57 results about "Insufficient Sample" patented technology

Apparatuses and methods for determining the concentration of an analyte in a physiological sample are provided. The subject apparatuses include at least one light source, a detector array, means for determining whether a sufficient amount of sample is present on each of the plurality of different areas, and means for determining the concentration of the analyte based on the reflected light detected from those areas determined to have sufficient sample, where areas having insufficient sample are not used in analyte concentration determination. The subject methods include illuminating each area of a test strip, obtaining reflectance from each of the different areas, determining which areas have sufficient sample based on detected light therefrom and deriving analyte concentration from the areas determined to have sufficient sample, where areas determined not to have sufficient sample are not used in the derivation. Also provided are kits for use in practicing the subject methods.

While deep learning is used for classification, multi-modal information with rich samples and classification contribution variability of each modality are considered, and the problem of insufficient samples is solved by using a semi supervision method. Data of different modalities of a hyperspectral image is sent into a deep neural network, the semi supervision method is used and a large number ofunlabeled samples are utilized, and the deep neural network based on self-encoding is used for feature learning. All labeled and unlabeled data are sent into the self-encoding deep neural network tocarry out learning, similar networks are designed for different modalities, a respective initialization parameter is obtained through self-encoding reconstruction, and hidden attributive classification of labeled samples is obtained through a clustering method. For the unlabeled data, a deep characteristic is calculated through a multi-target deep network, then a similar marked sample is searchedbased on a clustering label, and finally, labels of the unlabeled samples are predicted according to the label information of the labeled samples.

The invention relates to a hot-rolled strip steel surface defect detection method based on a generative adversarial network, which comprises the following specific steps: (1) extracting an industrialfield hot-rolled strip steel surface defect image, and carrying out image preprocessing; and (2) constructing a generator model and a discriminator model of the generative adversarial network GAN, namely adding a condition label vector c into the input of a generator for outputting a classification image; introducing pixel loss Lp into generator training to improve the quality of the generated image; arranging a discriminator branch and a multi-classification branch in the discriminator, so that a multi-classification function is realized, and the classification precision is improved; (3) optimizing the constructed generative adversarial network parameters by using a PSO (Particle Swarm Optimization); and (4) combining the generated image and the real image into a hot rolled strip steel surface defect sample set. According to the method, the problem of insufficient sample data can be solved, the defect image feature extraction speed and accuracy are improved, and a new effective methodis provided for hot-rolled strip steel surface defect detection.

The invention provides a method for human motion recognition base on a generated antagonism network. The method firstly designs a step-by-step generation recognition network model, constructs a classifier on the basis of the antagonistic network, and realizes image generation and classification functions. Secondly, the structure similarity is introduced into the discriminator to improve the quality of the generated image by adding constraints. Finally, the image is generated and recognized in the human motion image database which accords with the daily life. The invention solves the problem oflow recognition rate under the condition of insufficient samples by combining the natural generation and recognition of images. In the aspect of image expansion and recognition, the method has the characteristics of natural sample expansion, high recognition rate and strong robustness.

The invention belongs to the technical field of radar signal processing, and discloses a method for sparse reconstruction of a conformal array clutter covariance matrix. The method comprises the following steps: setting a conformal array as a uniform semicircular array composed of N array elements; determining the position and the array element pattern of each array element, determining the beam steering vector of the conformal array, and getting the array pattern of the conformal array; selecting an observation range gate l of clutter data, getting space-time steering vectors respectively corresponding to Nc clutters on the lth observation range gate and the clutter power corresponding to the lth observation range gate; and getting clutter data of a to-be-detected unit. Based on the prior knowledge of carrier aircraft information and array information and the data of the to-be-detected unit, a clutter plus noise covariance matrix is sparsely reconstructed using the method disclosed by the invention. The method avoids the problems of insufficient samples and sample contamination.

The invention belongs to the technical field of medical image processing, particularly relates to a medical image synthesis method, classification method and device based on an adversarial neural network, and aims to solve the problem that the accuracy cannot meet the requirements due to insufficient training samples in a brain disease classification task. The medical image synthesis method includes the steps: constructing a cyclic generative adversarial model comprising a category loss calculation function; training the cyclic generative adversarial model based on a first feature image set and a second feature image set; and when the sample classification loss satisfies a condition, taking an image generated by the cyclic generative adversarial model as sample data. According to the medical image synthesis method, the category loss is added on the basis of the Cycle-GAN network model, so that the synthesized brain image is more real, and the sample size is increased by two times in anunsupervised mode, and the problem of insufficient sample size in the brain disease classification process by using a deep learning method is solved, and the classification accuracy can be improved.

The invention provides a bullet screen text classification method, a bullet screen text classification device, equipment and a storage medium. The method comprises the steps: obtaining an imbalance training data set with a pre-marked category, and dividing the training data set into a sufficient sample and an insufficient sample; training the sufficient samples by adopting a textCNN model; carrying out model training on the insufficient samples by adopting an SVM classifier; inputting a text to be tested into the trained textCNN model, and outputting classification probabilities of various categories in sufficient samples; and if the output classification probability is smaller than a first preset threshold, inputting the to-be-tested text into a trained SVM classifier, and outputting a predicted category. According to the method, the classification models for different text scales are obtained through separate training according to the sizes of the training samples, then the two classification models are combined to be used for classifying the to-be-detected text, the problem of data imbalance of the training samples is solved, compared with single model training, the risk of over-fitting can be reduced, bias is reduced, and the recognition accuracy is higher.

A method for evaluating the reliability of a power module of a modular level converter, the method includes calculating junction temperature IGBT of IGBT and diode, junction temperature of diode and hot spot temperature of capacitor according to model data of IGBT module, model data of capacitor and MMC task profile, calculating junction temperature IGBT of IGBT and diode in steady state, calculating junction temperature of diode and hot spot temperature of capacitor, calculating junction temperature IGBT of capacitor, calculating junction temperature IGBT of IGBT and hot spot temperature IGBT, calculating junction temperature IGBT. The junction temperature of IGBT and diode is counted by rainflow counting method, and the low frequency thermal cycle list of the whole year is obtained. TheIGBT, diode and capacitor lifetime values are calculated by the preset lifetime model according to the low frequency thermal cycle list and the hot spot temperature of the capacitor. The Weibull lifedistribution is used to evaluate the reliability of the power module in the MMC. The electric thermal stress energy in the MMC is fully taken into account in the method. The method has good engineering applicability for evaluating the reliability of the MMC under different task profiles. At the same time, it overcomes the shortcomings of insufficient samples of life statistics data of DC engineering components.

The invention relates to a CNN-based automatic identification method of lepidopteron types. During preprocessing, background removing is carried out on a collected insect specimen image; a minimum bounding box of a foreground image is calculated; and an effective foreground area is obtained by cutting out. Feature extraction is carried out by using an Imagenet pre-training depth learning neural network model. And classification and identification are carried out based on two kinds of conditions; when samples are sufficient, parameters of a DCNN classification layer are trained based on a fine tuning network structure, thereby realizing end-to-end classification and identification; and when a sample data set is small and insufficient samples are provided for training DCNN parameters, an X<2> kernel SVM classifier for a small sample set is used for realizing classification and identification. The automatic identification method has advantages of simple operation, high identification precision, high fault tolerance, good time performance and the capability of improving lepidopteron type identification obviously.

The invention provides a BNSobol method for sensitivity analysis on the precision of a Helicopter fire control system. The method comprises the steps of performing global sensitivity analysis of the influence of error sources on the precision of the fire control system by combining a Bayesian network and a Sobol index; establishing the Bayesian network according to priori knowledge; learning related parameters in the network through Bayesian estimation to obtain a probability that the precision of the fire control system reaches a specified level under the condition of taking different values for the error sources; and finally processing a probability result by utilizing Sobol method variance decomposition to obtain sensitivity coefficients of the error sources. The invention proposes a new sensitivity analysis mechanism combining the Bayesian network and the Sobol index method, thereby providing reference and theoretical support for performing the sensitivity analysis on the precision of the Helicopter fire control system under the condition of insufficient sample quantity, and providing an uncertain quick quantitative analysis concept for other large-sized complicated systems.

The invention discloses an industrial process fault diagnosis method based on high order correlation and belongs to the field of industrial process monitoring. According to the method provided by theinvention, high order correlation information in measurement data of each sensor is obtained without supervision from an angle of representing learning through utilization of a stack sparse self-coding network, and three monitoring indexes SRE, M2 and C are provided for obtained high order correlation characteristics. In a hierarchical learning mode, according to the method, change of a small fault or an early fault in an industrial process is represented relatively finely, so detection of whether the fault occurs or not is facilitated. According to the provided monitoring indexes, whether process operation is kept in a control domain or not can be monitored and certain guidance for identifying a fault type is provided. In an unsupervised learning mode, influence of insufficient sample labels and imbalanced data in the industrial process is not influenced. The method has important practical significance in solving an industrial process monitoring problem in practice.

The invention discloses an ensemble learning method based on self-adaptive sample expansion. On the one hand, multiple weak classifiers are integrated through adopting a manner of bootstrap feature sampling and dynamic weighted voting, and the advantages of high classification accuracy and good repeatability of the ensemble learning method are inherited; and on the other hand, the method can realize self-adaptive expansion of samples through iteration classification and sample screening based on spatial dispersion and attribute similarity on the basis of a small number of ground survey samples, and the model under-learning problem caused by small/medium-sized samples in classification is solved. The ensemble learning method to which the scheme relates adopts a manner of self-adaptive sample expansion, can effectively solve the problem of insufficient samples in a remote-sensing classification process, and reduces manpower and time spent by researchers for obtaining samples at the sametime.

The invention discloses an Alzheimer's disease detection method based on data space transformation, and the method comprises the following steps: 1, modeling brain functional magnetic resonance imaging (fMRI) data of all data sets, and extracting features from a data model; 2, performing feature selection on the features extracted in the previous step; 3, mapping the selected feature data of all the data sets into the same subspace by using a data space transformation method; 4, training a machine learning classification model by using the data cross validation in the subspace, and adjusting parameters to obtain an optimal computer-aided diagnosis (CAD) model. According to the method, the problem of inconsistent distribution of different data sets is solved, available training samples fordisease diagnosis are increased, and the conditions of low accuracy and insufficient generalization ability of an auxiliary diagnosis model due to insufficient sample size are relieved; meanwhile, theAD auxiliary diagnosis accuracy based on the fMRI data is greatly improved by using feature engineering and a machine learning algorithm.

The invention discloses a multi-stress comprehensive satellite electronic product service life prediction method. The force thermoelectric comprehensive use environment of the satellite electronic product is fully considered; accumulating and competition relationships between fault modes and fault mechanisms; single-stress simulation analysis and multi-stress cumulative damage analysis are carriedout; the life prediction of the satellite electronic product is completed by combining a theoretical model and a simulation analysis result; compared with the prior art, the method has very high engineering practicability, so that the service life prediction work in the past is limited under certain engineering development conditions such as insufficient sample size and test data, the service life prediction of the product can be carried out, and the efficiency of service life prediction of the satellite electronic product can be greatly improved; time and cost for service life evaluation ofsatellite electronic products can be effectively saved, service life prediction efficiency is greatly improved, and economic benefits are high. Important reference is provided for service life and reliability evaluation of satellite electronic products, and the method can be popularized and applied to service life prediction and analysis work of electronic products in other fields.

The invention relates to a liver pathological image sample enhancement method based on random transformation, which comprises the following steps: 1) carrying out block division on a liver pathological image to obtain a plurality of image small blocks; 2) performing random transformation on each image small block to form an extended sample; wherein the random transformation comprises more than one of horizontal mirror image overturning, vertical mirror image overturning, cutting, brightness adjustment, saturation adjustment and hue adjustment; and 3) inputting the extended sample into a deep learning model to train the liver pathological image, and performing corresponding enhancement on the liver pathological image to obtain an enhanced sample of the liver pathological image. According to the method, original pathological samples can be effectively expanded, the problems of insufficient sample number and non-uniform sample distribution are solved to a certain extent, the requirement of a large sample size of a deep learning model is met, the situations of over-fitting and insufficient generalization ability of the trained model can be effectively avoided, and the reliability of an auxiliary analysis result is improved.

The invention provides a double-branch abnormity detection method based on crowd beharivor priori knowledge. The method comprises the steps of extracting interaction information of crowds in a video by utilizing a social force model; learning abnormal scores for different time slices in the video by using a multi-instance learning method; capturing global dependence of the video features by utilizing an attention model; and combining the original video with the crowd interaction information video corresponding to the original video by using a double-branch model. According to the method, priori information of abnormal behavior judgment of human beings is fully considered; a sufficient number of normal and abnormal samples are used for learning normal and abnormal modes of crowd behaviors,so that anomaly detection can recognize the crowd behaviors in a video on a certain semantic level, the problem of performance loss caused by insufficient samples and background interference of crowdsin the video can be well solved and adapted, and the method has higher robustness; the method does not need a data label which is accurate to the fragment level, and even if the training object is the fragment of the video, only the label of the video level is needed.

The invention relates to the field of natural language processing, and particularly relates to a training method and device of a sequence labeling model. The method is used for effectively training asequence labeling model under the condition of insufficient sample data volume. The method comprises the following steps of: training a sequence labeling model based on the sample training statement set; obtaining first loss information, determining an adversarial disturbance factor according to the model parameters; and obtaining second loss information based on the sample training statement setadded with the adversarial disturbance factor, adjusting model parameters of the sequence labeling model based on target loss information obtained through calculation of the first loss information andthe second loss information, performing iterative training, and determining that a convergence condition is satisfied. Thus, by adding the anti-disturbance factor, different loss information can be obtained based on one sample training statement, the generalization ability of the sequence labeling model obtained through training is higher, the precision is higher, introduction of unnecessary noise interference is avoided, and resource consumption is reduced.

The invention provides a nested named entity recognition method and system, electronic equipment and a readable medium. The nested named entity recognition method includes the steps: marking all textsin a corpus based on a preset text marking method to obtain a mark set, wherein the mark set comprises texts and corresponding named entities, and at least one text corresponds to multiple named entities; based on a preset clustering method, clustering the mark set according to each named entity to obtain a cluster set, the cluster set comprising a text and a named entity uniquely corresponding to the text; and based on a preset named entity recognition model with adaptive data enhancement, respectively identifying named entities in each cluster set. A nested named entity recognition problemis converted into a non-nested named entity recognition problem, so that the influence of named entity nesting on the recognition effect is reduced; the data enhancement degree is gradually improved according to the training effect; the data enhancement use intensity is controlled at the optimal level; and the training effect is improved so as to adapt to the nested named entity recognition task under the condition of insufficient samples.

The invention discloses a station caption detection method, a station caption detection device and a readable storage medium. The method comprises the following steps: obtaining a station caption dataset, and grouping the station caption data set to obtain a station caption training set; constructing a multi-loss fusion twin neural network, and training the constructed multi-loss fusion twin neural network based on the station caption training set to obtain a trained multi-loss fusion twin neural network; and detecting a station caption to be detected through the trained multi-loss fused twinneural network. According to the method, a twin neural network framework is constructed, so that the influence of insufficient samples on a training network is well eliminated. The unknown new typesof sensitive station captions can be better detected.

The invention provides a method of analyzing insulating material performance by using three-parameter Weibull distribution to process flashover voltages. The method comprises the steps of collecting n flashover voltages, arranging the voltages in an ascending order, building a voltage vector U and assigning failure ordinals; establishing a three-parameter Weibull distribution flashover voltage probability model; calculating the corresponding accumulated flashover probabilities of the flashover voltages under the failure ordinals; fitting the scale parameter, the shape parameter and the position parameter in the three-parameter Weibull distribution flashover voltage probability model; analyzing insulating material performance by using the three-parameter Weibull distribution flashover voltage probability model. The three-diameter Weibull distribution is employed to process flashover voltage data and calculate the flashover probabilities and thus is closer to the actual condition compared with that based on two parameters, is better in fitting effect and higher in accuracy; the failure level concept is employed for calculating the accumulated flashover probabilities and a medium rank formula is used for correcting the failure level, so that the problem of result deviation caused by insufficient samples is solved; the calculation is simple and the operating speed is high; the method can be used for flashover voltage prediction and the insulating design standard is determined according to the flashover probability requirement.

The invention relates to the technical field of visual relation detection, in particular to a feature coding model, a training method of a visual relation detection model and a detection method. The training method of the feature coding model comprises the steps of obtaining an initial feature coding model; acquiring sample data; inputting each piece of sample data into the initial feature codingmodel; extracting a guide graph from the visual common sense data based on the category; and training the initial feature coding model according to the guide graph, and adjusting the conversion matrixto update the target feature of each target area to obtain the target feature code of each target area. On one hand, the defect of insufficient sample data is overcome by utilizing a guide graph related to the category in the visual common sense. Therefore, enough sample data support can be provided when the target features are coded again, on the other hand, it is guaranteed that relationship perception is introduced when the target features are coded, conditions are provided for subsequent visual relationship detection, and then the accuracy of visual relationship detection can be improved.