48results about How to "Reduce training error" patented technology

The invention provides a posting predication system based on a neural network technique. The system adopts phase-space reconstruction for realizing non-linear time sequence analysis and adopts a Browser / Server structure based on a J2EE platform. The system structurally includes a data pre-processing module, a predication analysis management module and predication modeling and simulation interface software and has characteristics of high convergence rate and low training error. The predication precision of established models is good. However, selection of a range of neural network training samples should be paid attention. Sample quantity can be reduced appropriately for predication of posting concerning to emergencies and samples of too early time may not be applied. In application, damage of models due to abnormal values should be paid attention and the abnormal values should be checked and adjusted, so that the model predication precision can be guaranteed. The established models adopt quantitative analysis and acquire a certain precision. Besides, the perception performance is good.

The invention discloses a grounding grid corrosion rate level prediction method which comprises the following steps: (1) inputting training sample data; (2) randomly sampling training samples according to a bootstrap sampling principle in a Bagging algorithm, forming training sample bootstrap subsets with the number of M, and constituting training sample bootstrap subset data sets; (3) structuring a weak classifier model according to a k-nearest neighbor (KNN) algorithm, sequentially training the training sample bootstrap subsets with the number of M, and obtaining weak classifiers with the number of M; (4) structuring a strong classifier model according to an Adaboost algorithm; (5) inputting to-be-tested sample data, predicting a grounding grid corrosion rate level, obtaining a predicting result, and displaying the predicting result through a displayer. The grounding grid corrosion rate level prediction method has the advantages of being novel and reasonable in design, convenient and fast to use and operate, high in predicting precision, capable of achieving an accurate prediction to the grounding grid corrosion rate level by means of a small amount of data samples which are measured in the prior art, low in implementation cost, strong in practicability and high in value of popularization and application.

The invention discloses a fault type identification method of a transmission line based on a convolutional neural network. A convolutional neural network algorithm belongs to a deep learning algorithm. The deep learning algorithm is applied to the field of fault type identification of transmission lines, manual extraction of fault features is not needed for fault type identification, and a conventional line fault type identification method based on an artificial intelligence algorithm needs extraction of fault features in advance, so the invention simplifies the structure of fault type identification. The invention improves the identification efficiency of the line fault type identification, and in the application of the line fault type identification algorithm based on the deep learning,a plurality of parameters will cause the algorithm to be different in the training process, and the invention intends to optimize the line fault type identification algorithm. The method reduces the error rate of line fault type identification, and different activation functions can make the training error completely different. The method uses different activation functions to train the line faulttype identification, and finds the optimal activation function.

The invention discloses an inter-frame difference and convolutional neural network fusion-based ship video detection method. The method comprises the four parts of preprocessing a video; obtaining anROI of each frame and extracting low layer features; obtaining high layer features of each frame of image by using a modified VGG16 network; and predicting a ship saliency map of the ROI of each frameand extracting a ship target. A relationship between continuous video frames is fully utilized; the interference of a background is reduced; a moving ship is accurately located; a ship moving regionis obtained; and compared with ship image saliency detection only using the low layer features, the method not only can be directly applied to the ship video detection but also reduces the situation of incomplete ship detection, has higher adaptability to a complex inland river moving ship scene, has higher detection precision, solves the problem of inaccurate inland river ship target saliency detection, and has extremely high practical application values.

The invention discloses an integrated framework method for optimizing extremity learning machine by using a genetic algorithm. According to the method, during determining the fitness of each individual in the genetic algorithm, random drawing sample examples from a training set are used as a validation test set, so that the generalization of a trained network can be effectively improved; after the completion of the iteration, an extremity learning machine population with smaller training error is maintained in the genetic algorithm, and then based on the characteristics of the extremity learning machine, excellent individuals with smaller training errors and smaller weight output ranges are selected for integration. The method makes full use of the characteristic of fast training speed of the extremity learning machine, can optimize individuals of the extremity learning machine by using the framework of the genetic algorithm and less iteration times, and formulates a corresponding choice mechanism according to the theory of the extremity learning machine; the extremity learning machine individuals with the smaller training errors and the smaller weight output ranges are selected for network integration, so that in an acceptable training time range, generalization and network stability are remarkably improved.

The invention discloses a super-resolution reconstruction method based on a dense neural network and a two-parameter loss function. The method comprises a data preparation stage, a network structure construction stage, a model training stage, and an image reconstruction stage. The idea of a dense convolutional neural network structure is applied to the super-resolution reconstruction of a single-frame image, shallow features such as texture and contour in the network structure are fully utilized, and local features and global features of the image are combined to perform super-resolution reconstruction. On the above basis, the two-parameter loss function is adopted to perform more accurate optimization training on the network structure, and the algorithm effect is improved.

The invention relates to a method for predicting drilling axial force of a carbon fiber reinforced composite material. The method comprises the following steps: step 1, establishing a drilling three-dimensional finite element model of CFRPs; step 2, modeling and solving the weight index of each input parameter in the drilling model by adopting an extreme random forest regression algorithm based onmachine learning; and step 3, establishing a multilayer correction neural network model considering CFRPs material characteristics and input parameter initial weight indexes, and predicting the maximum drilling axial force. According to the method for predicting the drilling axial force of the carbon fiber reinforced composite material, the method can effectively solve the problems that the material failure standard adopted in a finite element model is greatly different from the material failure of an actual material, the overall prediction process is more time-consuming, the neural network model ignores the material parameters of the composite material, and the established training model is greatly different from the actual test result.

The invention provides a human face recognition method based on multiple-feature space sparse classifiers. The human face recognition method includes the following steps that original training samples, namely X1....XN, are projected onto an Eigenface feature space, a Laplacianface feature space and a Gabor feature space respectively to form a sub-dictionary OE, a sub-dictionary OL and a sub-dictionary OG; the genetic algorithm is used for carrying out joint optimization and training on the three sub-dictionaries to obtain a sub-dictionary NE, a sub-dictionary NL and a sub-dictionary NG; the sub-dictionary NE, the sub-dictionary NL and the sub-dictionary NG are used for training the sparse classifier SRCs. Each sparse classifier carries out sparse representation on a sample to be tested and obtains residual errors corresponding to the ith training sample, wherein the residual errors are RiE, RiL and RiG respectively and then, the mean value of the residual errors corresponding to the ith training sample are calculated. A category corresponding to the minimum value of the residual error mean value E[Ri] is the category which the human face sample to be tested belongs to. According to the human face recognition method based on the multiple-feature space sparse classifiers, the adopted dictionary training method can select a sample with the best separating capacity from each sub-dictionary, so that the human face recognition accuracy based on the sparse classifiers with the dictionaries is improved.

The invention discloses a Chinese lip language recognition method based on deep learning. The method comprises the following steps: S1, acquiring a lip sequence image; S2, extracting features of the lip sequence image from the obtained lip sequence image; S3, inputting the extracted features of the lip sequence image into a bidirectional long-short-term memory network to perform time-space featuresequence learning, training the features of the learned lip sequence image, and training the features of the learned lip sequence image to a lip language recognition model; and S4, decoding and recognizing the extracted features of the lip sequence image according to the trained recognition model from the features of the learned lip sequence image to the lip language, and recognizing a lip language result. The mouth positioning accuracy is high, the calculation method is simplified, and the problem of information decline is better solved by adopting the bidirectional LSTM network.

The invention belongs to the field of artificial intelligence, and relates to risk minimization and weighted least square theory. The invention discloses a regularized rapid neural network leaning method (Regularized Extreme Learning Machine, RELM). The method includes the following steps: (1) revising cost function, weighing empirical risks and structure risks, adjusting the ratio between the weighing empirical risks and the structure risks through parameters, and acquiring the best compromise of the two risks in the end; (2) revising model training, in order to reduce interference of outliers to models acquiring an anti-inference model, and adopting a method that error weighing different samples; (3) outputting weight calculating, first outputting initial value of the weight by using model training without the weight, acquiring each connecting weight and hidden layer threshold value of the neural network model by using the ELM technique, and finally acquiring the outputting value of the weight model.

The invention provides a kernel function combination-based PSO-LSSVM fluctuating wind speed prediction method, which comprises the following steps: using an ARMA model to simulate and generate a vertical space point fluctuating wind speed sample, dividing the fluctuating wind speed sample of each space point into a train set and a test set, and performing normalizing treatment on the train set and the test set; establishing a PSO-LSSVM model based on a combination kernel function; transforming the fluctuating wind speed train sample to a kernel function matrix by using the combination kernel function after PSO optimization, and mapping to a high dimensional feature space; obtaining a nonlinear model of the fluctuating wind speed train sample, and predicting the fluctuating wind speed test sample by using the model; comparing the test sample and a fluctuating wind speed result predicted by using the PSO-LSSVM of the combination kernel function, and calculating an average error and a root-mean-square error of the predicted wind speed and an actual wind speed and related coefficients. The invention ensures the prediction accuracy of the fluctuating wind speed.

The invention requests to protect a remote sensing image fusion method based on adaptive multi-scale residual convolution and a storage medium. The method comprises the steps: S1, carrying out the down-sampling of corresponding multiples on a low-resolution multispectral image and a high-resolution panchromatic image according to a Wald protocol, and constructing a data set needed by a model; s2, adaptively matching the multispectral image and the panchromatic image, and stacking the feature map as input; s3, constructing a multi-scale deep residual convolutional neural network by means of the property of the convolutional neural network; s4, training the network by adopting an Adam optimization algorithm to obtain a trained deep convolutional neural network model; and S5, inputting a multispectral image and a panchromatic image to be fused into the trained deep convolutional neural network model to obtain a fused multispectral image with high spatial resolution. According to the invention, the deep convolutional network is utilized to more fully extract the spatial features of the image, and the fusion effect of the multispectral remote sensing image fusion method is improved.

The invention provides a lane line detection method based on a semi-supervised generative adversarial network. The lane line detection method comprises the steps of: S1, constructing the generative adversarial network, and establishing a training set, a verification set and a test set of the generative adversarial network; S2, pre-training the generative adversarial network through utilizing labeled data in the training set; S3, performing real training on the generative adversarial network by using the labeled data and the unlabeled data in the training set, and adjusting hyper-parameters ofthe generative adversarial network in a real training process through using the verification set; S4, after the real training is finished, evaluating the generalization ability of the generative adversarial network through using the test set, and if the generalization ability reaches a preset standard, entering S5; and S5, inputting an actual street image into a generator network subjected to realtraining to obtain an actual lane line of the actual street image, and superposing the actual lane line on the actual street image to complete lane line detection. According to the lane line detection method, the lane line identification precision can be improved.

The invention discloses a white cell segmentation method based on multi-feature nonlinear combination. The method includes extracting the gray scale of white cells, color and texture gradient information of CIE Lab space and spectrum information of the white cells; combining in a nonlinear manner; processing the combined information in manners of orientated watershed transform and ultra metric boundary mapping OWT-UCM, and acquiring algorithm segmented results; performing corresponded parameter adjustment of the algorithm segmented results and expert segmented results; determining parameters until acquiring a predetermined segmentation result, and finally using the parameters for segmenting other white cells. The method has the advantages that the single feature noise influence is reduced, multiple pieces of feature information are combined in the nonlinear manner, the smaller training errors can be acquired as compared with that of the linear combination method, and the better segmentation result can be acquired; meanwhile, the combination information can be adjusted in an iterative manner, parameters of the network can be adjusted, a target can be closer gradually according to the predetermined result, and the target unavailability and inappropriate adjustment are avoided.

The invention discloses a neural network optimization method, called as a neural network optimization method of a lifted proximal operator machine (LPOM), and relates to the technical field of deep learning. The neural network optimization method disclosed by the invention has the advantages that existing first-order or second-order derivative information does not need to be used for directly optimizing a neural network, instead, a neural network is converted into a new LPOM optimization problem, and then the LPOM optimization problem is solved by virtue of an alternate minimization method. Byadopting the method disclosed by the invention, a layered structure of the neural network is eliminated in a solving process; solving can be carried out by virtue of an alternant iteration process; asolving method can be converted into a random algorithm in some degree, so that smaller calculation amount is maintained during iteration; and for a neural network more than 20 layers, training errorstill can be stably reduced by adopting the method disclosed by the invention.

The invention relates to a semi-supervised learning-based twinborn extreme learning machine classification data processing method, and belongs to the technical field of data mining and processing. According to the method, a semi-supervised learning algorithm for carrying out classification by adoption of two non-parallel classification surfaces on the basis of a random feature mapping mechanism isused for combining technologies of popular regularization, random feature mapping and the two non-parallel classification surfaces, so that defects, on problems of cross data and the like, of singleclassification surfaces are solved, relatively strong robustness is ensured when singular points exist, and the problem that the past algorithms cannot satisfy the generalization ability and the calculation efficiency requirement on few labeled samples at the same time is solved. The method is capable of sufficiently mining information in unlabeled data under the condition that less labeled data exists, is particularly suitable for fault diagnosis in the newly-developing technical field of high-speed rails, draught fans and the like, is high in calculation speed, is basically capable of carrying out real-time judgement, and is high in classification correctness.

The invention discloses a monocular vision-based ternary image calibration method. The method comprises the steps of performing further adjustment on a genetic algorithm, and enabling the adaptive genetic algorithm to have relatively high universality in each time period of population evolution by correcting a hybridization probability Pc and a mutation probability Pm; secondly, on the basis of referring to a radial basis neural network algorithm, improving a radial basis neural network by applying the improved genetic algorithm, thereby further reducing a fault classification problem training error of the radial basis neural network algorithm and realizing better training convergence; and finally, reconstructing an image through a super-resolution image reconstruction mathematic processing method, thereby remarkably improving the definition of the obtained low-resolution image.

The invention discloses a method and a system for optimizing extreme learning machine integrated learning based on a lion group algorithm. The improved lion group algorithm is applied to integrated learning of an extreme learning machine; the characteristics of high precision, fast convergence and fast extreme learning training speed of the lion group algorithm are fully utilized; a global optimalsolution is searched by using a lion group algorithm to follow a currently searched optimal value so as to reduce the number of iterations and optimize an extreme learning machine individual; a corresponding selection mechanism is formulated according to relevant theories of the extreme learning machine, the extreme learning machine with a small output weight and a small training error is selected to be used for integrating the network, the network stability and the generalization capability are remarkably improved within acceptable training time, and the method is a new method with practicalapplication value.

Provided is a monocular visual pattern processing method, comprising further adjusting a GA, and allowing an AGA to obtain better universality in various phases of group evolution through correcting a Pc and a Pm; based on a radial basis function neural network algorithm, employing an improved GA to improve a radial basis function neural network to further reduce fault classification problem training errors of the radial basis function neural network algorithm and obtain better training convergence; and finally reconstructing an image according to a super-resolution image reconstruction mathematical processing method to substantially improve the definition of an obtained low-resolution image.

The invention proposes an SAR image super-resolution method based on combined optimization, and the method is used for solving a technical problem that a conventional SAR image super-resolution method is poor in image restoration effect. The method comprises the following steps: inputting a high-resolution SAR image and a low-resolution SAR image in the same scene; carrying out the segmentation of the high-resolution SAR image and the low-resolution SAR image; cutting a training set into blocks; carrying out the feature extraction of high-low resolution image block sets; carrying out the clustering of a training image set, and obtaining K-class image blocks; carrying out the iterative optimization of the K-class image blocks, and obtaining K mapping matrixes; cutting a test low-resolution SAR image into blocks; carrying out the feature extraction of the test image blocks; selecting a most appropriate mapping matrix for each test image block, carrying out the reconstruction of each test image block, carrying out the clustering of the reconstructed image blocks, and obtaining the high-resolution SAR image. The method is accurate in reconstruction, is low in manual noise, and can be used for providing more accurate information for the subsequent interpretation and target recognition and detection of an SAR image.

The invention provides a joint training method of sub-dictionaries in multiple characteristic spaces and for face recognition. The joint training method of the sub-dictionaries in the multiple characteristic spaces and for face recognition comprises the steps that original training samples{X[1] to X[N]} are projected in the Eigenface characteristic space, the Laplacianface characteristic space and the Gabor characteristic space to form the sub-dictionary OE, the sub-dictionary OL and the sub-dictionary OG respectively, and joint optimization training is carried on the three sub-dictionaries by using a genetic algorithm to obtain the sub-dictionary NE, the sub-dictionary NL and the sub-dictionary NG. According to the joint training method of the sub-dictionaries in the multiple characteristic spaces and for face recognition, a sample, with the highest distinguishing capacity, in each sub-dictionary can be selected, and thus accuracy of face recognition on the basis of a sparse classifier using the sub-dictionaries can be improved.

The invention belongs to the technical field of distributed energy generation control, in particular to a wind power generation power prediction method based on a parallel wavelet neural network, comprising the following steps: 1) acquiring condition parameters of a wind farm and wind power output data; 2) constructing a training data set and a test data set of a prediction sample; 3) constructingand training a parallel neural network model; 4) after completing the training of the parallel neural network model, putting into use. The invention shortens the training time of the neural network,not only retains the advantages of the traditional wavelet neural network in carrying out the time series prediction, but also solves the shortcomings that the parameters of the wavelet neural networkare difficult to be determined, and improves the prediction accuracy.

The embodiment of the invention discloses a neural network training method and device. According to the embodiment of the invention, the method includes: obtaining a native neural network with a to-be-determined connection relationship wherein the native neural network comprises at least two network layers, and each network layer comprises at least one training node; acquiring node connection strategy information of a current training node in a current network layer; determining a target connection node corresponding to the current training node from the training nodes of the target network layer based on the node connection strategy information; establishing a connection relationship between the current training node and the target connection node to obtain a connected neural network; performing forward transmission on the sample data in the connected neural network to obtain a prediction output result; and based on the prediction output result, updating weight information of each training node in the connected neural network so as to determine the trained neural network. According to the scheme, the efficiency of neural network training can be improved, and the over-fitting phenomenon in neural network training can be improved.

The invention discloses a method for identifying fault types of transmission lines based on a convolutional neural network. The convolutional neural network algorithm belongs to a type of deep learning algorithm. By applying the deep learning algorithm to the field of fault type identification of transmission lines, the fault type The identification does not need to manually extract fault features. In the past, line fault type identification based on artificial intelligence algorithm needs to extract fault features in advance. This invention simplifies the structure of fault type identification; improves the identification efficiency of line fault type identification, and the line fault type based on deep learning In the application of the recognition algorithm, numerous parameters will cause the algorithm to vary greatly in the training process, and the present invention intends to optimize it; reduce the error rate of line fault type recognition, and different activation functions will make the training errors completely different. The present invention uses Different activation functions are used to train it to find the optimal activation function.

The invention discloses an end-to-end underwater image restoration method based on ambient light perception, and mainly solves the problem of poor color cast correction and sharpness processing effects during underwater image processing in the prior art. According to the scheme, the method comprises the following steps: respectively constructing an ambient light sensing network and a restoration subject network by using a Pytorch framework, and respectively constructing training sets B and C of the two networks; adopting an adaptive moment estimation algorithm to train an ambient light sensing network and a restoration subject network respectively by using B and C, inputting a to-be-processed image Ic into the trained ambient light sensing network, and outputting an ambient light value Ac; and inputting the Ac and the Ic into the trained restored subject network, and outputting a clear image Jc. The contrast of underwater images with different degradation degrees is improved, color cast can be effectively corrected, the peak signal-to-noise ratio, the structural similarity, the chromatic aberration formula, the non-reference image space quality evaluation and the underwater color image quality evaluation are all superior to those in the prior art, and the method can be used for clearness processing of the underwater images.

The invention provides a photovoltaic power prediction method. The photovoltaic power prediction method comprises the steps of determining historical meteorological data and historical output power ofa photovoltaic power station; preprocessing the historical meteorological data and the historical output power; constructing a neural network model based on the preprocessed historical meteorologicaldata and historical output power; training the neural network model; judging whether network parameters in the neural network model fall into local optimum or not; if the network parameters in the neural network model fall into local optimum, performing heuristic search and adaptive adjustment on the network parameters; determining meteorological data prediction information of the next predictionperiod, and obtaining output power prediction information of the next prediction period through the model obtained through training. After the network parameters in the neural network model fall intolocal optimum, heuristic search and adaptive adjustment are performed on the network parameters. Therefore, the neural network model has the capability of jumping out of the local optimum, and the training error can be continuously reduced.

The embodiment of the invention relates to the technical field of big data, and particularly provides a method, device, equipment and system for determining the number of service sites in an area. Themethod comprises the following steps: obtaining index data corresponding to a specified area, wherein the index data represents factors influencing the business volume of the business site; inputtingthe index data into a region classification model to obtain a region category corresponding to the specified region, wherein the region classification model is obtained by training according to indexdata corresponding to a region and a region category corresponding to the region, the index data corresponding to the region is obtained based on division of a target region, the region category corresponding to the region is determined according to the number of service sites included in the region, and the number of service sites corresponding to different region categories is different. By utilizing the embodiment of the invention, the number of the service sites in the region can be accurately determined, and therefore the site selection accuracy is improved.

The invention discloses an eye fundus image-based cardiovascular and cerebrovascular disease occurrence type and risk prediction method and system, computer equipment and a storage medium. The method comprises the following steps: S1, acquiring a sample data set of cardiovascular and cerebrovascular diseases, wherein the sample data set comprises fundus images of previous patients and occurrence conditions of cardiovascular and cerebrovascular diseases of three-year and five-year follow-up visit; s2, taking the fundus images in the sample data set as input data, taking occurrence conditions of cardiovascular and cerebrovascular diseases of follow-up visit of three years and five years as data labels, inputting the data labels into a deep learning model Inception-ResNet-V2 for feature extraction, training and verification, and constructing a cardiovascular and cerebrovascular occurrence type and risk prediction model based on the fundus images; s3, the eye fundus image of the patient to be detected is input into the constructed prediction model, a prediction result is output, and the prediction result is disease category classification and occurrence risk level classification of cardiovascular and cerebrovascular diseases of the patient in the next three years and five years. The heart and cerebral vessel occurrence type and risk are predicted based on the eye fundus image, and the method is simple, convenient, high in prediction accuracy and good in effect.

The invention discloses a three-dimensional CAD model intelligent classification method based on an improved depth residual network, and the method comprises the steps of representing a model througha plurality of views of a three-dimensional CAD model, carrying out the data enhancement on the views to obtain a multi-view data set, and further obtaining a training data set through the gray scaletransformation and the normalization processing; importing the training data set into an improved deep residual network for training, and storing the trained network; when classification is needed, obtaining and preprocessing one or more views of any size of the new model, further calling the trained network to recognize the input views, and finally obtaining the category with the highest score after the average value of all view recognition results as the category to which the model belongs. According to the present invention, by adopting the improved deep residual network, the intelligent classification of the three-dimensional CAD model is achieved, and the method has the advantages of being flexible and convenient in classification input, high in classification accuracy, good in practicability, high in intelligent degree and the like.

The invention discloses a grounding grid corrosion rate level prediction method which comprises the following steps: (1) inputting training sample data; (2) randomly sampling training samples according to a bootstrap sampling principle in a Bagging algorithm, forming training sample bootstrap subsets with the number of M, and constituting training sample bootstrap subset data sets; (3) structuring a weak classifier model according to a k-nearest neighbor (KNN) algorithm, sequentially training the training sample bootstrap subsets with the number of M, and obtaining weak classifiers with the number of M; (4) structuring a strong classifier model according to an Adaboost algorithm; (5) inputting to-be-tested sample data, predicting a grounding grid corrosion rate level, obtaining a predicting result, and displaying the predicting result through a displayer. The grounding grid corrosion rate level prediction method has the advantages of being novel and reasonable in design, convenient and fast to use and operate, high in predicting precision, capable of achieving an accurate prediction to the grounding grid corrosion rate level by means of a small amount of data samples which are measured in the prior art, low in implementation cost, strong in practicability and high in value of popularization and application.