126results about How to "Fully learn" patented technology

The invention relates to a deep learning-based question and answer matching method. The method comprises the following steps of: 1) sufficiently learning word orders and sentence local features of a question text and an answer text by utilizing two underlying deep neural networks: a long short-term memory network LSTM and a convolutional neural network CNN; and 2) selecting a keyword with best semantic matching on the basis of a pooling manner of an attention mechanism AM. Compared with existing methods, the method has the advantages of being in low in feature engineering workload, strong in cross-field performance and relatively high in correctness, and can be effectively applied to the fields of commercial intelligent customer service robots, automatic driving, internet medical treatment, online forum and community question answering.

The invention provides an intelligent device failure diagnosis method based on a support vector machine. The method includes the steps that preprocessing operation is conducted on device data; a failure diagnosis case knowledge base is built; failure diagnosis is conducted on the support vector machine; the failure information is obtained, and troubleshooting guide is conducted. By means of the intelligent device failure diagnosis method based on the support vector machine, the failure feature of a device is highlighted to the maximum degree, the situations that the device data are incomplete and imprecise are reduced, the method provides the possibility for building a precise and reliable failure diagnosis model, the problem of aging of the diagnosis model along with the runtime of the device is solved, the misdiagnosis rate of the failure diagnosis model is reduced, and the correct rate and speed of the device failure diagnosis are increased to the maximum degree.

The invention discloses an SVM classification model-based equipment fault diagnosing method. The method comprises the following steps: performing a preprocessing operation on equipment data; constructing a fault diagnosis case knowledge base; performing fault diagnosis on a support vector machine on the basis of an SVM classification model; acquiring fault information and performing maintenance guide. According to the support vector machine-based intelligent equipment fault diagnosis method, the fault features of equipment are highlighted to the maximum degree; the situations that the equipment data are incomplete and inaccurate are reduced; the possibility is provided for constructing an accurate and reliable fault diagnosis model; the problem that the diagnosis model ages along with the operating time of the equipment is solved; the misdiagnosis rate of the fault diagnosis is reduced; the accuracy and the speed of the equipment fault diagnosis are greatly improved.

A method for reconstructing a super-resolution image, including: 1) reducing the resolution of an original high-resolution image to obtain an equal low-resolution image, respectively expressed as matrix forms yh and yl; 2) respectively conducting dictionary training on yl and yhl to obtain a low-resolution image dictionary Dl; 3) dividing the sparse representation coefficients αl and αhl into training sample coefficients αl_train and αhl_train and test sample coefficients αl_test and αhl_test; 4) constructing an L-layer deep learning network using a root-mean-square error as a cost function; 5) iteratively optimizing network parameters so as to minimize the cost function by using the low-resolution image sparse coefficient αl_train as the input of the deep learning network; 6) inputting the low-resolution image sparse coefficient αl_testas the test portion into the trained deep learning network in 5), outputting to obtain a predicted difference image sparse coefficient {circumflex over (α)}hl_test, computing an error between the {circumflex over (α)}hl_test.

The invention discloses a rolling bearing fault identification method under variable working conditions based on ATT-CNN, and relates to a rolling bearing fault identification technology. The problemthat the generalization ability of an existing rolling bearing fault recognition method under variable working conditions is limited to a certain extent for a complex classification problem is solved.The method comprises the following steps: firstly, mapping vibration data to a nonlinear space domain through a convolutional neural network (CNN), and adaptively extracting rolling bearing fault characteristics under variable working conditions by utilizing the characteristic that the CNN has invariance on micro displacement, scaling and other distortion forms of an input signal; Secondly, an attention mechanism (ATT) thought is put forward to be fused into a CNN structure, and the sensitivity of bearing vibration characteristics under variable working conditions is further improved; And meanwhile, more abundant and diverse training samples are obtained through a data enhancement method, so that the network can be learned more fully, and the robustness is improved. The proposed fault diagnosis model based on the attention mechanism CNN (ATT-CNN) can realize multi-state recognition and classification of the rolling bearing under variable working conditions, and compared with other methods, higher accuracy can be obtained.

The invention provides a multi-tag entity based on a depth neural network and a tagging strategy. Relation joint extraction method based on annotation strategy can effectively avoid the error in the named entity recognition subtask will propagate to the relational classification subtask and ignore the interaction between the two subtasks. At that same time, the invention useS Tree-GRU, as the coding layer, can make the model learn the syntactic information of the whole sentence more fully, and help to identify the relationship between the two entities and the type of the relationship. In addition, because of the complexity of the massive text itself, an entity may be included in a variety of relationships, the use of multi-tag classifiers, a good solution to the above problem. The method of the invention obtains good results in different data fields, and can efficiently, accurately and intelligently extract information with practical value and research significance from massive text data.

The invention relates to a multi-modal remote sensing image high-rise feature fusion classification method based on deep learning. The method includes the steps of preprocessing a multi-mode remote sensing data set, manufacturing a training set, a verification set and a test set of a high-rise feature extractor to construct the model parameters of the high-rise feature extractor; constructing a model of the multi-mode high-rise feature extractor to extract high-rise features of a plurality of modal remote sensing images, extracting and storing the high-rise features of the multi-mode remote sensing images to train, verify and test a multi-classifier, a high-rise feature fusion algorithm being used for making a training set, a verification set and a test set of the classifier, constructingthe classifier model parameters for classifying newly generated data sets. Remarkable effect can be achieved in the remote sensing image ground feature classification through deep learning. The methodis suitable for classification of complex and marine remote sensing images and has the characteristics of being precise, rapid, efficient and safe, so that the utilization rate and the utilization value of the remote sensing images are improved.

The invention relates to a multi-scale nasopharyngeal tumor segmentation method based on CNN. Includes collecting MRI image data of nasopharyngeal region of several cases with nasopharyngeal tumor; Performing artificial edge labeling on the lesion area of the MRI image data collected in the previous step as label data layer by layer; Performing standardized preprocessing on the label data obtainedin the previous step and converting the label data into a two-dimensional data set; A CNN-based multi-layer two-dimensional convolution neural network is constructed and trained by using the two-dimensional data set in the previous step. For the MRI image data of nasopharyngeal region to be segmented, medical images of the same region and the same mode are collected, and the collected images arestandardized. The MRI image data of nasopharyngeal region to be segmented is segmented automatically by the network model. The invention can realize automatic segmentation of nasopharyngeal tumor, andcan obtain higher precision compared with mainstream network.

The invention discloses an autonomous underwater vehicle (AUV) track predicating method based on marine environment information. According to the method, the current track data of an AUV are acquired in real time, the current track data of the AUV are input into an AUV track predicting model for prediction, and then the predicated track of the AUV is obtained. The predicating method of the AUV track predicating model comprises the steps of forming a training sample of the AUV track predicating model, conducting principal component analysis on input data to obtain simplified input data, determining calculating parameters and establishing the AUV track predicating model of the BP neural network, and conducting BP neural network training on the simplified input data to obtain the weight values and threshold values of all layers of the AUV track predicating model. According to the method, measured AUV position data and marine environment data serve as the input of the track predicating model together, so that data are more complete, and then track predicting precision is improved.

The invention discloses a network intrusion detection method and device based on ensemble learning, and the method comprises the steps: collecting network intrusion detection data, extracting features, carrying out the preprocessing, building a training data set, selecting attack type data with a smaller data size, and increasing the data size of the attack type data; training a plurality of learners for each data type in the training data set, and then fusing the learners together in an ensemble learning mode to form an ensemble learning model corresponding to each data type; setting an optimal classification threshold for each data type to minimize the cost of misclassification; and respectively inputting to-be-detected data into the ensemble learning model of each data type, and obtaining the data type to which the to-be-detected data belongs according to the output result of the ensemble learning model and the classification threshold. According to the method and the device, the problems of low detection accuracy, high false alarm rate and high false alarm rate of attack data of attack types with small data volume in the prior art can be effectively solved.

The invention discloses a PID optimization control method of a four-rotor aircraft. The method includes the following steps that power performance modeling is carried out on a PID controller; the PID controller is designed based on a power performance model; parameters of the PID controller are optimized through the particle swarm optimization algorithm; the parameters of the PID controller are further optimized through the combination of an improved particle swarm optimization algorithm and the genetic algorithm. According to the method, the improved particle swarm optimization algorithm is adopted to optimize the parameters of the PID controller, the speed and position of a particle are changed through comprehensive learning of surrounding particles, and better performance can be easily achieved through sufficient learning. A first particle can be updated to an optimal position after comprehensive learning of the surrounding particles; to avoid local optimum, especially for multi-peak functions which are prone to local optimum, the particles are recombined through selection, crossing and mutation of the genetic algorithm, and the prematurity phenomenon of the particles is avoided.

The invention discloses a multistep prediction method for effective parking space occupation rate of a parking lot. The multistep prediction method comprises the following steps of: 1) determining the time sequence of the effective parking space occupation rate of the parking lot; 2) based on the time sequence of the effective parking space occupation rate, setting a multistep prediction step length N; 3) predicting the effective parking space occupation rate for previous n steps; 4) obtaining a new time sequence, then reconstructing the phase space of the new time sequence to obtain a d-dimensional phase space; 5) predicting later N-n steps to the d-dimensional phase space obtained from the step 4); and 6) combining the prediction value of the later N-n steps obtained from the step 5) and the prediction value of previous n steps obtained from the step 3), thus obtaining the prediction result of final N steps. The invention provides a combined prediction method of wavelet neural network-maximum Lyapunov exponent method according to different characteristics of earlier stage and later stage of multistep prediction on the effective parking space occupation rate, the prediction coverage time range is increased, and the precision and the stability are improved.

The invention belongs to the field of image processing, computer vision, deep learning and semantic image segmentation, and particularly discloses a method for segmenting high-resolution combined viewdermoscopy images on the basis of global cavity convolution. The method includes constructing high-resolution combined view feature extraction networks and semantic segmentation networks on the basisof cavity convolution; carrying out training by the aid of cross entropy and jaccard approximation coefficient combined loss functions; carrying out data enhancement and post-processing during prediction. The method has the advantages that sufficiently comprehensive context information can be sieved by combined view by the aid of the global cavity convolution, high-resolution images can be retained, accordingly, sufficiently detailed information can be captured, and the dermoscopy images can be accurately segmented.

The invention discloses a multi-scale CNN-BiLSTM non-coding RNA interaction relationship prediction method with introducing attention, and belongs to the field of bioinformatics and deep learning. Themethod comprises the following steps: (1) proposing a coding mode k-mers suitable for a gene sequence; (2) using a multi-scale convolution kernel for replacing a single-scale convolution kernel, so that topic features with different lengths between sequences are captured, the feature diversity is enriched, and the model prediction performance is improved; carrying out down-sampling on each convolved feature map by using a plurality of pooling windows with different scales, so as to avoid ignoring potential effective information; (3) fusing a BiLSTM model on the basis of the CNN, so that long-distance information dependence between sequences can be better processed, and feature information is fully learned; and (4) introducing an attention mechanism, and distributing different weights to different words in the text vector by using the attention mechanism to distinguish the importance of the information, so that the attention mechanism pays more attention to the key information, and thepurpose of enhancing learning is achieved.

The invention discloses a method of predicting ship navigation meteorological factors according to marine meteorological information. A sample data set is firstly obtained through ship navigation database screening query and data normalization; then, an alternating sparse auto-encoders (ASAE) deep learning regression prediction model is built, based on a large amount of historical ship navigation sample data sets, a least square method, an L2 weight decay constraint and a KL contrast sparsity constraint are adopted to define a cost function, and alternating unsupervised learning and supervised learning are combined for parameter tuning; and finally, to-be-predicted ship navigation data are substituted to the ASAE prediction model with the solved optimize4d parameters for meteorological factor prediction. The method of the invention can build a function relationship between the ship speed and diversified marine meteorological conditions in the ship navigation, the complicated nonlinear problem in meteorological factor calculation can be solved, the meteorological factor prediction precision is improved, and an important role is played in energy-saving emission-reduction navigation optimization.

The invention discloses a GCAM-based high-resolution SAR image airport runway area automatic detection method and system. The GCAM-based high-resolution SAR image airport runway area automatic detection method comprises the steps of performing down-sampling on a high-resolution SAR image to generate a medium-resolution image; inputting the medium-resolution image into a geospatial context attention mechanism network GCAM to extract a runway area; and performing coordinate mapping on the extracted runway region to obtain an airport runway region detection result of the final high-resolution SARimage. Experiments show that compared with DeepLabV3 +, RefineNet and MDDA networks, the method is high in precision and short in consumed time, geographic space information of an SAR image airport can be fully learned, and high-precision, rapid and automatic extraction of a high-resolution SAR image airport runway area can be achieved.

The invention discloses a foreign exchange transaction method and system based on a deep enhancement learning algorithm. The method comprises the steps of (1) establishing an enhanced learning Double-DQN model for a forex trading scene, (2) constructing a deep neural network model and training the enhanced learning Double-DQN model, and learning a Q function, (3) loading trained model parameters,using training data sets of different features, repeating step (2) to perform superposition training and parameter fine-tuning, adjusting the period of model update according to needs, regularly carrying out superposition training on the data sets of different features, and continuously training the model by using data sets of new features such that the model has better scalability and robustness.According to the method and the system, the model can make effective forex trading actions in different data environments, different training sets are used to perform superposition training on the model, and so that the model can be more robust in the context of complex forex data streams.

The invention provides a scene-mode foreign language translation and learning system, including a pair of navigation glasses. A translator of a service end timely translates real-time returned information into language used by a user. Meanwhile, the learning system has a scene identification function, timely returns to-be-learned foreign contents of different types according to different learning scenes, and perform expansion according to the returned to-be-learned text, so as to provide more foreign language information related to the test for the user, and to enable the user to make full use of fragments of time to perform scene-mode foreign learning.

The invention provides a small-target human body detection method based on balanced sampling and nonlinear feature fusion, and the method comprises the steps: carrying out the fusion and enhancement of the features of all scales in a training process, and enabling the scale where a small-target human body is located to obtain sufficient human body semantic information; using an instance balance sampling strategy for positive samples during subsequent loss calculation, it is guaranteed that human bodies of different scales can be trained in the same way, an intersection-parallel ratio balance sampling strategy is used for negative samples, it is guaranteed that enough samples are difficult to participate in training, and finally the classification capacity of the network is improved; and aknowledge distillation training method is adopted in a training strategy, so that the model size is compressed while the precision is ensured, and the reasoning speed is increased.

The invention discloses a hypothetical semi-supervised learning based open domain question answering method. The method includes the following steps: (1) extracting articles related to a question froma corpus by using an information retrieval technology; (2) assuming that articles that come with a given question answering training set are unique positive labels and the articles extracted for thecorpus are negative labels; (3) constructing a deep learning model, learning features of the positive label by training an article scorer, and training a reader to select correct answers from the articles; (4) carrying out article relevance ranking, and sending the first n articles with high relevance to the scorer and re-labeled the articles according to the scores; (5) repeating the step 3 and step 4 until the model converges; and (6) finishing the model training and applying the open domain question answering. The method can greatly improve the quality of article extraction and the accuracyof answers in a conventional open domain question answering system without relying on additional manual annotation and external knowledge.

The invention discloses a road surface garbage sensing method for intelligent road sweeping. The method comprises the steps: establishing and marking a garbage image database; using a data enhancementmethod which comprises geometric transformation and color transformation of the image; randomly scaling, cutting and arranging the images; expanding a data domain by utilizing a generative adversarial network; positioning and recognizing large garbage and small garbage existing on the road surface by adopting target detection and density estimation combined sensing; after a rectangular frame andlabels of the garbage are obtained through target detection, converting the rectangular frame into a density map form, and assigning different density weights according to different labels; combiningthe density map obtained by conversion with a density map generated by a density estimation algorithm to obtain a final pavement garbage density image; calculating candidate cleaning points, and inputting the candidate cleaning points into a path planning module; based on the obtained garbage distribution information, inputting the garbage distribution information the path planning module, and adjusting the driving path. Intelligent sweeping of the sweeper is achieved, and high practical value is achieved.

The invention provides a SAR image change detection method based on maximum edge metric learning, which solves the problem that the SAR image change detection is susceptible to speckle noise and the traditional metric cannot measure the sample difference information well, and has the implementation steps comprising: constructing all samples by inputting the remote sensing image before and after the change; constructing training samples containing all the boundaries; constructing positive and negative constraint pairs using the training samples; by using the positive and negative constraint pairs as input, establishing a structured support vector machine model to obtain a mapping matrix, and decomposing the mapping matrix; mapping all the samples to the feature space by using the decomposedmapping matrix, and performing SAR image change detection classification on all the samples in the feature space. The invention is highs in classification accuracy, and still maintains a high-precision classification effect especially in the case of low time complexity, which maintains good boundary information while suppressing noise. The invention is used for SAR image change detection.

The invention discloses an unbalanced data classification method based on mixed sampling and machine learning. The method comprises the steps of step 1, generating a training set; step 2, for a few types of sample sets P in the training set, copying P to generate P ', using P and P' to synthesize PP ', adopting an smote algorithm to generate S on the basis of the PP', and P, P 'and S form PP' S at the same time; step 3, for the majority of types of sample sets N in the training set, randomly undersampling without putting back to obtain t Ni; step 4, repeatedly executing the step 2 for t timesto obtain t different PP 'Si, and synthesizing Ni and the corresponding PP' Si into a new training set to obtain t subsets; step 5, training to generate t classifiers Hi; and step 6, integrating t Hito obtain a final classifier H, and utilizing the classifier H to complete classification of the unbalanced data set. According to the method, the attention of few types of samples is improved, and meanwhile information of multiple types cannot be excessively lost; The possibility of over-fitting and over-generalization is reduced; The training effect is good, overfitting is not prone to occurring, and the training speed is high.

The invention relates to a photovoltaic power station power prediction method and system based on a recurrent neural network. The method comprises steps of acquiring historical output power data and weather forecast data recorded by a photovoltaic power station; performing data processing to obtain a historical output power data time sequence and a corresponding historical meteorological data time sequence, and performing normalization processing and segmentation to form a sample data set; and constructing and training a recurrent neural network model. Collecting output power data in a period of time, performing data processing, and inputting the data into the recurrent neural network model; and outputting a prediction result by the recurrent neural network model, and obtaining corresponding output power data as a photovoltaic power station power prediction value. According to the method, the photovoltaic power prediction model based on the recurrent neural network model is trained by combining the historical data of the photovoltaic power station and the NWP weather forecast data, the photovoltaic power generation power in the next 24 hours is predicted, and the prediction precision is improved.

The invention relates to a tissue image recognition method and device, a readable medium and electronic equipment, and relates to the technical field of image processing. The method comprises the steps: obtaining a tissue image collected by an endoscope, carrying out the preprocessing of the tissue image to obtain a target image, carrying out the recognition of the target image through a pre-trained recognition model, so as to determine the target type to which the tissue image belongs, wherein the recognition model is obtained through joint training with a preset comparison identification model according to a preset sample image set, the sample image set comprises a first number of labeled sample images with known types and a second number of unlabeled sample images with unknown types, and the first number is smaller than the second number; and comparing the structure of the recognition model to be the same as the structure of the recognition model, comparing neuron parameters of the recognition model, determining according to the neuron parameters of the recognition model, and if the target type indicates that the tissue image is an effective type, performing specified processing on the tissue image. The practicability and accuracy of the recognition model can be improved.

The invention discloses a face retrieving method based on an L1 norm neural network. A neural network can be obtained by means of a cost function of an L1 norm and a pore training sample, and an image matched with a test image is retrieved in a retrieval database through the neural network. Due to the fact that parameters in the neutral network obtained on the basis of the cost function of the L1 norm are determined, the phenomenon that the parameters are redefined according to different test images cannot occur for each retrieval, the retrieval speed is high, and the consumed time is less; meanwhile, the neutral network determined on the basis of the L1 norm is the neural network for deep learning, the powerful learning capacity is achieved, and test image characteristic information can be fully learned, so that the accuracy is very high. The invention further provides a face retrieving device based on the L1 norm neural network, and the technical effect can be achieved.

The invention relates to a dynamic vision sensor-oriented brain-like gesture sequence identification method. The method comprises the following steps of address event expression data capture, space-time convolution, pulse pooling, space-time feature learning and gesture sequence identification by using a learned DoubleSTS model. The method has the characteristics of high noise robustness, high precision, high efficiency, rapid convergence, time sensitivity, brain imitation and the like, and each gesture in the gesture sequence can be successfully recognized.

The invention discloses a model training and a prediction method and a device based on the model training. The method comprises the following steps: identifying whether the layer sub-model is the lastlayer sub-model according to each layer sub-model of the model to be trained which includes at least two layer sub-models in turn; If not, inputting each sample data in the training set containing positive or negative sample tags into the layered sub-model to train the layered sub-model; And updating the sample data in the training set by using the sample data whose confidence level of the corresponding positive sample data in the training set is greater than the confidence level threshold value corresponding to the layer sub-model based on the confidence level of each sample data corresponding to the positive sample data in the training set output by the layer sub-model after training; If yes, inputting each sample data in the training set containing positive or negative sample tags intothe layered sub-model, and training the layered sub-model to improve the prediction accuracy of the model.

The invention discloses a federated product recommendation method, device and equipment and a computer storage medium. The method is applied to a recommendation terminal. The method comprises the following steps of when a product recommendation request is received, extracting target user data corresponding to the product recommendation request, acquiring candidate set product data, inputting the target user data and the candidate set product data into a collaborative memory network model, acquiring a feature representation set, the collaborative memory network model being obtained by executinga longitudinal federation process based on user feature information and product feature information and conducting iterative training on an initial analysis model, and generating a product recommendation set corresponding to the target user data according to the feature representation set. According to the method, the collaborative memory network model is constructed through longitudinal federation, the model is fully learned while data leakage is reduced, and the recommendation accuracy is improved.