Waveform-Prediction Augmentation and Deep Manifold Learning Enable Imbalanced Fault Diagnosis in Rotating Machinery Authors: Yu Tian, Shunsheng Guo, Yibing Li, Xin Zhang, Li Jiang Published: 2026 ABSTRACT Fault diagnosis of rotating machinery is essential for ensuring the reliable and safe operation of industrial equipment. However, imbalanced training data often bias intelligent diagnostic models toward majority classes, resulting in insufficient representation of minority faults. In addition, the high dimensionality and redundancy of vibration signals make it difficult to extract discriminative features and preserve meaningful neighborhood structures, thereby degrading diagnostic performance. To address these challenges, a waveform-prediction augmentation and deep manifold learning model for imbalanced fault diagnosis is proposed for rotating machinery. First, minority-class fault signals are augmented through time-series prediction using a nonlinear autoregressive neural network (NARNN), thereby alleviating the class imbalance at the data level. Subsequently, a deep manifold feature mapping (DMAP) is proposed to extract discriminative fault features. In DMAP, a stacked autoencoder (SAE) is employed to perform preliminary feature extraction on the training samples and obtain deep latent feature representations. The deep features extracted by the SAE are further processed using a dynamic time warping (DTW)-assisted manifold learning method for dimensionality reduction. In this process, DTW is used to calculate pairwise dissimilarities between deep representations and construct the neighborhood graph, thereby better preserving the local neighborhood relationships and topological structure of the deep features and producing more discriminative low-dimensional representations. Finally, a k-nearest neighbor classifier is adopted for fault classification. By integrating data augmentation, deep feature learning, and manifold dimensionality reduction, the proposed model effectively improves the fault diagnosis performance in imbalanced scenarios. Experiments on the bearing and planetary gearbox datasets demonstrate that, under an imbalance ratio of 0.1, the proposed method improves diagnostic accuracy by 10.2 and 1.6 percentage points, respectively. The comparative results further confirm its effectiveness in alleviating the diagnostic difficulties caused by class imbalance. 1. Introduction Rotating machinery is a fundamental component of modern industry and has been extensively applied in manufacturing, transportation, aerospace engineering, and wind power systems. Bearings and gears are key elements in rotating machinery. Their degradation or failure directly affects the operational stability and reliability of the entire system and may cause severe economic losses or accidents. Therefore, condition monitoring and fault diagnosis of rotating machinery are of great practical significance. In recent years, with the rapid development of machine learning and deep learning techniques, intelligent fault diagnosis (IFD) has attracted increasing attention and has gradually become an active research topic [1,2,3]. However, traditional IFD methods require sufficient and well-balanced training data to achieve satisfactory performance. In most cases, fault samples are usually much scarcer than normal samples due to the low occurrence probability of faults and the high cost of destructive experiments. This imbalance limits the ability of existing methods to adequately characterize minority fault classes, causing the classifier to be biased toward majority classes and degrading the diagnostic performance for minority samples in practical scenarios [4]. Moreover, the insufficient extraction of fault-discriminative information further weakens the generalization ability of the model [5]. Therefore, improving the diagnostic ability of intelligent models in imbalanced scenarios remains a critical challenge in rotating machinery fault diagnosis. To alleviate the adverse influence of class imbalance, various imbalanced fault diagnosis methods have been developed. These approaches are commonly divided into three categories: data-level, feature-level, and classifier-level methods. Feature-level methods improve the feature representation of minority fault classes through feature optimization, ensemble learning [6], and transfer learning [7]. Classifier-level methods, such as cost-sensitive learning [8] and boundary modulation [9], improve imbalanced diagnosis by modifying the training objective or optimization strategy of the classifier. As the most direct strategy, data-level methods aim to modify the composition of the training set to alleviate the imbalance. Cao et al. used multi-step downsampling to capture multi-scale temporal representations from raw vibration signals to improve small-sample fault diagnosis performance [10]. Dong et al. improved minority sample representation through density-aware oversampling [11]. Ma et al. proposed an abductive reasoning and physical model guided multi-scale synthetic minority oversampling technique (ARPMG-MS-SMOTE) framework that uses multi-scale synthetic minority oversampling technique (SMOTE) to synthesize structurally coherent minority bearing fault samples from physics-informed latent features, thereby improving fault diagnosis under extreme class imbalance and limited faulty data [12]. Wang et al. proposed an model-data-driven sample generation (MDDSG) framework that uses a threshold-guided generative adversarial network (GAN) to generate high-quality synthetic bearing fault samples, thereby alleviating data imbalance and improving diagnostic robustness under scarce fault data [13]. However, downsampling may discard useful fault information, while many existing sample-level data expansion methods increase the risk of overfitting and are often insufficient to characterize the nonlinear dynamic evolution of vibration signals. Moreover, the temporal dependencies and dynamic evolution patterns inherent in fault vibration signals are not fully considered. To address this issue, data augmentation based on time series prediction provides a feasible way to extend minority vibration signals while preserving their temporal continuity. Previous studies have shown that nonlinear autoregressive neural networks (NARNNs) has strong predictive capability for nonlinear and non-stationary periodic signals [14]. Wang et al. used an NARNN to predict future degradation trends from the extracted vibration-signal time-series features [15]. Gong et al. applied an NARNN to predict future trends of aging features from historical data, thereby providing extended feature inputs for accurate battery aging trajectory prediction [16]. Gidom et al. investigated an optimized NARNN model for short-term wind speed forecasting, demonstrating its strong capability in predicting stochastic wind-speed time-series data with competitive accuracy and lower computational cost [17]. More directly, Zhou et al. proposed an imbalanced fault diagnosis method where NARNN is successfully used to predict and expand minority fault samples, thereby balancing different health-condition datasets before continuous wavelet transform (CWT)-based time–frequency transformation and convolutional neural network (CNN) classification [14]. Therefore, by predicting the subsequent evolution of minority fault signals, NARNNs can generate additional samples with temporal continuity, thereby alleviating the imbalance problem while retaining time-series-related fault information. However, the augmented high-dimensional vibration samples still contain redundant and nonlinear information. Therefore, extracting compact and discriminative fault features from these samples remains a further challenge. Conventional dimensionality reduction methods, including principal component analysis (PCA) [18] and linear discriminant analysis (LDA) [19], have been commonly adopted for feature representations. PCA projects high-dimensional data onto directions with maximum variance, while LDA aims to find a discriminative subspace that improves class separability. However, these methods are essentially linear and may be insufficient to describe the complex nonlinear structures embedded in vibration signals. To capture the underlying geometry of complex high-dimensional data, manifold learning methods have been increasingly applied to fault feature extraction, such as local linear embedding (LLE) [20], Laplacian eigenmaps (LE) [21], locality preserving projection (LPP) [22], Isometric feature mapping (Isomap) [23], and local tangent space alignment (LTSA) [24]. Among these methods, uniform manifold approximation and projection (UMAP), proposed by McInnes et al. [25], has received considerable attention because it can preserve local neighborhood information while retaining part of the global data structure. Specifically, UMAP builds a fuzzy topological graph in the original feature space and learns a compact low-dimensional embedding through optimization, showing favorable performance in visualization and dimensionality reduction tasks. On this basis, McConville et al. developed the (Not Too) Deep Clustering framework (N2D), in which an autoencoder is first used to learn latent representations, followed by UMAP to uncover the underlying manifold [26]. This demonstrates the advantage of combining deep feature learning with nonlinear manifold representation. Chen et al. introduced adaptive weighted distance manifold learning (AWDU) into bogie bearing fault diagnosis, where adaptive weighted distance and adjacency relationships were used to guide the low-dimensional embedding of high-dimensional vibration features [27]. However, most existing manifold learning methods rely on Euclidean distance to construct neighborhood graphs, which may not adequately capture the nonlinear structure of vibration-signal features and may therefore lead to inaccurate neighborhood connections. Dynamic time warping (DTW) [28] provides a flexible dissimilarity measure by capturing nonlinear correspondences between signal representations and has been widely applied to fault diagnosis. Gao et al. proposed dynamic time warping Gaussian spatiotemporal self-attention network (DGSSANet) for fault diagnosis in intermittent industrial processes, where DTW is used to transform unequal-length sensor signals into equal-dimensional warping matrices and enhance the modeling of temporal similarity [29]. Sun et al. introduced a two-stage DTW-based fault detection method for rotating machinery under variable speeds, where a fixed-speed reference signal is estimated to guide DTW in determining optimal warping paths for signal regularization, and order-domain DTW distances are further used for fault type identification [30]. Li et al. proposed semisupervised high-order graph diffusion network (SHGDN) for cross-domain rotating machinery fault diagnosis, in which FastDTW is used to measure the similarity of discriminative subsequences called Shapelets and construct a dynamic graph with semantic perception ability [31]. Therefore, DTW-derived similarity can better characterize nonlinear relationships between signal representations and support the construction of a more reliable neighborhood graph, thereby facilitating the preservation of local topological structure during low-dimensional mapping. Based on the above analysis, two limitations can be identified in existing imbalanced fault diagnosis studies:(1)Most data augmentation methods mainly improve class distribution by increasing the number of minority class samples, but they pay insufficient attention to the temporal evolution characteristics of fault signals.(2)The high dimensionality and redundancy of vibration signals pose challenges to discriminative feature extraction and neighborhood construction. Moreover, different dissimilarity functions affecting neighborhood construction are insufficiently considered. (1)Most data augmentation methods mainly improve class distribution by increasing the number of minority class samples, but they pay insufficient attention to the temporal evolution characteristics of fault signals. Most data augmentation methods mainly improve class distribution by increasing the number of minority class samples, but they pay insufficient attention to the temporal evolution characteristics of fault signals. (2)The high dimensionality and redundancy of vibration signals pose challenges to discriminative feature extraction and neighborhood construction. Moreover, different dissimilarity functions affecting neighborhood construction are insufficiently considered. The high dimensionality and redundancy of vibration signals pose challenges to discriminative feature extraction and neighborhood construction. Moreover, different dissimilarity functions affecting neighborhood construction are insufficiently considered. Aimed at the two key issues, this paper proposes a waveform-prediction augmentation and deep manifold learning for imbalanced fault diagnosis in rotating machinery. An NARNN is first used to augment minority-class fault signals. An SAE then extracts compact latent representations from the augmented samples. Subsequently, DTW is incorporated into UMAP to achieve similarity measurement during manifold construction. The main contributions of this paper are summarized as follows. (1)An NARNN-based signal augmentation strategy is developed for minority fault classes. By exploiting time-series prediction, this strategy expands minority-class samples and reduces class imbalance, while retaining the temporal dynamic information of vibration signals. An NARNN-based signal augmentation strategy is developed for minority fault classes. By exploiting time-series prediction, this strategy expands minority-class samples and reduces class imbalance, while retaining the temporal dynamic information of vibration signals. (2)A deep manifold feature mapping (DMAP) method is developed by combining an SAE and DTW-enhanced UMAP. The SAE is used to extract compact latent features, while DTW is introduced as a dissimilarity function to construct a neighborhood graph during manifold construction. A deep manifold feature mapping (DMAP) method is developed by combining an SAE and DTW-enhanced UMAP. The SAE is used to extract compact latent features, while DTW is introduced as a dissimilarity function to construct a neighborhood graph during manifold construction. (3)An integrated imbalanced fault diagnosis framework is established by combining data-level augmentation, deep feature learning, manifold dimensionality reduction, and KNN classification. The framework improves the representation and recognition ability of minority fault categories. An integrated imbalanced fault diagnosis framework is established by combining data-level augmentation, deep feature learning, manifold dimensionality reduction, and KNN classification. The framework improves the representation and recognition ability of minority fault categories. (4)Experiments are conducted on bearing and gearbox datasets under different imbalance ratios. The results indicate that the proposed method achieves better diagnostic performance than several representative methods, demonstrating its effectiveness for imbalanced rotating machinery fault diagnosis. Experiments are conducted on bearing and gearbox datasets under different imbalance ratios. The results indicate that the proposed method achieves better diagnostic performance than several representative methods, demonstrating its effectiveness for imbalanced rotating machinery fault diagnosis. The remainder of this paper is structured as follows. Section 2 introduces the principles of the NARNN, SAE, and UMAP. Section 3 describes the proposed fault diagnosis framework in detail. Section 4 reports the experimental validation on bearing and gearbox datasets. Section 5 concludes the paper and discusses future work. 2. Theoretical Basis 2.1. Nonlinear Autoregression Neural Network A nonlinear autoregression neural network (NARNN) is a nonlinear extension of the traditional autoregressive model. It replaces the linear mapping in the autoregressive model with a nonlinear function approximated by a feedforward neural network. Therefore, the NARNN has been widely used for time-series forecasting tasks involving complex dynamics. For a univariate time series, NARNN predicts the current output based on several previous outputs. Its general form can be written as:(1)x(t)=f(x(t−1),x(t−2),x(t−3)…x(t−n)) where x(t) represents the output at time t, f(·) denotes the nonlinear mapping function learned by the neural network, and n is the delay order. Therefore, the current output is related to the outputs of the preceding n time steps. Figure 1 shows the basic topology of the NARNN model. The network mainly consists of an input layer, hidden layers, and an output layer. In this configuration, x(t) represents the input signal of the NARNN, and 1:10 denotes the selected delay range. The value 10 below the hidden layer indicates that ten hidden neurons are used. W and b represent the connection weights and biases, respectively. The detailed structure of the NARNN is presented in Figure 2. Let xi denote the sampled data point of the original signal. Given the connection weight wij between the input node and the hidden neuron, together with the threshold aj, the output of the j-th hidden neuron Hj can be calculated as:(2)Hj=f(∑i=1nwijxi+ai),j=1,2,….,l where i is the index of the input data, l denotes the number of hidden neurons, and f(·) is the activation function of the hidden layer. Specifically, xi represents the i-th input datum, wij denotes the connection weight from the i-th delayed signal to the j-th hidden neuron, and aj is the threshold of the j-th hidden neuron. According to the hidden-layer output Hj, the overall network output O can be obtained through a linear operation, which is expressed as(3)O=f(∑j=1lHjwj+b) where wj represents the weight between the j-th hidden neuron and the output neuron, and b denotes the threshold of the output neuron. 2.2. Stacked Autoencoder A stacked autoencoder (SAE) is a deep feature learning model constructed by stack-ing multiple autoencoders layer by layer. Its fundamental principle is to learn representative latent features from raw data through unsupervised reconstruction. A single autoencoder is mainly composed of an encoding module and a decoding module. For an input sample x, the encoding module transforms it into a latent representation h through a nonlinear mapping:(4)h=f(Wx+b) where W denotes the weight matrix and b denotes the bias vector; f(·) is the nonlinear activation function. The decoding module then reconstructs the input from the latent representation:(5)x^=g(W′h+b′) where g(·) is the decoder activation function and x^ denotes the reconstructed output. The training objective is to minimize the difference between the reconstructed output and the original input. Here, the mean squared error (MSE) is used to measure the reconstruction loss:(6)L=1N∑i=1Nxi−x^i22 As illustrated in Figure 3, the latent representation learned by each autoencoder is used as the input to the subsequent layer, forming a hierarchical feature extraction framework. Through successive encoding operations, the original high-dimensional data are gradually compressed into a low-dimensional latent space, while the decoder imposes reconstruction constraints to preserve essential information. Owing to its nonlinear representation capability, the SAE can effectively capture the complex structural characteristics embedded in vibration signals and generate compact yet informative deep feature representations for subsequent fault diagnosis. 2.3. Uniform Manifold Approximation and Projection Uniform manifold approximation and projection (UMAP) is a manifold learning algorithm introduced by McInnes et al. in 2018 [25]. The method was developed for scalable dimensionality reduction and is built on concepts from Riemannian geometry and algebraic topology. Unlike t-SNE, which mainly emphasizes local structure preservation, UMAP is able to maintain neighborhood relationships while retaining certain global structural information. Therefore, it has been widely used for feature visualization and low-dimensional representation learning. Its relatively high computational efficiency also makes it suitable for processing large-scale data. As a dimensionality reduction algorithm based on a k-nearest neighbor graph, UMAP mainly consists of two steps: graph construction and graph layout. In graph construction, a k-nearest neighbor graph is first generated using a selected distance metric to describe the local relationships between samples. The neighborhood graph is then transformed into a weighted fuzzy simplicial set, which is used to describe the manifold information contained in the high-dimensional data. In the graph layout stage, UMAP constructs a corresponding fuzzy simplicial set in the low-dimensional space and optimizes the embedding by reducing the discrepancy between the high-dimensional and low-dimensional graphs. This discrepancy is usually measured by a cross-entropy objective function. During the optimization process, samples with high similarity are encouraged to remain close to each other, whereas samples with low similarity are separated in the embedding space. The objective function is commonly optimized using stochastic gradient descent (SGD). 3. The Proposed Method This section presents the proposed imbalanced fault diagnosis model based on waveform-prediction augmentation and deep manifold learning. As shown in Figure 4, vibration signals from different types of rotating machinery are first collected using traditional sensors such as accelerometers. Then, an NARNN is employed to expand the minority fault signals, and the balanced vibration signals are further normalized and segmented. Meanwhile, a stacked autoencoder is employed to extract deep fault features from the preprocessed samples. Subsequently, DTW is introduced to calculate pairwise distances between the extracted features and construct a neighborhood graph for manifold approximation and projection, yielding low-dimensional representations for fault classification. Finally, the obtained features are visualized and fed into a classifier to complete fault state identification. A detailed description of the main stages is illustrated below. The flowchart of the proposed method. 3.1. Data Balancing and Preprocessing Assume that the original vibration signal dataset contains C classes, where the length of the normal-state signal is M0 and the length of the c-th fault signal is Mc. An NARNN is employed to augment the imbalanced fault data and the length of the normal state signal is taken as the balancing reference. The c-th fault signal can be expressed as(7)Sc={s1,s2,…,sMc} According to Equation (1), the NARNN establishes a nonlinear autoregressive mapping relationship based on historical time-series signals, and the prediction process can be expressed as(8)s^t=f(st−1,st−2,…,st−p) where p denotes the time-delay order, f(·) is the nonlinear mapping function learned by the NARNN, and s^t represents the predicted signal. The balanced fault signal S˜c is obtained by concatenating the original fault signal with an extended sequence of length M0 − Mc:(9)S˜c={s1,s2,…,sMc,s^Mc+1,…,s^M0} Then, Z-score normalization is applied to the balanced data of each class to minimize the effect of amplitude variations on model training and distance measurement. The normalized signal s¯i is given by(10)s¯i=si−μiσi where μi and σi denote the mean and standard deviation of the corresponding sample, respectively. Subsequently, a sliding window with length L and step size r is used to segment the normalized signals. The sample set X={x1,x2,…,xN} (X∈RN×L) and the corresponding label set Y={y1,y2,…,yN} (Y∈RN×1) are obtained. Finally, the dataset is partitioned into the training set Xtr={x1,x2,…,xNtr}∈RNtr×L, the test set Xte={x1,x2,…,xNte}∈RNte×L according to a predefined ratio, Ntr and Nte are the number of training and testing samples, respectively. 3.2. Deep Manifold Feature Mapping After augmentation, the vibration signals of all classes maintain a consistent signal length. However, the normalized and segmented samples still contain noise and redundant information, which may degrade the stability and reliability of fault diagnosis, when they are directly fed into the subsequent classifier. To address this issue, a deep manifold feature mapping module is proposed to extract discriminative low-dimensional fault features from the balanced vibration samples. It can be seen in Figure 4 that the proposed module is implemented through three stages: deep feature extraction using an SAE, DTW-based neighborhood construction, and feature mapping and dimensionality reduction with manifold structure preservation. 3.2.1. Stage 1: Deep Feature Extraction Using an SAE First, the balanced training samples are processed by the SAE encoder for deep feature extraction. The SAE consists of an encoder and a decoder. The encoder contains three fully connected layers and maps each vibration sample to a 100-dimensional latent feature vector. ReLU activation functions are used in the hidden layers, while a sigmoid activation function is employed in the output layer of the decoder to reconstruct the input signal. The SAE is trained in an unsupervised manner by minimizing the discrepancy between the input samples and their reconstructed outputs. In the training process, Adam is selected as the optimizer, with the learning rate, batch size, and maximum number of epochs set to 0.001, 32, and 100, respectively. To avoid overfitting, an early-stopping strategy is introduced. Specifically, if the validation loss fails to decrease for five consecutive epochs, the training procedure is terminated, and the model parameters with the minimum validation loss are retained for subsequent feature extraction. After training, the decoder is removed, and only the encoder is retained to map the balanced training samples into a low-dimensional latent feature space. The obtained deep latent features are denoted as(11)H={h1,h2,…,hNtr}∈RNtr×Lh where Lh denotes the dimensionality of the SAE-extracted latent features. In this study, Lh is set to be 100. These compact representations reduce redundancy information and provide the input for the subsequent DTW-based neighborhood construction. 3.2.2. Stage 2: DTW-Based Neighborhood Construction Dynamic time warping (DTW) is a path-based distance measure that quantifies sample dissimilarity by identifying the minimum-cost alignment path through a local cost matrix. It first constructs a Euclidean distance matrix and then applies dynamic programming to identify the optimal warping path by backtracking through this matrix. In the proposed framework, DTW is used to calculate pairwise distances between the SAE-derived features. The resulting distances are then used to identify neighboring samples and construct the neighborhood graph for subsequent manifold embedding. Let X={x1,x2,…,xn} and Y={y1,y2,…,ym} be two time-series signals with lengths n and m, respectively, where i and j index the elements of X and Y, respectively. First, an n × m local cost matrix d is constructed. Each element d(i,j) denotes the local matching cost between xi and yj, which is calculated using the Euclidean distance as follows:(12)d(i,j)=(xi−yj)2 Then, the cumulative cost matrix is computed recursively using Bellman’s recursion. Therefore, the DTW distance between X and Y can be obtained from the minimum cumulative matching cost as follows:(13)D(i.j)=d(i,j)+minD(i−1,j)D(i,j−1)D(i−1,j−1)(14)DTW(X,Y)=D(n,m) where D(i,j) denotes the minimum cumulative matching cost from the starting point (1,1) to the current position (i,j), and D(n,m) represents the final DTW distance between the two sequences. Figure 5 illustrates the computational differences between DTW and Euclidean distance. The central grid represents the local cost matrix, where dark red and light blue regions indicate high and low local costs, respectively. The red dashed diagonal denotes the fixed one-to-one correspondence used by Euclidean distance, whereas the blue curve shows the minimum cumulative-cost path obtained by dynamic programming. Unlike the fixed diagonal correspondence, DTW provides a more flexible matching scheme by allowing non-diagonal transitions, causing the two distance measures to produce different pairwise distances and neighborhood relationships for the same feature representations. Direct pairwise DTW calculations can become computationally expensive for large-scale vibration datasets. Therefore, FastDTW is considered an optional approximation for efficiently computing pairwise dissimilarities between SAE-derived features, particularly when extending DMAP to large-scale datasets or online monitoring applications [32]. Given two latent feature vectors hi={hi1,hi2,…,hiLh} and hj={hj1,hj2,…,hjLh}, their FastDTW distance is defined as(15)dij=FastDTW(hi,hj)=Dij(Lh,Lh) where Dij(Lh, Lh) is the terminal element of the cumulative cost matrix, representing the minimum accumulated matching cost corresponding to the optimal warping path from the starting point (1,1) to (Lh, Lh). The recursive calculation of the cumulative cost matrix is given in Equations (5) and (6). The similarity matrix is obtained as (16)DFastDTW=0d12…d1Ntrd210…d2Ntr⋮⋮⋱⋮dNtr1dNtr2…0 where the element dij denotes the DTW distance between sample hi and sample hj. Based on this distance matrix, the k-nearest neighborhood Nk(x^i) of each sample is constructed. 3.2.3. Stage 3: Manifold Feature Embedding For any sample hi and its neighboring sample, the directed fuzzy membership degree is defined as(17)μ(x^i,x^j)=exp−max(0,d(x^i,x^j)−ρi)σi where d(x^i,x^j) is the DTW distance between point xi and xj; ρi is the local connectivity constraint used to ensure that points connect to at least one data point with an edge of weight 1; and σi is the normalization factor that controls the smoothness of the neighbors. The calculations of ρi and σi are derived from the following:(18)ρi=min{d(xi,xij)|1≤j≤k,d(xi,xij)>0}(19)∑j=1kexp(−max(0,d(xi,xij)−ρi)σi)=log2(k) Since the neighborhood relationship between samples is directional and usually asymmetric, a fuzzy set operation is adopted to symmetrize the directed fuzzy membership degrees, thereby obtaining an undirected weighted neighborhood graph:(20)Bij=μ(x^i,x^j)+μ(x^j,x^i)−μ(x^i,x^j)μ(x^j,x^i) where Bij denotes the fuzzy adjacency weight between samples in the high-dimensional space. The fuzzy adjacency weights of all samples constitute a high-dimensional weighted graph G = (V,E,B), where V is the sample set, E is the edge set, and B is the edge weight matrix. In the graph layout stage, assume that the representation of samples in the low-dimensional space is Z={z1,z2,…,zNtr}∈RNtr×d, where d is the embedding dimensionality. The fuzzy membership degree between samples in the low-dimensional space is defined as(21)νij=1+a‖zi−zj‖2b−1 where zi−zj denotes the distance between two points in the low-dimensional space, and a and b are hyperparameters used to control the scale and density distribution of the low-dimensional embedding. These two parameters are determined by curve fitting according to the preset parameters dmin and s, dmin denotes the minimum distance between sample points in the low-dimensional space, and s represents the distribution scale of the low-dimensional embedding. During the calculation, the target distance decay function ϕ(r) is first constructed. When the distance between two low-dimensional points is smaller than dmin, a high similarity is maintained; when the distance is larger than dmin, the similarity gradually decays as the distance increases:(22)ϕ(r)=1,0≤r