133 results about "Fitting Problems" patented technology

A method and system for detecting and estimating leaks in an industrial boiler whereby the method and system formulate the leak detection problem as a least squares fitting problem, where one or more of the fitted parameters estimate leak flows. The method and system create a representation that incorporates a leak model component, a process model component and a noise model component into the representation. This invention provides a variety of leak flow estimates and associated significance tests that have a much higher statistical efficiency and are sensitive to leaks with a wider range of leak growth rates than prior approaches.

The invention discloses a wireless positioning method and system based on depth learning and aims to solve a problem of non-high indoor positioning precision. The method comprises steps that S1, wireless signal intensities received by all sampling points are acquired; S2, a depth learning algorithm is utilized to input the acquired wireless signal intensities and coordinates to a depth learning neural network to generate a depth learning neural network with a weight; and S3, wireless signal intensities of users are acquired and are inputted to the depth learning neural network with the weight to determine the positions of the users. The method is advantaged in that indoor positioning precision can be improved, and an over-fitting problem in a full connection layer of the neural network can be avoided.

The invention relates to an inversion method of nuclear magnetic resonance two-dimensional spectrum. The method includes firstly, extracting and estimating noise, namely extracting noise in CPMG (Carr-Purcell-Meiboom-Gill) echo string of acquired data and estimating standard deviation of the noise; secondly, compressing data, namely generating an inversion kernel and performing truncated singular value decomposition and reconstruction by sequence of a nuclear matrix to complete data compression; and thirdly, fitting the data, namely subjecting fitting problem of the compressed data to regularization, performing iterative solution to regularization factors and inversion spectrum by newton method with non-exact one-dimensional search so as to obtain the inversion spectrum. Execution efficiency of two-dimensional inversion algorithm and resolution of two-dimensional spectrum are increased greatly, and the inversion method is well robust.

The invention discloses a dam image crack detection method based on transfer learning. The method comprises the following steps of collecting the dam crack image, preprocessing the image data set to fill the data set through a generated antagonism network GAN; using a pre-training model MobileNet that does not retain a top layer and a full-connection layer to extract image features, splicing a Flatten layer after MobileNet, splicing the full connection layer whose activation function is ReLU after the Flatten layer, and splicing the Full Connection Layer whose activation function is Sigmoid asthe output layer; frozening the first K depth decomposable convolution structures in MobileNet, and fixing the relative weights of these K depth decomposable convolution structures; training the model and only updating the weights of the unfrozen network layer in the process of model training; using the trained model to detect the dam cracks in the image. The method of the invention solves the over-fitting problem under the condition of small data set, and improves the prediction performance and the running speed through the migration learning idea.

The utility model discloses a method and equipment for constructing a classification model based on a convolutional neural network. The method comprises a step of convolution. A first stage of training is carried out on a training sample in a random convolution mode so as to acquire a convolution template value for a convolution operation, thereby constructing the classification model comprising the convolution template value. The first stage of training carried out on the training sample in the random convolution mode further comprises that connection between elements in a characteristic pattern of the current convolution layer and elements in a characteristic pattern of an upper layer which is adjacent to the current convolution layer is interrupted in a random mode based on a predetermined probability threshold value as for at least one current convolution layer. According to the invention, the number of weights used when the sample is trained can be reduced, and an over-fitting problem is relieved, thereby improving a generalization ability of the convolutional neural network.

The invention discloses a meta-learning algorithm based on stepwise gradient correction of a meta-learner, and the algorithm comprises the steps: firstly, obtaining training data with noise marks anda small amount of clean unbiased metadata sets; establishing a meta-learner, namely a teacher network, on the metadata set relative to a classifier, namely a student network established on the training data set; and carrying out united updating of student network parameters and teacher network parameters by using random gradient descent; obtaining a student network parameter gradient update function through a student network gradient descent format; feeding the network parameters back to the teacher network, and updating the teacher network parameters by using metadata to obtain a corrected student network parameter gradient format; and then updating the student network parameters by using the correction format. Accordingly, the student network parameters can achieve better learning in thecorrection direction, and the over-fitting problem of noise marks is weakened. The method has the characteristics of easiness in understanding, realization, interpretability and the like of a user, and can be robustly suitable for an actual data scene containing noise marks.

The invention discloses a handwritten character computer identification method. The method is characterized in that the method is based on a secondary convolution neural network structure model, the model has 9 layers, the model comprises an input layer, 5 hidden layers, a full connection layer and an output layer, convolutional layers and pooling layer are staggered to form the hidden layers, the model has a Dropout layer behind the full connection layer, and the method comprises the specific steps: pre-training a filter, inputting a picture data set for training, activating a function using a Relu, outputting a characteristic vector and the like. The method is advantageous in that a convergence speed can be accelerated through the method, an over-fitting problem is solved through the method, an accumulation error is reduced through the method, and the recognition rate is improved through the method.

A system and method of reconstructing an image through solution of a data fitting problem, wherein the data fitting problem is not susceptible to efficient solution as a whole, is disclosed, which may comprise gathering, via a computing device, k-space image data, selecting a data fitting problem solution algorithm for the k-space data, decomposing the data fitting problem solution into a plurality of sub-problem solutions each susceptible to efficient solution separately for the k-space data, obtaining, via the computing device, the plurality of sub-problem solutions for the k-space data, and reconstructing, via the computing device, the image based upon a weighted average of the plurality of sub-problem solutions for the k-space data. The image may be a magnetic resonance image. The data fitting problem may comprise a minimization of a plurality of linear terms of a least square data fitting solution. The image may comprise under-sampled k-space image data.

The invention discloses an esophageal cancer pathological image automatic discrimination device based on a convolution neural network and a discrimination method thereof. The device comprises an imageacquisition module, an image processing module, a data storage module, a migration learning module, a network training module and a discrimination module. The screening method of the invention comprises the following steps: 1, an image acquisition module collects pathological images and constructs an image database of pathological slices of esophageal cancer; 2, each pathological image database is expanded through an image processing module; 3, the expanded pre-training network pathological image data set is used to complete the migration learning; 4. on the basis of the acquired convolutional neural network structure, the network is trained with the expanded pathological image data set of esophageal cancer and the weights are fine-tuned to get the discriminant network model, and the intelligent discriminant is realized with the discriminant module. The invention overcomes the over-fitting problem in the depth learning process caused by the labeled esophageal cancer pathological imagedata set as a training sample due to the lack of large-scale disclosure, and improves the recognition rate.

The invention provides a network flow prediction method and device based on a cognitive network. The method comprises the following steps of: carrying out least square method processing on an input signal X(t); outputting prediction sample data Y(t); carrying out wavelet transformation on the Y(t); decomposing the Y(t) into components with different frequency compositions; carrying out wavelet transformation on a coefficient sequence {D1(k), D2(k), ...... DL(k), AL(k)} at the k moment; training the network with the component {D1(k), D2(k), ...... DL(k)} as input of an Elman network and a wavelet coefficient {D1(k+T), D2(k+T), ...... DL(k+T)} at the k+T moment as output; training the network with the component of {AL(k)} as input of a linear network and {AL(k+T)} as output; training the network with each trained wavelet component {D1(k+T), D2(k+T), ...... DL(k+T), AL(k+T)} as input of a BP network and the original flow time {f(k+T)} at the k+T moment as the network output; obtaining the prediction output; introducing an LMS (Least Mean Square) algorithm to pre-process the input sample aiming at advantages and disadvantages of the traditional flow model and prediction method; inputting the input sample to a WNN (Wavelet Neural Network) prediction model, therefore, the over-fitting problem in the traditional model is solved, and a more accurate model and prediction are provided for the network flow.

The invention discloses a deep learning human face identification method based on weighting L2 extraction. According to the method, firstly, the human face feature vector is extracted through various-convolution-kernel convolution, then, a weighting L2 extraction method is utilized for carrying out dimensionality reduction on the feature vector, and then, a local average normalizing processing method is adopted for normalizing the feature vector, so a layer of network in the deep learning is formed, the same method is used for building three layers of deep leaning networks, in addition, the three layers of deep learning networks are subjected to cascade connection for forming a layered three-layer deep learning network, and finally, a support vector machine classifier is utilized for carrying out human face training and identification. The deep learning human face identification method has the advantages that the weighting L2 extraction method is provided for realizing the feature dimensionality reduction, the over fitting problem in the training and the single feature problem in the traditional L2 extraction are solved, the feature vector dimensionality reduction is effectively realized, meanwhile, the human face identification performance can be improved, higher grade of features can be effectively extracted, the stability is high, and the identification performance is high.

The invention discloses a handwritten character-oriented one-stage automatic identification and translation method. The method mainly comprises a text identification method and an end-to-end identification and translation method. According to the method, an attention mechanism is used for replacing an RNN structure in the CRNN, so that calculation can be parallelized, and the calculation cost is reduced; in the training process of the Transformer model, random replacement is carried out on input of a decoder, the situation of prediction errors in the prediction process is simulated, and the over-fitting problem is relieved; according to the method, an end-to-end recognition and translation model is provided, an end-to-end model is trained by using a transfer learning-based mode, a recognition result does not need to be given explicitly, and the picture content is directly translated. The method has the following advantages: 1, the training speed of the text recognition model is greatlyimproved; 2, the decoder input is randomly replaced in the training stage, so that the generalization ability of the recognition model is greatly improved; and 3, the translation accuracy of the end-to-end recognition and translation model is higher than that of the two-stage model.

The invention discloses a mean-line-based blade front and back edge fitting and section line smooth reconstruction method. The method is used for solving the technical problem that an existing method is poor in practicability. According to the technical scheme, on the basis of a fitting problem of front and back edge measuring points, a restrained least square fitting method is built for fitting of front and back edges meeting the design requirement. Different from unrestraint least square method front and back edge fitting, the method adopts a mean line as constraint conditions for front and back edge fitting, the restrained least square method is built, a Gauss-Newton method is used for iterative solving, finally, a fitting front and back edge curve approximating the designed front and back edges is obtained, front edge radius errors are lowered to 0.007% from original 0.670%, and the back edge radius errors are lowered to 0.062% from original 1.018%. The fitting front and back edge and the blade back and blade basin achieve smooth connection, and the smooth section line curve is built. The method has the advantages of being high in calculation precision, high in convergence rate and effective in constraint.

The invention relates to a tyre building machine combined roller and a method thereof, aiming to solve the fitting problem caused in the process of rolling a tyre shoulder externally circular conical surface by adopting a movable finned roller and improve the flexibility in the process of automatically adjusting the press roll initial interval in an on-line manner. The combined roller is provided with a rack installed on a base; a first roller device, a second roller device, a third roller device and a fourth roller device are respectively connected on the rack by a respective bracket; the second roller device has two groups of rollers 2 and roller cylinders 2, and two groups of roller ball splines 2 and driving devices for driving the roller 2 to do transversely horizontal displacement which are all arranged at the front end of the roller device 2; the roller 2 comprises at least two rollers which are in fastened connected with each other and are sheathed on an intermediate shaft via bearing sleeves; adjacent rollers are made of different materials and are in gapless section contact; and the intermediate shaft is connected with the robber bracket arranged on the driving device.

The invention relates to a method for automatically evaluating errors of three-dimensional geometrical shapes and belongs to the technical field of precise measurement. Based on the evaluation method, three-dimensional measurement is performed on a certain functional surface of a workpiece with a coordinate measuring machine, three-dimensional coordinate values of measured points are obtained, the coordinate values are analyzed and calculated to directly obtain the error evaluation result, parameters of the shape of the measured surface do not need to be provided, and measuring factors do not need to be constructed manually. The three-dimensional geometrical shapes can be divided into planes, spherical surfaces, cylindrical surfaces, columnar surfaces, revolution surfaces, helicoidal surfaces and composite surfaces according to the inherent translation and rotation invariance and characteristic. The invariance number and the characteristic vectors of a measured object are calculated according to the coordinate values of the measured points, the types of the measured shapes are recognized, an error evaluation mathematical model meeting the minimum region evaluation criteria is further built, a fitting problem is solved to obtain the optimum fitting factors, and then the error evaluation result can be figured out. The initial values of the parameters required by solving the fitting problem can be obtained through calculation of the coordinate values of the measured points and the recognized characteristic vectors. Through the evaluation method, the errors of the shapes of measured workpieces can be reflected comprehensively and really, human intervention is reduced, and the shape error evaluation process is made intelligent and easier to operate.

The invention provides a DDQN-based autonomous guidance maneuver decision-making method for an unmanned aerial vehicle, which is an unmanned aerial vehicle autonomous guidance maneuver decision-makingmethod based on combination of a priority sampling double-depth Q learning algorithm and a Markov decision-making process; the double-Q learning algorithm is introduced to improve the iteration modeof the deep Q learning algorithm, and the training efficiency is improved. A priority sampling method is adopted to promote rapid convergence of the algorithm, and the diversity of historical data isbetter utilized; the unmanned aerial vehicle can realize autonomous guidance maneuvering decision making according to the external flight environment state, and completes autonomous guidance maneuvering decision making under a fixed target point to effectively improve the flight autonomy of the unmanned aerial vehicle. According to the method, the over-fitting problem existing in the DQN algorithmis solved, the offline training efficiency of the autonomous guidance maneuvering decision-making method of the unmanned aerial vehicle is greatly improved, the autonomy of the unmanned aerial vehicle in the flight process is enhanced, and the task execution efficiency of the unmanned aerial vehicle is improved.

The invention relates to a three-dimensional measurement accelerometer error non-singularity estimation method in an external field environment, in particular to a three-dimensional accelerometer error non-singularity estimation method in a field or an external field environment, wherein no special devices are needed. According to the method, first, all posture information needing to be collected for estimating the error of a three-dimensional accelerometer in an external field is determined; second, an accelerometer error model is built, and unknown parameters needing to be estimated are made to be clear; third, the unknown parameter estimation problem of the three-dimensional accelerometer error model is converted into a fitting problem of geometric model parameters of an ellipsoid; fourth, in parameter estimation of the three-dimensional accelerometer error model, an ellipsoid fitting algorithm avoiding singular points is adopted; fifth, an ellipsoid correcting model is adopted in parameter estimation of the three-dimensional accelerometer error model, and the problem that scale factors and installation error matrixes are not unique is solved. The three-dimensional measurement accelerometer error non-singularity estimation method is applicable to systems such as a small unmanned aerial vehicle, a robot and a handheld terminal with a low-cost, small-size and low-parameter-stability accelerometer as a sensor.

The invention relates to an intelligent photovoltaic array fault diagnosis method based on an optimal rotating forest. Firstly, the data of photovoltaic electrical characteristics under various operating conditions are collected and mapped to obtain the overall fault characteristics. Secondly, the ReliefF feature selection algorithm is used to rank the importance weights of the fault features, andthe most important fault features are obtained. Then, the input variables of the base classifier are obtained by using the improved rotating forest algorithm. Furthermore, the limit learning machinereplaces the decision tree in the original rotating forest algorithm to overcome the over-fitting problem and obtain the optimal model parameters by traversal method. Furthermore, the optimal trainingmodel of rotating forest fault diagnosis is obtained by combining the improved rotating forest algorithm with the limit learning algorithm to train each sample in the training set. Finally, the training model is used to detect and classify the fault of PV arrays. The method of the invention has high classification accuracy and is an effective alternative scheme for photovoltaic fault diagnosis.

The invention discloses a gesture recognition method based on a convolutional neural network and an anti-convolution neural network. For certain specific types of gestures, a training sample is generated by using the anti-convolution neural network, so that an over-fitting problem when a model is trained can be solved. In addition, the convolutional neural network is used for gesture recognition,so that the recognition accuracy and the efficiency can be effectively improved, and the real-time gesture recognition can be achieved. According to the gesture recognition method, the convolutional neural network and the anti-convolutional neural network are combined and applied to gesture recognition, compared with a traditional recognition method, the neural network has the advantages that convolution can be directly carried out on the neural network and image pixels, image features are extracted from the image pixels, and the processing mode is closer to the processing mode of a human brain vision system.

The invention belongs to the technical field of image rain removal, and provides a single-image-oriented rain removal method based on a cascaded hole convolutional neural network, which is used for solving the problem of restoration of a single image shot in rainy days. The method comprises the following steps: firstly, modeling rainwater, and dividing a rain image into a rainwater region layer, arainwater layer and a background layer; extracting a rainwater region layer image from an input image through cascaded multi-channel convolutional neural networks with different void ratios, obtaining a rainwater layer image through convolution, and obtaining a rain-removed background layer image through convolution and summation of the input image. Details of different scales of the image are effectively extracted through the cascaded hole convolutional neural network, the network adopts a residual network structure to increase the network depth, and the over-fitting problem is avoided; an evaluation experiment is carried out on a public data set, and the experiment shows that compared with a single-image rain removal classic method, the peak signal-to-noise ratio (PSNR) can be improvedby 2-8, and the image structural similarity (SSIM) can be improved by 0.04-0.22.

The invention, which belongs to the technical field of medical image processing, discloses a mouse CT image kidney segmentation method based on a random forest and a statistic model. The method comprises: on the basis of a training sample, establishing a high-contrast-ratio organ mean value model and a low-contrast-ratio organ mean value model respectively; estimating the location of a kidney in a target image; extracting features of the training sample and the target image; and training a random forest and completing target segmentation. The feature expression for the CT image is constructed and thus the random forest can segment the CT image precisely, so tat problems of large data volume, complicated random forest calculation, and too low speed for the CT sequence image are solved. Meanwhile, an over-fitting problem caused by the statistic model is solved and a model can be established by using a few of samples; and kidney segmentation in the CT image is realized. The method having advantages of high precision, fast speed, and capability of being free of manual intervention has the great reference application value in fields like medical image segmentation.

The invention discloses an oil-well pump running condition recognition method and an oil-well pump running condition recognition device. The method comprises the following steps: inputting target characteristic vector of a target grayscale image to RWELM, carrying out RWELM operation to output a running condition, and training the RWELM by taking the RWELM as a model for oil-well pump running condition recognition and using an indicator diagram and the running condition as training data, wherein structural risk minimization theory is introduced to RWELM in the training process, namely a hidden layer is regulated by virtue of regulating parameters; therefore, an over-fitting problem of a conventional extreme learning machine is solved; and by replacing general hidden layer excitation function through a wavelet function, a problem of local optimum of the extreme learning machine is solved, and the method has the advantages of high diagnosis speed and high accuracy. The device disclosed by the invention, which is embedded in oil-well pump running condition recognition equipment or system, can discover the running condition in real time, thus providing a basis for running condition maintenance of an oil-well pump.

The invention provides a rock bust danger level prediction method based on a local weighting C4.5 algorithm and relates to the technical field of rock burst prediction. The method includes the steps of firstly, adopting an MDLP method for conducting discretization on continuous attribute data in sample data, then adopting a local weighting method for selecting a training set and calculating the weight of samples, utilizing the weight of the samples to calculate an information gain ratio of each attribute, and selecting sample attributes as root nodes of a C4.5 decision tree and splitting attributes of other branch nodes according to the information gain ratios; finally, adopting the weight of the samples to substitute the sample number to conduct pessimistic pruning on the created decisiontree, and correspondingly achieving prediction of rock burst dangers and the like in a predicted area. According to the provided rock bust danger level prediction method based on the local weightingC4.5 algorithm, the defect is overcome that the preference selection values have too many attributes when information gain is adopted for selecting node splitting attributes in an ID3 algorithm; an over-fitting problem is avoided, and the prediction accuracy of a model is high.