652 results about "Single model" patented technology

An article is disclosed which provides for powering or charging the battery of a variety of different electronic devices, with the article being in two parts. The first part has components such that it is "universal" for use with all of the variety of different devices, and the second part has components which are specific to the operation of only a single model, type or brand of such device or of closely related devices which operate with the same voltage and plug-compatibility requirements of the specific device. In particular, the article with both parts (modules) present will operate a specific model, type or brand of device. However one can disconnect the two parts from each other and connect a different second part for a different device to the same first part, thereby being able to operate two different devices with substitution of only one part (the second) rather than requiring substitution of the entire article when a difference device must be powered or its battery charged. The article is particularly useful for powering or charging the batteries of devices such as cellular telephones, computers and the like from house, office or other building electrical systems or from vehicle electrical systems.

A model construction module (MCM) for constructing a probabilistic model (PM) of a wireless environment (RN) in which a target device (T) communicates using signals that have a measurable signal value (x), such as signal strength. The model construction module forms several submodels (611-631) of the wireless environment (RN). Each submodel indicates a probability distribution (F1-F3) for signal values at one or more locations (Q1-QY) in the wireless environment. The module combines the submodels to a probabilistic model (PM) of the wireless environment (RN), such that the probabilistic model indicates a probability distribution for signal values at several locations in the wireless environment. Alternatively, the model may insert new locations to a single model based on a combination of existing locations. The combination of submodels or existing locations comprises combining the inverse cumulative distribution functions of the submodels or existing locations.

A system and method are provided for processing audio and speech signals using a pitch and voicing dependent spectral estimation algorithm (voicing algorithm) to accurately represent voiced speech, unvoiced speech, and mixed speech in the presence of background noise, and background noise with a single model. The present invention also modifies the synthesis model based on an estimate of the current input signal to improve the perceptual quality of the speech and background noise under a variety of input conditions. The present invention also improves the voicing dependent spectral estimation algorithm robustness by introducing the use of a Multi-Layer Neural Network in the estimation process. The voicing dependent spectral estimation algorithm provides an accurate and robust estimate of the voicing probability under a variety of background noise conditions. This is essential to providing high quality intelligible speech in the presence of background noise.

The invention discloses a method and a device for building a classification forecasting mixed model. The method includes: dividing a sample data set into data sets of different types according to data characteristics, performing data cleaning for the data sets and performing variable selection after data cleaning is finished to generate variable sets of different types, and adopting at least one classification forecasting single model for each variable set to build the classification forecasting mixed model. Through the method and the device, the classification forecasting mixed model is respectively built after data subdivision, and accuracy of classification forecasting is improved.

The invention discloses a fast rendering method and device for a virtual scene and a model. The method comprises steps of firstly obtaining the rendering request of the virtual scene and the model to be rendered and the standard material library; creating a read and write file including the corresponding relation among the scene parameter, the model and the material according to the rendering request, and selecting a material corresponding to the model to be rendered from the pre-established standard material library after loading; setting and adjusting the scene parameter according to the rendering request, rendering the model according to the selected material and adjusting the material parameter to complete the rendering satisfying the specified rendering request; when a single model corresponds to a plurality of materials, directly subjecting the plurality of materials to a three-dimensional presence UV mapping on the surface of the three-dimensional model; and generating a high light effect and its corresponding material by the hand-drawn trajectory according to the rendering request, and rendering the model's high light effect. The method has the advantages of high efficiency, openness, automation, integration of tools, light weight, sustainability, and the like on the basis of greatly improving the efficiency of virtual scene and model rendering.

The invention discloses a reading understanding model method based on paragraph internal reasoning and joint question answer matching, and the method comprises the following steps: S1, constructing avector for each candidate answer, the vector representing the interaction of a paragraph with a question and an answer, and then enabling the vectors of all candidate answers to be used for selectinganswers; S2, carrying out experiment. According to the model provided by the invention, paragraphs are firstly segmented into blocks under multiple granularities; an encoder is used for summing the intra-block word embedding vectors by utilizing neural word bag expression; then, a relationship between blocks with different granularities where each word is located through a two-layer forward neuralnetwork is modeled to construct a gating function, so that the model has greater context information and captures paragraph internal reasoning at the same time. Compared with a baseline neural network model such as a Stanford AR model and a GA Reader, the accuracy of the model is improved by 9-10%. Compared with a recent model SurfaceLR, the accurcay is at least improved by 3% and is about 1% higher than that of a single model of the TriAN, and in addition, the model effect can also be improved through pre-training on an RACE data set.

The invention discloses a multi-model fused short text classification method. The multi-model fused short text classification method comprises a learning method and a classification method. The learning method comprises the following steps: carrying out word segmentation and filtration on short text training data to obtain a word set; calculating the IDF value of each word; calculating the TFIDF values of all the words and constructing a text vector VSM; and carrying out text learning on the basis of a vector space model, and constructing an ontology tree model, a keyword overlapping model, a naive Bayesian model and a support vector machine model. The classification method comprises the following steps: carrying out word segmentation and filtration on a to-be-classified short text; generating a text vector on the basis of the support vector machine model; respectively classifying by using the ontology tree model, the keyword overlapping model, the naive Bayesian model and the support vector machine model to obtain single model classification results; and fusing the single model classification results to obtain a final classification result. According to the method disclosed in the invention, multiple classification modes are fused and the short text classification correctness is improved.

The invention discloses a junk mail filtering method based on support vector machine (SVM). The steps are as following: 1) analyze the mail and extract the message relevant to title, text and character set; 2) carry out divided syncopation to the extracted text message content; 3) make statistics of word frequency in mail and utilize TF-IDF formula to map the mail text to vector; 4) utilize LibSVM to train the mail sample and obtain support vector machine model; 5) utilize support vector machine model to classify new mail and obtain the probability value of junk mails; 6) utilize threshold value adjustment to guarantee a lower level of false positive rate of normal mails to junk mails and ultimately judge whether mails are junk mails. The invention utilizes the advantage of highest single model classification accuracy of the support vector machine, improves the correctness of junk mail filtering, according to the text feature and activity feature and at the same time, also effectively solves the problem of unequal miscarriage cost in junk mail filtering.

The invention discloses a predictive control method and system based on a multi-model generalized predictive controller. While process disturbance is inhibited, a preset desired output value is enabled to track the optimum set value track, and dynamic characteristics of a system are distinguished in parallel by adopting a plurality of fixed models and a plurality of adaptive models so that an actual output value and an optimum input control quantity of the system can be obtained. The invention also provides a predictive control system which is of a DRTO (Dynamic Real-Time Optimization) dual-layer structure, and adopts a multi-mode generalized predictive controller to replace an existing single-model generalized predictive controller. The predictive control method has the following beneficial effects of well matching the actual process characteristic in the production, reducing the system cost consumption, increasing the system economic benefit, improving the system transient state performance and the system regulating capacity when a system model parameter hops, being capable of effectively eliminating the interference of disturbance to system output, and reducing the influence of inconsistency of models of an optimization layer and a control layer in the DRTO dual-layer structure to the economic benefit.

The invention relates to an Internet porn image detection method based on a deep convolution nerve network. The method comprises the following steps of acquiring a porn image and a normal image through a manual calibration method and carrying out pretreatment and enhancement on the images so as to acquire an effective square training image; sending the acquired effective image into a deep convolution nerve network so as to train the network; verifying a network model on a verification set, adjusting a training set according to a result and continuously training the deep convolution nerve network; repeating the last step till that detection accuracy on the verification set reach an expected object or a network loss function begins convergent; testing the trained network on the training set. The method in the invention has the following advantages that the porn image detection method based on the deep convolution nerve network is provided and the method can be used to rapidly detect almost all the types of porn images through a single model; in an actual test, detection accuracy in the invention reaches above 98.6%.

The invention discloses a traffic state estimation device and method based on data fusion. The traffic state estimation device comprises a pattern classifying unit, a model selecting unit and a data fusing unit, wherein the pattern classifying unit is used for classifying traffic data of one or more roads into one of a plurality of preassigned patterns; the model selecting unit is used for selecting a corresponding neural network model from a plurality of neural network models according to the classified patterns; and the data fusing unit is used for inputting traffic data into the selected neural network model for carrying out data fusion and estimating a traffic condition of one or more roads. According to the invention, a multi-pattern classifying and multi-pattern modeling mechanism is introduced, aiming at each road, different patterns are respectively independently modeled, thus the precision of the single model can be improved, and different data source types of the traffic data, different characteristics and different factors of influencing the traffic conditions are fully utilized to obtain a more realistic and accurate fusion result and traffic condition estimation.

The invention discloses a random convolutional neural network-based high-resolution image scene classification method. The method comprises the steps of performing data mean removal, and obtaining a to-be-classified image set and a training image set; randomly initializing a parameter library of model sharing; calculating negative gradient directions of the to-be-classified image set and the training image set; training a basic convolutional neural network model, and training a weight of the basic convolutional neural network model; predicting an updating function, and obtaining an addition model; and when an iteration reaches a maximum training frequency, identifying the to-be-classified image set by utilizing the addition model. According to the method, features are hierarchically learned by using a deep convolutional network, and model aggregation learning is carried out by utilizing a gradient upgrading method, so that the problem that a single model easily falls into a local optimal solution is solved and the network generalization capability is improved; and in a model training process, a random parameter sharing mechanism is added, so that the model training efficiency is improved, the features can be hierarchically learned with reasonable time cost, and the learned features have better robustness in scene identification.

The invention discloses a regional convolutional neural network-based method for gesture identification and interaction under egocentric vision. The method comprises the following steps of S1, obtaining training data; S2, designing a regional neural network which is used for gesture classification and fingertip detection while being used for hand detection, and ensuring that an input of the neural network is a three-channel RGB image and outputs of the neural network are top left corner coordinates and top right corner coordinates of an external connection matrix of a gesture region, gesture types and gesture skeleton key points; and S3, judging the gesture types, and outputting corresponding interactive results according to different interactive demands. The invention provides a complete method for gesture identification and interaction under egocentric vision. Through single model training and partial network sharing, the identification speed and accuracy of the gesture identification under the egocentric vision are increased and improved.

A method for constructing a model representative of a heterogeneous medium from data expressed in different time scales with application to hydrocarbon characterization. The method first estimates sequentially, from the data expressed in each different time scale, parameters or physical quantities of the model, described for each different time scale. Second, a scale factor allowing conversion of a model described in a time scale into a model described in another time scale is determined. This determination is carried out by minimizing the dissimilarity between a parameter estimated in a time scale and this parameter estimated in the other time scale. Finally, estimation of a single model is performed, by taking simultaneously an account of the data expressed in the various time scales, and using the scale factor found previously for the time scale conversions.

The invention discloses a machine learning anti-fraud monitoring system based on transaction data. The system comprises a management platform, an ETL module, a sampling engine, a stream processing engine, a training engine, a prediction engine and a decision engine. The stream processing engine rapidly extracts and calculates characteristics of the huge original transaction data through streamed big data processing, the representative characteristics are obtained from the huge original transaction data, and information in the data is sufficiently extracted. In the model training module, various machine learning models and ensemble learning frameworks optimized for the capital loss ratio and the black sample recall ratio are used, and a composite model optimized for an indicator is obtained. The over-fitting and unstable defects due to a single model are overcome, and the stability and the generalization ability of the model are improved; according to a preset update time, the model training module automatically obtain the latest data and trains the model again, accordingly the model keeps the effectiveness all the time, and the problem of model inefficiency due to fraud variation is avoided.

The invention provides a Chinese named entity recognition model and method based on double neural network fusion, which belong to the field of named entity recognition. The problem that an existing single model often has insufficient feature representation is solved. The model comprises a Bert embedding layer used for converting a sentence from a character sequence to a dense vector sequence, anda Bi _ LSTM layer with a self-attention mechanism, wherein the Bi _ LSTM layer learns the implicit representation of the words from the context in the whole process, processes sentence layer information, and obtains the preceding and following text information with long-distance dependence characteristics; the model further comprises a stacking a DCNN layer which combines wider context informationinto a mark for representation, extracts local information of characters, and obtains preceding and following text information with wide local features, and a CRF decoding layer which is used for decoding dual-model output into sequence marks and explicitly outputting named entities through labels marked by the sequence marks. The model has the effect of enhancing the capability of implicitly acquiring context representation among character sequences of the model.

The invention relates to a binocular depth estimation method based on a depth neural network, which comprises the following steps: 1) preprocessing the input left and right viewpoint images to enhancedata; 2) constructing a multi-scale network model for binocular depth estimation, wherein the model comprises a plurality of convolution layers, an activation layer, a residual connection, a multi-scale pooling connection and a linear upsampling layer; 3) Designing the loss function to minimize the results in the continuous training process, so as to obtain the optimal network weights; 4) inputting the image to be processed into the network model to obtain the corresponding depth map, and repeatedly repeating the above steps until the network converges or reaches the training times. The invention adopts the idea of unsupervised learning, and only the left and right viewpoint images obtained by the binocular camera are used as network input. The adaptive design of the network sets the internal and external parameters of the camera as a single model parameter, so it can be applied to multiple camera systems without modifying the network.

The invention discloses a multi-model integrated flood forecasting method based on propagation time clustering analysis and belongs to the technical field of hydrologic forecasting. The method includes adopting a derived dynamic time warp matching method for floor process similarity analysis; estimating flow propagation time of each station in the upstream and the downstream; decomposing a sample into a plurality of clusters by performing clustering analysis on the flow propagation time; respectively building an SVM regression model for each sub-flow sequence to simulate the flood forming process; combining the sub-models into a comprehensive model. Comparing a comprehensive forecasting result acquired by the method with a model forecasting result acquired by a single model under conventional conditions and on the basis of flow clustering, the comparison result shows that the comprehensive model is better in comprehensive performance.

Visual map items may each be constructed and placed in position using logic defined by a map view component corresponding to each visual item, where that logic may depend on one or more values populated into parameter(s) of the map view component. Some of those parameter values may correspond to known map model parameter values. Others, however, may have been solved for using a model that defines analytical relationships between the map model parameters. In one embodiment, which of the map model parameters are input variable, and which are output model variables, may not be predetermined. Accordingly, a solver might be prepared for multiple solve operation paths even using a single model. The map view composition process may be entirely data-driven, and may include a mechanism for canonicalizing input data, and binding canonicalized input data to the model parameters.

The invention discloses a target area detection method based on deep learning, and belongs to the technical field of computer vision. The method mainly adopts a retinanet detection network. The RetinaNet is essentially a network structure composed of two FCN sub-networks of resnet + FPN +. ResNeXt50 and densenet 169 are respectively adopted by the backbone to replace the previous resnet. An FPN layer and a loss function of the retnanet network are modified, and finally model fusion is carried out. The target detection method combines the advantages of a current mainstream target detection method and already solves a series of practical problems. According to the algorithm, an experiment is carried out under the cococo2017, and the performance is very good. And the method is better than a single model under retinanet and a result obtained when the model is not improved. In addition, the method has good performance on other data sets.

The invention relates to a Mask-RCNN-based cervical cell smear image segmentation method and system, and the method comprises a data set construction step of carrying out the preparation and labelingof a training data set, a verification data set and a test data set, and carrying out the normalization and preprocessing of the data set; b, constructing and training a model, constructing an image segmentation model based on Mask-RCNN, training the model by using the training data set, and verifying an image segmentation result of the model by using the verification data set; and a model verification step of testing the model by using the test data set, and evaluating a segmentation result by using a similarity coefficient. According to the method, the deep neural network model trained by utilizing a large amount of data can be used for modeling and abstracting information contained in the large amount of data, so that the cells and the cell nuclei in the cervical cytology smear image can be positioned, detected and subjected to the instance segmentation through a single model.

The invention relates to a method for forecasting short-term wind speed of a wind farm based on hybrid neural network. The method comprises the following steps: S1, determining an input variable and an output variable of a hybrid neutral network forecasting model according to a preset forecasting time interval; and S2, forecasting the wind speed according to the hybrid neutral network forecasting model to obtain corresponding wind speed forecasting value. The invention also relates to a system for forecasting short-term wind speed of the wind farm based on the hybrid neural network. The system comprises a variable determination module for determining the input variable and output variable of the hybrid neutral network forecasting model according to the preset forecasting time interval; and a forecasting module for forecasting the wind speed according to the hybrid neutral network forecasting model to obtain the corresponding wind speed forecasting value. The method and the system provided by the invention have advantages of high computation speed and high reliability, solve the technical problem completely depending on a physical forecasting model and overcome the disadvantage of large forecasting error fluctuation based on a single model.

Visual items may each be constructed and placed in position using logic defined by a view component corresponding to each visual item, where that logic may depend on one or more values populated into parameter(s) of the view component. Some of those parameter values may correspond to known model parameter values. Others, however, may have been solved for using a model that defines analytical relationships between the model parameters. In one embodiment, which of the model parameters are known, and which are unknown, may not be predetermined. Accordingly, a solver might be prepared for multiple solve operation paths even using a single model. The view composition process may be entirely data-driven, with the solve and / or the visual items implemented using spreadsheets.

A method for predicting air quality with the aid of machine learning models includes: (A) providing air pollution data to perform an eXtreme Gradient Boosting (XGBoost) regression algorithm for obtaining a XGBoost prediction value; (B) providing the air pollution data to perform a Long Short-Term Memory (LSTM) algorithm for obtaining an LSTM prediction value; (C) combining the air pollution data, the XGBoost prediction value and the LSTM prediction value to generate air pollution combination data; (D) performing an XGBoost classification algorithm to obtain a suggestion for whether to issue an air pollution alert; and (E) performing the XGBoost regression algorithm on the air pollution combination data to obtain an air pollution prediction value. Two layers of machine learning models are built, and a situation where prediction results are too conservative when a single model does not have enough data can be improved.

Visual items may each be constructed and placed in position using logic defined by a view component corresponding to each visual item, where that logic may depend on one or more values populated into parameter(s) of the view component. Some of those parameter values may correspond to known model parameter values. Others, however, may have been solved for using a model that defines analytical relationships between the model parameters. In one embodiment, which of the model parameters are known, and which are unknown, may not be predetermined. Accordingly, a solver might be prepared for multiple solve operation paths even using a single model. The view composition process may be entirely data-driven, and may include a mechanism for canonicalizing input data, and binding canonicalized input data to the model parameters. The view composition framework may operate the same regardless of the domain.

The invention relates to a screening method for near infrared spectrum wavelength and Raman spectrum wavelength. The method comprises that acquired near infrared spectrum or Raman spectrum and data of concentration of corresponding composition to be detected are divided into a training set, a check set and a prediction set; a PLS model is established by using original spectrum and the concentration of the composition to be detected to obtain a real PLS model coefficient; the concentration of the composition to be detected is ordered randomly and a great number of PLS models are established by using the vectors of the concentration of the composition to be detected and original spectrum matrices; according to the models, the times that a single model coefficient is larger than the real PLS model coefficient are respectively calculated to obtain a corresponding probability value; the wavelength that the probability value is less than a threshold value is reserved; an optimum model is established by using the reserved wavelength to predict the concentration of the composition to be detected of a sample in the prediction set. By adopting the method, the invention has the advantages that the wavelength containing spectral information can be accurately extracted, the quantitative analysis model is simplified, the prediction accuracy of the quantitative analysis model is improved and the new wavelength screening technique is provided for the multivariate calibration analysis of the near infrared spectrum and the Raman spectrum.

Different subsets of multiarray resistivity measurements are inverted using a corresponding resistivity model. The models are selected based at least in part on the resolution of the measurements. Results of the inversion are combined to give a single model consistent with the resolution and depth of penetration of each subset of the multiarray data.