101 results about "Variable screening" patented technology

The invention discloses a bearing state noise diagnosis algorithm facing network variable screening and characteristic entropy fusion. A sensor collects bearing operation noise signals. The noise signals are segmented according to a time sequence and form a sample set. A time-frequency domain characteristic of a sample is extracted so as to acquire a time-frequency-domain one-dimensional characteristic row vector. An average influence value algorithm is adopted to realize first characteristic variable screening so as to acquire a sensitive characteristic set, and through calculating a characteristic entropy of the sensitive characteristic set, characteristic secondary screening and dimensionality reduction are performed on an average influence value similarity characteristic so as to acquire a final characteristic set. A PSO or GA optimization support vector machine is used to carry out training and establish a fault diagnosis model so as to determine a bearing fault type and output aresult. In the invention, complementarity of a characteristic average influence value and the characteristic entropy based on a network in characteristic selection and characteristic classification isused; and a disadvantage that the characteristic selection and a neural network classification algorithm are mutually isolated in bearing noise diagnosis is overcome so that a time-frequency domain characteristic index well reflects a bearing operation state and a classification network characteristic.

A continuously variable screening tool is used to generate a unique texture pattern for every color in a multicolor image. The continuously variable screening tool allows a single colorant version of the image to be generated with less information loss than typically suffered in the multicolor to single color transformation process. The continuously variable screening tool is generated by blending patterns from a set of reference screens. The reference screens are associated with selected reference colors in, for example, a machine independent color space. A calculated screen is generated through a weighted blend of reference screens located near the arbitrary color in the machine independent color space. Typically, the weights depend on the distance the arbitrary color is from each of the reference colors. The screens consist of arrays of threshold values. Each threshold value is associated with a dot position and an image pixel. Where a pixel value exceeds the associated threshold value, provision is made to place a mark in the associated dot position. An image processor operative to carry out the method comprises a continuously variable screening tool generator. An embodiment of the image processor includes a reference screen storage device and a screen blender.

The invention provides a method for construction of a discrimination model for discriminating daqu quality, the method comprises the following steps: 1, information extraction; 2, variable screening and optimization; 3, construction of the discrimination model; 4, verification of the model; and 5, discriminating; and the method is applied to discriminating of daqu quality in distiller's yeast. The method uses 454 high throughput sequencing technology, and solves the problem that the first generation sequencing technology cannot comprehensively and accurately reflect the composition information of daqu microbes; by use of partial least squares method for variable screening and optimization of the microbes and meanwhile combination with quadratic discriminatory analysis, the daqu quality discriminating evaluation is constructed, and is applied to discriminating of daqu quality grade, the method for discriminating finished product qu and out-taken yellow qu and white qu grade is constructed, and the method has the advantages of simple operation and use, higher discriminating accuracy rate, and the like.

The invention discloses a pre-loan risk analysis method, and the method comprises the steps: obtaining first return data of a user, building a variable pool based on the first return data, and the variable pool comprising time slice variables of multiple dimensions; performing time calibration on the time slice variables in the variable pool to obtain calibrated time slice variables after time calibration; screening out a target time slice variable meeting a preset condition from the calibration time slice variables, and performing grouping variable combination on the target time slice variable to obtain a variable short list combination; carrying out in-mold variable screening on the variable short list combination to obtain a model group; and obtaining each variable index, obtaining a prediction model meeting each variable index in the model group, and carrying out pre-loan risk analysis according to the prediction model. The invention further discloses a pre-loan risk analysis device and equipment and a readable storage medium. According to the method, the information dimension is enriched, the situation of causal mismatching is prevented, and the accuracy of a risk analysis result is improved.

The invention discloses a large chemical process distributed modeling method based on CCA-PLS. All process variables are collected in the large-scale chemical process, output variables are determined according to a process quality index, feature components of the process variables are extracted by the adoption of a canonical correlation analysis method, the largest association coefficients between each output variable and all input variables and corresponding main axis vectors are calculated according to the feature components, independent input variables and interaction input variables of each subsystem are selected according to the absolute values of components of the main axis vectors, and accordingly a large system decomposition is achieved. Modeling is carried out on the subsystems by means of a PLS algorithm after subsystem division, components which enable the covariance between the input variables and the output variables to be the largest are extracted, and a subsystem model is obtained through a regression modeling technology. The large chemical process distributed modeling method has the advantages that only input and output data in the process are used, input variable screening of the subsystems is carried out by means of the canonical correlation analysis principle, the model dimension number is lowered, the model structure is simplified, subsystem modeling is carried out through the PLS modeling algorithm, difficulties in calculation caused by a large number of collinear variables in actual application are eliminated, and modeling precision is high.

The invention discloses a method for detecting the content of trace elements in milk powder on the basis of laser-induced breakdown spectroscopy. The method comprises the following steps of making a sample to be tested; building original spectrum data of the sample; testing the content of the trace elements of the sample; amplifying the original spectrum data into a wavelet coefficient carrying more time domain or frequency domain information through high density wavelet transform; performing variable screening by using an improved random leapfrog algorithm; thus screening out variables closely relevant to the tested elements to complete the spectrum pretreatment; combing the correction set spectrum data after the treatment with the measured trace element content; building a correction model by using a partial least squares regression method; obtaining the optimum prediction model of the trace element content in the milk powder. The method has the advantages that the defects of requirement of a great number of samples, long detection pretreatment time, complicated experiments and the like of the exiting method in trace element detection can be avoided. The goal of detecting the content of the trace elements in the milk powder in modes of high speed, great quantity and simple operation is achieved.

The invention discloses a leave-multiple-out cross validation (LMOCV) method of a quantitative structure and activity relationship (QSAR) model of an organic pollutant. In the method, a uniform design method is integrated with LMOCV; related coefficients of uniform design optimized LMOCV (UDOLMOCV) serve as variable screening ending standards; meanwhile, the model is subjected to sample internal cross validation by using the UDOLMOCV during validation in the model; higher sample volatility is provided by using predictive capacity judgment indexes; validated samples are uniformly distributed in sample spaces by infrequent sampling validation of the sample; and the sample which is selected every time has quite high representativeness. The validated sample which is obtained by the method has quite high representativeness of sample distribution, so the defect of singleness of sample selection of a Monte Carlo method is overcome. The invention can be used for validating the stability and predicative capacity of the QSAR model and can be used for discovering and determining instability factors in the model.

The invention discloses a data processing method and a server. The method comprises the following steps: receiving request data, and obtaining feature variables to be processed from the request data, wherein the feature variables to be processed have first feature attributes; obtaining a derivative strategy, according to the derivative strategy, performing intersection combination of combinable features on the feature variables to be processed to obtain derivate variables, wherein the derivative variables have second feature attributes, and the second feature attributes are closer to target expected demands compared to the first feature attributes; obtaining a screening strategy, and according to the screening strategy, screening the derivative variables to obtain candidate target variables; and adding the candidate target variables into a variable pool used for storing the feature variables to be processed so as to perform a next round of variable derivation and variable screening according to the derivative strategy and the screening strategy until the target expected demands are observed.

The invention discloses a boiler NOx prediction method based on MI-LSTM. The method comprises: ranking the importance of auxiliary variables and selecting variables by using mutual information (MI) "minimum redundancy maximum correlation" as a criterion; using a selected auxiliary variable set as an input of a long short term memory (LSTM) model; establishing an MI-LSTM model for NOx emission prediction, and searching for an optimal parameter for the LSTM model by using multilayer grids. In the process of variable selection, a sequence forward selection method is used to determine the optimalinput feature set and the optimal model parameter with the model prediction accuracy as a goal, thereby effectively reducing the information redundancy between the input variables, reducing the modelcomplexity, and improving the prediction accuracy and generalization capability of the model.

The invention discloses an application of an improved partial least squares regression method to microorganism culture medium optimization. According to the application, a cross validation technologyand a variable screening technology are integrated to reduce potential overfitting problem of PLSR, namely, a proper number of potential variables are extracted by carrying out cross validation through a leave-one-out method, and a steady principle of coefficients is decided by utilizing a variable projection importance technology-based variable screening method and a regression model so as to establish simple and steady models with high prediction precision. According to the application, the VIP regression model is capable of effectively establishing the simple and steady models which are high in prediction precision and can be further popularized.

The invention provides a combination optimization-based near-infrared non-invasive blood glucose detection wavelength variable screening method. The method comprises the following steps of firstly obtaining transmissivity spectral data of human blood glucose detection by adopting LED near-infrared light sources with different wavelengths; secondly selecting an optimal variable group in groups by using a continuous projection algorithm, a genetic algorithm and a gradual selection algorithm for multiple pieces of wavelength variable data by taking a root-mean-square error as an index in combination with linear regression modeling; thirdly performing scoring on each group of variables through a weighted scoring method; and finally recombining three groups of variables, sorting comprehensive scores in sequence from high to low, and selecting first few variables with highest scores as final auxiliary variables. According to an algorithm, a most effective wavelength combination is extracted through weighted combination optimization for the problem of an over-fitting phenomenon of a model due to excessive wavelength variables in regression modeling; and the variables are accurately selected, so that the model can be greatly simplified, the calculation complexity of the model is lowered, and the prediction capability and the real-timelines and robustness of non-invasive blood glucose detection are improved.

The invention discloses a method for predicting the temperature of a furnace core dead stock column of an ironmaking blast furnace based on a multi-element linear regression algorithm, and belongs tothe technical field of metallurgy information processing. According to the method, the temperature target value DMTgoal of the furnace core dead stock column is calculated, data is processed, Pearsoncorrelation analysis is carried out on the processed data sample, and condition variables are preliminarily selected according to the result of correlation analysis. Pearson correlation analysis is carried out on each condition variable, and mutually independent condition variables are selected as possible according to the correlation analysis result to establish a model. Condition variable is then screened by a least square method and an AIC-based variable screening criterion, the fitting degree and the regression coefficient of the primary multi-element linear regression equation are checked, and a multi-element linear regression model is obtained. According to the method, the multi-element linear regression algorithm is provided for the first time to predict the temperature of the furnace core dead stock column, the temperature of the furnace core dead stock column in the next five days can be predicted in a high-precision mode, and the early warning function of the temperature of the furnace core dead stock column can be realized.

The invention relates to a method for rapidly predicting macromolecule energy gap. The method is based on a first principle in combination with a support vector machine, and comprises the following steps: looking up the structure of macromolecule and an experiment value of an energy gap from literatures; computing the energy gap according to different first principle methods, and selecting an optimal method; looking up 4 module macromolecule structures of repetitive units from the literatures and optimizing the 4 module macromolecule structures; computing the energy gap of the 4 module macromolecules; generating a descriptor through Dragon software, and dividing data into a modeling set and a testing set; performing variable screening through maximum correlation minimum redundance algorithm, and establishing a model through the support vector machine; rapidly predicating the energy gap of the macromolecules to be detected through the established model. The method is based reliable literature data and modeling methods, and the predicating model for the macromolecule energy gap has the advantages of being convenient, rapid, low in cost and free of pollution.

The invention discloses a wavelength selection method based on PCA modeling feedback load weighting. Based on PCA algorithm, according to the spectral detection data of different frequencies, Establishing and optimizing the metrology analysis model, weighting and combining the PCA load vector with modeling coefficient feedback, measuring the information contribution for each wavelength variable, and then choosing the information wavelength set with higher signal-to-noise ratio, can effectively reduce the number of wavelengths involved in modeling and reduce the complexity of the model; the wavelength combinations selected can be combined with various simple statistical algorithms such as linear discrimination or multiple linear regression to complete the qualitative or quantitative analysis. This method can improve the work efficiency of spectral information variable screening, and can be applied to the near infrared, infrared, ultraviolet and other frequency bands of spectral dimension reduction and rapid detection. It provides theoretical basis and technical support for the development and application of small special spectral instrument, and is expected to be widely used in thefield of hyperspectral image analysis.

The invention discloses a method for predicting deformation of a proximity structure caused by a shield through a random forest fused with an SVM. The method comprises the steps of: collecting data corresponding to main factors according to the main factors affecting the deformation and settlement of a building; establishing a random forest model, training the data according to the random forest model, and performing importance measurement to obtain an optimal feature set; and performing dimensionality reduction processing on the optimal feature set, and inputting the optimal feature set intoa support vector machine model for training to obtain a building settlement prediction result. The invention further discloses a device for predicting the deformation of the proximity structure causedby the shield through the random forest fused with the SVM. According to the method, random forest feature selection is utilized, variables having small correlation with a predicted value can be eliminated from excessive influence factors, key variables for modeling are screened out to obtain an optimal variable combination, the dimensionality of the support vector machine training model is reduced, the prediction precision is improved, a prediction result closer to the reality is obtained, and more stable and more accurate prediction of building deformation settlement is realized.

The invention provides a road rage driver dangerous behavior characteristic variable screening method based on a random forest algorithm. Based on original driving simulator experimental data, a driver behavior data set and a driving anger state set are constructed, dangerous feature recognition is conducted on driver behaviors in the road rage state through a random forest, and the method has practical significance in the aspects of designing a driving anger early warning system subsequently and the like.

The invention provides an agricultural product quality analysis method and an analyzer. The agricultural product quality analysis method comprises the following steps: acquiring spectral data and measuring reference value data; wherein the data is divided into a correction set and an external verification set; arranging the spectral data and the reference value data into a data matrix, and sampling the data matrix by using a Monte Carlo sampling method; carrying out key variable selection on a data matrix sampling result by adopting a variable selection algorithm, carrying out statistics on the selected frequency of each variable, and sorting; carrying out statistical stability on the high relative frequency variables, screening out stable key variables and establishing a mathematical model; substituting the acquired spectral data of the to-be-detected agricultural product into the mathematical model, and analyzing the quality of the agricultural product according to an operation result. In order to realize the application of the method, an agricultural product quality analyzer is developed, and a result is predicted and output according to stable key variables. According to the method, Monte Carlo sampling parameters are optimized, the operation cost is reduced to the maximum extent while a stable key variable screening result is obtained, and the working efficiency is improved.

The invention discloses a non-native permanent resident population identification method based on spectral clustering. The method comprises steps that step (1), the residence data information of handset users is collected; step (2), for the step (1), multiple user residence behavior characteristic variables are constructed; step (3), variable screening for the variables of the step (2) is carried out, and dimensionless processing is further carried out; and step (4), a spectral clustering method is employed to carry out clustering analysis, and a model is established; and the residence data information of the step (1) comprises the residence day number, residence duration and ECI switching data. Statistics of the residence day number, residence duration and the ECI switching data of the users in a present city can be realized according to the collected residence data information, permanent residents and non-permanent residents are different in residence characteristics, so the residence characteristic variables are formed through processing the residence information, the two types of users can be distinguished through employing the clustering algorithm, and the identification result is more scientific.

A competitive adaptive reweighting key data extraction method for Raman spectrum analysis of insulating oil is characterized by comprising the following steps: 1, obtaining spectral data by actual measurement; 2, processing all original data, and extracting a PLS subset from actually measured spectral data through a partial least square method; 3, screening out singular samples by using Monte Carlo interaction verification; and 4, screening variables of the Raman spectrum of the insulating oil for multiple times by adopting a CARS method. According to the Raman spectrum key variables screenedby the method, unnecessary wavelength points in the Raman spectrum can be removed, and the colinearity of Raman spectrum data is removed. Due to limit by the Raman signal detection limit of an aging characteristic substance in the oil, characteristic peaks of all aging characteristic substances are difficult to observe by directly detecting the insulating oil, redundant information is removed by utilizing a variable screening means, key variables are obtained, and the relationship between the key variables and the characteristic peaks is analyzed, and thus, effective support can be provided for extraction and discrimination of spectral aging characteristic quantity.

The invention relates to a method for measuring whey protein in a dairy product on the basis of data-driven Raman spectrum. According to the method, the optimal variable combination of a to-be-measured substance is accurately extracted from complicated and fluctuant translation invariant wavelet coefficients with a variable screening method, the variable combination is subjected to spectral reconstruction with the adoption of translation invariant wavelet transform, the best time domain/frequency domain resolution ratio is obtained while spectral interference such as matrixes and the like areeffectively stripped, and following modeling analysis is facilitated. Besides, according to the method, by analyzing reconstructed spectrum data features and inherent law of unknown dairy product samples, separate modeling is only performed on a single unknown dairy product sample, so that a target ideal model is dynamically approached, and the uncertainty of actual dairy product sample products is effectively overcome. Finally, a data-driven model of alpha-lactalbumin and beta-lactoglobulin in the dairy product sample is finally constructed and converted into corresponding whey protein content, nondestructive testing of the whey protein in the dairy product is further realized, and field detection of related dairy products can be conveniently performed.

The invention discloses a lost member customer group data prediction method and device and a model training method and device, and belongs to the technical field of machine learning and artificial intelligence. The method comprises the following steps: counting and analyzing member loss data to obtain predefined member loss sample data; performing feature construction operation including data preprocessing and variable screening on the member loss sample data to obtain feature variable data; and inputting the feature variable data into a preset machine learning classification model for training to obtain a trained machine learning classification model, with the machine learning classification model being used for predicting a lost member customer group. By combining big data mining, feature data construction and machine learning, intelligent prediction of lost user data is achieved, the prediction precision is high, the learning and generalization capacity is high, and powerful data support is provided for user loss early warning.

The invention discloses a fault prediction method based on a pile gripper hydraulic system and the pile gripper hydraulic system. The fault prediction method comprises the following steps: data acquisition: acquiring real-time data of a pile gripper; variable screening: variables related to the hydraulic system are screened out from the collected real-time data to serve as input, and whether the hydraulic system breaks down or not serves as output; splitting a data set: splitting the data set into a training set and a test set; constructing a neural network: constructing a deep neural network model according to the data set; data set training: importing a training sample into a neural network model for training; data prediction: importing a test sample into the neural network model and outputting a prediction result; and outputting a judgment result. The invention aims to provide the fault prediction method based on the pile gripper hydraulic system and the pile gripper hydraulic system, which can predict possible hydraulic faults.

The invention discloses a long-term rainfall prediction model construction method, a long-term rainfall prediction method and a long-term rainfall prediction device. The construction method comprises the following steps: acquiring a sample set; screening explanatory variables in the sample set based on an error discovery rate control method of multi-hypothesis testing and a random forest model to obtain predictive factors influencing precipitation in corresponding months of the next year; and carrying out random forest modeling according to the prediction factors influencing the precipitation in the corresponding month of the next year and the precipitation in the corresponding month of the next year, and training to obtain a long-term precipitation prediction model of the precipitation in the corresponding month. According to the long-term rainfall prediction model construction method, the long-term rainfall prediction method and the long-term rainfall prediction device provided by the invention, variable screening can be optimized from an experience-dependent method to a data-dependent method, the problem of false positive error rate of a random forest method during empirical variable screening is improved, and the accuracy and reliability of model prediction can be effectively improved.