93 results about "Neural network regression" patented technology

A method and system for determining fractional flow reserve (FFR) for a coronary artery stenosis of a patient is disclosed. In one embodiment, medical image data of the patient including the stenosis is received, a set of features for the stenosis is extracted from the medical image data of the patient, and an FFR value for the stenosis is determined based on the extracted set of features using a trained machine-learning based mapping. In another embodiment, a medical image of the patient including the stenosis of interest is received, image patches corresponding to the stenosis of interest and a coronary tree of the patient are detected, an FFR value for the stenosis of interest is determined using a trained deep neural network regressor applied directly to the detected image patches.

The invention discloses a method for predicting the remaining service life of complex equipment based on a combined depth neural network. The main steps are as follows: acquiring multi-sensor data ofcomplex equipment; Obtaining effective measurement data by feature selection; obtaining A plurality of slice samples by preprocessing; Establishing the neural network regression model which combines the attention mechanism and depth neural network; The slice samples and their corresponding labels are inputted into the neural network regression model to train the neural network regression model offline. inputting The slice samples of multi-sensor data to be predicted into the trained neural network regression model, and the remaining service life of complex equipment is obtained. Considering the data characteristics of the multi-sensor signal, the invention fully excavates the local characteristics and the time sequence information in the data, has high prediction accuracy and wide applicability, and can be widely applied to various pieces of complex equipment.

A method of obtaining one or more components from an image may include normalizing and pre-processing the image to obtain a processed image. Features may be extracted from the processed image. Neural-network-based regression may then be performed on the set of extracted features to predict the one or more components. These techniques may be applied, for example, to the problem of extracting and removing bone components from radiographic images, which may be thoracic (lung) images.

The invention provides an abnormity detection method based on deep learning in complex environment. An object space-time characteristic extracted through a convolution neural network regression method is input into an LSTM model, and then motion trajectories of multiple objects in the complex environment are tracked; non-linear space-time actions of adjacent individuals are captured in a case of irregular movements of the multiple objects, dependence of the motion trajectories between the adjacent individuals is evaluated, and future motion trajectories of the individuals are predicted; and abnormity detection is completed according to abnormity probabilities of the future motion trajectories of the individuals. The method can reduce the false detection rate of images. In the prior art, a space-time characteristic of a single object is mainly detected without considering a mutual interference condition existing between the motion trajectories of the adjacent individuals in the complex environment. According to the LSTM model, the dependence between the several individuals is evaluated, and the future motion trajectories of the objects are predicted by using a coding and decoding framework, so an accurate result can be obtained when abnormity detection is performed on movements of the multiple objects.

The invention relates to a training test method of a BP neural network regression model and is applied to an oak laser cutting system to predict notch widths. The method mainly comprises steps that a,data acquisition, data sets of N experiment samples are acquired, and M sets of experiment data are totally included; b, data pre-processing; c, data grouping; d, optimization searching of super parameters of a BP neural network and initialization; e, first-time training of the BP neural network; f, second-time training of the BP neural network; and g, training accomplishment of the BP neural network, the notch widths of the oak laser cutting system under different parameters are predicted. The method is advantaged in that over-concentrated and over-sparse data conditions are trained, and thetraining effect of the BP neural network is improved.

The invention discloses an image quality evaluation method based on a combination neural network and a classification neural network. In the training stage, an objective reality quality image of a distortion image obtained by adopting a full-reference image quality evaluation method is adopted as supervision, a normalized image of the distortion image is trained to obtain a combination neural network regression training model for different distortion types; a classification label of the distortion image is adopted as supervision, and the normalized image of the distortion image is trained to obtain a classification neural network training model; in the testing stage, the normalized image of the distortion image to be evaluated is input into the classification neural network training model,and a distortion type is obtained; according to the distortion type, the normalized image is input into the corresponding combination neural network regression training model to obtain an objective quality evaluation prediction quality map, and adopting a saliency map for performing weighing pooling on the objective quality evaluation prediction quality map, and obtaining an objective quality evaluation prediction value. The method has the advantage that the correlation between the objective evaluation result and subjective perception is effectively improved.

The embodiments of the invention disclose a hand key point detection method, a gesture recognition method and related devices. The hand key point detection method comprises the steps of: acquiring a hand image; inputting the hand image into a pre-trained thermodynamic diagram model to obtain a thermodynamic diagram of hand key points; inputting the thermodynamic diagram and the hand image into a pre-trained three-dimensional information prediction model to obtain hand structured connection information; and determining three-dimensional coordinates of the hand key points in the world coordinatesystem according to the hand structured connection information and two-dimensional coordinates in the thermodynamic diagram. According to the embodiments of the invention, the two-dimensional coordinates and the hand structured connection information are successively predicted through the two models so as to calculate the three-dimensional coordinates of the hand key points relative to the three-dimensional coordinates returned directly through a deep neural network; the calculation amount of each model is small; the method is suitable for the mobile terminal with the limited calculation capacity; due to the small calculation amount and the short detection time of the hand key points, the hand key points are detected in real time, and gesture recognition can be easily applied to the mobile terminal.

The invention discloses a vehicle dynamic weight measuring method based on neural network regression. The vehicle dynamic weight measuring method comprises the following steps: (1), data preprocessing: collecting original data of vehicle dynamic wheel weight by using a weighing sensor and preprocessing the original data; (2), data feature extraction: acquiring the characteristics of the data related to a static weight measuring scene; (3), model training: using the data characteristics as input, and training a dynamic weight measuring model by using a neural network regression algorithm; and (4), vehicle weight measuring: calculating the weight of a vehicle by the dynamic weight measuring model according to vehicle dynamic wheel weight data obtained in real time. The invention further discloses a system adopted by the vehicle dynamic weight measuring method. The vehicle dynamic weight measuring method provided by the invention has the advantages that the nonlinear relationship betweenthe dynamic wheel weight data and accurate vehicle weight can be deeply excavated, so that the accuracy of dynamic vehicle weight measuring is improved.

The invention provides a method for predicting shaft power of an industrial extraction condensing steam turbine, which bases on a thermodynamic model of the extraction condensing steam turbine. Considering that the steam turbine is affected by environmental temperature, condensed water flow, temperature, and other unknown factors in practical industrial application process, the influences of changes of parameters such as quality of cooling water, steam inlet quality of main steam, extraction pressure and the like on the extraction quality and the discharging quality of the extraction condensing steam turbine are introduced according to practical industrial data; by adopting a neural network regression method, the practical condensing pressure and the extraction temperature of the extraction condensing steam turbine in industrial application can be worked out; subsequently the practical extraction and discharging enthalpy value of the extraction condensing steam turbine is obtained by the calculation according to industrial standard IAPWS-IF97 of water and steam; subsequently the practical shaft power output of the steam turbine is calculated according to the thermodynamic method; therefore, the direct estimating on the entropy efficiency of the steam turbine is avoided, the prediction precision on the shaft power of the industrially-applied steam turbine is improved, and foundation and basis are provided for the optimizing and the rebuilding and the like of a public engineering system.

A method of obtaining one or more components from an image may include normalizing and pre-processing the image to obtain a processed image. Features may be extracted from the processed image. Neural-network-based regression may then be performed on the set of extracted features to predict the one or more components. These techniques may be applied, for example, to the problem of extracting and removing bone components from radiographic images, which may be thoracic (lung) images.

The invention discloses a detection method of pipeline leakage. The detection method of the pipeline leakage solves the problem that an existing method is low in detection accuracy. The method comprises the following steps that stress wave signals of each monitoring point in a pipeline with known leakage conditions are collected, characteristic values are extracted, a training sample is constructed to obtain a sample characteristic value; the sample characteristic value is used as an input signal of a support vector machine classifier, and the optimal classification surface function of the support vector machine for judging the leakage situation is obtained according to the actual leakage situation of each monitoring point; and the sample characteristic value is used as an input signal ofa neural network regression device, and a neural network model and a leakage point model position which are calculated on the position of the leakage point according to the actual position of the leakage point is obtained. The detection method realizes the accurate judgment and calculation of the leakage situation of the pipeline and the position of the leakage point.

The invention discloses a coal mine gas prediction method based on deep learning. An observable data set oriented to big data analysis is established according to actual production data collected on site; data preparation of high-dimensional gas data is carried out; the method comprises the steps of preprocessing measurement misalignment and missing data, preprocessing a time sequence, normalizingsample data, reducing dimensions and the like. The method comprises the following steps of: selecting a deep neural network-DNN (Deep Neural Network) as a gas data sensing model; based on Keras, a distributed deep learning framework is established, a plurality of machine learning algorithms are integrated, then an automatic machine learning engine is created, a model is trained, testing is completed, and intelligent prediction of coal mine gas is achieved by applying the deep neural network regression model. By means of the method, a more accurate risk pre-judgment basis can be provided for the mining and tunneling process of the coal mine.

The invention discloses a high-precision indoor positioning method based on joint vision and wireless signal characteristics, and the method comprises the steps: in the offline stage, a to-be-positioned indoor site is collected, a WiFi fingerprint database and an image database are constructed, and the scene information of an environment is obtained; and in the online stage, a mobile terminal collects WiFi fingerprint data and image data in real time, performs coarse positioning on the collected WiFi fingerprint data, determines a potential area of a user, and then adopts a method based on deep neural network regression for the image data of the coarse positioning area to complete prediction of an accurate positioning position. According to the indoor positioning method, the wireless signal characteristics and the visual characteristics are fused so that the positioning error is further reduced while the calculation resources are reduced and the calculation complexity is reduced, and the high-precision indoor positioning is realized.

The invention discloses a flexible optical network time domain equalization method and system based on a composite neural network, and belongs to the field of optical fiber communication systems, andthe method comprises the steps: (1) preprocessing a received signal transmitted by a flexible optical network; (2) calculating an amplitude distribution histogram of the preprocessed received signal;(3) inputting the amplitude distribution histogram into a first-stage multi-task neural network classifier, and outputting transmission parameters of the flexible optical network; (4) setting a weightand an offset parameter of a second-stage neural network regression device according to the transmission parameter of the flexible optical network; (5) carrying out time domain equalization on the preprocessed received signal by adopting a second-stage neural network regression device, wherein the number of input neurons of the first-stage multi-task neural network classifier is the same as the number of groups of amplitude histograms, and the number of output neurons of the first-stage multi-task neural network classifier is the same as transmission parameters of the flexible optical network. The time domain equalization method and system disclosed by the invention are wider in application range.

The present invention provides an evaluation method and device of automotive fuel economy. The method comprises a step of establishing a neural network regression calculation model and obtaining a fuel economy probability sequence according to a vehicle speed data sequence, an instantaneous fuel consumption data sequence, an engine rotation speed data sequence, an acceleration data sequence, the neural network regression calculation model, and a fuzzy probability mapping model, and a step of comparing each fuel economy probability in the fuel economy probability sequence with a preset probability threshold orderly, and evaluating the automotive fuel economy according to a comparison result. Through the method and the device, the automotive fuel non-economic process in an automotive driving process can be effectively identified, the automotive fuel economy is evaluated according to the automotive driving data corresponding to the automotive fuel non-economic process, and an automotive driving experience is improved.

The invention relates to a prediction method for coal-fired unit denitration control system inlet nitrogen oxide. The method comprises the steps of collecting the concentration of inlet nitrogen oxide; pre-processing the data; performing on-line sequential extreme learning machine learning; sending a new nitrogen oxide concentration collection value into an input end of a predication model built in the third step and calculating an output weight of the next moment; using the obtained output weight as the input of a single-implicit strata feedforward neural network regression model of the on-line sequential extreme learning machine to obtain the next prediction value; returning the next prediction value obtained in the fifth step to the fourth step. The prediction method employs the online extreme learning machine, so that the calculation speed is high, only output weight needs to be updated and calculation time is greatly saved; a recursion formula is added on the basis of an off-line extreme learning machine and new output weights are obtained according to new data, so that the online learning capability is implemented, the calculation time is short and the prediction precision and the generalization ability are better than those of a neural network.

The embodiment of the invention relates to a method for simulating and rebuilding ECG lead data based on neural network algorithm, which comprises the following steps: obtaining ECG monitoring data ofthe monitored person; the ECG monitoring data comprises the lead data of at least one limb lead and the lead data of at least one chest lead; a multi-neural network regression prediction model is trained for the reconstruction of multi-lead ECG signals based on the neural network machine learning algorithm; the independent variable of the multi-neural network regression prediction model is the lead data of at least one limb lead and the lead data of at least one chest lead; the dependent variable is the lead data of the remaining unknown leads except at least one limb lead and at least one chest lead; the multi-neural network regression prediction model comprises a weight coefficient and a bias coefficient, and the weight coefficient and the bias coefficient are determined by the resultsof the training of the neural network machine learning algorithm; the lead data of the remaining unknown leads is predicted according to the weight coefficient and the bias coefficient obtained by thetraining.

A yield prediction apparatus is provided. The yield prediction apparatus may include at least one processor coupled to at least one non-transitory computer-readable medium. The at least one processor may be configured to receive a first variable associated with operating characteristics of a semiconductor device, perform a simulation for the operating characteristics of the semiconductor device, perform a neural network regression analysis using a result of the simulation to determine a first function for the first variable, and predict a yield of the semiconductor integrated circuit based on an advanced Monte Carlo simulation. An input of the advanced Monte Carlo simulation may include the determined first function.

The invention discloses a human face rendering method based on a Hermite interpolation neural network regression model, and belongs to the technical field of a realistic graphics real time rendering technology. The human face rendering method based on a Hermite interpolation neural network regression model includes the steps: human face area dividing, face radiancy parameter precomputation, sample data acquisition, construction and trainning of a Hermite interpolation neural network regression model, and final rendering. The human face rendering method based on a Hermite interpolation neural network regression model introduce a regression analysis theory into the human face rendering process, uses the Hermite interpolation neural network to construct a learning model, uses the sample set to train, and determines the weight matrix between each hidden layer neuron so as to effectively excavate the non-linear association between the physical attribute and he geometrical characteristic attribute of the visible points in each subarea of the face. By means of the nonlinear mapping, the human face rendering method based on a Hermite interpolation neural network regression model can quickly map the characteristic attribute of each point on the surface of the face into the color value of the point in the given lighting condition. The human face rendering method based on a Hermite interpolation neural network regression model can effectively reduce the computing scale, and can preferably realize real-time rending of realist graphics of a human face.

The invention discloses a well deviation angle prediction method based on an ensemble learning algorithm. According to the method, an SVR learner model, a neural network regression algorithm learner model, a random forest regression algorithm learner model and a Gaussian regression algorithm learner model in machine learning are employed to learn a learning sample, consisting of drilled borehole trajectory data, a drilling mode, a bottom drilling tool structure parameter and the like, of a certain well, the above four learner models are trained so as to separately predict the well deviation angle of a blind area at the bottom of the well, and then linear regression is carried out on a training result and a target value to obtain a final prediction result. Verification results of actual drilling data show that the method is high in prediction precision, effectively reduces the error of predicting the well bottom well deviation through a traditional constant curvature extrapolation method, and improves the accuracy of predicting the well deviation angle.

The invention relates to a proton therapy monitoring method, device and system based on a neural network. The proton therapy monitoring method based on the neural network comprises the steps that according to a three-dimensional body model built in advance, a tumor lesion area is determined; after a proton beam shines the tumor lesion area, positive electron nuclide is measured in preset time to obtain distributed information of the positive electron nuclide; according to the distributed information, an image reconstruction algorithm is adopted to build a PET image; the PET image is input to a built neural network regression model and a neural network classification model, and dose distribution and range of the proton beam are determined; the position of a prague peak is determined according to the dose distribution and range; whether the position of the prague peak and the dose distribution conform to the preset proton therapy requirement or not is detected; and if not, beam out parameters of the proton beam are adjusted. By adopting the technical scheme of the proton therapy monitoring method based on the neural network, the influence of breathing and exercising of human body organs on the measurement can be reduced, the range of protons and the precision of the dose distribution are improved, and the accuracy of proton therapy is improved.

The invention discloses a wireless ad hoc network performance prediction method based on an improved BP neural network regression algorithm. The method simultaneously relates to the field of wirelesscommunication networks and machine learning. A traditional BP neural network regression algorithm is improved, three network performance indexes, namely throughput, time delay and packet loss rate, ofthe wireless ad hoc network in the time-varying environment are predicted respectively through the improved algorithm, and the convergence rate of network parameters is effectively increased on the premise that the prediction performance of an original algorithm is guaranteed. According to the method, an empirical data set is constructed by combining an actual task scene and three MAC protocols (CSMA / CA, DTDMA and ESTDMA), and each piece of data can represent one task scene; and the traditional BP neural network is improved, so that the convergence rate of network parameters is improved. Thebasic idea of the method is that features are extracted by analyzing actual task information to construct an empirical data set; an amplification function is introduced into a BP neural network parameter offset calculation formula to improve the parameter convergence rate, and an improved algorithm is used to learn an empirical data set to obtain a learning model; and calling the learning model topredict the network performance for the new task.

An apparatus for processing an audio signal to obtain control information for a speech enhancement filter has a feature extractor for extracting at least one feature per frequency band of a plurality of frequency bands of a short-time spectral representation of a plurality of short-time spectral representations, where the at least one feature represents a spectral shape of the short-time spectral representation in the frequency band. The apparatus additionally has a feature combiner for combining the at least one feature for each frequency band using combination parameters to obtain the control information for the speech enhancement filter for a time portion of the audio signal. The feature combiner can use a neural network regression method, which is based on combination parameters determined in a training phase for the neural network.

The invention provides a land water reserve prediction method and equipment based on a neural network algorithm. The land water reserve prediction method comprises the following steps: acquiring spatial resolution of surface parameter information, land water reserve information and land water reserve information; carrying out resampling for reducing the spatial resolution on the surface parameterinformation to obtain first surface parameter information; constructing a neural network regression model based on the first surface parameter information and the land water reserve information; obtaining target surface parameter information in to-be-predicted time, and performing resampling for reducing spatial resolution on the target surface parameter information to obtain second surface parameter information; and inputting the second surface parameter information into the neural network regression model, and determining land water reserve information corresponding to the target surface parameter information within to-be-predicted time. Compared with the prior art, the method can achieve the precise prediction of the land water reserve information in a historical period, and obtains theland water reserve dynamic change data in a long time sequence.

A space-time load prediction method based on a graph neural network and regional gridding relates to the technical field of power system power distribution network planning, and comprises the following steps: step 1, feature engineering: selecting data features; 2, constructing a network topology, and fusing the feature information in the step 1; 3, transmitting the feature information of each power supply unit based on the topological graph in the step 2; 4, predicting the load of the power supply unit based on the network topological graph obtained in the step 2 and the power supply unit information obtained in the step 3; and step 5, based on the previous steps, dividing grids, and carrying out unit load power supply grid load prediction. According to the method, a to-be-predicted region is divided into a plurality of grids, and a neural network load prediction model is used for a load structure to obtain load prediction results of the whole city at different time and regions; a load prediction model is established through a gridding technology, a graph neural network, regression prediction and other methods, power grid topological structure information is fused, and more accurate prediction is provided for a power distribution network planning load prediction task of a power system.

The invention relates to a congestion traffic flow traceability analysis method, which comprises the following steps of S1, constructing a deep neural network multi-classification model based on automatic vehicle identifier data and vehicle road source data of vehicles in a congestion area, and obtaining space sources of the vehicles; and S2, constructing a deep neural network regression model based on the space source of the vehicle and the data of the automatic vehicle identifier to obtain a time tracing result of the vehicle. Compared with the prior art, the method has the advantages that the source information of the traffic flow in the congestion area is considered, so that the capacity of relieving congestion from the network level is achieved, and a new research view angle for relieving congestion is provided; compared with a traditional machine learning algorithm, the reasoning accuracy can be obviously improved.

The invention provides an image coding processing method and device, and the method comprises the steps: carrying out the downsampling of an original image frame, and obtaining a target image frame; determining a target residual error of an image block of the target image frame; inputting the target residual error into a pre-trained target residual error network model to obtain the probability ofeach quantization parameter corresponding to the target residual error output by the target residual error network model, and determining the quantization parameter of which the probability is greaterthan a predetermined threshold value as a target quantization parameter; generating a quantization parameter table of the original image frame from the target quantization parameter according to thecorresponding position of the image block and the original image frame; and image coding is performed on the original resolution image and the quantization parameter table, so that the problem of subsequent image coding errors caused by inaccurate determined optimal quantization parameters due to the fact that a neural network regression device is trained to map a plurality of features of textureinformation of an extracted image block to determine the optimal quantization parameters in related technologies can be solved.

The invention discloses a plantar pressure image registration method based on deep learning. According to the invention, a cascade convolutional neural network regression model is used to estimate the registration parameters of a plantar pressure image, the to-be-registered plantar pressure image uses a main shaft algorithm to obtain an initial registration parameter, and further the rotation angle of the to-be-registered image is reduced to a certain range, so that the subsequent secondary registration is facilitated. The designed cascade convolutional neural network framework is divided into a coarse adjustment network and a fine adjustment network, a source plantar pressure image is generated through different transformation parameters, and the generated data set is used for model training, and then the to-be-registered image is input into the trained coarse adjustment network model and the trained fine adjustment network model in sequence, and finally, the results outputted by the coarse adjustment network model and the fine adjustment network model are superposed and combined to obtain a final plantar pressure image registration parameter, so that the efficiency of optimizing the registration parameter is remarkably improved.

The invention discloses a college student dormitory allocation method based on a machine learning algorithm. The invention belongs to the crossing field of social behavioristics, data science and system science. According to the method, a prediction model of a student score change trend is established mainly through historical score data and dormitory data of college students and through a classical algorithm based on machine learning, including a BP neural network, Logistic regression, local linear regression and a support vector machine, and a data model suitable for the actual situation ofeach college is screened out through model precision comparison; dormitory states are defined according to student score classification, and based on the score change trend prediction model, a conversion score under each dormitory state is calculated; constraint conditions are reasonably set according to actual conditions, the maximum conversion score serves as a target function, the number of dormitories in each state under the maximum target function value is solved through a CPLEX optimization solver, and optimal allocation of the dormitories is achieved.