86 results about "Kernel extreme learning machine" patented technology

The invention discloses a kernel extreme learning machine flood forecast method based on sparse self-encoding. The method comprises the following steps that (1) medium and small river flood data are selected, and the data are collated and cleaned; (2) the appropriate forecast factors are selected and samples are collated, and the sample data are preprocessed; (3) unsupervised learning is performedon the original sample data through a multilayer sparse self-encoder, and the optimal network layer parameters are trained; and (4) the sample data through multilayer sparse self-encoding learning act as the input of a KELM model, an SAE_KELM model is constructed and the corresponding result is predicted and evaluated. The SAE method is fused on the basis of the KELM model, a deep network model is constructed, and the abundant intrinsic information between the complex data can be learnt by increasing the number of layers of the model. The "essential" characteristics of the original hydrological data can be learnt by the SAE_KELM model, the learnt characteristics have more essential depiction for the data, and the mapping relationship between the characteristic value and the target value can be better fit by the KELM model.

The invention discloses a method for predicting the remaining service life of a rolling bearing integrated with the KELM (Kernel Extreme Learning Machine), and belongs to the technical field of the bearing service life prediction. The method is used to solve the problem that the prediction of the remaining service life of the rolling bearing has difficulty in prediction and low prediction accuracy. The method firstly extracts features of a vibration signal based on the variational mode decomposition, introduces a new similarity dimension reduction method for features dimension reduction, and further extracts the features-CEF (Cyclic Enhancement Features) with strong monotonicity, similarity, and stability. Multiple KELM models are constructed through that the CEF extracted by the multiplebearings is used as the input of the KELM, the ratio of the current service life to the whole life, p, that is, the life percentage is used as the output. A prediction model integrated with KELM is constructed by combining the random forest to obtain a current prediction result p value. The CEF of the test bearing is input into the prediction model, the current p value is predicted, and the secondorder exponential smoothing method is used for fitting to predict the RUL of the bearing. The experimental verification shows that the proposed prediction method has higher prediction accuracy than other literatures.

The invention discloses a bearing fault diagnosis method combining an improved sparse filter and a KELM. The method comprises the following steps: embedding a Min-Max regular term into an original sparse filter to obtain an improved sparse filter. The Min-Max regular term can describe the internal structure information of the original data, and promotes the similar samples to be close to each other and promotes the samples of different classes to be separated from each other, thereby generating discriminative characteristics. Feature discrimination mainly lies in that class label information is used in construction of the Min-Max regular term, and a pseudo label is used for replacing a real label to guide the construction of the Min-Max regular term. Vibration signals of different operation conditions of a rolling bearing are collected to serve as a training set, the training set is used for training an improved sparse filter model and a kernel extreme learning machine model to obtainmodel parameters, and therefore establishment of a fault diagnosis classification model is completed, and the diagnosis model can accurately identify the rolling bearing fault.

The invention discloses a gas path fault diagnosis method and system for aeroengine dynamic process. The method includes the steps of establishing a feature extraction model based on a multi-layer kernel extreme learning machine; adopting a hidden Markov model based on time series modeling for fault identification. The method solves the problem that the conventional data-based engine fault diagnosis uses time series measurement data to diagnose the fault with low accuracy in the existing aeroengine gas path fault diagnosis in the dynamic process, is suitable for the engine dynamic fault diagnosis in consideration of the degradation of gas path components and the redundancy of sensor parameters, and has a positive promotion effect on engine health management and maintenance cost reduction.

The invention discloses a random forest classification method used for coronary heart disease data classification and based on a kernel extreme learning machine and parallelization. Sampling with replacement is performed on a coronary heart disease sample set by using a Bootstrap method so that different coronary heart disease data training subsets and test subsets are generated to be used for base classifiers; the kernel function of the hybrid kernel form is used as the kernel function of the kernel extreme learning machine so that the influence of the kernel type on the performance of the classification model can be reduced; model training is performed on the kernel extreme learning machine by using the coronary heart disease data training subsets and performing testing is performed on the base classifiers by using the test subsets, cyclic judgment is performed by using the mode of sorting and particle swarm optimization and the optimized new base classifiers are regenerated and thebase classifiers having poor classification performance are eliminated and substituted so that the objective of enhancing the overall classification performance can be achieved; and the random forestmodel is formed and then the classification result is selected by using a relative majority voting method.

The invention relates to a transformer fault diagnosis method, which comprises the following steps of S1, obtaining the concentration of dissolved gas in transformer oil and sample data correspondingto a fault conclusion, performing preprocessing, and generating a training sample set and a test sample set; s2, establishing a kernel extreme learning machine prediction model by adopting the generated training sample set; s3, optimizing kernel function parameters and penalty coefficients of the kernel extreme learning machine by adopting a crisscross algorithm in the model training process; andS4, inputting the test sample into a trained kernel extreme learning machine for prediction to obtain a transformer fault diagnosis result. According to the transformer fault diagnosis method, the problem that transformer fault data encoding and kernel extreme learning machine parameter selection are difficult is effectively solved; meanwhile, the local optimization problem of a traditional BP neural network is avoided, the method can be applied to scientific research and engineering application in the related fields of transformers, the recognition speed is high, the recognition rate is high,and the diagnosis precision of transformer faults is greatly improved.

The invention discloses an intelligent decomposition control planning method for paths of a carrying robot in an intelligent environment. The method includes the following steps: step one, constructing a global map three-dimensional coordinate system for carrying areas of a carrying robot, and obtaining a walkable area coordinate in the global map three-dimensional coordinate system; step two, obtaining a training sample set; step three, constructing a global static path planning model of the carrying robot; and step four, obtaining an optical path in real time, and then finishing a transportation task. A path planning model is established through construction of a kernel extreme learning machine optimized by a wolf pack algorithm, a global optimal solution can be quickly found in an intelligent environment, and the problem of local optimum of conventional path planning is avoided.

The invention discloses a satellite power source main bus-bar current interval prediction method. According to the method, based on a prediction model trained by an optimized kernel extreme learning machine, a prediction interval is determined by using a proportionality coefficient method, and the parameters of the proportionality coefficient method are optimized by using a differential evolution algorithm. The method specifically includes the following steps that: satellite power source main bus-bar bus current data are preprocessed, noise data can be removed, and normalized data can be obtained; the parameters of the kernel extreme learning machine are optimized by adopting the differential evolution algorithm; the optimized kernel extreme learning machine is adopt to construct an initial prediction model; comprehensive indexes for evaluating the quality of the prediction interval are given, the prediction interval is determined through adopting the proportionality coefficient method, and the satisfaction degree of the prediction interval is evaluated; and the prediction proportionality coefficient of the interval is optimized through using the differential evolution algorithm, so that an optimal satellite power source main bus-bar current prediction interval can be obtained. The satellite power source main bus-bar current interval prediction method of the invention is based on complicate satellite power source main bus-bar current data, and has the advantages of higher prediction accuracy and better effect.

The invention discloses a fault diagnosis method for a centrifugal pump rotor system, particularly relates to a fault diagnosis method for the centrifugal pump rotor system based on variational mode decomposition and a kernel extreme learning machine, and belongs to the field of rotary machine fault diagnosis. The method mainly comprises the steps that S1, acquiring vibration acceleration signals(normal, rotor misalignment, rotor imbalance, bearing inner ring faults, bearing outer ring faults and bearing rolling body faults) of a centrifugal pump rotor system in a normal state and a fault state, and obtaining a time domain signal sample set; S2, carrying out variational mode decomposition on the obtained time domain signal sample set to obtain an intrinsic mode function component; S3, solving an energy value, a waveform factor, a pulse index, a margin coefficient, a peak factor and a kurtosis value for each intrinsic mode function component; S4, constructing a feature matrix, and normalizing the processed data; S5, training the fault diagnosis system of the centrifugal pump rotor system by adopting the training sample; S6, inputting the test sample or the real-time sample into thecentrifugal pump rotor system fault diagnosis model to diagnose the fault.

The invention relates to wireless locating. For improving the locating precision, reducing the sample data dimension and increasing the locating speed, a locating prediction model is obtained. The technical scheme adopted by the invention is an improved kernel extreme learning machine locating method. The method comprises the steps of firstly, obtaining training data by adopting a method for performing multi-time measurement in a same position; secondly, dividing the data measured in the same position into a sample sub-space, extracting features of the sample sub-space, and replacing originaltraining data with the features of the sample sub-space; meanwhile, improving a kernel extreme learning machine algorithm by utilizing related theories of matrix approximation and matrix extension; and finally, training the obtained processed training data by utilizing an improved kernel extreme learning machine to obtain the locating prediction model, and performing position estimation by using the obtained locating prediction model to achieve the purpose of locating. The method is mainly applied to the wireless locating occasion.

The invention is suitable for a field of information technology and provides a model parameter optimizing method and device. The method includes obtaining sample data and performing standardizing treatment on the sample data; obtaining the optimal penalty coefficient C and the optimal kernel scale Gamma of a kernel extreme learning machine through combining with a Moth optimization algorithm of chaos theory by adopting the sample data subjected to standardization; according to the sample data subjected to standardization, the optimal penalty coefficient C and the optimal kernel scale Gamma, constructing a target classification predication model. The invention solves a problem that the optimal penalty coefficient C and the optimal kernel scale Gamma cannot be obtained by utilizing a grid searching method in the prior art and is beneficial to improvement of effect of the constructed model in classification and predication of determined problems.

The invention proposes a short-period load prediction method for a microgrid based on SPSS and RKELM, and the method comprises the steps: (1), carrying out the online data collection, and periodicallyupdating a historical database; (2), carrying out the preprocessing of historical data, and extracting load sample features; (3), constructing an offline load prediction model; (4), screening a historical sample similar to a to-be-predicted point precursor load as an online training sample through SRC (Spearman Rank Correlation); (5), calculating a load prediction value at a future moment according to the online training sample and the offline load prediction model. The method employs a rapid RKELM (Reduced Kernel Extreme Learning Machine), a chaos particle swarm optimization algorithm and the SRC, and achieves the building of a prediction model comprising offline parameter optimization and an online load. Through the periodic updating of model parameters, the method guarantees the timeliness of an algorithm, reduces the complexity of online prediction and calculation, reduces the storage quantity of historical data, reduces the calculation cost, and can achieve the more accurate prediction of the short-period and super-short-period loads of the microgrid.

The invention discloses a kernel extreme learning machine based quick traffic sign detecting method and belongs to a field of image signal processing and mode recognition. The method includes reading an original sample image; utilizing a BING based objectness method for producing an area that may contain traffic signs; extracting HOG features of the candidate area and sending the features to a kernel extreme learning machine classifier; and obtaining a final detection result. According to the invention, a traditional slide window scanning method is abandoned. The BING algorithm is used for reducing search space and improving detection speed. A traditional ELM algorithm has a single hidden layer structure and has huge boundedness in complicated signal analysis. The invention adopts KELM (Kernel Extreme Learning Machine) for classification detection. The kernel extreme learning machine improves the stability of a learning model and enhances the generalization performance, improves the detection performance and keeps an advantage of low time consumption of ELM (Extreme Learning Machine).

The invention discloses an emotion recognition method and system based on electroencephalogram signals. The method comprises the following steps: acquiring to-be-identified multi-channel electroencephalogram signals that are electroencephalogram signals from multiple channels when a to-be-identified person watches videos capable of stimulating different emotions; carrying out feature extraction onthe to-be-identified multi-channel electroencephalogram signals based on a discrete wavelet transform algorithm to obtain electroencephalogram features, including frequency band entropies and frequency band energy, of all channels; according to a minimum redundancy maximum correlation algorithm, performing feature selection on the electroencephalogram features to obtain electroencephalogram feature selection signals; and classifying the electroencephalogram feature selection signals by adopting a kernel extreme learning machine algorithm to obtain an electroencephalogram signal emotion recognition result. The emotion recognition precision can be improved.

The invention discloses a kernel extreme learning machine-based virtual extension method for leaked gas monitoring concentration data, and relates to the technical field of dangerous chemicals. The extension method comprises the steps of firstly selecting position point coordinates Xs and Ys and concentration data of a monitored space region S1 as a training sample set, wherein coordinate values are input values of a network, and the concentration data serves as an output value of the network, so that the network is constructed and trained; and secondly determining coordinates (XPn, YPn) according to space positions S2-S1 of virtual monitoring points needed to be extrapolated or interpolated, wherein n is a predicted point number, the predicted point number forms input values in a predicted sample set together with the coordinates in the training sample set and is input to the trained network, the output value of the network is a to-be-predicted target value, namely, gas concentration data of all monitoring points of a virtually extended space S2, and data on an initial monitoring surface S1 is kept unchanged. According to the method, the source characteristic inverse computation precision is effectively improved without adding the monitoring points; and moreover, the workload is reduced and the working efficiency is improved.

The invention provides an online predicting method for silicon content of blast furnace molten iron. According to a blast furnace silicon element transferring mechanism, a parameter which affects of silicon content of the molten iron is selected from blast furnace operation parameters as an input variable of a predicting model. A Pearson correlation analysis method is used for determining a lag time length between the input variable and the silicon content of the molten iron. Then standardizing processing is performed on the sample data and the predicting data of the input variable. The influence of different dimensions to model predicting accuracy is eliminated. A kernel extreme learning machine is utilized for predicting the silicon content of the molten iron at a next time point. A sliding window updating method is used for performing online updating on training set data. A genetic algorithm is introduced for optimizing the key parameter of the kernel extreme learning machine model.The online predicting method according to the invention is suitable for long-time online predicting to the silicon content of the blast furnace molten iron. An actual testing result proves a fact that the predicting method according to the invention has relatively high predicting precision. The method facilitates advanced understanding of the silicon content of the blast furnace molten iron by blast furnace operators, thereby adjusting the operation parameters in time.

The invention discloses a gait recognition method based on an extreme learning machine. The method comprises the steps of preprocessing, feature extraction and classification recognition, and specifically comprises the following steps that the preprocessing is used for obtaining a standard moving target contour sequence with the uniform size, and the steps are included and comprises: 1-1, extracting a moving target contour sequence; 1-2, image standardization; The feature extraction is used for obtaining gait feature parameters with good characterization, and comprises the following steps: 2-1, gait period extraction; 2-2, extracting an action energy diagram; 2-3, dimension reduction is carried out through two-dimensional principal component analysis; Classification recognition is carriedout by adopting a kernel extreme learning machine (KELM); 3-1, constructing a kernel extreme learning machine neural network model; 3-2, training a neural network of the kernel extreme learning machine; 3-3, performing classifying and identifying. The action energy diagram extracted by the method contains more dynamic and static information, a complex image processing process is not needed, the extraction mode is simple, and the method has very good characterization.

The invention relates to the technical field of hyperspectral image classification, and discloses a hyperspectral image classification method based on guided filtering and a kernel extreme learning machine, and the method comprises the following steps: S1, carrying out the normalization preprocessing of hyperspectral image data; s2, performing principal component analysis on the normalized hyperspectral image data to obtain a plurality of different principal component images; s3, performing guided filtering processing on the input image through the guided image, and extracting space-spectral information; s4, selecting 5% of the images after guided filtering processing as a training set, and taking the remaining images as a test set; s5, performing linear weighting on the space spectrum kernel and the spectrum kernel in a linear weighting mode to form a weighting kernel; s6, the weighted kernel is sent to a kernel extreme learning machine, then the trained kernel extreme learning machine is used for carrying out classification prediction on the test sample, and the hyperspectral image classification method has the advantages that the classification precision is very high, and the time cost spent in the whole classification process is very low.

The invention discloses a method for predicting a terrorist attack based on the stochastic subspace. The method comprises a first step of establishing a training set, and maintaining a linear classifier Wt by an online learner; a second step of counting the number z of terrorist attack events happening in a country in the next month of the month in a GTD database; step 3: randomly selecting s groups of feature subsets from overall features of a given terrorist attack data set using a stochastic subspace method and generating s base classifiers in an integrated classification algorithm of a kernel extreme learning machine; a fourth step of putting the S groups of feature subsets into the kernel extreme learning machine for learning to obtain output results; a fifth step of integrating the outputs of the s base classifiers to obtain a final classification result; and a sixth step of performing model application: inputting the value of an independent variable for each record in a test set to obtain the value of a predictive variable, that is, the probability of a terrorist attack event happening in the next month.

The invention discloses a road traffic jam early warning method and system. The method comprises steps of carrying out feature classification according to obtained multi-source traffic data, constructing a corresponding feature membership function, and obtaining a first fuzzy weight; constructing an artificial membership function for the multi-source data by adopting an expert evaluation method, and calculating a second fuzzy weight; according to a fusion fuzzy weight obtained after the first fuzzy weight and the second fuzzy weight are fused, performing fuzzy weighted average of the feature membership function, performing ambiguity resolution of the obtained weighted average membership function of different feature quantities, and obtaining multi-source fusion traffic data; constructing aroad traffic jam model by adopting a kernel extreme learning machine group algorithm, and calculating an optimal road traffic jam index; and obtaining current multi-source traffic data, predicting acurrent congestion index, and carrying out early warning on whether a current road is congested or not by comparing the current congestion index with the optimal road traffic congestion index; and constructing a man-machine hybrid enhanced intelligent multi-source data fusion system for exerting the group intelligence of road participants.

The invention relates to an organic pollutant migration numerical model substitution method based on a multi-core extreme learning machine, and the method comprises the following steps: building an underground water organic pollution multiphase flow migration numerical model according to observation data, determining pollution source characteristics, aquifer parameters and the value range of each variable, which have relatively high contribution degree to the spatial and temporal distribution of pollutants, in the model; preparing a training sample set; training a single-kernel extreme learning machine substitution model of the multiphase flow numerical model; establishing a nonlinear programming optimization model of kernel function key parameters and kernel function combination weights; and solving the optimization model by using a genetic algorithm, identifying an optimal kernel function parameter and a combination weight, and constructing a genetic evolution multi-kernel extreme learning machine intelligent substitution model of the numerical model. The problem of substitution of the underground water organic pollution multiphase flow numerical model is solved, the calculation efficiency of pollutant transport simulation prediction is improved, and an efficient solution is provided for underground water pollution source characteristic and pollutant transport parameter inversion identification.

The invention provides a fault detection method and system based on a global preserving unsupervised kernel extreme learning machine. The fault detection method based on a global preserving unsupervised kernel extreme learning machine comprises an offline modeling step and an online monitoring step. The online monitoring step comprises the following sub-steps: normalizing test data, which is the condition data in a non-linear working process; calculating the kernel vector of the test data according to a kernel function, and centralizing the mean value of the kernel vector in a feature space toobtain a test kernel vector; extracting a low-dimension feature information matrix of the test data from the test kernel vector according to the global preserving unsupervised kernel extreme learningmachine, and calculating the monitoring statistics of the test data; and judging whether there is a fault in the non-linear industrial process according to whether the monitoring statistics of the test data exceeds the control limit, so as to achieve the purpose of real-time detection of a process fault.

The invention discloses a swarm optimization kernel extreme learning and sparse representation mechanical fault identification method, which is used for improving the efficiency and precision of mechanical fault identification. The invention provides a kernel extreme learning and sparse representation mechanical fault recognition method capable of effectively improving the fault recognition precision by combining the advantage of high efficiency of a kernel extreme learning machine and the advantage that sparse representation captures intrinsic characteristics of signals through dictionary redundancy. The bee colony optimization algorithm is integrated into the kernel extreme learning machine method, and the optimal model parameters of the kernel extreme learning machine are obtained through the optimization algorithm to further improve the performance of the recognition model. An input mechanical signal sample is firstly subjected to fault identification by utilizing a swarm-optimizedkernel extreme learning machine, and an input sample which cannot reach an expected identification result is subjected to secondary identification by adopting a sparse representation method, so thatrapid and accurate fault identification is realized. The method is suitable for mechanical fault identification.

The invention discloses a random forest classification system based on a kernel extreme learning machine and parallelization, and the system comprises a single-unit module and a parallelization module. The single-unit module is mainly used for the processing of a non-mass data set, and comprises four submodules: a data extraction module, a model training module, a model evaluation model, and a model optimization module. According to the technical scheme of the invention, the system can support the processing of mass data, and also can achieve the classification performances of data.

The invention discloses a hyperspectral image classification method and system based on cascade spatial-spectral feature fusion and a kernel extreme learning machine, and the method comprises the steps: carrying out the normalization preprocessing operation of a hyperspectral image, and dividing a data set into a training set and a test set; performing convolution on an input hyperspectral image by using a space-spectrum attention residual neural network to respectively obtain space and spectrum information; performing space-spectrum feature extraction on data of a test set by using the trained network, calculating an output weight matrix of a hidden layer of a kernel extreme learning machine while training the neural network, and then inputting the extracted features and the output weight matrix into the kernel extreme learning machine, thereby achieving the purpose of classifying hyperspectral images. According to the method, spectral attention information and spatial attention information of the hyperspectral image are fully utilized, and deep feature extraction can be performed on the hyperspectral remote sensing data, so that rapid and accurate classification is realized.

The invention discloses a pedestrian detection method fusing depth perception features and a kernel extreme learning machine. The method comprises the following steps: 1, constructing a DAGnet neuralnetwork comprising two parts; 2, training a DAGnet convolutional neural network by using the preprocessed sample to obtain a DAGnet model; 3, obtaining a depth perception feature vector by using a DAGnet model; 4, training the kernel extreme learning machine by using the depth perception feature vector to obtain a pedestrian recognition model; 5, performing generalization performance estimation onthe kernel extreme learning machine; 6, learning a second-level feature map and a GVBS saliency detection algorithm by using a DAGnet model to obtain a saliency map of the test image, and marking anapproximate area of a pedestrian in the test image; 7, scanning the approximate region by using a multi-scale sliding window to obtain a depth feature vector of the region where the window is located;and 8, identifying whether the area contains pedestrians or not by utilizing a pedestrian identification model. The method can obtain better detection performance, effectively improves the precisionand speed of pedestrian detection, and is better in robustness.

The invention belongs to the technical field of mechanical fault diagnosis, and particularly relates to a reciprocating compressor fault diagnosis method based on improved RCMDE, which comprises the following steps that machine body surface time domain vibration signals of a reciprocating compressor under different operation conditions are acquired, and an initial vibration signal is processed by adopting parameter optimization variational mode decomposition to obtain an intrinsic mode function component; according to a kurtosis-correlation coefficient criterion, a group of components containing the most abundant information amount is selected, and a fault signal is reconstructed; the nonlinear behavior of the vibration signal after noise reduction is quantitatively analyzed through the improved fine composite multi-scale dispersion entropy, and a fault feature vector is formed; input features are selected in a dimensionality reduction mode through a kernel principal component analysis method and input into a kernel extreme learning machine to be classified and recognized, and the operation state and the fault type of the reciprocating compressor can be distinguished. Through improved fine composite multi-scale dispersion entropy analysis, nonlinear behaviors of vibration signals are quantitatively described, feature vectors are formed, and fault types can be diagnosed more accurately.

The invention discloses a short-term wind power generation prediction method and device, equipment and a storage medium, and relates to the field of new energy power generation prediction, and the method comprises the following steps: obtaining main meteorological characteristics affecting the generated power, and combining the historical data of the main meteorological characteristics with the historical data of the generated power to obtain a target data set; setting and initializing parameters of the multivariate universe algorithm and the kernel extreme learning machine; dividing the target data set into a training set and a test set, and training a kernel extreme learning machine by using the training set; running a multivariate universe algorithm to obtain an optimal combination of penalty parameters and kernel parameters in the kernel extreme learning machine; and substituting the optimal combination into a kernel extreme learning machine, and testing the test set by using the kernel extreme learning machine with the optimal combination to obtain a prediction result. According to the scheme, the parameters of the kernel extreme learning machine are optimized by using the multivariate universe algorithm, the problems of local optimal solution and the like can be avoided, and the prediction precision of the model is improved.

The invention discloses an electric locomotive idling online identification method based on fuzzy entropy and a kernel extreme learning machine. The method comprises a multi-scale fuzzy entropy feature extraction module, an optimal kernel extreme learning machine model module and an idling online identification module. Compared with the prior art, the method has the positive effects that idling is identified from the angles of signal feature extraction and machine learning classification, features capable of remarkably representing the idling state / adhesion state are extracted from accurately measured locomotive wheelset speed signal features, and the features are classified by a classifier to realize idling online identification; compared with a traditional idling recognition method adopting input with errors, the online recognition effect is more excellent, more accurate and quicker; compared with a traditional method, the method is higher in adaptability to complex operation conditions and operation environments of the electric locomotive, has more excellent recognition precision compared with a traditional idling recognition method, and effectively solves the problem that a threshold value is difficult to set during online recognition in the traditional method.

The invention belongs to the technical field of mechanical operation state trend prediction, and particularly relates to a method for predicting the stability of an angular contact ball bearing retainer. The method comprises the following steps: S100, calculating the instability of the retainer; S200, constructing a kernel extreme learning machine KELM model; S300, training the KELM model; And S400, predicting the instability of the retainer by using the trained KELM model. A KELM model is trained through the instability, obtained through calculation of an angular contact ball bearing dynamicmodel ADORE, of a retainer. and if not, the stability of the retainer under the action of other rotating speeds and loads can be predicted by adopting the trained KELM model, and calculation does notneed to be carried out by adopting a dynamic model, so that the calculation time is saved, and the analysis efficiency is improved.