429 results about "Data dimensionality reduction" patented technology

The invention relates to an improved parallel channel convolutional neural network training method. The improved parallel channel convolutional neural network training method comprises steps that characteristic extraction of data of the convolutional neural network is carried out through utilizing direct connection and convolutional channels to acquire characteristic matrixes; the characteristic matrixes are merged, and data dimension reduction is further carried out; the convolutional neural network is trained, and a loss value of network training at present is calculated; error items and a weight gradient of each layer are calculated; whether the network is in a convergence state is determined according to the loss value, if not, an initialization parameter of the convolutional neural network is adjusted according to the weight gradient, and re-training is further carried out; if yes, the network training result is outputted. The method is advantaged in that data circulation in the network can be guaranteed through introducing the direct connection channel, a problem of gradient instability during deep convolutional neural network training is solved, and deeper networks can be trained; through maximum pooling and mean value pooling, characteristic matrix dimensions of two times of characteristic extraction can be made to be consistent, and advantages of two pooling methods are integrated.

The invention discloses an effective micro-expression automatic identification method which comprises the steps of micro-expression frame sequence preprocessing, micro-expression information data study and micro-expression identification. The method for micro-expression frame sequence preprocessing comprises the steps that frames of obtained micro-expression sequences are detected, data of an image of each frame are extracted so that graying processing can be conducted on the data, and all the micro-expression sequences are interpolated into the frame of the unified number through the linear interpolation method. The method for micro-expression information data study comprises the steps that the micro-expression sequences obtained in the preprocessing stage are written in a tensor mode, then, the intra-class distance of the same class of micro-expressions is minimized in a tensor space through the discriminating analysis method of tensor expression and the between-class distance of different classes of micro-expressions is maximized, so that data dimension reduction is achieved, and characteristic data are ranked in a vectorized mode according to a class discriminating capacity descending order. A nearest neighbor classifier is used for micro-expression identification. Compared with the methods of MPCA, GTDA, DTSA and the like, the effective micro-expression automatic identification method has the advantages of being high in rate of identification, low in computer performance requirement and easy to achieve.

The invention relates to a streaming time series data dimensionality reduction and simplified representation method based on piecewise linear representation. The method comprises the following steps that: S1: presetting data segments and compression parameters; S2: carrying out data scanning on streaming time series data in a slide window way, and entering a streaming data buffer zone; S3: judging whether the fitting error of an initial segmented data segment exceeds an ME_ES (Maximum Error for Entire Segment) or not, carrying out reservation if the fitting error of the initial segmented data segment exceeds the ME_ES, and marking the initial data segment as "inseparable", and if the fitting error of the initial segmented data segment does not exceed the ME_ES, carrying out secondary optimal segmentation; and S4: moving the data segment which is marked as "inseparable" in the streaming data buffer zone out of the streaming data buffer zone, judging whether the streaming time series data to be processed is in the presence or not, if the streaming time series data to be processed is in the presence, returning to the S2, and otherwise, ending. By use of the method, data dimensionality reduction execution efficiency is guaranteed to a high limit, the fitting accuracy of data simplified representation is optimized to a certain range, and accuracy and the execution efficiency of data representation can be improved.

The invention relates to a data mining improved type K mean value clustering method based on linear discriminant analysis, namely an LKM algorithm. Firstly, an LDA is adopted to conduct linear dimensionality reduction on an original n-dimensional data set A to obtain a one-dimensional data set Y, then a k mean value clustering algorithm is adopted to conduct clustering analysis on the data set Y after dimensionality reduction, and final results are output. The method that data dimensionality reduction is combined with the K mean value clustering method is adopted, and defects of the k mean value clustering algorithm on high-dimensional data are overcome through the data dimension reduction technology. The aims of lightening dimensionality curses and eliminating other uncorrelated attributes in high-dimensional space are achieved through data dimension reduction. Meanwhile, the performance of the k mean value clustering algorithm for processing the high-dimensional data is also improved, and the correlated defects of the k mean value clustering algorithm are overcome.

The invention belongs to the image processing field and provides a super-pixel level information fusion-based hyperspectral image classification method and system. The objective of the invention is to solve the problems of high computational complexity, low classification accuracy and large inter-band redundancy of an existing classification method. The method includes the following steps of: filter generation; Gabor feature extraction; hyperspectral image super-pixel segmentation; super-pixel feature data calculation; super-pixel spatial coordinate calculation; data dimensionality reduction; hyperspectral feature data generation; and multitasking support vector machine classification.

The invention provides a method for reducing dimensions of high spectroscopic data based on a factorial analysis model, comprising the following steps: (1) reading in the high spectroscopic data; (2) establishing the factorial analysis model for dimension reduction of the high spectroscopic data; (3) calculating average of data, covariance matrix and correlation matrix; (4) calculating proper value and standard eigen vector of the correlation matrix of the data; (5) carrying out solving on factor loading matrix by a parametric estimation method; (6) calculating covariance of special factors and communality of data variables in the factorial analysis model; (7) calculating the biggest factor loading spin matrix based on variance; (8) calculating factor scores by a least square method based on variance; and (9) obtaining eigen dimensionality characterizing high spectroscopic data, thereby realizing dimension reduction of the high spectroscopic data. The method is an automatic method for dimension reduction of the high spectroscopic data, which can effectively remove relativity of wave bands of the high spectroscopic data and improve separability of different categories of ground objects.

The invention discloses a data dimension reduction method based on neighborhood rule. The method includes the following steps: firstly, a spherical neighborhood of present sample points is established by using the geometric spherical-modelling theory and all the sample points contained in the spherical neighborhood are adopted as candidate neighbor points, thus not only preserving the effectivity of the dimension reduction capability when data sets are sparse but also getting the advantages of low-sensitivity to isolated points and good stability of the preserved topological structure; then a data relevance matrix more matching semantics can be obtained by relevance measurement based on route clusters to update the candidate neighbor points in the spherical neighborhood and optimize the regular neighborhood space of the present sample points, thus improving the phenomenon that the dimension reduction of sample sets provided with folded curved faces is apt to suffer the integrated-structure distortion in case of heterogeneous data distribution. The experiments on different sample sets demonstrate that the method provided by the invention is available and effective.

The present invention provides a liquid crystal hyperspectral calculation imaging measurement device of three-dimensional encoding. The device comprises a front-end lens 2, a wave band selection and splitting module 3, a space encoding module 4, a collimating lens 5, an area-array detector 6, a data storage module 7 and a calculation reconfiguration module 8; and based on the three-dimensional encoding, the measurement device performs projection measurement of the three-dimensional spectral data of an object consisting of two-dimensional space information and one-dimensional spectral information in the random encoding information, and performs dimensionality reduction of the hyperspectral data at the data collection phase to obtain the compressed hyperspectral data with the selected central wavelength. Compared to the traditional hyperspectral imaging system, the liquid crystal hyperspectral calculation imaging measurement device and method of three-dimensional encoding can realize the compressing sampling on the space, and can perform spectrum selection at the data collection phase so as to realize the data dimensionality reduction, avoid data redundancy, reduce the data volume, improve the information utilization and facilitate rear-end transmission and storage.

The invention discloses a multi-index fused hyperspectral remote sensing image dimensionality reduction method. Dimensionality reduction can be efficiently and rapidly achieved through a band selection method. At first, subspace decomposition is carried out on obtained hyperspectral remote sensing images, the entropy and the standard distance between mean values of a band and correlation coefficients between bands are fused by means of a method based on Choquet fuzzy integral in each subspace, and then band selection is carried out according to unified indexes to achieve data dimension reduction.

The present invention provides a health state assessment method based on physiological data and psychological data, a health state assessment device, and a system for rehabilitation. The health stateassessment method includes the following steps that: physiological data and psychological data related to health states are acquired, the physiological data are normalized, and data dimensionality reduction and / or data cleaning is performed on the psychological data; data weighting and value assignment are performed on the normalized physiological data and the psychological data which have been subjected to the data reduction and / or data cleaning; and health state assessment is carried out according to physiological and psychological comprehensive data which have been subjected to the weighting and value assignment. The health state assessment method is in line with the actual situation of each patient; the evaluation result of the method is more accurate; and rapid and effective rehabilitation can be achieved.

The invention discloses a hyperspectral image classification method based on extended morphology and active learning, which solves the problem that the prior art can not fully excavate the hyperspectral image space information and leads to low classification accuracy. The steps are as follows: 1) inputting hyperspectral image data; 2) reducing the dimension of the data, extracting spectral features, and obtaining spatial features through morphological section transformation; 3) fusing the space spectrum features to divide the train and test sample sets; 4) classifying that SVM by using the train sample set; 5) performing an active learning cycle, wherein sample marks are selected by that MCLU criterion and the AP clustering, and training and testing sample set are updated; 6) performing SVM classification by using a new training sample set until the number of training samples reaches a preset number, and obtaining the final classification result. The invention combines the morphological characteristics of multiple structural elements with active learning, makes full use of space spectrum information, and improves classification accuracy on the premise of small samples.

The invention discloses a residual service life prediction method for large-scale equipment based on multi-parameter feature fusion. The method comprises the steps: obtaining multiple sensor time sequence parameters of large-scale equipment in a laboratory through a large-scale online monitoring system; performing regression analysis on the multi-parameter continuous values by using a ReliefF algorithm, and obtaining a parameter type with relatively large correlation with the equipment state through feature weight screening; performing data dimension reduction and feature extraction on the screened parameters based on a principal component analysis method, and obtaining a health index representing the operation state of the large-scale equipment through weight fusion; constructing an HMM model based on an expectation maximization algorithm, taking the health indexes as a training set for model training, and finding a grading model used for evaluating the equipment health state corresponding to the current health index; calculating an exponential likelihood value through a Viterbi algorithm to obtain a health index nearest to the likelihood value, and predicting an exponential difference by using a weighted average method to obtain a health state fitting curve; and calculating a residual service life prediction value of the large-scale equipment.

The invention discloses a data dimension reduction method based on neighborhood similarity and mainly solves the problem that an existing method only uses the Euclidean distance to measure a sample neighborhood structure, so that recognition results are non-ideal when a data structure is not balanced. The data dimension reduction method comprises the following realization steps: (1) inputting and normalizing data and randomly initializing a basis matrix and a coefficient matrix; (2) calculating a diagonal covariance matrix of a sample; (3) calculating KL (Kullback-Leibler) divergence through the diagonal covariance matrix; (4) calculating neighborhood sample similarity through the KL dispersion degree; (5) calculating a neighborhood class label distribution matrix of the sample; (6) calculating neighborhood class label similarity through the neighborhood class label distribution matrix; (7) calculating neighborhood similarity through the neighborhood sample similarity and the neighborhood class label similarity; (8) applying to iterative criterions according to the neighborhood similarity to obtain the basis matrix and the coefficient matrix after dimension reduction. The data dimensionality reduction method is high in accuracy rate and can effectively perform feature extraction and dimensionality reduction on data and be used for data and image processing.

The invention relates to the technical field of using electroencephalogram for identification. The invention provides a method capable of more comprehensively reflecting and analyzing information of electroencephalogram, extracting effective electroencephalogram feature parameters with obvious individual variation from the information and realizing the aim of identification. Therefore, the technical scheme of the invention is as follows: the resting electroencephalogram identification method based on bilinear model comprises the following steps: using an electrode cap worn on the head of a subject to collect the original resting electroencephalogram signals; processing the original resting electroencephalogram signals; establishing a composite model with of linear and nonlinear components; adopting main components to analyze PCA and perform data dimension reduction; and performing identification based on a support vector machine. The invention is mainly used for identification.

The invention discloses a method for carrying out producing area tracing on radix salviae miltiorrhizae by combining determination of the content of microelements and rare-earth elements with multivariable statistical analysis. According to the method disclosed by the invention, difference analysis is carried out on all elements of all samples by adopting one-way analysis of variance. When p is less than 0.05, element data of different producing areas has obvious difference. Main component analysis is used for data dimension reduction analysis and linear judgment analysis is used for estimating whether radix salviae miltiorrhizae medicinal materials from the different producing areas have remarkable difference or not. The invention provides the method for carrying out producing area tracing on the radix salviae miltiorrhizae; and the method has the advantages of sensitivity, rapidness and high judgment accuracy.

The invention discloses a video stream image-based block signal conversion and target detection method and a block signal conversion and target detection device. according to the method, a video imageis subjected to grid segmentation, and the signal fluctuation value of each region at a video monitoring moment is represented by means of the variance values of all pixel values in the above segmented block region of segmented two-dimensional sub-images. In this way, the data dimensionality reduction treatment of a two-dimensional image to a one-dimensional signal is completed. An integral videoimage is converted into a plurality of one-dimensional signal waveforms to be detected. For segmented sub-images in each region of the image, corresponding one-dimensional signal data is obtained. Asa result, through the display method of real-time waveforms, the noise reduction treatment is carried out by means of the smoothing and filtering method during the digital signal processing process.The weak change of the background in sub-regions is excluded. When a foreground object or an abnormal condition occurs, the one-dimensional signal shows a strong fluctuation change. As a result, the abnormal fluctuation existing in the video stream can be obviously judged. In this way, whether the foreground target appears in the video detection area or not can be judged.

The invention discloses a deep learning method for flow abnormity detection of an intelligent power grid server. According to the method, on the basis of the high-dimensional time sequence data information for monitoring the flow of a power grid server, an Encoder-Dector deep neural network based on two layers of LSTM neurons is utilized to construct the flow normal state estimation, and based onthe estimated normal flow behavior, the distribution condition of the residual error between the normal flow behavior and the measured value is analyzed, and finally a confidence interval is given asa control limit for judging the abnormal behavior, thereby realizing the detection of the abnormal flow behavior of the smart grid server. According to the method, the nonlinear relation between variables can be dynamically approximated in a self-adaptive mode, the manual intervention is not needed for selecting variable characteristics, the data dimension reduction is not needed, the certain universality is achieved, and the method has very important scientific significance and application value for preventing severe faults.

A method and system for identifying representative data trends using sketches. A sketch is a lower dimensional vector used to represent higher dimensional data. The properties of sketches include data dimensionality reduction, sketches synthesized from other sketches, and the distance between sketches comparable to the distance between the data the sketches represent. Exemplary embodiments include identifying relaxed periods and average trends.

The invention discloses a state evaluation method of a communication network of an intelligent substation based on clustering and neural network. The method comprises the following steps of: introducing the weight obtained by an analytic hierarchy process into a standard Euclidean distance space algorithm for data dimensionality reduction; then, dividing a network abnormal state into five classesby using a clustering method; taking a classification result as an ideal output basis of a neural network model training sample; and finally, building a neural network model based on nine evaluation indicators, and carrying out state evaluation on the communication network of the intelligent substation. The state evaluation method uses an evaluation model combining the clustering and the fuzzy neural network; therefore, the state can be effectively evaluated; the interaction between influencing factors can be well characterized; and more real-time and accurate evaluation results can be obtained.

The invention provides an unsupervised clustering method and device for power system operating conditions, wherein the method comprises the following steps: obtaining power system power flow cross sections to be analyzed at different times, and constructing power flow vectors corresponding to each power flow cross section, wherein all power flow vectors constitute power flow vector spaces; using at-distribution random nearest neighbor embedding algorithm for data dimensionality reduction in the power flow vector space; using a hierarchical clustering algorithm for clustering analysis of the reduced dimension power flow vector space to obtain the clustering result of the operation conditions of the power system to be analyzed. In the clustering analysis of power system operating conditions, in addition to considering trend information, power network topology information is also used, the t-distribution stochastic nearest neighbor embedding algorithm is utilized to reduce the dimensionof power flow vector can effectively reduce the computational complexity while preserving the local structure of the original data. The hierarchical clustering algorithm is utilized to cluster the reduced dimension power flow vector can perform better in dealing with the complex distribution of samples.

The invention relates to a time series data classification method based on multi-level shape. The method comprises the following steps: 1) preprocessing time series data: carrying out data dimension reduction processing on an original time series by using an SAX method; 2) obtaining an initial sub-sequence of the time sequence: extracting a sub-sequence set in the time sequence by a window slidingmethod, and indirectly controlling the extraction length of the sub-sequence by changing and adjusting the size of a window; step 3) discovery and extraction of a multi-level Shapelet candidate set:filtering and merging the candidate set through the proposed multi-level Shapelet framework, and selecting the Shapelet with large information gain as the candidate set; and (4) carrying out Shapeletconversion and constructing a classifier. According to the invention, the method is adopted, an efficient multi-level shapelet candidate set filtering model is provided, the number of Shapelet candidate sets is effectively reduced, the Shapelet sets with high classification capacity are rapidly screened, and then effective classification of time series data is achieved through an ELM classifier.

A mineral identification method based on full-spectrum-segment hyperspectral remote sensing data is disclosed. The method comprises the following steps of (1) reading hyperspectral data in different wave band ranges; (2) carrying out minimum noise separation on an image and carrying out data dimensionality reduction; (3) calculating an information entropy of a pixel in a minimum noise separation result obtained in the step (2), setting a threshold value and extracting the pixel with a small information entropy; (4) corresponding the pixel extracted in the step (3) to an original image according to a pixel position, acquiring spectral characteristic parameter, and comparing with and marking the spectral characteristic parameter of a mineral spectral curve in a spectral library; (5) inputting a marked sample into a learning device and training the learning device to obtain a mineral identification result of each single wave band range; and (6) based on a main body majority voting method, fusing each wave band range identification result and completing full-spectrum-segment mineral identification. By using the method of the invention, higher identification accuracy can be acquired when prior information of an identification area is less, and full-spectrum-segment data can be used to make an identification result be comprehensive and accurate.

The present disclosure relates to a motion trajectory representation method and device, and belongs to the technical field of data processing. The method includes: acquiring k position data corresponding to the motion track of the target object; determining the access situation of the motion track to preset n hotspot areas according to the k position data, and the hotspot area is calculated according to several historical motion tracks The obtained geographical area whose visit popularity is higher than the preset threshold; determines the trajectory feature vector corresponding to the movement trajectory according to the access situation, and the trajectory feature vector is used to represent the characteristics of the movement trajectory. This disclosure solves the problem that the large amount of data brought by directly using position data to represent the motion trajectory is not conducive to the calculation and processing of the subsequent analysis of the motion trajectory; In order to simplify the representation, the effect of data dimensionality reduction is achieved, and it is convenient for the subsequent analysis and calculation of the motion trajectory.

The invention discloses a hyperspectral remotely-sensed data dimensionality reduction method based on improved hierarchical clustering. The hyperspectral remotely-sensed data dimensionality reduction method comprises the following steps of: selecting hyperspectral remotely-sensed image data to be analyzed, wherein the hyperspectral remotely-sensed image data comprise L wave bands; calculating a spectral distance between every two wave bands by a security identification (SID) algorithm to acquire a clustering center extracted for setting a spectral distance matrix and a number k of wave bands to be selected; performing clustering analysis on the image data by a hierarchical clustering method based on a similarity distance matrix; acquiring k clustering center data to finish a feature extraction process; and selecting the most representative wave band from each clustering center to acquire k wave bands to finish a feature selection process. By the method, the dimensionality reduction efficiency can be improved; and the data information loss caused by the conventional hyperspectral remotely-sensed data dimensionality reduction method is reduced.

The invention provides a pedestrian motion mode recognition method based on a deep hybrid model. The method comprises: acquiring data of four mobile phone placement positions and seven daily pedestrian motion modes by using an acceleration sensor, a gyroscope and a magnetometer which are arranged in the smart phone; automatically extracting features by adopting a convolutional neural network (CNN), then performing dimension reduction processing on the extracted features through a principal component analysis method, and inputting a processed result into a hybrid model for recognition in an XGBoost learning mode. According to the method, the CNN serves as a trainable feature extractor capable of automatically obtaining features from input, PCA is used for reducing the dimension of high-dimensional feature data to reduce the calculated amount, the XGBoost serves as a recognizer on the top layer of the network to output a result, and high reliability of feature extraction and classification is effectively guaranteed.

The embodiment of the invention discloses a post-loan risk assessment method and a device, and a storage medium, and the method comprises the following steps: carrying out the preprocessing of storedloan data, including data dimension reduction, data re-classification, data combination and data cleaning, and the loan data after classification comprises static data and dynamic data; performing feature engineering on the preprocessed dynamic data to extract corresponding feature engineering features; constructing an evaluation model based on a deep learning network, and training the evaluationmodel by adopting expert features and feature engineering features in the financial model; and performing data evaluation based on the evaluation model and outputting a final approval list. Accordingto the method, the efficiency of post-loan data analysis can be improved, the precision of a risk assessment model is improved, the predictable times of default are increased, and finally, the post-loan risk can be effectively controlled.

A data dimension reduction method based on data global-local structure preserving projections includes the following steps: 1), creating a sample set and calculating Euclidean distances among sample pairs; 2), diving territory of each sample point to acquire neighboring points and non-neighboring points; 3), creating a neighboring-right matrix and a non-neighboring-right matrix of the sample set; 4), creating target functions corresponding to data global structure preservation and data local structure preservation, and constructing optimization problems; 5), transforming the optimization problems into generalized eigenvalue problems, and creating a projection matrix according to feature vectors acquired by solution; 6), projecting the sample set to obtain dimension reduced data. According to the method, global and local structures of high-dimensional data are maintained during dimension reduction, the dimension reduced data can describe basic features of original high-dimensional data completely and accurately, and accordingly, data dimension reduction effect is improved.

The invention discloses an application method of a machine learning classification model in adolescent autism auxiliary diagnosis. The method is characterized by being implemented according to the following mode. The method comprises the steps of 1, establishing a model training method; 2, constructing a model evaluation index; 3, performing characteristic engineering of the autism auxiliary diagnosis system; 4, performing data dimension reduction processing; 5, carrying out feature selection; and 6, carrying out model training and result analysis. According to the invention, a machine learning method is introduced into the field of autism research; the high efficiency and the reliability brought by the method are greatly helpful to the auxiliary diagnosis of autism. The application fieldof the invention can be embodied in: (1) disease diagnosis and treatment, (2) smoking addiction, network addiction and network game addiction, (3) cognition and other health fields, and the like.

The invention discloses a risk driving scene rapid extraction and grading method for intelligent network connection automobile testing. The risk driving scene rapid extraction and grading method comprises the steps that (1) carrying out intelligent network connection automobile driving time sequence data normalization processing; (2) carrying out dimension reduction processing on the intelligent networked automobile driving time sequence data; (3) performing discretization and symbolization processing on the intelligent networked automobile driving time series data; and (4) carrying out rapidextraction and hierarchical processing on the intelligent networked automobile risk driving scene. According to the risk driving scene rapid extraction and grading method for the intelligent networkedautomobile test, various risk driving scenes in the intelligent networked automobile test working condition can be rapidly extracted and graded, and a theoretical method support can be provided for comprehensive construction and refining optimization of a driving scene library in the subsequent intelligent networked automobile test stage.