39results about How to "Reduce data volume requirements" patented technology

The invention provides a face attribute recognition method based on the multi-task deep learning and relates to the face attribute recognition technique in the field of computer vision. The method comprises the steps of preparing an image data set; subjecting each image in the image data set to face detection one by one; detecting face key points in all detected faces; aligning each face with a standard face image according to the face alignment method based on detected face key points to form a face image training set; calculating an average face image in the training set; constructing a multi-task deep convolutional neural network, and training network parameters after subtracting the average face image from each face image in the face image training set so as to obtain a convolutional neural network model; detecting faces and face key points in a to-be-recognized test image, and aligning each face in the above image with the standard face image based on the face key points; placing the standard face image into the constructed convolutional neural network model after subtracting the average face image from the standard face image and conducting the feedforward arithmetic operation so as to obtain a result.

The invention belongs to the technical field of speech translation, and relates to a BERT embedded speech translation model training method and system and a speech translation method and device.The training method comprises the following steps: collecting model training data; pre-training a BERT model by utilizing a source language in the training data, taking the pre-trained BERT model as a machine translation model coding layer, training a machine translation model by utilizing paired source language and target language texts, and obtaining a plurality of machine translation models by setting the number of decoding layers in the machine translation model; training a speech recognition model by using the speech translation data of the source language pairs; and taking the trained speech recognition model coding layer as a speech translation model coding layer initialization parameter, weighting the outputs of the plurality of machine translation models in an entropy weighting mode to train a speech translation model, and completing speech translation model training in combination with a model loss function. The recognition performance of the speech translation model is improved, and then the speech translation efficiency and quality are improved.

The invention relates to the field of traffic management and control, in particular to a road segment travel time prediction method and a device based on SVM and Kalman filter. The training set data and the test set data composed of the set travel time information and the corresponding travel time are obtained, and the support vector regression machine is trained by the support vector machine regression algorithm. Obtaining an initial predicted travel time matrix according to the support vector regression machine and the test set data, constructing a state equation according to the intersection time delay and the road condition in the set travel section information, and constructing an observation equation according to the predicted travel time, the weather condition and the state equationcorresponding to the continuous sampling time before any time; according to the Kalman filter algorithm to solve the observation equation corresponding to the actual prediction time at any time, by selecting a continuous set of sampling time before any time corresponding to the prediction travel time, real-time change of the coefficient matrix of the Kalman filter, so that the entire prediction model is more reasonable and more accurate.

The invention discloses a quiescent voltage stability margin prediction method based on a Tri-Training-Lasso-BP network. The quiescent voltage stability margin prediction method applies technologies such as a neural network, semi-supervised training and ensemble learning to prediction of the quiescent voltage stability margin of the power system, proposes an online prediction method based on the Tri-Training-Lasso-BP network, and is formed by a three-body training method (Tri-Training), a least absolute shrinkage and select operator Lasso method, and a BP (back propagation) neural network. Themethod is characterized in that the method is composed of a Lasso method and a BP (back propagation) neural network. The quiescent voltage stability margin prediction method can reduce the requirements on the data quantity and quality of the training set, gives play to the advantages of mass data collected in the daily operation process of a power system, improves the generalization capability and prediction precision of a network, reduces the manual intervention, and irons out the problems that a conventional method is difficult to achieve the online real-time prediction of the voltage stability margin, needs a large number of training samples, and is liable to cause the over-fitting.

The invention discloses a Lesion image generation method, device and computer-readable storage medium, which relates to the computer field, in particular to the medical image processing field. The method comprises the following steps: obtaining a lesion medical image; Segmentation of the medical image to obtain a segmented image; Extracting the pathological tissue from the pathological medical image; restoring The medical images of the corresponding organs after the pathological tissues were extracted. grouping The images of pathological tissue extracted from the same organ together to form aset of pathological tissue images. grouping Non-pathological medical images of the same organ into a set of non-pathological medical images; misfusing The image set of pathological tissue and the non-pathological medical image set to generate new pathological medical image of the organ. The invention also provides a lesion image generating device and a computer-readable storage medium. A large numb of lesion medical images are quickly generate by using that original lesion medical images, and the efficiency of obtaining the lesion medical image is greatly improved.

The invention discloses a Chinese civil aviation air traffic control speech recognition method and system. The method comprises the following steps: acquiring voice feature data, wherein the voice feature data is time sequence feature information extracted based on a voice signal; and inputting the voice feature data into a trained acoustic model to obtain a recognition result which represents air traffic control Chinese term characters corresponding to the voice signal; wherein the acoustic model comprises a TRM module, a BiGRU module, a full connection layer and a CTC module which are connected in sequence, the TRM module comprises a multi-head self-attention layer, a first residual connection and layer standardization layer, a feed-forward layer and a second residual connection and layer standardization layer which are connected in sequence, the BiGRU module comprises a bidirectional gating circulation unit network, the CTC module comprises a connection time sequence classification layer, and the acoustic model is obtained by training an air traffic control instruction term voice sample with a Chinese character label. The method has the advantage of high recognition accuracy.

The invention discloses a method for predicting the corrosion rate of a submarine crude oil pipeline based on a PCA-ABC-SVM model. The method comprises the following steps: obtaining detection data, including an actual corrosion rate value, of the submarine crude oil pipeline to be evaluated; according to the obtained detection data, establishing a seabed crude oil pipeline corrosion index system through a PCA algorithm; according to the seabed crude oil pipeline corrosion index system, obtaining a trained corrosion rate prediction model through an SVM algorithm and an ABC algorithm; and substituting detection data into the trained corrosion rate prediction model to obtain a corrosion rate prediction result. The method can solve the problems that a seabed crude oil pipeline corrosion rate prediction model in the prior art has large basic data requirements and is insufficient in prediction reliability, is high in analysis speed, high in accuracy and high in reliability, and can provide scientific basis and technical support for seabed crude oil pipeline corrosion failure risk early warning.

The invention discloses a method for detecting and judging an electroencephalogram signal caused by love impulse. The method comprises the following steps of: establishing a portrait photo material library; collecting electroencephalogram signal data Va; preprocessing the electroencephalogram signal data Va; dividing the electroencephalogram signal sections into two groups; performing channel dimension reduction processing; finding an effective brain region capable of distinguishing two states of love impulse and no love impulse and a component characteristic Fij of the effective brain region;and judging whether the electroencephalogram signal is caused by love impulse or not when the testee is stimulated by adopting a self-defined core algorithm.

The invention provides an active learning self-iteration image classification method and system. The technical scheme of the method comprises a model training step: carrying out the manual marking ofthe type of a target in a training sample, inputting the training sample which has been manually marked into a target detection model and a first-degree learning model, and carrying out the first-degree learning of the target; training the target detection model and the metric learning model; and an inference calculation step: using the trained target detection model and the metric learning modelto identify and classify the target in a to-be-detected image. The method solves the problems that an existing deep learning image classification method is poor in effect and needs a large amount of annotation cost during multi-class classification.

The embodiment of the invention provides a user identity association method and device, and the method comprises the steps: carrying out the polling of an API of a first platform through employing a plurality of APP identities according to the seed ID of a first preset number of first platforms, and obtaining a second preset number of user IDs and a polling record; starting crawler operation of the second platform, scanning the polling record to obtain a corresponding URL, and obtaining an association ID of the URL pointing to the second platform and a non-association ID of the URL not pointing to the second platform in the polling record; extracting features of the association ID and the non-association ID to obtain a first feature vector for training a binary classification model; and obtaining the first platform user ID and the second platform user ID, obtaining a feature vector after feature extraction, and inputting the feature vector into the binary classification model to obtainan identity association result. According to the user identity association method and device provided by the embodiment of the invention, the effective features are extracted from the personal data of the user, the user identity association in the multi-source social network is realized, and the computing resources are saved under the condition of ensuring higher accuracy.

The embodiment of the invention discloses a face recognition method and equipment , and the method comprises the steps: obtaining a first face image and a second face image, enabling the first face image to be a current face image obtained by a camera, and enabling the second face image to be a stored reference face image; judging whether the mode of the first face image is the same as the mode ofthe second face image or not; if the mode of the first face image is different from the mode of the second face image, respectively mapping the first face image and the second face image to a cross-modal space to obtain a first sparse face feature of the first face image in the cross-modal space and a second sparse face feature of the second face image in the cross-modal space; and performing face recognition on the first face image according to the first sparse face feature and the second sparse face feature. According to the embodiment of the invention, the cross-modal face recognition speed can be improved.

The invention provides an estimation method for the distribution of the plug-and-play critical conditions of distributed power sources. The estimation method comprises the following steps that: load distribution on a power distribution line is acquired, a function relationship between the plug-and-play critical conditions of the connection of the distributed power sources to the power distributionline under a preset voltage deviation constraint condition and the connection positions of the distributed power sources is determined according to the load distribution; plug-and-play critical conditions under a preset power flow reverse transmission constraint condition and a function relationship between the plug-and-play critical conditions and the connection positions under a preset line capacity constraint condition are obtained; and the distribution of the plug-and-play critical conditions is obtained according to a function relationship distribution curve corresponding to the three constraint conditions. The method has the advantages of low requirement for data size, small calculation amount, and simple calculation within an error allowable range.

The invention provides an equipment model identification method, device and system, and belongs to the technical field of data processing. The method comprises: determining a first number of candidate physical addresses with relatively high similarity with a target physical address from a database according to the obtained target physical address; then determining, from equipment models corresponding to the first number of candidate physical addresses, the equipment model with the highest occurrence frequency as the equipment model of the target terminal equipment. Therefore, even if the equipment model corresponding to the target physical address is not stored in the database, the equipment model of the target terminal equipment can be determined according to the similarity of the physical addresses, so that the success rate of identifying the equipment model is effectively improved, and the requirement on the data volume stored in the database is reduced.

The invention belongs to the technical field of artificial intelligence image processing, and particularly relates to a one-dimensional sequence dimension-raising clustering method and system. The system comprises an electrocardiosignal collection module, a preprocessing module, a denoising module, a Gram transformation module, an unsupervised Kmeans clustering module, and an output result module.The electrocardiosignal collection module imports one-dimensional electrocarddiosignals, stores the one-dimensional electrocardiosignals in a database, and gives footnotes on the one-dimensional electrocardiosignals from X0 to Xn according to a time sequence of importing the one-dimensional electrocardiosignals into the database; the preprocessing module is used for scaling the time sequence X =x1, x2,..., xn of the one-dimensional electrocardiosignals; the denoising module is used for screening and denoising the one-dimensional electrocardiosignals; the Gram transformation is used for converting the one-dimensional electrocardiosignal into a two-dimensional electrocardiogram image; According to the method, recognition of electrocardiosignals is converted into an image classification problem from a time domain problem, the similarity in data is measured by conducting inner product on a time sequence, the similarity is converted into a Gram matrix, and the dimension increasing processfrom one dimension to two dimensions is completed.

The invention discloses a human body characteristic parameter prediction method based on semi-supervised learning. The method comprises the following steps of constructing a data set, wherein the dataset comprises a labeled data set based on a real human body and a label-free data set based on a virtual human body; preprocessing the image of the data set; training a semi-supervised model by usingthe data set, and constructing a stable mapping model between the input image and the human body characteristic parameters; and processing the to-be-detected input image by using the semi-supervisedmodel, and predicting to obtain human body characteristic parameters. Only a small amount of real labeled human body data is collected. A large amount of unlabeled human body data is generated by means of the simulator. A stable semi-supervised model mapping model can be established by means of a small amount of labeled human body data, and human body characteristic parameters are accurately predicted.

The invention provides a question and answer data processing method, a device, a storage medium and electronic equipment, and belongs to the technical field of data processing. The method comprises the following steps: collecting historical conversation data of a target user and user tag data of the target user; generating a question sequence about historical consultation questions of the target user according to a time sequence of the historical conversation data; according to the user label data of the target user, searching a to-be-recommended question with the access permission of the target user in a consultation question library; and predicting the problem sequence through a pre-trained deep learning model, and determining a target recommendation problem of the target user in the to-be-recommended problems according to a prediction result. According to the method, the interested questions of the user can be determined, and the dependence of the recommendation method on the data volume of historical data of the user is reduced.

The embodiment of the invention discloses an image processing method and a related device. The method comprises the steps of obtaining a first real face image; importing the first real face image intoa three-dimensional face model creation engine, and establishing a first three-dimensional face model matched with the face image; adjusting at least one model parameter of the first three-dimensional face model according to a model parameter adjustment strategy, and generating a plurality of two-dimensional virtual face images of the first user and a face feature value of each two-dimensional virtual face image; taking each two-dimensional virtual face image and the corresponding face feature value as sample data, training a preset two-dimensional face model, and obtaining a trained first two-dimensional face model; importing the first real face image into a first two-dimensional face model to obtain a face feature value of the first user; and adding a face feature value of the first user in a preset face feature value template data set of the first user. According to the embodiment of the invention, the data size requirement of the real face photo is reduced, and the image processing efficiency and accuracy are improved.

The invention discloses an NGS targeted capture method based on a dark probe technology and application of the NGS targeted capture method in differential depth sequencing. The invention provides a targeted capture high-throughput sequencing method for simultaneously detecting exons and SNP sites in a target region range of a whole genome. The method comprises the following steps: designing two groups of original probes capable of capturing all exons and all SNP sites in the target region according to a nucleotide sequence of the target region; one of the two groups of original probes has a labeled form and an unlabeled form at the same time, and the other group only has a labeled form; matching and combining the two groups of probes according to different proportions of the labeled probes, and hybridizing with a genome library to be detected to obtain a capture library; and performing high-throughput sequencing. Compared with a standard exon sequencing method, the dark probe capture method disclosed by the invention has the advantages that the sequencing coverage degree is controlled according to the requirements of different areas under the same single-tube reaction, and the data utilization rate of NGS is remarkably improved. The dark probe method disclosed by the invention can be used as a high-cost-performance alternative scheme of the WGS.

The invention provides a steel and iron material fatigue performance prediction method based on transfer learning guided by a mechanical theory, and relates to the technical field of steel and iron material design and machine learning application. According to the method, a mechanical theory mechanism is introduced into machine learning, and the small sample problem of material high-cost attribute prediction is solved. The relationship among the steel grade components, the process and the target performance is established on the basis of mechanical theory guidance. According to the method, aiming at obtaining the target performance with high cost, the transfer learning model for accurately predicting the target performance can be established by utilizing the high correlation between the target performance and the source performance, namely based on the guidance of the mechanical theory and only utilizing dozens of groups of target performance data. According to the method, the data size requirement of machine learning for high-cost target performance is remarkably reduced, the high-cost target performance evaluation and prediction efficiency is remarkably improved, and finally the new material research and development rate is improved.

The embodiment of the invention relates to the technical field of data processing, and discloses a score prediction method. The method comprises the steps that: scores of a plurality of users for a plurality of items are obtained, wherein the scores of the plurality of users for the plurality of items comprise the score of a target user for at least one non-target item, and the score of at least one non-target user for a target item; a distance matrix is generated according to the scores of the plurality of users for the plurality of items, wherein the distance matrix comprises distances between the plurality of users and the plurality of items calculated according to the scores of the plurality of users for the plurality of items; and the distance matrix is input into a pre-trained deep neural network, so that a predicted score matrix can be obtained, wherein the predicted score matrix at least comprises a predicted score of the target user for the target item. According to the scoreprediction method, the recommendation method, the processing device and the storage medium, the accuracy of model prediction scores can be improved.

The invention discloses an optical channel fault diagnosis method based on transfer learning. The method comprises the following steps: acquiring optical network performance, alarm, log and topological data of a certain training area, and constructing a data sample set required by model training; extracting training state samples, training state features and training state relationships from the data sample set; selecting a transfer learning mode, and inputting the extracted training state samples, training state features and training state relationships for model training to obtain an optical channel state diagnosis training state model; selecting optical network performance, alarm, log and topological data of a reasoning area, and performing health state labeling on optical channel data as a reasoning state sample; loading the reasoning state sample into an optical channel state diagnosis training state model, and training an optical channel state diagnosis reasoning state model; and calling the newly generated optical channel state diagnosis reasoning state model to obtain the health state of the analysis object optical channel. The invention also provides a corresponding optical channel fault diagnosis system based on transfer learning.

The invention discloses an equipment type identification method combining electric power fingerprint knowledge and a neural network. The method comprises the following steps: S1, acquiring voltage and current sampling data when equipment is used; S2, setting a time interval, and segmenting the data obtained in S1; S3, converting the data obtained in the step S2 into common electrical characteristic quantities; S4, inputting the electrical characteristic quantity obtained in the S3 into a knowledge extraction model to obtain electric power fingerprint knowledge points of the equipment; S5, encoding the electric power fingerprint knowledge points obtained in the step S4, and splicing the electric power fingerprint knowledge points with the electric characteristic quantities obtained in the step S3 to obtain a total characteristic vector; And S6, inputting the total feature vector obtained in the step S5 into a trained neural network to obtain the equipment type. Compared with a traditional load identification method, the invention has the advantages that a machine learning method and a knowledge-driven method are organically combined, the requirement for the data size can be greatly reduced, the convergence data and judgment speed of the model can be increased, and the accuracy of the model can be improved.

The invention relates to the technical field of electromechanical equipment life prediction, in particular to an electromechanical equipment life prediction method and system based on data driving, and the method comprises the steps: obtaining a data set of multiple pieces of equipment; abnormal data are screened out; performing period division according to the distribution of the abnormal data on the time axis; performing cycle combination on cycle division results of different devices to obtain a training set and a test set; inputting the training set into a selected data driving model for training; inputting the test set into the trained data driving model for testing; and after the test is finished, predicting the service life of any electromechanical equipment according to the obtained data driving model. According to the method provided by the invention, the abnormal data are screened, the data of multiple devices are subjected to cycle division and then are recombined to obtain the training set and the test set, the model is trained and tested, the life prediction is completed, and the data are integrated in the data division stage, so that the model is more comprehensive and accurate.

The invention provides a pulmonary nodule detection method and device based on a neural network and image processing equipment. The method comprises the steps of S1, carrying out the recognition and segmentation of a pulmonary nodule in a CT image based on a 3D DCNN network frame, and obtaining candidate pulmonary nodule data, S2, constructing eigenvalue vectors of candidate pulmonary nodule data in the CT image to obtain a sample data set, and S3, dividing the sample data set into a training set and a test set according to a proportion, performing classification training on the training set by using a pre-constructed GRU model, inputting the test set into the trained GRU model for detection, and outputting a pulmonary nodule detection result. According to the method, after the CT lung image is automatically detected and segmented, the focus area is extracted, the feature value is constructed, the neural network is used for modeling, the candidate pulmonary nodule area is detected, false positive is reduced, and the accuracy of automatic detection of the lung image is improved.

The embodiment of the present invention provides a method and system for classifying medical record data, which preprocesses the original medical record data to obtain a data set that can match the classifier; then, randomly divides the data set after feature selection into a specified number After dividing each data block into a training set and a test set, each data block is input into the corresponding classifier respectively; then, based on the TPE algorithm, all classifiers are predicted by weighted voting, and run After the specified number of TPE algorithms, the number of classifiers with the smallest verification error is selected as the optimal number of classifiers; finally, the optimal number of classifiers is input into the deep cascade forest model, and the deep cascade forest model is run successively until it satisfies Preset accuracy to obtain the optimal classification result of medical record data. The technical solutions of the embodiments of the present invention have low requirements on the amount of data, are easy to train, and have the advantages of strong adaptability.

The invention discloses a settlement monitoring method based on a D-InSAR technology and image weighted stacking. The method comprises the steps: selecting a plurality of common main images from an SAR image database, carrying out the baseline estimation of the processed SAR images, generating a connection diagram, processing the corrected SAR images, generating a plurality of interference diagrams, performing filtering processing on all interference diagrams and calculating coherence of the interference diagrams, performing phase unwrapping on each interference diagram after filtering processing, removing the interference diagrams with poor coherence from the unwrapped interference diagrams, performing orbit refining and re-de-flattening processing on the interference diagrams after removing, performing weighted stacking on the interference diagrams; and finally, carrying out phase-to-deformation processing on the phase diagram, and carrying out geocoding to generate a deformation diagram. According to the invention, the advantages of the D-InSAR technology and weighted stacking are well combined, the precision of ground subsidence monitoring can be improved, the requirement for a data size is lowered, and therefore long-time and large-range time sequence deformation monitoring can be performed on an area.

The invention discloses an electricity sale quantity prediction method and system based on waveform characteristics. The method comprises the following steps: acquiring a historical temperature-time function change curve and a historical electricity sale quantity-time function change curve; according to a temperature-time change curve of four days before a to-be-predicted time period in a preset period, determining a temperature characteristic value of the temperature-time change curve; determining the electric quantity similarity of the electric quantity sale-time change curve according to the electric quantity sale-time change curve of the first four days of the to-be-predicted time period in the preset period; according to the temperature characteristic value and the electric quantity similarity, determining the change rate of the electricity sale quantity; and calculating the predicted electric quantity according to the change rate of the electric quantity sale. By means of the mode, the requirement for the data size can be reduced, the implementation time is short, the calculated amount is reduced, and the cost of power enterprises can be reduced.