59results about How to "Accurate and Efficient Prediction" patented technology

The invention discloses a crowd density estimation and people counting method based on a full convolutional network. A training data preparation module, a prediction model designing module, a prediction model training module and a real-time detection module are included; the full convolutional network comprises a deep layer convolutional neural network and two shallow layer convolutional neural networks; the deep layer convolutional neural network is used for processing scenes close to cameras of the crowd, human face and body features are obtained, and the maximum pooling operation is adopted; the shallow layer convolutional neural networks are used for processing the scenes far from the camera of the crowd, and obtaining body contour information and adopts the average pooling operation.The model adopts the full convolutional network and is applicable to input images of any size; due to the mode of deep layer and shallow layer network combination adopted by the model, the model can be migrated to different application scenarios; the system can efficiently and accurately predict the crowd density and the crowd quantity.

The invention discloses a photovoltaic power generation output power tracking algorithm based on a genetics algorithm improved RBF-BP neural network. By building an RBF-BP neural network, an error absolute value between predicted output and expected output of photovoltaic power generation output power is taken as the fitness, and then a genetics algorithm is adopted for selecting, intersecting and mutating data acquired by photovoltaic power generation equipment in order to find out an individual corresponding to the optimal the fitness. The photovoltaic power generation output power tracking algorithm disclosed by the invention combines the advantages that an RBF neural network is high in rate of convergence, good in heap sort performance and the BP neural network is high in self-learning and self-adaptive capabilities, and has the characteristics of better generalization performance, higher rate of convergence, higher prediction precision and the like.

The subject invention leverages the conditional Gaussian (CG) nature of a continuous variable stochastic ARMAxp time series model to efficiently determine its parametric gradients. The determined gradients permit an easy means to construct a parametric structure for the time series model. This provides a gradient-based alternative to the expectation maximization (EM) process for learning parameters of the stochastic ARMAxp time series model. Thus, gradients for parameters can be computed and utilized with a gradient-based learning method for estimating the parameters. This allows values of continuous observations in a time series to be predicted utilizing the stochastic ARMAxp time series model, providing efficient and accurate predictions.

The invention relates to the information processing field and particularly relates to an enterprise-industry classification system based on automatic information screening. According to the system, an industry classification neural network model is constructed by combining a circulating neural network with a threshold control method, and the automatic classification judgment of secondary industries of enterprises is realized according to business scope information and name information of the enterprises. According to the system, features of text data are automatically extracted by virtue of a deep learning technique and a GRU circulating neural network, the automatic information screening and filtering of the business scope based on a company name can be realized by adding a threshold controlled neural network, and key information is automatically screened from different types of the secondary industries which are difficultly differentiated, so that the efficient and precise prediction of the types of the secondary industries is realized. The deficiency that a circulating neural network is independently used is remedied, and meanwhile, the advantage of the neural network that the features are automatically extracted without manual intervention is developed.

The subject invention leverages the conditional Gaussian (CG) nature of a continuous variable stochastic ARMAxp time series model to efficiently determine its parametric gradients. The determined gradients permit an easy means to construct a parametric structure for the time series model. This provides a gradient-based alternative to the expectation maximization (EM) process for learning parameters of the stochastic ARMAxp time series model. Thus, gradients for parameters can be computed and utilized with a gradient-based learning method for estimating the parameters. This allows values of continuous observations in a time series to be predicted utilizing the stochastic ARMAxp time series model, providing efficient and accurate predictions.

The invention discloses a method for predicting the distribution of a thin sand body. The method includes the following steps: 1. using well logging information to identify a sand body, determining the pattern of longitudinal growth of the sand body from a geological section; 2. conducting phase rotation at -90 DEG on an original seismic section, such that the thin sand body corresponds to seismic wave amplitude response characteristics; 3. projecting a typical well-sand group to the seismic section, determining the position of the thin sand body to the seismic section; 4. longitudinally conducting continuous stratal slice on a seismic data body, observing the change characteristics of the seismic wave forms of the slices, determining a response characteristic slice of the thin sand body; and 5. using the response characteristic slice to determine the distribution range of the thin sand body on a plane. The method can accurately and efficiently predicts distribution and thickness of the thin sand body in a sand body.

The invention relates to a multi-unmanned aerial vehicle task assignment conflict resolution method under communication delay constraint. A method for setting a comparison threshold value and carryingout priority ranking on unmanned aerial vehicles predicting potential assignment conflicts with extremely low communication cost is designed for the problem that assignment conflicts may be caused bycommunication delay in a distributed structure by taking multi-unmanned aerial vehicle task assignment in a complex battlefield environment as a research background; and then communication in the formation is started, and an assignment scheme of each unmanned aerial vehicle is compared to determine a currently executable optimal task, so that the purpose of conflict resolution is achieved. The conflict resolution mechanism can accurately and efficiently predict and dispatch assignment conflicts, and has feasibility and reasonability. Due to the fact that dynamic threshold setting is adopted,the method has good applicability to different battle environments.

Methods and systems of active discovery and collaborative collision avoidance are described for aircrafts. Specifically, an aircraft predicts firstly the four-dimensional information of its short future flight path based on existing information, then shares the information actively among its adjacent aircrafts, realizes finally collaborative collision avoidance based on the future flight paths of adjacent aircrafts. The described methods contain the following steps: prediction and broadcast of short future flight path, broadcast reception and analysis and collision avoidance, and resumption of planned path and state. According to the methods, a system is also described to implement the above method and it contains seven corresponding modules. Since sharing the future information rather than the past information among aircrafts, the advantage of this intervention is strong real-time, low computation complexity, low cost, and more aircrafts accommodated in the same space.

The invention provides an overdue prediction model generation method and terminal equipment suitable for the technical field of data processing. The method comprises that user attribute data is screened via a learning model, and N types of user attribute data with highest relational degree of an overdue type label are obtained; users are grouped at random to obtain training sample matrixes in which the user number is at random, and the training sample matrix are trained to obtain sub prediction models respectively. and a voting coefficient of each sub prediction model is calculated on the basis of the number of included users, and a needed overdue prediction model is constructed. Thus, overdue of users can be predicted effectively and accurately.

The invention provides a production energy consumption prediction method and system based on a neural network, an electronic terminal and a storage medium, and the method comprises the steps of determining one or more production feature parameters related to a production technology, and selecting a production data complete set in a preset time period based on the generation feature parameters; screening out a production data subset related to production energy consumption from the production data complete set; normalizing the production data subset to generate model training data; and establishing a neural network prediction model based on the model training data, wherein the neural network prediction model is used for predicting production energy consumption data. According to the invention, the technical problems of vacancy, low precision, single influence factor and the like of the energy consumption prediction technology in the enterprise production process in the prior art can besolved, so that users are helped to accurately and efficiently predict energy consumption data, and the requirements of energy conservation and efficiency improvement of enterprises are met.

The invention discloses a reservoir feature prediction method based on deep learning. The method comprises the following steps: acquiring a logging data training set; constructing a convolutional neural network and a forward-backward long-short term memory neural network added with an attention layer, and performing sequence modeling by combining the convolutional neural network and the forward-backward long-short term memory neural network to generate a multimode Bi-LSTM model; inputting the training set into a multi-mode Bi-LSTM model, training the multi-mode Bi-LSTM model by adopting a joint training mode, and performing joint optimization on parameters of a convolutional neural network and a forward-backward long-short term memory neural network added with an attention layer; and inputting actual logging data into the trained multimode Bi-LSTM model, and performing prediction through the model to obtain a prediction result of reservoir characteristics. According to the method, the advantage that bidirectional long and short time memory can efficiently and accurately perform time sequence prediction is utilized, and the attention mechanism layer is added, so that the defect that the convolutional neural network processes data with sequence correlation is made up, and reservoir characteristics such as porosity and permeability of reservoirs with different depths can be accurately predicted.

The invention relates to an image processing method and device, a computer readable storage medium and a computer device. The method comprises the steps of obtaining a to-be-processed image and a faceimage included in the to-be-processed image; carrying out clustering processing on the face images, and clustering the face images meeting the face similarity condition into the same face category; determining a user attribute label corresponding to each face category according to the face image included in each face category; determining human face co-occurrence features among different human face categories according to the frequency of occurrence of the human face images belonging to different human face categories in the same to-be-processed image; and inputting the user attribute tags respectively corresponding to the different face categories and the corresponding face co-occurrence features into a pre-trained user relationship decision model to obtain a user relationship graph. According to the scheme provided by the invention, the mining efficiency of the user relationship graph can be improved.

The invention discloses an optical interconnect module coupling efficiency prediction method based on a neutral network with a momentum. The method comprises the steps that a finite element model of an optical interconnect module is established, the temperature-vibration load is applied, the influence factors which influence the optical interconnect module coupling efficiency are analyzed based on a single factor method, orthogonal experiments are conducted on the main factors which influence optical coupling so as to establish a plurality of experimental groups different in level, the experimental groups are subjected to simulation experiments, the optical interconnect module coupling efficiency is obtained, the neural network is trained by using the obtained multi-combination data as training samples, and the optical coupling efficiency can be accurately predicted by using the trained network. By means of the optical interconnect module coupling efficiency prediction method, the defects that a standard BP neutral network is slow in learning convergence, likely to become the local minimum and the like are avoided, it is achieved that the optical interconnect module coupling efficiency is efficiently and accurately predicted under the temperature-vibration combined load, and a scientific, effective and rapid means is provided for designing and manufacturing the high-coupling-efficiency and high-speed optical interconnect module in the practical engineering application.

A circuit health state prediction method and system based on an integrated deep neural network are provided and relates to a technique for predicting a power electronic circuit failure. The invention serves to identify and diagnose a health state of a simulation circuit based on historical data by using an integrated deep neural network, and the method includes: carrying out parameter aging simulation experiments for different devices; extracting a series of time domain features of output signals through a temporal transformation method, and establishing health indices of the devices based on an improved angular similarity; predicting a health state of the simulation circuit in degeneration by using CAE and LSTM-RNN; and predicting validity of the circuit health state prediction method by referring to relevant evaluation indices. The invention is capable of effectively predicting the health state of the simulation circuit and is highly accurate and easy to implement.

The invention discloses a regional sea surface temperature prediction method based on CNN-LSTM, and relates to the field of physical ocean, computer graphic image processing and deep learning. The ,method comprises three steps of training sample establishment, model construction and model algorithm adjustment: firstly, carrying out segmentation processing on regional sea surface temperature data by adopting a set-aside method, and setting a predicted time window; training the sea surface temperature training sample and establishing a sea surface temperature prediction model by adopting an algorithm based on combination of a convolutional neural network CNN and a long short-term memory neural network LSTM; and finally, adjusting and training parameters of the model by adopting a trial-and-error method according to the error of the model, and determining parameters of the prediction model, thereby realizing efficient prediction of the regional sea surface temperature. Practice proves that the method can extract the spatial features of the sea surface temperature through the CNN, and then extract the time sequence features through the LSTM, thereby improving the prediction precision and efficiency of the sea surface temperature, and expanding the application of the deep learning method in regional sea surface temperature prediction.

The invention discloses an error prediction and real-time compensation technology for a five-degree-of-freedom hybrid robot based on a neural network. The technology comprises the following steps of (1) pose error decomposition on the hybrid robot, (2) pose error measurement and prediction of a parallel mechanism, (3) series swivel error prediction, (4) joint error compensation of the hybrid robot, (5) compensation neural network training and (6) constructing of a joint error compensator. According to the technology, prediction and compensation are implemented by directly utilizing error measurement data of the robot, the robot tail end pose errors caused by robot geometric errors and non-geometric factors such as gaps, friction, temperature and gravity can be compensated at the same time, the technology belongs to a comprehensive error compensation method, and the compensation effect is obviously superior to that of a traditional error compensation method based on geometric error identification.

The invention provides an early warning method and system based on a holiday and festival flow prediction algorithm, and the method comprises the steps: firstly obtaining a traffic data set of a whole year and a whole time period, carrying out the preprocessing, fusing the surrounding environment data and weather data of a highway section with the preprocessed traffic data set, and forming a complete traffic data set; the method comprises the following steps: selecting actual traffic data, establishing an ST-GCN model for training by selecting a preset data granularity and adopting a holiday flow prediction algorithm to obtain an optimal ST-GCN model to predict a plurality of target results, establishing an evaluation index according to the actual traffic data and the plurality of predicted target results, evaluating the ST-GCN model by adopting a comparison model, verifying the ST-GCN model as the optimal model, and determining the actual traffic data according to the evaluation index. According to a plurality of target results predicted by the ST-GCN model which is verified to be the optimal model, in combination with the grade division standard of the expressway operation state, the road condition information is marked with different color information, and a picture is generated for display, so that the passing efficiency of the expressway is effectively improved, and the traffic flow trend during holidays and festivals is efficiently and accurately predicted.

The invention relates to a CNN-based time sequence prediction method and a model determination method. The CNN-based time sequence prediction method comprises the steps: acquiring historical time sequence data, determining periodic parameters according to periodic characteristics of the historical time sequence data, wherein the periodic parameters comprise periodic types and periodic durations corresponding to the periodic types; determining component data corresponding to the prediction time point in the historical time sequence data based on the prediction time point, the historical time sequence data, the period parameters and a preset cycle span, the component data comprising the data of the closest time period and the period data; and predicting the component data by adopting the determined CNN model to obtain a prediction result corresponding to the prediction time point. By adopting the CNN-based time sequence method, the subsequent time sequence information can be efficientlyand accurately predicted.

The invention relates to an engine combustion noise optimization prediction method and device and a storage medium, and the method comprises the steps: carrying out the combustion process simulation through constructing a finite element model of engine combustion, and obtaining an original sample space; performing a Latin hypercube sampling test in the original sample space, obtaining initial sample point data through sample generation and correlation control, and performing weight calculation according to a fuzzy analytic hierarchy process to obtain target sample point data; and generating a noise prediction approximation model according to the target sample point data and a sequence iteration response surface method, determining that a value of a multiple correlation coefficient of the approximation model is not less than a preset coefficient value, fitting a response relationship between the combustion parameters and the combustion noise, determining an optimal combustion parameter combination, and predicting the optimal combustion noise according to the optimal combustion parameter combination. According to the method, through reasonable sampling and weight distribution, combustion noise prediction with relatively high efficiency can be realized only by a small number of samples, and reference is provided for reasonable optimization.

A noise distribution prediction method for a wind power plant with a complex terrain mainly comprises the following steps: firstly, generating a wind turbine aerodynamic noise source on the complex terrain considering the influence of wake flow based on an engineering wake flow model; secondly, proposing a boundary ray grid method to improve the whole noise distribution prediction calculation efficiency; and finally, solving through a PE parabola noise propagation equation of the complex terrain, carrying out noise power logarithm superposition, and obtaining the noise distribution condition of the wind power plant of the complex terrain. The boundary ray grid method is innovatively provided, on the basis that the accuracy is good, the problem that the calculated amount of an existing noise prediction method on a complex terrain is large is solved, and the noise prediction precision and the calculation efficiency of the wind power plant are improved. Due to the large scale of the windpower plant and the trend of developing to complex terrains, the aerodynamic noise influence of the wind power plant becomes more and more serious, and the method has important application prospects for wind power plant noise prediction.

The invention discloses a feature extraction method based on airline ticket booking behavior data, which comprises the following concrete steps: 1, retrieving original data from airline database and carrying out data preprocessing; 2, analyzing that preprocessed data by a feature extraction system to obtain the discrimination information of a normal user and an abnormal user; Step 3, extracting features based on the discrimination information of the normal user and the abnormal user, wherein the features include temporal features, quantitative features and location features. The invention extracts features from time series, location information and IP information, and predicts and intercepts malicious orders with high efficiency and accuracy, so as to obtain higher prediction accuracy.

The invention discloses a method for improving the prediction precision of a Gaussian wake flow model on a wind turbine wake flow field. The method comprises the following steps: step 1) obtaining a thrust coefficient of a wind turbine; 2) obtaining a proportionality coefficient; 3) obtaining an initial standard deviation coefficient; 4) acquiring standard deviation; step 5) combining a Gaussian wake flow model to obtain the speed loss of a wake flow area; the invention aims to improve the precision of the obtained initial standard deviation coefficient by improving the obtaining means of the initial standard deviation coefficient from the aspect of improving the universality of the initial standard deviation coefficient, and solves the problems that the universality of the initial standard deviation coefficient obtained by adopting a conventional method is insufficient, therefore, the technical problem that the Gaussian wake flow model has great influence on the prediction precision of the wake flow area speed in different environments is solved.

The invention discloses a vibration prediction method of a manipulator based on GA-ELM. The method comprises the steps: modeling is performed, importing the model into ANSYS software is performed to analyze the transient dynamic characteristics, the maximum deformation of vibration under different input conditions is obtained, The samples of training set are obtained, normalized, GA-optimized ELMmodel is used to find the chromosomes with the minimum fitness value, as the optimal initial w and b, the optimal individual is assigned to the ELM model, the model is established, the samples of testset are compared, and the model precision is evaluated. The invention combines genetic algorithm (GA) and GA of extreme learning machine (ELM) through ANSYS simulation analysis. ELM model is used topredict the vibration of manipulator under different influence factors, which provides a theoretical basis for the optimization of manipulator design.

The present invention discloses a method and system for improving synthetic voice rhythm naturalness. The method comprises a step of receiving a text to be synthesized, a step of determining the basic synthesis unit sequence corresponding to the text, wherein, the basic synthesis unit sequence comprises one or more basic synthesis units, a step of determining whether each basic synthesis unit is weak reading or not, a step of obtaining the synthesis parameter model corresponding to the basic synthesis unit, carrying out weak reading processing on the synthesis parameter model corresponding to the basic synthesis unit if the basic synthesis unit is weak reading, and obtaining an updated synthesis parameter model, a step of generating the synthesis parameter model sequence corresponding to the basic synthesis unit sequence, and a step of generating continuous voice according to the synthesis parameter model sequence. By using the method and the system, the naturalness of continuous synthetic voice can be simply and effectively improved.

The invention relates to a user feature optimization method and device in a user feature group, a medium and electronic equipment, and belongs to the technical field of machine learning application, and the method comprises the steps: initializing and generating a plurality of user feature combinations; inputting each user feature combination and the optimization target into a first machine learning model to obtain an evaluation score; obtaining a plurality of user feature combinations in a first score range and a plurality of user feature combinations in a second score range; obtaining a userfeature difference between the user feature combination of each first score range and the user feature combination of the second score range; inputting the plurality of user feature differences and the optimization target into a second machine learning model to obtain a prediction optimization user feature combination; and obtaining a target optimization user feature combination. According to theinvention, the user feature combination corresponding to the optimization target is predicted through the machine learning model according to the distinguishing features of part of the user feature combinations, so that the high efficiency and accuracy of obtaining the target user feature combination are ensured.

The invention discloses a human hand trajectory prediction and intention recognition method based on multi-feature fusion. The method comprises the following steps: acquiring face and shoulder key point data of a person; acquiring palm track data; inputting the face and shoulder key point data of the person into a support vector machine to obtain face orientation modal information; inputting the palm trajectory data sequence into an SG filter, and eliminating trajectory data fluctuation to obtain smooth trajectory data; carrying out parallel fusion on the two types of modal information to obtain multi-modal fusion information; and inputting into the LSTM network, and outputting a predicted trajectory of the palm. According to the method, the face orientation features are extracted by using part of the face key points, the face orientation features are further fused with the human arm trajectory data, the moving trajectory of the human arm in the space and the final arrival position of the human arm are predicted, and the moving trajectory is efficiently and accurately predicted.

The invention relates to an energy consumption equation construction method and apparatus, a computer device and a storage medium. The method comprises the steps of obtaining an energy consumption variable, the energy consumption prediction target precision and energy consumption prediction sample data in an application scene; constructing an initial energy consumption equation according to the energy consumption variable and the energy consumption prediction target precision; adjusting equation coefficients of the initial energy consumption equations by adopting a genetic algorithm, and obtaining equation coefficients corresponding to the initial energy consumption equations when respective equation adaptive values are minimum according to the energy consumption prediction sample data; and substituting the obtained equation coefficient into the initial energy consumption equation to obtain a target energy consumption equation with the minimum equation adaptive value. In the whole process, the target energy consumption equation is accurately constructed by adopting the genetic algorithm based on the variable corresponding to the current application scene, and the energy consumptioncan be simply and accurately predicted directly through the accurately constructed target energy consumption equation in the subsequent energy consumption prediction process. In addition, the invention further provides an energy consumption prediction method and device, computer equipment and a storage medium.

The invention discloses a coupling reaction yield intelligent prediction method based on an attention convolutional neural network. The method comprises the steps of data acquisition, model construction and yield intelligent prediction. The method comprises the following specific implementation steps: 1) calculating and extracting feature descriptors of compounds by utilizing chemical related software, and performing subsequent processing by taking the feature descriptors as original data of a training set and a test set; 2) importing the feature descriptor data into a convolutional neural network, and fusing an attention mechanism into a convolutional neural network model; 3) training the acquired data by using a built attention convolutional neural network model, and storing the model when the value of a loss function MSE of the model is minimum; 4) enabling a user to adjust model parameters by himself / herself to achieve the optimal prediction effect; and (5) loading the trained model, and carrying out intelligent prediction on test data. According to the coupling reaction yield intelligent prediction method, a chemical owner can be assisted to rapidly predict the coupling reaction yield, and the chemical synthesis process is greatly accelerated.

The invention provides a psychological state prediction method and device and an electronic device, and relates to the technical field of online education. The method comprises the steps of obtainingthe interaction event information of a learner and an online education system; generating a learning interaction situation according to the timestamp corresponding to the interaction event information; inputting the learning interaction situation into a pre-trained deep learning network model; and predicting the psychological state corresponding to the learning interaction situation, so that the relationship between the interaction behavior of the learner and the psychological state of the learner can be well embodied through the method, and accordingly the psychological state prediction of the learner becomes more accurate and efficient, and an instructor is helped to better understand the psychological state of the learner.

A method for predicting the yield of manufacturing semiconductor devices includes steps of: acquiring defect data of semiconductor devices to be predicted, wherein the semiconductor devices to be predicted include finished semiconductor devices and semi-finished semiconductor devices, and the defect data indicates a defect type and location of at least one defect of the semiconductor devices; inputting the defect data into a pre-trained yield prediction model, wherein the yield prediction model includes a neural network structure and a classification structure, the neural network structure is used to extract defect feature vectors from the defect data, and the classification structure is used to output classification results of qualified or unqualified yield according to the defect feature vectors; and determining, by the yield prediction model, classification results of qualified or unqualified yield of the semiconductor devices.