81results about How to "Solve overfitting" patented technology

The invention discloses a graphic pattern text detection method based on deep learning. The method includes firstly, training a depth convolution self-encoding network by combining graphic pattern text samples, than adopting marked samples, and classifying through a sparse dictionary; extracting graphic pattern texts from a sample library, rotating, shifting and transmitting, and combining the graphic pattern texts with pure background graphics; adopting a combined sample seat, establishing a depth convolution self-encoding network, and learning characteristic templates in layered training and entire optimizing manners; performing characteristic extraction on the characteristic templates acquired by deep network learning according to the acquired marked samples; sampling the extracted characteristics in the size of the original graphics, adopting single blocks as identifying units, and training the sparse dictionary and a classifier; after training, performing multi resolution decomposition on the graphics to be processed, utilizing the characteristic templates to extract characteristics, and utilizing the sparse dictionary to acquire results in a classified manner.

The invention relates to a soft sensing method for load parameters of a ball mill. The method is that a hardware supporting platform is used to obtain vibration signals, vibration sound signals and current signals of a ball mill cylinder to soft sense ball mill internal parameters (ratio of material to ball, pulp density and filling ratio) characterizing ball mill load. The method comprises the following steps that: the vibration, the vibration sound, the current data and the time-domain filtering of the ball mill cylinder are acquired, time frequency conversion is conducted to the vibration and the vibration sound data, kernel principal component analysis based nonlinear features of the sub band of the vibration and the vibration sound data in frequency domain are extracted, nonlinear features of the time domain current data are extracted, feature selection is conducted to the fused nonlinear feature data and a soft sensing model based on a least squares support vector machine is established. The soft sensing method of the invention has the advantages that the sensitivity is high, the sensed results are accurate, the practical value and the popularization prospect are very good, and the realization of the stability control, the optimization control, the energy saving and the consumption reduction of the grinding production process is facilitated.

The invention provides a recommendation model training method and device and a recommendation method and device. The training method of the recommendation model comprises the steps of determining first initial parameter values of training parameters in a to-be-trained recommendation model, wherein the first initial parameter values are target parameter values obtained after initial parameter values are iteratively updated by a click rate estimation model trained in advance; acquiring user characteristics of at least two sample users and attribute characteristics of at least two sample application programs; generating a positive sample of that the sample users purchase the exposed sample application programs and a negative sample of that the sample users do not purchase the exposed sample application programs based on the user characteristics and the attribute characteristics; and on the basis of a sample set comprising at least one positive sample and at least one negative sample and the first initial parameter value, training a to-be-trained recommendation model to obtain the recommendation model, and outputting the exposure conversion rate of each sample user on each sample application program by the recommendation model.

The invention provides an object-oriented high-resolution remote-sensing image classification method. The method comprises the steps of S1, conducting segmentation processing on images to be processed to obtained a plurality of subimage objects; S2, obtaining feature information of subimage objects; and S3, classifying subimage objects according to the obtained feature information, wherein images to be processed are high-resolution remote-sensing images, the feature information of subimage objects comprises spectral information, shape information and texture information of subimage objects. According to the method, on the basis of object-oriented classification, a classification method combining probabilistic latent semantic analysis and a support vector machine is introduced, the problem that 'the same features with different classifications' and 'the same classifications with different features' are not high in identification ratio in the prior art is solved, the classification precision of high-resolution remote-sensing images is greatly improved, advantages of latent semantic analysis (LSA) and advantages of probabilistic latent semantic analysis (PLSA) are combined, and the problems of overfitting and local optimum which are caused by random initialization are effectively solved.

The invention discloses a migration convolution neural network method for face expression recognition, which comprises the following steps: S1, acquiring a face expression image data set, dividing thedata set into a training set, a verification set and a test set; 2, cascade that migration network and the convolution neural network to construct a migration convolution neural network model; The input of MCN model is facial expression image data, and the output is facial expression category. 3, train that migration convolution neural network model by using a training set, and optimize the trained migration convolution neural network model by using a verification set; S4, the facial expression recognition accuracy of the test set is tested by the optimized migration convolution neural network model. The invention utilizes the migration convolution neural network to recognize the facial expression, and solves the problems that the small data set cannot be converged and over-fitted on a large number of networks.

The invention provides an automatic blog writer interest and character identifying method based on a support vector machine. The automatic blog writer interest and character identifying method includes building an interest classified training sample set and a character classified training sample set at first; respectively processing the two training sample sets by a Chinese morphology analyzer to obtain a candidate interest feature item set and a candidate character feature item set; then analyzing the two candidate feature item sets by the aid of a statistics method; building an interest classified feature item set and a character classified feature item set; displaying the interest classified training sample set and the character classified training sample set into vector forms by the two feature item sets; and finally respectively using two groups of training interest classifiers and character classifiers. The classifiers are used for identifying interests and characters of other writers. By the aid of the automatic blog writer interest and character identifying method, the interests and the characters of the writers can be accurately identified, the method is applied to various personal services based on information of the writers, service providers can sufficiently know users, service quality is improved, and the method has an extremely high practical value.

The invention provides a bladder tumor detection method based on MaskRcnn, At least that method comprise: a few pictures of medical bladder are collected as training samples, An image processing method is used to solve the problem of over-fitting caused by inadequate medical bladder training pictures which generate unbalanced sample data. The picture has a specific rectangular recognition frame which contains bladder position marker or tumor position marker or has both bladder and tumor position marker. A medical picture is characterized by utilizing a brand-new paramete ResNet debugged according to that invention, and a predicted ROI region is generate; The predicted ROI region represents the possible location of the bladder and tumor; The training speed and testing speed are effectivelyimproved by weighting ResNet operation. The coordinate translation and scale scaling of the region of interest are carried out by using the edge of the border. A target rectangular frame is obtained by using a non-maximum resistance. The invention uses artificial intelligence method, namely deep learning, which greatly reduces the repetitive manual operation of the traditional method and improvesthe speed and efficiency. An instanteous function is achieved.

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 belongs to the field of knowledge maps, particularly relates to a knowledge map long-tail relation completion method, system and equipment based on an attention mechanism, and aims at solving the problem that a traditional relation completion model generates overfitting on long-tail relation prediction due to the fact that the number of long-tail relations is small. The method of thesystem comprises the steps of obtaining a knowledge graph to be completed, constructing the knowledge graph to be completed into a first knowledge graph and a second knowledge graph according to a relationship type between entities of the knowledge graph, obtaining entity vector representation fusing neighborhood information in the first knowledge graph to serve as first representation, performing vectorization representation on each entity in the second knowledge graph according to the first representation, constructing a support set and a query set, obtaining the relationship type label ofeach entity pair in the query set through a preset long-tail relationship prediction method of multiple network types, and complementing the relationship between the entities. According to the method,the neighborhood information of each entity in the knowledge graph is fused, so that the problem of overfitting during long-tail relation prediction is avoided.

The invention discloses a training method of an image classification network, an image classification method and device and a server. The training method comprises the following steps: preparing a data set of pictures with labels as input in advance; Constructing corresponding hierarchical neural network structures according to different classification levels; And carrying out hierarchical training on each hierarchical neural network structure to obtain a parent class corresponding to the maximum probability value and probability values of the input pictures under the parent class belonging todifferent subclasses. According to the method and the device, the technical problem of overfitting caused by redundancy of a full connection layer due to very large data set classification data is solved. Through the training method provided by the invention, the phenomena of low network training speed and over-fitting of the network caused by too many full connection layer parameters are solved.According to the image classification method provided by the invention, hierarchical training is adopted, so that a subclass classification result can be more accurately obtained on a classificationresult of a parent class, and accurate classification is realized.

The invention discloses an encrypted traffic identification method oriented to category imbalance. The method comprises the following steps: S1, acquiring a data set; S2, balancing the data set: processing the original experimental data set by adopting an improved SMOTE algorithm based on density estimation; S3, data preprocessing: reading a data stream, cutting off data, and carrying out normalization processing; S4, optimizing the feature set: automatically extracting features through a variational automatic encoder model, identifying features commonly used in the field through network traffic, and obtaining the optimal feature set by using a feature selection method based on a tree model; S5, identifying flow: inputting the optimal feature set into a random forest CGA-RF classifier algorithm improved based on a genetic algorithm, and identifying target encrypted flow; and S6, analyzing the obtained index result, and optimizing an encrypted traffic identification method. The method is high in recognition rate and low in false alarm rate, and is suitable for encrypted traffic recognition of data set category imbalance and difficult feature extraction.

The invention discloses an encrypted traffic identification method based on ensemble learning. The encrypted traffic identification method is characterized by comprising the following steps of: (1) collecting a data set; (2) preprocessing the data; (3) balancing the data set; (4) automatically extracting features; (5) identifying the flow; (6) and analyzing an obtained index result, selecting proper parameters, and optimizing a algorithm. According to the encrypted traffic identification method, the problem of model under-fitting or over-fitting caused by imbalance of sample categories is solved, the recognition rate is high, the false alarm rate is low, and the encrypted traffic identification method is suitable for encrypted traffic recognition of category imbalance and difficult featureextraction of the data set.

The invention provides a reinforcement learning training method and a decision-making method based on reinforcement learning, and the method comprises the following steps: obtaining a plurality of groups of historical state data; inputting each group of historical state data into a reinforcement learning model to obtain preliminary decision data; inputting each group of historical state data and the preliminary decision data into a pre-established Bayesian neural network model to obtain a state variable quantity and a reward value, the state variable quantity being a difference value between current state data and next state data; and updating model parameters of the reinforcement learning model according to each group of historical state data, the corresponding preliminary decision data,the state variation and the reward value. By implementing the method, the training sample size of the reinforcement learning model can be increased, the reinforcement learning effect is improved, andthe accuracy of a dynamic decision planning result is improved.

The invention discloses a method of dynamically predicting a middle and short term area pediatric outpatient amount based on a time sequence. A middle and short term pediatric outpatient amount in an area is dynamically predicted accurately so that a scientific basis is provided for solving problems of area health resource programming, optimization configuration and the like. Simultaneously, through introducing different disease outpatient amount distribution, abnormal states, such as sudden infectious diseases and the like, can be determined so that spreading of the infectious diseases in groups can be rapidly prejudged. The method is realized based on a time sequence prediction model. The method comprises the following steps of step1, sequence smoothness inspection and processing; step2, event variable identification; step3, parameter estimation and inspection; step4, optimal sequence length identification; and step5, dynamic prediction. The above steps are mainly used for selecting an optimal sequence prediction model, a time variation influence is eliminated, problems of overfitting of time sequence prediction and a poor extrapolation effect are effectively overcome and accurate prediction is performed.

The invention discloses a deep reinforcement learning algorithm based on time sequence difference, which is used for solving a hybrid flow shop scheduling problem of a related parallel machine, combines a convolutional neural network with TD learning in reinforcement learning, and performs behavior selection according to input state features; and therefore, the scheduling decision process of the actual order response type production and manufacturing system is better met. A scheduling problem is converted into a multi-stage decision problem, a convolutional neural network model is used to fit a state value function, manufacturing system processing state characteristic data is input into a model, a time sequence difference method is used to train the model, a heuristic algorithm or an allocation rule is used as a scheduling decision candidate behavior, and a reinforcement learning reward and punishment mechanism is combined, thus selecting an optimal combined behavior strategy for each scheduling decision. Compared with the prior art, the algorithm provided by the invention has the advantages of strong real-time performance, high flexibility and the like.

The present invention provides a QRS wave group classification method based on a support vector machine. The method comprises the following steps: a) selecting an appropriate QRS wave group, and performing preprocessing of the data; using a support vector machine to perform training a classification model between each two classes of data; reading QRS wave group data required predicting the type of the waveform; c) performing preprocessing of the data, wherein the preprocessing comprises normalization processing and data dimension algorithm processing; d) performing prediction of the type of the waveform through adoption of the classification model; and e) using a voting method to realize identification and classification of a plurality of forms of QRS wave groups, and obtaining a prediction result. Through random selection of a training set, the QRS wave group classification method based on the support vector machine performs repeat training and test and finally obtain an optimal classification model so as to realize identification of different forms of the QRS wave groups.

The invention discloses a vehicle re-identification method and system based on a double attention mechanism. The method comprises the following steps: constructing a convolutional neural network for vehicle feature extraction; constructing channel attention parts for paying attention to different channels; constructing a granularity attention component used for paying attention to the feature suitable granularity; randomly selecting multiple types of vehicles in each training batch, and randomly selecting multiple images in each type of vehicles to construct batch images; carrying out real-time data enhancement on the batch images and then inputting the batch images into a convolutional neural network; constructing a cross entropy loss function and a triple loss function after smooth regularization of the batch labels, and adding the cross entropy loss function and the triple loss function to obtain an integral loss function; and carrying out feature extraction on the trained convolutional neural network, calculating Euclidean distances between the features, and reordering the distances to obtain a vehicle re-identification result. According to the invention, the fine-grained features of the vehicle image can be better obtained, and the precision and stability of the model are improved.

The invention provides an image training set generation method and a model training method based on image data enhancement, and the method comprises the steps of constructing an initial image training set, and then training an initial target detection model; re-detecting the original training sample image and then outputting a detection output image; comparing a detection frame formed on each detection output image with a labeling frame on the original training sample image, and cutting error target image blocks selected by the detection frames; cutting a rejected image block from the selected original training sample image, selecting a wrong target image block, and compensating the wrong target image block at the original position of the rejected image block to form a negative sample image; and mixing into an initial image training set to form an image training set. According to the invention, the problem that a pattern training set expanded by using a data enhancement method in the prior art can improve the generalization ability of a trained target detection model but cannot improve the recognition accuracy of the target detection model to a specific target object is solved.

The embodiment of the invention discloses a training method, device and equipment of an image generator and a storage medium, and relates to the technical field of artificial intelligence such as deeplearning, big data processing, data enhancement and computer vision. The method comprises the following steps: respectively inputting codes of a first modal image and a second modal image of the sameobject into an initial model of an image generator to obtain a first image and a second image; respectively inputting the first image and the second image into a target network model to obtain a feature value of the first image and a feature value of the second image; and determining a loss function based on the feature value of the first image, the feature value of the second image and the typeof the target network model, and adjusting the parameters of the image generator to obtain a trained image generator, thereby providing an image generator capable of generating a specific modal imagein the first modal image and / or the second modal image. The problem of overfitting caused by too few training samples of a specific modal image recognition model is solved, and the recognition precision of the model is improved.

The invention discloses a hyperspectral image classification method based on a cross-grouping spatial-spectral feature enhancement network. The hyperspectral image classification method comprises thefollowing steps: 1) cross-grouping spectral features; 2) extracting multichannel grouping spectral characteristics; 3) performing spatial feature cross grouping; 4) extracting grouping space features;5) performing spectral space channel information interaction; and 6) performing hyperspectral image pixel classification. According to the method, spectral spatial feature information is utilized toperform cross grouping and feature extraction operation on spectral features and spatial features respectively, so that the correlation between adjacent spectra can be effectively weakened; channel self-attention and pixel position self-attention operations are adopted to enhance features obtained by cross grouping, information interaction and fusion are carried out on spatial features and spectral features, the fused features are used for classification, and the network classification performance can be improved.

The invention discloses an image segmentation method based on a weight variation expectation maximization criterion. The method comprises: first of all, extracting characteristic information of an image to be segmented, then describing the distribution of the image characteristic information by use of an expansion Gauss mixture model, based on the weight variation expectation maximization criterion, estimating the variation distribution of parameters of the expansion Gauss mixture model, after estimation is finished, obtaining a probability generated by each class to be divided from each pixel point, finally carrying out determination, and taking a sequence number corresponding to the maximum value in the probability values about each class of each pixel point as a class finally distributed to the pixel point so as to finish an image segmentation process. According to the invention, the segmentation quality and effect of a color image can be effectively improved, and the segmented image has quite good smoothness. By using the method, the over-fitting and under-fitting problems easily occurring in a conventional segmentation method based on a maximum likelihood criterion can be avoided.

The invention discloses an image colorization method and system and a computer readable storage medium. The method comprises the following steps: A, converting an image to be colorized from an RGB color space to a YUV color space, and separating Y channel data; B, copying the Y channel data, and constructing two-channel data together with the Y channel data; C, using the data of the two channels as the input of a depth convolution auto-encoder to predict UV channels respectively, wherein the depth convolution auto-encoder is formed by connecting a plurality of rejump layers; D, combining the Ychannel data with the UV channel data predicted in the step C to construct a complete YUV color space image; and E, converting the YUV color space image into an RGB color space image to obtain a final colorized image. According to the method, the problems of model gradient disappearance and over-fitting can be better solved, a better coloring effect and better image definition are achieved, artifacts generated by the image can be effectively reduced, and the color saturation can be effectively enhanced.

The invention discloses a character-based hierarchical text sentiment analysis method and system, and the method comprises the steps: carrying out the preprocessing of given text data, wherein the preprocessing comprises designing a character set, dividing sentences in a text, and obtaining a text representation in a digital form based on the character set; establishing a character-level neural network model: inputting the preprocessed text data into the character-level neural network model, sequentially passing through a model embedding layer, a convolutional neural network layer and a decoding layer, and extracting and outputting a feature vector of each sentence in the text; and establishing a sentence-level neural network model: taking output of the character-level network as input, and outputting probability distribution of sentiment classification of the text through a recurrent neural network layer, an attention layer and a decoding layer in sequence. The initial features of thetext are extracted from the character level, the sentence level network contains the time sequence information, the network can tend to sentences beneficial to the sentiment analysis result, and theaccuracy and robustness of the model are improved.

The invention discloses a dynamic safety assessment method based on an extreme gradient boosting decision tree. The method comprises the steps: 1, establishing an original sample set; 2, forming an intermediate sample set; 3, forming an efficient sample set; 4, performing offline training on the extreme gradient boosting decision tree by utilizing the efficient sample set, and constructing a dynamic safety evaluation model based on the XGBoost decision tree; 5, due to the influence of various operation factors of the power system, updating the efficient sample set through time domain simulation, and updating aDSA model by using the updated sample set; 6, based on the real-time measurement data of the synchronous phasor measurement unit, performing online DSA on the power system by using the DSA model. The invention aims to provide a dynamic safety assessment method based on an extreme gradient boosting decision tree. According to the method, online DSA can be efficiently and accuratelycarried out on the power system, the robustness of the evaluation model is good, and the method can adapt to the change of a network topology structure.

The invention discloses a data center energy efficiency optimization method based on transfer learning in the field of data mining and machine learning. The method comprises a model training method and a model reasoning decision-making method. A basic model is pre-trained based on Base Model by using data samples of all units; according to the method, hidden layer parameters learned by a basic model are migrated to List-wiseModel of each unit, and fine adjustment is carried out by using a relatively low learning rate, so that the problem of missing of part of unit samples is solved, and the generalization ability of the model is improved; by designing a multi-task learning model and adding rank constraint to multi-target loss, the problems of overfitting and model high variance caused by noise are solved, and a sample pair is selected in a random and periodic sliding window mode to improve the convergence rate of the model; the optimal unit control parameters are obtained by using theoptimal linear search of the energy consumption prediction task, the control parameters are sorted by using the rank prediction task sorting model, and the optimal control parameters are comprehensively selected, so that the accuracy and robustness of the optimal control parameters are improved.

The invention discloses an error remote estimation method and system for an intelligent electric meter, a terminal and a storage medium. The method comprises the following steps: acquiring transformer area and user file data and electric quantity data of the intelligent electric meter; preprocessing the electric quantity data of the intelligent electric meter, and taking the preprocessed electric quantity data as electric quantity sample data; according to the electric quantity sample data, the transformer area and user archive data and a transformer area electric energy loss calculation method, constructing a standardized perception matrix and a measurement vector group; and calculating the metering error of each intelligent electric meter in the transformer area by adopting a self-adaptive parameter regularization orthogonal matching pursuit algorithm for the standardized sensing matrix and the measurement vector group. The calculation efficiency in the electric power metering process can be improved, and a guarantee is provided for monitoring the error of the electric energy meter in a running state in real time.

The invention relates to a training method of a vehicle logo classification model, a vehicle logo recognition method, device and apparatus. The vehicle logo classification model training method comprises the steps of obtaining a plurality of training sample images, and fusing any two training sample images in the plurality of training sample images according to a preset fusion proportion to obtaina fused sample image; inputting the fused sample image into a vehicle logo classification model for classification processing to obtain a vehicle logo classification result; respectively calculatinga vehicle logo classification result and a first classification loss and a second classification loss of category labels of any two training sample images, and fusing the first classification loss andthe second classification loss according to a fusion proportion to obtain a fusion loss; and adjusting model parameters in the vehicle logo classification model according to the fusion loss, and circularly training until the vehicle logo classification model is converged. Discrete samples can be continuous, the smoothness in neighborhoods is improved, and the problem of overfitting is solved; andmeanwhile, the model training efficiency is improved.

The invention discloses a log data enhancement method and system and a log data classification detection method and system. The method comprises the steps of proprocessing collected log data; merging the preprocessed log data; constructing a generative adversarial network model, and training the generative adversarial network model by utilizing the log data subjected to the data merging processing; generating a log data sample according to the trained generative adversarial network model; and performing data combination based on the log data sample and the log data subjected to the data merging processing to form an enhanced log data set. The problem of unbalanced log data samples in the field of network space security can be effectively solved, the problems of similarity and overfitting caused by lack of representative samples and the problem of classification detection of unknown network threats are solved, and the purposes of log data enhancement and expansion and classification detection are achieved.

The invention relates to an online public opinion early warning method based on an LSTM model. The method is mainly characterized in that the method is based on an online public opinion early warningindex system and a long-short-term memory network model; the invention provides an index system comprising five dimensions of theme attributes, propagation and diffusion and the like. According to themethod, a Re-LSTM (Regulation Long Short Term Memory, long and short term memory network) model is provided on the basis, thus constraining the input weight of each unit in the network by using a regularization method, and replacing a tanh activation function with a softsign function. The method is reasonable in design, and compared with other classic models, the model constructed this time can improve the early warning accuracy and can better avoid the problems of gradient disappearance and overfitting.

The invention discloses a transformer multi-source partial discharge mode identification method based on a parallel feature domain. The method comprises the steps: extracting multi-source partial discharge signals to construct a multi-source partial discharge time domain signal set and a time-frequency domain signal set; combining a plurality of single neural network automatic encoders into two stacked encoders; selecting a sigmoid function as an activation function between the stacked encoder network layers, and obtaining an activation value of the next network layer by using the activation function; adding a regularization item to a loss function for adjusting each layer of network parameters in the stacked encoder; setting an optimization solution method of the loss function as a near-end guide random sub-gradient algorithm; adding softmax as a classification layer of the neural networks; and carrying out parallel training on the multi-source partial discharge time domain signal setand the time-frequency domain signal set by using the stacked encoder, wherein feature matrix data has corresponding labels, and a classification result is compared to finely adjust network parameters. The method has the advantages of high classification precision, strong generalization ability of a deep learning model and the like, and is suitable for occasions such as multi-source partial discharge mode identification of a transformer.