【2018-L-J-J】McInnes L,Healy J,Melville J.Umap:Uniform manifoldapproximation and projection for dimensionreduction[J].arXiv preprint arXiv:1802.03426,2018.

[0205] 【2021-Y-H-C】Wang Y,Huang H,Rudin C,et al.Understanding How DimensionReduction Tools Work:An EmpiricalApproach to Deciphering t-SNE,UMAP,TriMap,and PaCMAP for Data Visualization[J].J.Mach.Learn.Res.,2021,22(201):1-73.

[0206] 【2008-DM-G】Laurens V D M,Hinton G.Visualizing Data using t-SNE[J].Journal of Machine Learning Research,2008,9(2605):2579-2605.

[0207] 【2002-FR-MI】Bach F R,Jordan M I.Kernel independent component analysis[J].Journal of machine learning research,2002,3(Jul):1-48.

[0208] 【2004-S-S】Lafon S S.Diffusion maps and geometric harmonics[M].YaleUniversity,2004.

[0209] 【2008-T-H】Lin T,Zha H.Riemannian manifold learning[J].IEEEtransactions on pattern analysis and machine intelligence,2008,30(5):796-809.

[0210] 【2019-E-MK】Amid E,Warmuth M K.TriMap:Large-scale dimensionalityreduction using triplets[J].arXiv preprintarXiv:1910.00204,2019.

[0211] 【2021-A-B-H】Narayan A,Berger B,Cho H.Assessing single-celltranscriptomic variability through density-preserving datavisualization[J].Nature Biotechnology,2021,39(6):765-774.

[0212] 【1964-JB】Kruskal J B.Multidimensional scaling by optimizing goodnessof fit to a nonmetric hypothesis[J].Psychometrika,1964,29(1):1-27.

Claims

1. A method for nonlinear feature extraction of medical Internet of Things (IoT) data, characterized in that: It includes the following steps: Step 1: Acquire medical IoT data and perform mean removal to obtain raw data so that the data features of the raw data are better clustered around zero, thereby improving the execution efficiency of model building; at the same time, divide the raw data into training set X, validation set V and test set T. Step 2: Construct an anomaly detection model suitable for medical IoT data. The specific execution steps of the anomaly detection model are as follows: Step 2-1, using a sparse, binary random projection matrix M m×n data of sample number n is embedded from the original feature space X d×n into a high-dimensional transition space Y m×n , the dimension m of the transition space is higher than the dimension d of the original feature space; the specific steps of step 2-1 are as follows: Step 2-1-1, constructing random projection matrix M m×n , random projection matrix M m×n Each row element of M is randomly selected as 1 in 6 positions and 0 in the rest, that is, it satisfies the constraint: Step 2-1-2, let x = (x1, x2, ..., x d ) T , x∈X d×n x represents a data object in the original feature space; y = (y1, y2, ..., y3) m ) T y∈Y m×n y represents the new data object generated after spatial transformation; Step 2-1-3, the process of dimensional projection, is as follows: y = Mx (2) Where M represents the random projection matrix M m×n ; Step 2-2: Use PCA to extract principal component information and transition the data from the high-dimensional transition space Y. m×n Projected onto the low-dimensional feature space Z k×n The nonlinear projection vector z is obtained, achieving linear dimensionality reduction; Step 3: Use the training set X and the validation set V to perform Bayesian hyperparameter optimization on the anomaly detection model to obtain the optimal hyperparameter set X. best Optimal hyperparameter set X best Including the optimal dimensionality increase dimension m and the PCA dimensionality reduction dimension k; the specific steps for Bayesian hyperparameter optimization of the anomaly detection model are as follows: S1, randomly initialize the hyperparameter set of the anomaly detection model: the dimension m of the up-dimensional projection and the dimension k of the PCA down-dimensionality reduction, and construct H from the initialized point set; S2, construct the surrogate model p based on the probability distribution of H; S3, combining the acquisition function of the surrogate model p, finds a new set of hyperparameters x. ′ ; S4. Evaluate the objective function f based on the validation set V and the new hyperparameter set; S5, {x ′ ,f(x ′ Add this to H and update the knowledge base; S6. Repeat steps S2-S5 until the maximum number of iterations is reached, and find the optimal hyperparameter set X. best = (m*, k*); Step 4, based on the optimal hyperparameter set X best The performance of the anomaly detection model was calculated using the test set T; Step 5: Select the best-performing anomaly detection model as the final detection model, and input the medical IoT data to be detected into the final detection model to obtain the corresponding nonlinear feature extraction results.

2. The nonlinear feature extraction method for medical IoT data according to claim 1, characterized in that: The specific steps of step 2-2 are as follows: Step 2-2-1, let y = (y1, y2, ..., y m ) T y∈Y m×n y represents the data object in the dimensional transition space; z = (z1, z2, ..., z k ) T , z∈Z k×n z represents a new data object in a low-dimensional feature space, and k <d<m; Step 2-2-2, the process of PCA retaining principal components is as follows: z = PCA(y) (3).

Images