104 results about "Dimensionality reduction algorithm" patented technology

The invention provides an ultra-high-dimensional data dimension reduction algorithm based on information entropy, belongs to the field of high-dimensional data preprocessing, and aims to solve the following problems existing in actual application of a conventional PCA algorithm: when data dimensionality (feature) is high enough, all data characteristic read values cannot be stored in memory in one time for analysis and calculation; a partitioning processing method, without depending on a cloud platform and a distributed type calculation platform, is used for experiment; however, the method is long in time consumption and cannot satisfy actual application requirement. On the basis, the thought of the information entropy is introduced for improving the PCA algorithm; the improved algorithm can be used for processing the ultra-high-dimensional data dimension reduction; and experiment results prove that the operating time consumption of the improved algorithm is shortened by 60 times compared with that of the partitioning processing algorithm when it is ensured that the same proportion of the original data information is reserved.

The invention belongs to the signal detection and depth learning technology field and discloses an unknown-class signal detection method based on a monitored learning classification model. The CNN istrained through a tagged data set, and the monitored learning classification model is established; the last SoftMax layer is dropped, and the classification model is utilized to carry out characteristic extraction; the characteristics extracted by the CNN are utilized, the t-SNE dimensional reduction algorithm is utilized to visually detect the classification effect and obtain low-dimensional characteristic data; the two-dimensional data after dimensional reduction is utilized, and frame detection accuracy is calculated through utilizing a distance-based calculation formula. The method is advantaged in that the CNN+t-SNE detection method is creatively proposed, the high-dimensional data is mapped to the low-dimensional space, the more intelligible result is acquired, calculation is convenient, added unknown-class signals can further have their own center to be gathered in one position, essential characteristics of the input data can be excellently extracted through the classificationmodel, and the unknown-class signals can be effectively detected.

The invention provides a short video recommendation model based on short video multi-modal features, and the model comprises the steps: 1, carrying out the feature extraction of short video title features through employing a TF-IDF method, and reducing the dimension of a short video title feature vector to k dimension through employing a PCA dimension reduction algorithm; 2, extracting 128-dimensional original features of the short video content, and adopting a PCA dimension reduction algorithm to reduce the dimension of the feature vector of the short video content to k dimension; 3, extracting 128-dimensional original features of the short video background music, and reducing the dimension of the feature vector of the short video background music to k dimensions through a PCA dimension reduction algorithm. According to the invention, the influence effects of the feature data of different modes on the user behaviors generated by the user are considered to be different; and learning the influence proportion of different modal data of the short video on the user by using a hidden Markov model, and mapping the multi-modal features of the short video to a unified vector space for fusion based on the influence proportion to obtain short video feature data represented by the multi-modal data features.

The invention belongs to the field of power generation scheduling of power systems and discloses a network province two-stage multi-power short-period coordination peak shaving method. On a network province two-stage coordination platform, the operation characteristics of different types of power sources such as water, fire, pumping storage and nuclear power can be given full play to on the basis of the daily load complementary characteristic of a plurality of provincial level power grids, and the peak shaving requirements of the multiple power grids are coordinately met. According to the technical scheme, an original problem is decomposed into a pumping storage sub-problem, a conventional hydropower sub-problem, a conventional thermal power sub-problem and a conventional nuclear power sub-problem which are independent of one another according to the types of the power sources, and a pumping storage heuristic search algorithm, a water and electricity hybrid dimensionality reduction algorithm and a thermal power load shedding peak shaving algorithm are adopted for solving the sub-problems; on the basis, system loads are used as correlation factors, the preceding algorithms and an internetwork electric power distribution quadratic programming method are coupled to form a coordination framework of the sub-problems, electric power distribution of power stations among the multiple power grids is adjusted through iteration, and the differential peak value load requirements of the multiple provincial level power grids are met. Through the method, the multiple power sources and the multiple power grids can efficiently and coordinately run, a regional power grid network debugging power source peak shaving effect can be played better compared with actual scheduling, and the method is efficient and practical.

The invention relates to a large-scale multi-view data self-dimension-reduction K-means algorithm and a system, and belongs to the technical field of information processing, and the method comprises the steps: 1, carrying out the normalization of data with different features, and enabling all data to be in a range of [-1, 1]; 2, initializing; 3, optimizing the algorithm; and 4, using a data set tooptimize the algorithm according to the algorithm until the algorithm is finally converged to obtain a final clustering result, measuring the clustering effect by using the interaction information entropy and the purity, and repeating the step 3 by selecting different initial values, and removing the average value of the result to complete the experiment. The relationship between the features andthe clustering targets is fully considered; information complementation among different views is utilized, self-dimension reduction of high-dimensional data is achieved by searching for an optimal subspace on a single view, a loss function is reconstructed through non-negative matrix factorization (NMF), the different views share the same clustering indication matrix, therefore, multi-view information complementation is achieved, and clustering of large-scale multi-view data is completed.

The invention relates to the technical field of tea quality monitoring, in particular to a method for monitoring the fermentation degree of black tea through a hyperspectral coupled nanocrystallization colorimetric sensor. Volatile substances in the fermentation process of black tea are captured by utilizing a nanocrystallization colorimetric sensing array, colorimetric array feature information is efficiently extracted by combining a hyperspectral image technology with dimension reduction algorithms such as principal component analysis and linear discriminant analysis, and an information fusion discrimination model with strong robustness and high accuracy is established by adopting algorithms such as partial least squares discrimination, multivariate linear discrimination, a support vector machine, an extreme learning machine, an artificial neural network and a deep belief network to realize rapid and accurate discrimination of the fermentation degree of black tea. The method has the characteristics that the analysis speed is high, the sensitivity is high, the cost is low, a sample does not need to be pretreated, and online nondestructive detection is facilitated.

The invention discloses a non-contact single step-counting method and system based on commercial Wi-Fi. The method comprises the following steps: making a repeating motion by a participant according to preset action requirements, and acquiring a CSI (Channel State Information) signal of the participant by virtue of the commercial Wi-Fi; defining that d(t) is the sum of mean absolute deviations ofall amplitude data in a time sliding window, solving a noise level in a time period t according to the d(t), and screening out a noise signal from the CSI signal; eliminating a static element of eachsubcarrier in the CSI signal by virtue of a Hanpur filter, and removing other noise by using a Savitsky-Gray filter; extracting a single signal from the CSI signal subjected to noise reduction by adopting a dimensionality reduction algorithm for factor analysis; performing time-frequency analysis, feature extraction and peak detection on the single signal, wherein the peak detection specifically refers to detection on four constraint conditions such as corrugation pitch, crest height, crest convexity and crest width. The system comprises a microprocessor, a WiFi transmitting end and a WiFi receiving end.

The invention discloses a network fault diagnosis method based on a wavelet neural network. The method comprises the following steps: S1, obtaining network data in a fault state and a normal state; S2, carrying out numeralization and normalization processing on the network fault data, and carrying out data dimension reduction by adopting a PCA dimension reduction algorithm; S3, creating a wavelet neural network model, selecting an improved grey wolf optimization algorithm, and taking optimized parameters as parameters of the wavelet neural network model; then taking the network fault data processed in the second step as input, adding momentum factors when parameters are reversely adjusted, and establishing a network fault diagnosis model through continuous training; S4, inputting real-time network state data, and judging whether the network has a fault; and S5, outputting a network fault diagnosis result and a fault category. The momentum factor is introduced, so that the local optimization capacity of the diagnosis model is improved; an improved grey wolf algorithm is adopted, initial parameters of a fault diagnosis model are optimized, and the randomness of initial parameter selection is avoided.

The invention relates to the technical field of electric power project construction and in particular to a cost analysis method for a power transmission line project. In an existing analysis method, atypical scheme can be rapidly screened when engineering parameters are relatively single, however, when the engineering parameters and the typical scheme are analyzed to have a great difference, thetypical scheme screened subjectively and technical characteristics of a to-be-analyzed project may have great difference, thus, the cost analysis based on the technical scheme deviates from a practical result. The invention provides the cost analysis method for the power transmission line project. The method comprises the following steps of establishing a basic database; preprocessing basic indexdata; constructing a project characteristic dimensionality reduction algorithm based on a principal component analysis; establishing a grey relational model; calculating a project investment of a representative scheme; and performing a cost comparison analysis on the to-be-analyzed project and the representative project investment.

The invention relates to a system for life evaluation and forecasting by repeatedly using a carrier and a method for life evaluation and forecasting by repeatedly using the carrier, and belongs to thetechnical field of spacecraft maintenance and guarantee. Specific to the problem of repeat use of the carrier through many complex severe environments, sensors distributed on the carrier are mainly used for collecting data of an aircraft component in real time; a weighting and dimensionality-reduction algorithm is applied to perform weighting and dimensionality-reduction on part component data required to be monitored, thus the index quantification difficulty of life evaluation and forecasting is solved; on this basis, the residual life of the part component is matched by combining with the inherent property data of the product and others; the sensor data updated in real time can update the fitted residual life curve, thus the forecasting result is more exact and reliable.

Embodiments of the present invention disclose an active anchoring positioning optimization control method and system. The method comprises: acquiring anchor cable tension data corresponding to each anchor cable in the platform, calculating anchor cable tension data required for each anchor cable since the platform reaches a desired position under the current environmental disturbance, and assigning the anchor cable tension data to the respective anchor machine controller to adjust the anchor cable tension of each anchor cable; detecting platform displacement change data in real time; comparingthe platform displacement change data with the expected displacement, and outputting a comparison result; and based on the comparison result, optimizing the tension threshold corresponding to each anchor cable under the current environmental disturbance, and assigning the optimized new tension threshold to the respective anchor machine controller for adjustment. According to the technical schemeof the present invention, the defect that only the approximate solution of the tension assignment can be obtained by using the relaxation method is overcome, the dimensionality reduction algorithm isused to strictly satisfy the equation, and the defect of the dimensionality reduction algorithm is solved; and by using the optimization scheme of the K-value weight in the adaptive adjustment evaluation function, the anchor equipment can actively adjust the tension, and the positioning ability of the ship is improved

The invention discloses a soil classification method based on a hyperspectral computing imaging system, thereby realizing rapid and accurate classification of different types of soil. According to theinvention, soil classification is carried out by using an LCTF-based hyperspectral computing imaging system; a compressed measurement value of a soil image is measured and reconstruction of an original spectral image is performed by using acompressed sensing theory, thereby improving the spectral resolution of the soil spectral image; on the basis of a spectral difference theory, differences between different soil types are enhanced by using spectral differences between all soil samples and all soil types as input features; and then classification is carried out by using a 3D-CNN, and advantages of the spectral images are utilized by using soil spectral image space and spectral information. According to a feature dimensionality reduction algorithm, the original spectral information is kept to the greatest extent, the spectral dimension is reduced, the training efficiency is improved, the type features of the soil are highlighted, and the classification accuracy is improved.

Embodiments of the present invention provide an Android application threat degree evaluation model establishing method, evaluation method and system. The model establishing method comprises: extracting feature data, including authority information and additional feature information of training samples, and constructing feature vectors according to the state of the feature data; using a clusteringalgorithm to cluster a feature vector set of the training samples according to the authority information, and dividing the feature vector set into different feature vector clusters to obtain a clustering algorithm model; using a data dimension reduction algorithm to perform feature selection on feature vectors in the feature vector clusters, and obtaining a corresponding feature selection dictionary; classifying feature vectors in the feature vector clusters by using multiple preset machine learning algorithms; obtaining a corresponding machine learning algorithm according to a classificationresult to obtain a classification algorithm model; and establishing an Android application threat degree evaluation model according to the preprocessing algorithm, the clustering algorithm model, thefeature selection dictionary and the classification algorithm model. According to the embodiments of the present invention, the classification accuracy and universality of the application threat degree evaluation are improved.

The invention discloses an influence evaluation system construction method and system and influence evaluation method. The influence evaluation system construction method comprises the following stepsof: obtaining a set quantity of history data of at least one influence index in at least one influence dimensionality; respectively converting the obtained data of each influence dimensionality intodecibel numbers; and constructing an influence index model through a dimensionality reduction algorithm according to the decibel numbers. According to the influence evaluation system construction method, comprehensive influences of media objects such as articles, events and media are scientifically and quantitatively assessed through basic data such as reading numbers, like numbers, forwarding numbers and comment numbers; besides ensuring mathematically optimum, the method does not depend on preferences of calculation formula designers; all the weights are calculated according to large-scale data, so that objectivity is provided; and the finally calculated network media influences, namely the influence indexes, are based on specific media with high influences, so that values of the influence indexes have clear physical meanings.

The invention discloses an Android malicious application detection method fusing multi-feature classification, belongs to the technical field of Android malicious software detection, and particularlyrelates to an Android malicious application detection method fusing multi-feature classification. The method comprises the following steps: extracting Android permission and static broadcast characteristics by adopting a static analysis method; extracting dynamic broadcast and sensitive API features by adopting a dynamic monitoring method; according to the method, and carrying out dimensionality reduction and decorrelation processing on the features based on the Pearson correlation coefficient dimensionality reduction algorithm. Through carrying out classification detection on the extracted features through the random forest algorithm, existing Android malicious application programs can be detected, and meanwhile Android malicious programs leaking privacy through a broadcast mechanism canbe effectively detected.

An embodiment of the invention discloses a voice dereplication method, a device thereof, a server and a storage medium, wherein the voice dereplication method comprises the steps of acquiring an MFCCcharacteristic matrix of target short voice by means of a Mel-frequency cepstral coefficients (MFCC) algorithm, and converting the MFCC characteristic matrix to a target image; based on a deep learning model and a characteristic dimension reducing algorithm, extracting the target image characteristic of the target image, and determining a target index of the target image characteristic; determining each historical image characteristic which corresponds with each historical short voice according to the target index, and determining whether the target short voice is a repetition voice by means of a repetition degree between each historical image characteristic and the target image characteristic. The voice dereplication method, the device thereof, the server and the storage medium overcome defects of ignorance to deep information of a voice content and rough evaluation to two voices with similar contents in an existing voice dereplication method, and realizes quick and effective dereplication processing on the voice data based on the voice content.

The invention provides an image recognition and classification method of a multi-classification support vector machine based on minimum reconstruction error search dimension reduction and particle swarm optimization which is used for image content recognition and classification. According to the method, firstly, optimal parameters of a dimension reduction algorithm are searched for through minimumreconstruction errors, then dimension reduction is conducted on samples in a handwritten digital data set through the optimized minimum reconstruction error search algorithm, key information is extracted from the samples, and the dimension is reduced to 17 dimensions. Secondly, a plurality of one-to-one support vector machine classifiers is constructed according to the number of sample categoriesin the data set, and the support vector machine classifiers are combined to indirectly realize multi-classification; and with the trained highest classification accuracy adopted as a target, global optimization is performed on the parameters of each support vector machine classifier by using a swarm intelligence optimization algorithm, so that a final image recognition and classification method with high recognition accuracy can be obtained.

The invention discloses a clustering seismic facies analysis method based on feature coding of a restricted Boltzmann machine, which comprises the following steps of: (1) processing actual seismic data, eliminating noise in a seismic signal and improving a signal-to-noise ratio; (2) based on the seismic sedimentology principle, forming formation slices through uniform segmentation along the top and bottom boundaries of a target layer; (3) introducing an unsupervised restricted Boltzmann machine dimension reduction algorithm based on the formation slices, and extracting potential seismic waveform information capable of expressing real target layer reservoir feature changes while reducing dimensions; and (4) for feature selection after dimension reduction, completing spatial clustering analysis of seismic waveform data, and forming a corresponding seismic facies diagram. According to the method, research is carried out on the aspects of denoising, feature extraction, unsupervised learning, semi-supervised learning and the like of the seismic pre-stack waveform based on deep learning, how to better generate the seismic facies diagram by using the extracted low-dimensional features is researched, and geology interpretation work is effectively helped.

The invention relates to a tissue blood oxygen imaging detection method based on two-stage spatial mapping. The tissue blood oxygen imaging detection method comprises the following steps of: constructing a tissue structure model before an operation; setting material types and optical parameters in tissues; performing reflection type optical imaging simulation to obtain high-dimensional full-spectrum information; obtaining a mapping data set of a low-dimensional RGB channel based on the spectral characteristics of an illumination light source and an imaging camera; by using a manifold dimensionality reduction algorithm, achieving dimensionality reduction of high-dimensional spectral data with the blood oxygen saturation as a main parameter, constructing a first kernel function, and achieving mapping of a data set obtained after dimensionality reduction and a low-dimensional RGB channel data set; obtaining an inverse operation mapping relation of the first kernel function based on an L1optimization strategy; and, imaging the actual to-be-detected tissue in the operation, and calculating in real time through the adjusted inverse operation mapping relation to obtain an estimated valueof the blood oxygen saturation. Compared with the prior art, the two-stage spatial mapping relation is utilized in the invention; the accuracy of blood oxygen parameter estimation can be effectivelyimproved; and meanwhile, the real-time performance of imaging detection is guaranteed.

The invention relates to a human body tracking method based on a self-adaptive kernel function and mean value shifting. The human body tracking method includes two stages, the first stage is a learning stage, a set of training samples of human body walking is firstly read, human body prospect shapes are mapped to be coordinates in a low-dimensional space through a dimensionality reduction algorithm, a low-dimensional human body shape space is obtained, the human body prospect shapes are then recovered through an interpolation reconstruction algorithm, and parameters, capable of mapping from a low dimension to a high dimension, of the interpolation reconstruction algorithm can be obtained. The second stage is a tracking stage, a human body optimum kernel shape in a video frame is searched for in the low-dimensional human body shape space, and the human body in the video frame is tracked by using a mean value shifting algorithm. Compared with the prior art, the human body tracking method improves the shape of the kernel function in a traditional mean value shifting algorithm, so that the shape of the kernel function is not fixed and changes in a self-adaptive mode according to changes of shapes of the tracked human body, histogram modeling and matching of the kernel function are further performed in the high dimension space, and therefore the performance of a human body tracking technology is improved.

The invention discloses an automatic visual analysis method and system for celestial body spectral data, and belongs to the technical field of astronomical spectral analysis and big data processing and analysis. The problems that many tedious screening, drawing and other work are needed during spectral analysis, and a large number of unknown types of spectrum clustering and outlier spectrum searching and analysis are lack of automation are solved. According to the technical scheme, the method comprises the following steps: acquiring spectral data and preprocessing the spectral data; reducing the preprocessed spectral data to two dimensions by using an SP-tSNE dimension reduction algorithm; projecting the data after dimension reduction to an interactive two-dimensional plane for visualization; carrying out frame selection on points on the interactive two-dimensional plane by using a mouse, and automatically displaying a mean value spectrum of spectrums corresponding to the points in the region by the system; a user marks an emitting ray and an absorbing ray on the spectrum according to own needs, and the position of the ray can be adjusted by adjusting the red shift; the system provides a function of storing the spectrum of the frame selection area; the method is applied to celestial body spectrum analysis.

The invention discloses an optimization method for steganalysis of a convolutional neural network. The optimization method comprises the following steps that the intra-class aggregation degree of samples of the same class is visualized; and dimensionality reduction is carried out on a feature set corresponding to each sample point by adopting a nonlinear t-sne dimensionality reduction algorithm until the feature set is observed under a two-dimensional visual condition; steganography detection is performed on variable coefficients of the sample set; in order to eliminate the influence of the dimension and the measurement scale on the measurement of the aggregation degree of the samples, the variation coefficient is selected to measure the discrete degree of the samples so as to reflect the feature learning ability of the CNN steganalysis algorithm. The value of the variable coefficient is positively correlated with the discrete degree of the sample, and the larger the discrete degree of the sample is, the larger the variable coefficient is; the more the samples are gathered, the smaller the variable coefficient is; the feature set is adjusted based on the variation coefficient posteriori; a feature screening layer is added to manually filter the features obtained by the algorithm, some so-called'bad 'features unfavorable for later classification are removed, the accuracy of the algorithm can be improved to a certain extent, and the measurement effectiveness of the variable coefficient can be verified again.

The invention relates to a working condition recognition method based on point switch action curve similarity characteristics. The working condition recognition method comprises the following steps: 1, selecting a reference template from historical action curves of a point switch; 2, constructing a pairing matrix; 3, calculating the distance di, j of each group of curve pairs in the Pn, m, and constructing an action curve distance matrix Dn, m; 4, performing dimension reduction on the Dn, m through a dimension reduction algorithm; 5, drawing a relative curve shape distribution diagram of a historical action curve C of the point switch; 6, clustering the Fn, 2: [f1, f2,..., fn] through a clustering algorithm; 7, adjusting clustering parameters, and executing the step 6 repeatedly until theshapes of the action curves in the step Sc are the same; and 8, marking the curve in the Sc by using c representing the working condition type to finish the identification of the working condition ofthe point switch. Compared with the prior art, the working condition recognition method has the advantages of effectively solving the problem that the action curve of the point switch cannot be further analyzed through machine learning and big data due to no data label, and being wide in application, visual in working condition change trend, high in efficiency and the like.

The invention discloses a radiotherapy plan assessment method and device based on unsupervised learning, and the method comprises the following steps: S1, obtaining a corresponding anatomical structure vector according to the data of n patients for any disease; s2, performing dimensionality reduction on the high-dimensional anatomical structure vector by using a dimensionality reduction algorithm; s3, clustering the data according to an anatomical structure by using a clustering algorithm to obtain k categories; s4, according to the k categories, analyzing and processing a clinically implemented high-quality radiotherapy plan result in the data corresponding to each category to obtain a plan scoring template corresponding to each category; s5, new data of any patient is calculated, and the category of the new data is judged; and S6, adopting the plan scoring template of the corresponding category to guide the design of the radiotherapy plan corresponding to the new data, and scoring the dose distribution result of the radiotherapy plan. According to the method, personalized evaluation of cases with different geometric structure complexity is realized, and the quality and efficiency of the radiotherapy plan are improved.

The invention discloses a multi-level deep fusion mining method for multi-mode cross-boundary big data of commercial vehicles. The method comprises the following steps: S1, collecting an original dataset of multi-mode cross-boundary big data of a vehicle; s2, performing data preprocessing on the collected original data set; s3, performing data mining on the preprocessed data by using a WEKA algorithm to extract feature keywords; s4, calculating the weights of the feature keywords and the similarity between the different feature keywords through a TF-IDF technology, and constructing a weight and similarity matrix; and S5, constructing a regression model based on the sample. According to the invention, via t-SNE dimension reduction, WEKA algorithm feature extraction and a TF-IDF algorithm,an analysis strategy of dimension reduction and feature extraction in sequence is adopted for high-dimensional data, effective fusion of multi-level deep fusion mining of cross-boundary big data is achieved, and the problems that due to a high-dimensional data set with complex data types and numerous data features, the fusion efficiency is low, and the working efficiency is not remarkably improvedare solved.

The invention provides a single cell data visualization method, system and device and a storage medium. The system comprises a user side, a system terminal and a data memory; the system terminal comprises a user verification module, a data quality module, a PCA dimension reduction module, a dimension reduction clustering module and a cell population marker gene display module. A preprocessing module is used for verifying login information of a user, calling processing data and providing data selection; the data quality module is used for detecting the data quality condition of each cell; the PCA dimension reduction module is used for performing dimension reduction calculation on the high-dimensional data; the dimension reduction clustering module is used for calculating cell clustering conditions obtained by different dimension reduction algorithms; and the cell population marker gene display module is used for calculating a differential gene table and a cell marker gene enrichment table between cell populations according to cell population clustering. By adopting the system, the requirement of a user on single cell data analysis can be basically met, the time consumption is greatly reduced, and the working efficiency is improved.

The invention provides a score clustering analysis method based on t-SNE. The method comprises the steps: importing original data, carrying out the t-SNE dimension reduction of high-dimensional scoredata, carrying out the K-Means clustering of the score data after the t-SNE dimension reduction, and obtaining a clustering result. According to the method, after the original data is preprocessed, the data in the high-dimensional space is subjected to dimensionality reduction by using the t-SNE algorithm, and then the original data is clustered by using the K-Means algorithm, so that the problemof non-ideal clustering effect caused by over-high data dimensionality is effectively solved. Due to the fact that the distribution characteristics of the high-dimensional data are completely reservedthrough the t-sne dimension reduction method, the clustering result of the high-dimensional data is obtained through reduction of the clustering result of the data after dimension reduction. The superiority of the dimensionality reduction algorithm in the student score analysis algorithm can be seen by comparing the results of dimensionality reduction before clustering and direct clustering of the high-dimensional data.