45 results about "Nonlinear dimension reduction" patented technology

A method of identifying and localizing objects belonging to one of three or more classes, includes deriving vectors, each being mapped to one of the objects, where each of the vectors is an element of an N-dimensional space. The method includes training an ensemble of binary classifiers with a CISS technique, using an ECOC technique. For each object corresponding to a class, the method includes calculating a probability that the associated vector belongs to a particular class, using an ECOC probability estimation technique. In another embodiment, increased detection accuracy is achieved by using images obtained with different contrast methods. A nonlinear dimensional reduction technique, Kernel PCA, was employed to extract features from the multi-contrast composite image. The Kernel PCA preprocessing shows improvements over traditional linear PCA preprocessing possibly due to its ability to capture high-order, nonlinear correlations in the high dimensional image space.

The invention belongs to the technical field of remote sensing image processing and in particular relates to a manifold dimension reduction method of hyperspectral images based on image block distance. According to the method, a novel distance measure, namely an image block distance measure is provided and is applied to neighborhood selection and lower dimension coordinate embedding in manifold learning, and a nonlinear dimension reduction method of novel hyperspectral remote sensing images is obtained. The physical properties of the hyperspectral images are utilized, the spectral information and space information of the images are combined, the local characteristics between data points can be better kept, and the characteristics of an original data set are well kept on the basis of reducing the image information redundancy to the greatest degree. The method has high applicability on various different hyperspectral data, and has important application values in detection and identification aspects of high-precision terrain classification and ground targets based on the hyperspectral remote sensing images.

The invention discloses a low-energy planetoid precise track transfer detection method for complex constraints, and belongs to the technical field of aerospace. The method comprises the steps that firstly, various complex non-consistent strong coupling constraints needing to be met by a track detection design task are determined, and the mapping relation between track design parameters and the various complex non-consistent strong coupling constraints is built; under a mass center rotating coordinate system, a detector dynamics equation is built; an initial value is provided through the builtlinear detector dynamics equation, a precise quasi periodicity track under an ephemeris model is obtained by adopting a nonlinear dimension reduction method and second-order differential coercion; onthe basis of the precise quasi periodicity track under the ephemeris model, a quasi manifold disturbance method is adopted for optimizing the obtained transfer track initial value; for the various complex non-consistent strong coupling constraints, the obtained transfer track initial value is corrected, and the precise low-energy transfer track is obtained. The method has the advantages of being high in efficiency, good in convergence and low in energy needed by transfer.

The invention relates to a dynamic adaptive adjustment method of virtual network resources based on nonlinear dimensionality reduction. The method comprises steps: data of an underlying physical network are acquired, and data of nodes or links, relative to real-time remaining resources in multiple adjacent time points, in the underlying physical network are obtained; dimensionality reduction processing is carried out on the acquired data, and a two-dimensional relationship distribution chart of the nodes or the links in the underlying physical network is obtained; the nodes or the links in the underlying physical network are clustered according to the two-dimensional relationship distribution chart; according to the clustering result, mapping is carried out again on the virtual network in the next period of time, and in the re-mapping process, a node cluster or a link cluster with a low utilization rate in the underlying physical network is preferably selected to carry out virtual resource mapping; and after running for certain period of time, the above steps are carried out again until the process of allocating the virtual resources is over. With the method, the execution efficiency of mapping algorithm can be effectively improved, the resource utilization rate is improved, and load balancing can be realized.

The invention discloses a graph layout method for a large-scale network, and the method comprises the steps of enabling each node in graph data to be expressed as a low-dimensional dense vector through a network embedding expression model based on machine learning, and constructing an embedding matrix of the graph data; and projecting the embedded matrix through an improved nonlinear dimension reduction algorithm to obtain a graph layout result of graph data in a two-dimensional space. The invention further discloses a graph layout device for the large-scale network. According to the invention, the calculation efficiency is higher, the required storage space is smaller, the local and global structure characteristics of the graph data can be maintained, meanwhile, the nodes with higher degree values in the graph data can be relatively dispersed from the neighbor nodes under the condition of maintaining local structure information, and the possible congestion or overlapping phenomenon can be effectively relieved.

The invention studies a feature fusion process of a rolling bearing vibration signal multi-view feature set, and provides feature fusion based on random forest feature selection and auto-encoder dimension reduction. A rolling bearing diagnosis method disclosed by the invention comprises the steps of: 1) extracting rolling bearing multi-view characteristics by using statistical characteristics anda time sequence signal spectrum analysis method; 2) outputting feature importance of high-dimensional features by using a random forest, and eliminating invalid features based on the feature importance to reduce feature dimensions; and 3) performing nonlinear dimensionality reduction on redundant features with the same feature importance in the feature set by using an auto-encoder so as to obtaina small-redundancy low-dimensional feature set capable of clearly expressing the bearing state difference.

The invention discloses a pure electric vehicle driving condition construction method, which comprises the following steps of performing data acquisition on a pure electric vehicle driving condition, dividing a test route into a plurality of short-stroke segments, and obtaining characteristic parameters of the pure electric vehicle driving condition from the plurality of short-stroke segments; carrying out nonlinear dimensionality reduction on the characteristic parameters of the driving condition of the pure electric vehicle through kernel principal component analysis, classifying the characteristic parameters after nonlinear dimensionality reduction through a hybrid clustering method, and screening a plurality of short-stroke fragments according to the classification result in combination with the condition duration weight of each category and the Pearson's correlation coefficient; constructing a plurality of candidate working conditions of the pure electric vehicle; and calculating and comparing the relative error value and the SAPD frequency value of the characteristic parameters in the plurality of candidate working conditions of the pure electric vehicle, and constructing the driving working condition of the pure electric vehicle. The working condition construction precision is higher, the actual driving characteristics of the electric vehicle can be better reflected, and the consistency of the obtained working condition curve and the actual working condition is higher.

The invention discloses a vehicle detection tracking method based on color feature nonlinear dimensionality reduction. The method comprises the steps of collecting the road vehicle driving video datathrough video shooting; extracting the LBP features; converting the RGB space into a CN space through color conversion, and extracting the color features; carrying out feature fusion on the LBP features and the color features; enabling the input characteristics to return to target Gaussian distribution, and training a related filter in an on line manner; adding a scale pool to enable the size of atracking frame to be matched with the target vehicle; calculating a correlation filter response value so as to determine a new position of the target vehicle; repeating the above steps until the video is finished; and connecting the central positions of the tracked vehicle obtained by each frame, and drawing the track of the tracked vehicle in the video. According to the invention, through the LLE dimension reduction and in combination with the LBP features, the accuracy of the tracking algorithm can be improved.

The invention discloses a quantum kernel principal component analysis method. The method comprises the steps: combining the truncation Taylor expansion and the Hamiltonian quantity simulation technology, and achieving the effective Hamiltonian quantity simulation of a general nonlinear kernel matrix; preparing a kernel matrix quantum state by adopting quantum arithmetic to serve as input of a quantum phase estimation algorithm, realizing the quantum characteristic value solution of the kernel matrix, and finally outputting a target quantum state subjected to quantum kernel principal componentanalysis by combining controlled rotation, solution operation and quantum sampling operation. The method solves the problem that an existing quantum principal component analysis algorithm cannot be matched with a nonlinear dimension reduction task, and exponentially accelerates an existing classical kernel principal component analysis algorithm by means of the unique parallel advantage of quantumcomputation.

The invention discloses a unified fault locating method for a comprehensive energy system. The method comprises the steps of firstly extracting typical characteristic quantities of energy subsystems of a data collection apparatus through data preprocessing and standardizing the typical characteristic quantities; secondly aggregating heterogeneous characteristic quantities into a high-dimensional matrix in space and time; and thirdly performing nonlinear dimension reduction on the matrix by utilizing Isomap and the like, and based on the value of a local sparsity coefficient and a node association relationship, performing fault identification and locating on the comprehensive energy system. Different characteristic quantities among a power system, a natural gas system and a thermal system are coupled into a comprehensive characteristic quantity in a unified way, so that the fault identification precision is improved; and massive data of the comprehensive energy system is fully utilized,the barrier and limitation of a single system are broken through, and unified fault locating of the comprehensive energy system is realized.

A computer implemented method and system for declaring arrhythmias in cardiac activity are provided. The method and system are under control of one or more processors that are configured with specific executable instructions. The method and system obtain far field cardiac activity (CA) signals for a series of beats and builds an N-dimensional data set from data values for features of interest from the CA signals. The method and system utilize a manifold structure to map the N-dimensional data set, through nonlinear dimensional reduction, onto an M-dimensional data set and declares an atrial fibrillation (AFL) episode based on a relation between the M-dimensional data set and one or more AFL classification criteria.

Embodiments of the invention disclose a visualization method, system and device for multi-dimensional network node classification, and a storage medium. The method is based on a network embedding technology of machine learning, the network embedding technology of machine learning is combined with a regularization mechanism and an attention mechanism to obtain a low-dimensional dense vector of each node in a multi-dimensional graph network, and the low-dimensional dense vectors form a low-dimensional embedded matrix. Based on a non-linear dimension reduction algorithm, the low-dimensional embedded matrix is projected to obtain a coordinate value of each node in multi-dimensional graph data in a two-dimensional space, and a classification result is presented by adopting label information of the nodes for color mapping and employing a visualization technology. According to the invention, low-dimensional embedding obtained in the embodiments of the invention fuses node close distance, node long distance and the attribute information of the node at the same time. The obtained low-dimensional embedded matrix is projected into a two-dimensional layout space based on the nonlinear dimension reduction algorithm, and the influence of various feature information in the original multi-dimensional graph network on node classification is visually displayed from a visual angle by adopting the visualization technology.

The invention provides a method for mining an illegal accident corresponding relation based on an LLE and K-means method. The method comprises the following steps: collecting data required by traffic violation and traffic accident correlation analysis; classifing traffic accidents by considering different indexes; selecting an illegal type and an accident type with the highest occurrence frequency as an illegal label and an accident label of a person respectively; counting illegal types-accident types, and building an illegal types-accident type matrix; determining three thresholds to screen traffic violation types; building a personnel-type correspondence matrix; performing standardization processing on the data by using a zero-mean standardization method; reducing the data from a high dimension to a low dimension by using an LLE nonlinear dimension reduction method; carrying out clustering analysis by using an improved K-mean value algorithm for the two different accident type classification modes respectively. According to the invention, the defect of large randomness in the traditional K-means algorithm is overcome, and a corresponding relation between the traffic violation type and the traffic accident type is further mined.

The invention relates to the field of artificial intelligence, is applied to the field of intelligent transportation, and discloses a vehicle braking method, device and equipment and a storage medium, which are used for improving the judgment efficiency and accuracy of judging whether a vehicle passes through a traffic intersection or not. The vehicle braking method comprises the steps of preprocessing a scene image of a road in front of a vehicle through a nonlinear dimension reduction algorithm to obtain a to-be-detected image; judging whether a traffic signal lamp image exists in the to-be-detected image or not by utilizing a neural network model, and if the traffic signal lamp image exists in the to-be-detected image, identifying the image color of the traffic signal lamp image; when the image color is yellow, calculating the vehicle passing speed according to the distance interval data, the current vehicle speed and the signal lamp conversion time period; and if the vehicle passing speed is larger than or equal to the standard speed threshold value, starting a braking system to brake the vehicle till the vehicle stops moving.

The invention provides a multi-angle mutual transformation method for a facial image. The method comprises the following steps of: representing the facial image in a column vector form of the gray level of pixels; transforming a problem into a problem that the weight of a point adjacent to a certain point in a high dimensional space in a local neighborhood embedding nonlinear dimension reduction theory is needed to be solved by taking a single-frame input facial image at a certain angle as one point in the high dimensional space and a training set facial image at the same angle as the point adjacent to the point; synthesizing the facial image at the target angle by using the solved weight and a training set facial image at the target angle; and representing the synthesized facial image at the target angle in the matrix form of the gray level of pixels. According to the multi-angle mutual transformation method for the facial image, an algorithm is simple; and the calculation speed and the synthesis effect of hair and facial edges are obviously superior to those of the prior art.

The invention relates to a small fixed-wing aircraft stall control method based on nonlinear dimensionality reduction, and belongs to the technical field of active flow control of small fixed-wing aircrafts. The attack angle of the wing is measured in real time, attack angle data are processed in real time through a nonlinear dimension reduction algorithm, whether the processed attack angle reaches a critical stall attack angle or not is judged, when the attack angle reaches the critical stall attack angle, the synthetic jet piezoelectric pump is controlled to work, low-momentum gas is sucked into a pump cavity from the tail edge of the wing through a flow guide pipeline, and the low-momentum gas is pumped into the pump cavity. And gas is sprayed to the surface of the wing through the flow guide pipeline to inhibit separation of a boundary layer on the surface of the wing, and when the attack angle is gradually reduced from the critical stall attack angle, the aerodynamic performance of the original wing section is kept. The method has the advantages that the measurement attack angle error is reduced, piezoelectric synthetic jet control has higher response speed so as to deal with the sudden stall condition of the aircraft in the air, speed compensation is carried out on the flight attitude under the stall attack angle, and stall of the aircraft is prevented.

The embodiment of the invention provides an urban complex electricity consumption prediction method and device, electronic equipment and a storage medium. The method comprises the steps of acquiring historical hour electricity consumption sequences of various types of electricity consumption entities in an urban complex; carrying out splitting and dimensionality reduction on the historical hour power consumption sequence according to preset calendar tags, and obtaining a first power consumption sequence for each calendar tag; carrying out nonlinear dimension reduction on the first electricityconsumption sequence in a time dimension to obtain a second electricity consumption sequence; inputting the second power consumption sequence into a preset prediction neural network for prediction toobtain a first power consumption prediction result based on the calendar label; and based on the first power consumption prediction result, calculating to obtain a power consumption prediction resultof the complex. And the prediction result of the electricity consumption of the urban complex can be more accurate.

The invention discloses an LED classification method based on manifold learning, and the method comprises the following steps: S1, obtaining an image comprising an LED, and converting the image into agray image; S2, performing image masking on the area, except for the edge part area of the LED fluorescent glue, of the grayscale image in the step S1 to obtain a mask image; S3, performing dimensionreduction processing on the mask image obtained in the step S2 to obtain dimension reduction data; and S4, transmitting the dimension reduction data obtained in the step S3 to a classifier for classification, and obtaining LED classification of good LEDs, LEDs with large glue amount and LEDs with small glue amount. According to the method, the LED fluorescent glue edge annular image is separated,the nonlinear dimension reduction algorithm is combined to change the distribution of glue amount characteristics, the glue amount characteristics of the LED are extracted, the interference of redundant information on characteristic extraction is reduced, and the classification accuracy of a classifier is improved; and meanwhile, in the dimension reduction algorithm, a conditional probability function used in an iterative process is optimized, so that the overall time consumption of the algorithm can be reduced.

The invention discloses a composite material defect detection method based on generated kernel principal component thermal image analysis, and belongs to the technical field of composite material thermal imaging nondestructive testing. The method comprises the following steps: step 1, acquisition of a thermal image data set of a composite material; step 2, amplification and preprocessing of thermal image data: establishing a spectrum normalization generative adversarial network to generate a thermal image; step 3, establishment of a kernel principal component analysis model: performing feature space mapping and projection matrix calculation; step 4, image reconstruction and defect visualization; and step 5, model performance evaluation. According to the method, a data amplification strategy based on a generative adversarial network and a nonlinear dimension reduction technology based on kernel mapping are adopted to analyze thermal image data with nonlinear characteristics; under the condition that the original thermal image data is less, generating data with the same distribution as the thermal image of the experiment record; a kernel principal component thermal imaging analysis model is adopted to solve the problem that defects and backgrounds are difficult to separate in thermal image analysis, and the visibility of the defects is improved.

The invention provides an improved nonlinear POD dimension reduction method based on a transient time sequence. Firstly, principal components are extracted from high-dimensional data, the obtained principal components are dimensionless, a dynamic vibration model after dimensionality reduction is obtained, a set of principal components are obtained from the transient process of the system through adimensionality reduction method, the first principal component and the second principal component form a projection space, and the original system is projected into the space. Compared with an original high-dimensional system, the method has the advantages of being simple, simple in model construction, small in calculation complexity, high in applicability and accuracy and capable of being widelyapplied to a multi-degree-of-freedom nonlinear rotor system. By comparing the obtained calculation result with the actual result, the obtained data error is within 5%.

The invention provides a behavior detection method. The behavior detection method comprises: extracting human body feature point coordinate information; enhancing the coordinate information content of the human body feature points to obtain kinematics information of human body postures; combining the human body feature point coordinate information and the kinematics information to obtain high-dimension information, and meanwhile, preseting a nonlinear dimension reduction algorithm to obtain low-dimension effective data subjected to redundancy elimination and low signal-to-noise ratio; and performing unsupervised clustering on the low-dimensional effective data, constructing convolutional neural network training to form a behavior recognition classifier, and outputting and displaying the motion behavior of the human body posture. After the attitude mode is obtained by fusing the relevance information of the attitude and the action and kinematics parameters and adopting an unsupervised algorithm, classifier training is carried out by using 1DCNN (one-dimensional convolutional neural network) and combining continuous frame time sequence information to obtain the action category, so that the effect of behavior recognition is achieved, and the problem that accurate behavior detection is relatively difficult is solved.

The invention relates to a NOx emission prediction method of a thermal power plant based on a generalized mutual entropy auto-encoder, in particular to a thermal power plant NOx emission prediction method based on a generalized mutual entropy gated stacked target related auto-encoder. The method comprises the following steps: acquiring thermal power plant data including NOx emission and related influence factors, preprocessing all data, taking the related factors influencing NOx emission as input, performing nonlinear dimensionality reduction and feature extraction through an auto-encoder, and establishing a model between the NOx emission influence factors and the NOx emission. The NOx emission prediction method is accurate and reliable in NOx emission prediction and has high practical engineering application value.

The invention provides a medium-short term load prediction method and device based on manifold learning, and belongs to the technical field of load prediction, and the method comprises the steps: carrying out nonlinear dimensionality reduction of a historical load data set by a local linear embedding method to acquire a low-dimensional manifold sequence; inputting the low-dimensional manifold sequence into the trained long-term and short-term memory neural network model to obtain a prediction sequence; and reconstructing the prediction sequence by adopting a manifold learning reconstruction method to obtain a load prediction value. The manifold learning method is adopted to reduce the dimension of the load data, and the manifold learning method can better mine the nonlinear characteristicsof the load. Meanwhile, the low-dimensional manifold sequence obtained after dimension reduction is predicted by adopting a deep learning method, so that the time sequence rule of the low-dimensionalmanifold sequence is better mined, and the load prediction precision is improved.

The invention relates to the technical field of image retrieval, in particular to a weather situation field rainfall similarity forecasting method based on image feature combination, which comprises the following steps of: extracting overall features, local features, isoline features and nonlinear dimension reduction features of a situation field, and performing similarity measurement on the extracted features, so that a plurality of features jointly determine a retrieval result; the rainfall similar forecast based on the situation field is realized, and the accuracy of the final result is improved. In addition, after being extracted, the low-dimensional features of the historical situation field are stored in a historical feature database, similarity measurement is directly conducted on the low-dimensional features when the similarity is calculated, and the retrieval speed is increased. The method can be used for precipitation analysis in the meteorological field.

The invention discloses an epilepsy prediction system based on feature channel fusion and deep learning, and the system employs a T-distribution random neighbor embedding algorithm (t-SNE) of a nonlinear dimension reduction algorithm to carry out the fusion of feature channel information of epilepsy electroencephalogram signals. Time domain and frequency domain information obtained through short-time Fourier transform calculation serves as features to be input into the deep residual contraction neural network, and epilepsy seizure is predicted by recognizing the epilepsy seizure interval and the epilepsy seizure early stage. According to the method, from the aspects of feature dimension improvement and classifier design, artificial feature extraction is not needed, the expression of feature information is improved, and a new method is provided for subsequently pushing epilepsy prediction to clinical application.

The invention relates to a data processing technology, and discloses a project data risk assessment method comprising the following steps: obtaining original project data and project characteristics, and selecting a pre-constructed score card system according to the project characteristics; calling a pre-constructed main label set in the label card system according to the original project data, and constructing a main label calculation model according to the main label set; performing calculation by utilizing the main label calculation model to obtain main label evaluation of the original project data; obtaining a feature vector by using historical project data, and constructing a scoring model according to the feature vector and the scoring card system; and performing nonlinear dimensionality reduction on the main label evaluation by using the scoring model to obtain a scoring result. In addition, the invention also relates to a block chain technology, and the scoring result can be stored in a node of a block chain. The invention further provides a project data risk assessment device, electronic equipment and a storage medium. The problems that risk scoring is not accurate enough and efficiency is low can be solved.

The invention relates to nucleation classifier based on local spline embedding. Training data and test data are selected; the dimension of the training data is reduced in a nonlinear way based on local spline embedding; the expanded form of the test data is deduced using a kernel method according to the obtained optimal nonlinear embedding of the training data, namely, nonlinear embedding of the test data on a low-dimensional manifold is obtained; and the dimension-reduced test data is classified using a linear support vector machine (SVM) algorithm. The invention overcomes the defect that good classification performance cannot be achieved for nonlinear classification. According to the invention, the dimension of high-dimensional tagged data is reduced using a nonlinear dimension reduction algorithm based on local spline embedding, and the features of the high-dimensional tagged data are extracted; then, new high-dimensional test data without tags is embedded; and finally, the new test data is classified using the SVM algorithm according to the characteristics of the data.

The invention discloses a deep network characterization method of rich structural information. Network characterization is performed from a comprehensive perspective by using rich multi-order structural information, for example, transition probability direction adjustment control parameters are imported for transition matrices of different orders, nonlinear dimension reduction processing is performed on the transition matrices of different orders by using stacking noise reduction autoencoder, and multi-order information is fused by using an attention mechanism to well improve the network characterization effect.