1084 results about "Hyper parameters" patented technology

The invention relates to a condition monitoring data stream anomaly detection method, in particular to a condition monitoring data stream anomaly detection method based on an improved gaussian process regression model. The problem that an existing method for processing monitoring data stream anomaly detection is poor in effect is solved. The method comprises the steps that firstly, the historical data sliding window size is determined; secondly, the types of a mean value function and a covariance function are determined; thirdly, the hyper-parameter initial value is set to be the random number from 0 to 1; fourthly, q data closest to the current time t are extracted; fifthly, the gaussian process regression model is determined; sixthly, prediction is conducted by means of the nature of the gaussian process regression model; seventhly, PI of normal data at the time t+1; eighthly, monitoring data are compared with the PI; ninthly, whether the real monitoring data need to be marked to be abnormal or not is judged; tenthly, beta (xt+1) corresponding to the monitoring value at the time t+1 is calculated; eleventhly, the real value or prediction value and the t+1 are added into DT; twelfthly, new DT is created. The condition monitoring data stream anomaly detection method based on the improved gaussian process regression model is applied in the field of network communication.

The invention discloses a pile-up noise reduction own coding network bearing fault diagnosis method based on particle swarm optimization. The bearing fault diagnosis method provides an improved pile-up noise reduction own coding network SADE bearing fault diagnosis method, SDAE network hyper-parameters, such as cyber hidden layer nodes, sparse parameters, input data random zero setting ratio, are selected adaptively by particle swarm optimization PSO, a SADE network structure is determined, top character representation of malfunction inputting a soft-max classifier is obtained and a classification of defects is discerned. The bearing fault diagnosis method has better feature in learning capacity and more robustness than feature of learning of ordinary sparse own coding device, and builds a SDAE diagnostic model having multi-hidden layer by optimizing the hyper-parameters of noise reduction own coding network deepness network structure with the particle swarm optimization, accuracy of the classification of defects is improved at last.

An improved system and method is provided for sparse Gaussian process regression using predictive measures. A Gaussian process regressor model may be construction by interleaving basis vector set selection and hyper-parameter optimization until the chosen predictive measure stabilizes. One of various LOO-CV based predictive measures may be used to find an optimal set of active basis vectors for building a sparse Gaussian process regression model by sequentially adding basis vectors selected using a chosen predictive measure. In a given iteration, a predictive measure is computed for each of the basis vectors in a candidate set of basis vectors and the basis vector with the best predictive measure is selected. The iterative addition of basis vectors may stop when predictive performance of the model degrades or no significant performance improvement is seen.

The invention relates to an urban road congestion degree prediction method based on time sequence traffic events. The method comprises the steps of: S1, acquiring historical traffic event data, real-time traffic event data and video monitoring data of an urban road section; S2, identifying traffic congestion forewarning events in the video data through the 3D CNN, and performing data space-time fusion according to the historical traffic events; S3, determining congestion degree classification labels, constructing a time sequence traffic congestion event data dictionary, and screening a training set, a verification set and a test set; S4, establishing an LSTM sequence data classification model, inputting the training set, and iteratively updating model parameters by utilizing a gradient descent method; S5, inputting the verification set into the model with updated parameters, optimizing and adjusting hyper-parameters, and selecting an optimal model; and S6, inputting the test set into the optimal training model, checking the effectiveness of the model, and carrying out road congestion prediction according to real-time traffic monitoring data. According to the method, a sequence dataclassification model is established by using LSTM, and the urban road congestion degree is predicted based on time sequence traffic events.

The invention relates to a reservoir prediction method based on a bidirectional recurrent neural network. The reservoir prediction method includes: generating well seismic data, generating well markeddata, generating a marked dataset, building and training the bidirectional recurrent neural network, storing a trained model, extracting seismic data corresponding to all geographic positions in a region, and performing prediction on the seismic data to acquire probability distribution of reservoirs below the whole region. A one-to-one corresponding relation between each sampling point and reservoirs and non-reservoirs is established, wherein input of each time step is n-dimension seismic data of each sampling point while output of the same is corresponding reservoir or non-reservoir marks. An optimal model is acquired by training marked samples and adjusting hyper-parameters, in this way, unmarked data can be predicted to finally generate a predicted seismic body of same dimension, and value of each sampling point is prediction probability between 0 and 1, so that the method has good effect on reservoir prediction.

The invention discloses a Gaussian process regression-based method for predicting state of health (SOH) of lithium batteries, relates to a method for predicting the SOH of the lithium batteries, belongs to the fields of electrochemistry and analytic chemistry and aims at the problem that the traditional lithium batteries are bad in health condition prediction adaptability. The method provided by the invention is realized according to the following steps of: I. drawing a relation curve of the SOH of a lithium battery and a charge-discharge period; II, selecting a covariance function according to a degenerated curve with a regeneration phenomenon and a constraint condition; III, carrying out iteration according to a conjugate gradient method, then determining the optimal value of a hyper-parameter and bringing initial value thereof into prior distribution; IV, obtaining posterior distribution according to the prior part; V, obtaining the mean value and variance of predicted output f' without Gaussian white noise; and VI, together bringing the practically predicted SOH of the battery and the predicted SOH obtained in the step V into training data y to obtain the f', then determining the prediction confidence interval and predicting the SOH of the lithium battery. The method provided by the invention is used for detecting lithium batteries.

The invention discloses a deep neural network multitask hyper-parameter optimization method. The method comprises: firstly, a data training set of each task being subjected to model training to obtaina multi-task learning network model; secondly, predicting all points in an unknown region, screening candidate points from a prediction result, finally evaluating the screened candidate points, adding the candidate points and target function values of the candidate points into the data training set, and establishing a model, predicting, screening and evaluating again; and so on, until the maximumnumber of iterations is reached, finally selecting a candidate point corresponding to the maximum target function value from the data training set, that is, the hyper-parameter combination of each task in the multi-task learning network model. According to the method, the Gaussian model is replaced by the radial basis function neural network model, and the radial basis function neural network model is combined with multi-task learning and is applied to the Bayesian optimization algorithm to realize hyper-parameter optimization, so that the calculation amount of hyper-parameter optimization isgreatly reduced. The invention further discloses an electronic device and a storage medium.

A data modeling platform includes a distributed modeling ensemble generator and a progress tracker. The distributed modeling ensemble generator preprocesses and models an input dataset according to a user listing of modeling types, modeling algorithms and preprocessing operations. The generator includes a plurality of model runners, one per modeling type, and a data coordinator. Each model runner operates with a changing plurality of distributed independent modeling services and generates a changing set of points in a hyper-parameter space defining hyper-parameters for the modeling algorithms and preprocessing operations. Each distributed modeling service uses a selected one of the hyper-parameter points and generates a validated score for that point. The data coordinator coordinates the operation of the model runners and provides the hyper-parameter points and their resulting scores to the progress tracker.

The invention provides a target detection method based on SSD improvement, and the method selects ResNet-101 to replace VGG-16 as a basic network of a whole model, and provides richer feature map information. A feature fusion strategy is adopted, so that the detection effect of multi-scale prediction of a network prediction layer on a small target is improved to a certain extent; an SE-block module is added to a classification branch and used for capturing global environment information of the feature map and outputting the feature map with channel weight, so that classification is more accurate; and a Centerness parallel to the classification prediction layer is added to suppress a low-quality bounding box so as to improve the detection precision. According to the method, anchor-frame-free detection is adopted, all hyper-parameters related to the anchor frame are avoided, the hyper-parameter amount is greatly reduced through the adopted prediction bounding box strategy, the network design complexity is reduced, and the training stage becomes very simple. According to the method, a loss function adopts a focalloss function, so that the model detection precision is improved while the detection speed is kept.

The use of machine learning for pattern recognition in magnetocardiography (MCG) that measures magnetic fields emitted by the electrophysiological activity of the heart is disclosed herein. Direct kernel methods are used to separate abnormal MCG heart patterns from normal ones. For unsupervised learning, Direct Kernel based Self-Organizing Maps are introduced. For supervised learning Direct Kernel Partial Least Squares and (Direct) Kernel Ridge Regression are used. These results are then compared with classical Support Vector Machines and Kernel Partial Least Squares. The hyper-parameters for these methods are tuned on a validation subset of the training data before testing. Also investigated is the most effective pre-processing, using local, vertical, horizontal and two-dimensional (global) Mahanalobis scaling, wavelet transforms, and variable selection by filtering. The results, similar for all three methods, were encouraging, exceeding the quality of classification achieved by the trained experts. Thus, a device and associated method for classifying cardiography data is disclosed, comprising applying a kernel transform to sensed data acquired from sensors sensing electromagnetic heart activity, resulting in transformed data, prior to classifying the transformed data using machine learning.

The invention discloses a deep-belief-network-based diagnosis method of a distribution network. The method comprises: original monitoring data of a distribution network are obtained; denoising and normalized modeling are carried out on the original data; hyper parameters of a distribution network fault diagnosis model are set; 70% of collection data are used as training samples to train the model;the rest of 30% of samples are used for testing the trained fault diagnosis model, wherein six outputted nodes are used for expressing probabilities of occurrences of two kinds of faults at three phases respectively and training is carried out again if the precision does not meet the requirement; the operation state of the distribution network is monitored by the model; and if a fault is caused,a fault type and a line are obtained. According to the invention, the deep learning theory is applied to the fault diagnosis of the distribution network; and changing characteristics of all collectiondata under various faults are learned automatically under the circumstances of complicated structure, many devices, and frequent data deficiency. Meanwhile, the good fault tolerance performance is realized; the fault diagnosis accuracy and timeliness of the distribution network are improved; and thus the distribution network can work stably and safely.

The invention discloses a weak supervision fine-grained image recognition method based on a visual self-attention mechanism. The method involves a student-model module, a teacher-model module and a classification-model module. The student-model and the teacher-model are combined through a Teacher-Student loop feedback mechanism based on Pairwise Approach sorting learning so as to form a self-attention region recommendation network, so that the relation between discriminative region positioning and fine-grained feature learning is enhanced, the discriminative region in the fine-grained image can still be accurately detected under the condition of lacking a target bounding box and a part marking point, and the recognition accuracy is promoted to be remarkably improved; meanwhile, a convolution layer is shared by the three modules, namely, the student-model, the teacher-model and the class-model, so that the model storage space is effectively compressed, the calculation cost is reduced, the method meets the real-time recognition task requirement, and the method is suitable for a large-scale real scene; and besides, a dynamic weight distribution mechanism is adopted in multi-task jointlearning to reduce the amount of artificially set hyper-parameters and enhance the robustness of the model, finally, the whole model is trained and learned in an end-to-end single-stage mode, and thenetwork optimization difficulty is reduced.

The invention discloses a monocular camera object pose estimation method based on deep learning. The method comprises the following steps of 1) generating a training set and a verification set according to the projection of an obtained three-dimensional image of an object in a two-dimensional space, object coordinates corresponding to the projection and a label file of the object; 2) learning thetraining set by using a cascade convolutional neural network model, and iterating hyper-parameters; and 3) testing the trained cascade convolutional neural network model by using the test set, and when the accuracy of the trained cascade convolutional neural network model is not less than a first preset threshold, carrying out object pose estimation by using the trained cascade convolutional neural network model. The invention discloses a monocular camera object pose estimation device based on deep learning. By applying the embodiment of the invention, the application cost of object pose estimation can be reduced.

Computationally efficient accelerated D2-clustering algorithms are disclosed for clustering discrete distributions under the Wasserstein distance with improved scalability. Three first-order methods include subgradient descent method with re-parametrization, alternating direction method of multipliers (ADMM), and a modified version of Bregman ADMM. The effects of the hyper-parameters on robustness, convergence, and speed of optimization are thoroughly examined. A parallel algorithm for the modified Bregman ADMM method is tested in a multi-core environment with adequate scaling efficiency subject to hundreds of CPUs, demonstrating the effectiveness of AD2-clustering.

The present invention discloses an estimation method of a quasi-stationary broadband array signal direction of arrival (DOA) based on block sparse Bayesian learning. An intra-frame correlation and aninterframe independence of a quasi-stationary broadband signal frequency spectrum are employed to set a corresponding a block sparse prior distribution model for signals, and a block sparse Bayesian model is employed to perform estimation of sparse signals, so that an estimation result with higher precision is obtained. Array receiving signals are subjected to appropriate framing processing, eachframe of the signal is subjected to Fourier transform, and each block sparse Bayesian model for each signal is established in a frequency domain; under an assumption of each frame of the signal is independent, information of all the frames is combined to establish a total Bayesian model, and hyper-parameter vectors are employed to control a sparsity of all the frames of signals to be reconstructed; and finally, the expectation maximization algorithm (EM) is employed to obtain an iterative update formula of the hyper-parameter vectors. The estimation method of a quasi-stationary broadband arraysignal direction of arrival based on block sparse Bayesian learning fully utilizes a short-time stability feature of quasi-stationary broadband array signals to establish a block sparse model, and therefore a higher DOA estimation precision can be obtained.

The invention provides a hypergraph convolutional network model and a semi-supervised classification method thereof. The method comprises the following steps: 1, carrying out sparse coding on sample data features of a non-Euclidean structure to form a sample representation coefficient matrix; 2, constructing a hyperedge according to the similarity of the samples, calculating a hyperedge weight, and constructing a hyper-graph model; 3, by means of a hypergraph theory, defining convolution operation on a hypergraph, and constructing a hypergraph convolution network model; 4, defining a semi-supervised learning method on the hypergraph convolutional network, designing a loss function, and predicting category labels of all samples by using category information of a small number of calibrationsamples; 5, respectively making a semi-given label matrix for training, verifying and testing, setting network hyper-parameters, training a network model, and learning a convolution kernel and a regularity factor parameter of the network according to a random gradient descent algorithm; and 6, for given data, predicting an unknown sample category by using the trained model to realize semi-supervised classification.

The invention discloses a fast sparse Bayesian learning based direction-of-arrival estimation method and mainly solves the problems of heavy computation and large location estimation error in the prior art. The method includes the implementation steps: (1) adopting antenna receivers to form a uniform linear array; (2) sampling space signals and computing an array covariance matrix R; (3) vectorizing R to obtain a sparse model vector y; (4) dividing space domain grids, and constructing an over-complete base phi(theta) according to the structure of the sparse model vector y; (5) establishing a sparse equation according to the sparse representation relation of the sparse model vector and the over-complete base; (6) defining a hyper-parameter vector alpha, and adopting a fast sparse Bayesian learning algorithm to solve the sparse equation; (7) drawing a magnitude spectrogram according to an optimal estimation value of alpha to obtain a direction-of-arrival value. By the method, estimation accuracy of target reconnaissance and passive location under the conditions of low signal to noise ratio and small snapshot number is improved, computational complexity is lowed, and the method can be applied to target reconnaissance and passive location.

A CT image pulmonary nodule detection method based on a 3D residual convolutional neural network includes a training process and a testing process. The training process includes the following steps: S1, preprocessing an original image, resetting the voxel spacing to (1, 1, 1), and converting the voxel spacing into voxel coordinates; S2, capturing 3D positive and negative samples from a CT image; S3, setting a maximum and a minimum, and standardizing the sample data; S4, constructing a 3D convolutional neural network; S6, setting training hyper-parameters, and importing the training hyper-parameters to a data training model in the form of mini-batch; and S6, saving the model after the model is fully trained. The testing process includes a step S7: preprocessing test CTs, sampling the test CTs one by one in the form of sliders, importing the test CTs to the model for calculation, selecting samples with high confidence, and deleting repeated samples through a non-maximum suppression algorithm. The method is of high accuracy, and can be used to analyze whether there is a nodule in an image and the specific position of the nodule in the image.

The invention relates to an unmanned aerial vehicle maneuvering target tracking method based on DDPG transfer learning, and the unmanned aerial vehicle maneuvering target tracking method carrying outtask decomposition, initializes an environment state, neural network parameters and other hyper-parameters, and carries out the training of a neural network. At the beginning of each round, the unmanned aerial vehicle executes an action to change the speed and the course angle, to obtain a new state, stores the experience of each round in an experience pool to serve as a learning sample, and continuously iterates and updates parameters of the neural network. And when the training is completed, the neural network parameters trained by the sub-tasks are stored, and are migrated to the unmanned aerial vehicle maneuvering target tracking network in the next task scene until the final task is completed.

The invention discloses a word vector model incremental study method. The super parameter used in the method comprises: a vector dimension, a number scope negative data, and a text window length; the word vector mode is updated dynamically to complete the vector model optimization and therefore complete the incremental study through initializing new words that appear in the additional text and sampling counterexample based on the historical glossary word-list; the technical scheme study can avoid to study repeatedly, historical data the calculating complexion degree is greatly reduced; the scheme can maintain the higher study efficiency so as to satisfy the efficiency requirement of the online system.

The invention discloses a reinforcement learning based anaphora resolution method, which comprises the following steps: data preprocessing: carrying out word segmentation, sentence segmentation, part-of-speech tagging, part-of-speech reduction, named entity identification, syntactic analysis and word vector conversion on text data to obtain candidate preceding words and analogy word related characteristics; constructing a neural network model: combining the characteristics of the word vectors and the relevant characteristics which can learn the fingering pairs and the relevant semantic information, better sorting and scoring the candidate preceding words and the fingering words, and finally obtaining an fingering chain; and using the trained model to carry out anaphora resolution, inputting text data, and outputting a resolution chain. According to the method, deep learning training is carried out by adopting a reward measurement mechanism for overcoming the defects of a heuristic lossfunction, the model effect is improved, hyper-parameter setting is automatically carried out for different language data sets, the necessity of manual setting is avoided, the practicability of the model is improved, and the application range is expanded.

The invention relates to a pedestrian re-identification method fusing random batch masks and multi-scale representation learning. The pedestrian re-identification method comprises the steps of constructing a pedestrian re-identification training network; performing network hyper-parameter adjustment according to preset training parameters to obtain a learning network; shielding multi-scale representation learning and random batch mask branches to obtain a test network, and inputting the test set into the test network to obtain a corresponding test identification result; judging whether the accuracy of the test recognition result is greater than or equal to a preset value or not, if so, inputting the actual data set into the learning network, and otherwise, retraining the network; and finally, shielding multi-scale representation learning and random batch mask branches to obtain an application network, and inputting the query image into the application network to obtain a correspondingidentification result. Compared with the prior art, the method has the advantages that a random batch mask strategy, multi-scale representation learning and loss function joint training are used, moredetailed discrimination features of pedestrian images can be captured, and local important suppressed features are extracted.

The invention provides a training method of multi-moving object action identification, comprising the following steps of: extracting the movement track information of each moving object from video data; layering the movement track information of the moving objects; modeling for the movement mode of the multi-moving object action on each layer; carrying out characteristic description on the model of the movement mode by synthesizing the overall and local movement information in a video, wherein the characteristic at least comprises a three-dimensional hyper-parameter vector for describing the movement track by using a gaussian process; and training a grader according to the characteristic. The invention also provides a multi-moving object action identification method which identifies the multi-moving object action in the video by utilizing the grader obtained by using the training method. In the invention, the movement track of an object is represented by using the gaussian process from a probability angle, and a model is established for a multi-people action mode from three granularity layers, and the characteristics are extracted, which makes the representation of the multi-people action more practical.

The invention relates to a distributed acceleration method and system for a deep learning training task. The method comprises the following steps: (1) building a distributed GPU training cluster; (2)adopting a swap-in swap-out strategy, and adjusting a minibatch size on a single GPU working node in the distributed GPU training cluster is adjusted; (3) adjusting the learning rate according to theminibatch size determined in the step (2); And (4) carrying out deep learning training by adopting the hyper-parameters minibatch size determined in the steps (2) and (3) and the learning rate. On thepremise that the training accuracy is not affected, the communication time is greatly compressed simply and efficiently by reducing the number of times of parameter updating communication between clusters, and compared with a single GPU mode, the cluster expansion efficiency can be fully improved in a multi-GPU mode, and acceleration of the training process of the ultra-deep neural network modelis achieved.

The embodiment of the invention discloses a data processing method, a data processing device and a related device, for increasing the parameter solving speed and the parameter solving accuracy of an hLDA (hierarchical latent Dirichlet allocation) model. The method disclosed by the embodiment of the invention comprises the following steps of: sending the global initial global statistical information to various slave nodes, merging the local statistical information received from the slave nodes to obtain the new global statistical information, calculating the probability distribution between documents and themes and the probability distribution between themes and words according to the new global statistical information if the Gibbs sampling performed by the slave nodes is over, establishing and maximizing a likelihood function of a text set according to the calculated probability distributions to obtain new hLDA hyper-parameters, and calculating and outputting the probability distribution between the documents and the themes and the probability distribution between the themes and the words according to the new hLDA hyper-parameters if the iteration for solving the hLDA hyper-parameters is converged.

The invention provides a pre-stack seismic data retrieval method and system. The pre-stack seismic data retrieval method includes: acquiring pre-stack seismic data of the current stratum; selecting an advantageous frequency band range from the pre-stack seismic data; extracting angle gather seismic data and seismic wavelets from the advantageous frequency band range; determining angle gather reflection coefficient according to a convolution model, the angle gather seismic data and the seismic wavelets; determining hyper-parameters according to relevant vector machines and the angle gather reflection coefficient; determining a super-resolution sparse pulse reflection coefficient according to the hyper-parameters, the angle gather reflection coefficient and the Bayes criterion; comprehensively analyzing the super-resolution sparse pulse reflection coefficient to obtain layer distribution of the current stratum. By the pre-stack seismic data retrieval method and system, resolution and accuracy of retrieval results are improved, and accuracy in identifying stratum relation and property of the thin layer and the thin-interbed lithologic oil-gas reservoir is improved.

The invention provides an unmanned aerial vehicle flight state prediction method and system based on LSTM. The method of the invention comprises the following steps: step 1, constructing an action label dictionary; 2, collecting flight state information and motion data of the unmanned aerial vehicle; 3, preprocessing the collected flight state information of each unmanned aerial vehicle; 4, forming a data set in the form of a data matrix; 5, randomly dividing the data set into 70 percent of the training set and 30 percent of the verification set; obtaining the improved LSTM model of a long-short-term memory network with variance by training. 6, tuning the super parameters of the model to obtain a final model by using a verification set; 7: collecting the UAV flight state information of theUAV at the current time which needs to be predicted, and inputting the information into the final model after preprocessing and filtering to obtain the UAV flight motion prediction result. The technical proposal of the invention solves the problem that the existing control model cannot predict the flight action according to the collected flight state data.

The invention relates to a method for realizing super resolution for an image and belongs to the computer vision field. The method includes the following steps of: A1, data preprocessing: a certain number of high-resolution natural images are adopted to form a data set, a certain number of image blocks are extracted from the data set, Bicubic interpolation down-sampling and up-sampling in three times are carried out on the image blocks, and low-resolution images can be obtained; A2, network structure design: a designed convolutional neural network has 4 layers altogether; A3, hyper parameter selection: parameters such as network learning rate, learning momentum and batch_size are determined; and A4, network training and super parameter optimization: the convolutional neural network of all images in the training set from low-resolution images to corresponding high-resolution images is trained, and after any one image is inputted into the trained network, a high-resolution image can be obtained, so that the super resolution of the image can be realized.

The invention discloses a convolutional neural network image denoising method based on multi-scale convolution groups and parallelism. The method specifically comprises the following steps of 1, preparing a training set, selecting an appropriate data set as an original image in the training set, preprocessing the original image, simulating the real noise by adopting the Gaussian white noise, and adding the real noise into the original image as a noise image corresponding to the original image; 2, constructing a network model, and constructing a network model in combination with a convolution network mode of the multi-scale convolution group and the parallelism; 3, setting the hyper-parameters, a loss function and an optimization algorithm of the network according to the network model constructed in the step 2; 4, performing network training, and using the constructed network model in the step 2 to train a single-noise training set and a multi-noise training set respectively to obtain aplurality of network models corresponding to the training sets; and 5, testing the network performance. According to the method, more image contours and texture details can be reserved while the noise is eliminated.

The invention discloses a U-shaped cavity full-convolution integral segmentation network identification model based on a remote sensing image. The model comprises a data preprocessing module, a modeltraining module and a model evaluation module. The model is characterized in that the data preprocessing module is used for performing data preprocessing operation on a remote sensing image to obtaina data set, and performing equal-ratio sampling on the data set to generate a training set, a verification set and a test set; the model training module is used for establishing a U-shaped cavity full-convolution integral cut network model, training parameters of the U-shaped cavity full-convolution integral cut network model by utilizing data of the training set, performing model learning and updating network weights, adjusting hyper-parameters involved in the U-shaped cavity full-convolution integral cut network model by utilizing the difference between the data of the verification set and the recognition effect, and judging the convergence degree of the U-shaped cavity full-convolution integral cut network model to achieve the purpose of deep training. The model can be widely applied toremote sensing image ground objects of different scales.