270results about How to "Avoid the curse of dimensionality" patented technology

A cityscape image semantic segmentation method based on multi-feature fusion and Boosting decision forest includes the following steps of carrying out super-pixel segmentation on images, carrying out multi-feature extraction, carrying out feature fusion and carrying out training learning and classification recognition. The method effectively integrates 2D features and 3D features and remarkably improves recognition rates of targets. Compared with the prior art, segmentation results are consistent, connectivity is good, edge positioning is accurate, a Boosting decision forest classification mechanism is introduced, and stability of target classification is guaranteed.

Disclosed is a multi-camera system target matching method based on a deep-convolution neural network. The multi-camera system target matching method based on the deep-convolution neural network comprises initializing multiple convolution kernels on the basis of a local protective projection method, performing downsampling on images through a maximum value pooling method, and through layer-by-layer feature transformation, extracting histogram features higher in robustness and representativeness; performing classification and identification through a multi-category support vector machine (SVM) classifier; when a target enters one camera field of view from another camera field of view, performing feature extraction on the target and marking a corresponding target tag. The multi-camera system target matching method based on the deep-convolution neural network achieves accurate identification of the target in a multi-camera cooperative monitoring area and can be used for target handoff, tracking and the like.

The invention provides an emotion classification model training and textual emotion polarity analysis method and system. The emotion classification model training method comprises the steps that data are acquired from a corpus so that original data are obtained; the original data are preprocessed so that preprocessed data are obtained; word vectors are extracted from the preprocessed data through a neural network model; the word vectors are fused according to preset fusion rules so that sentence vector characteristics are generated; and an emotion classification model is trained according to the sentence vector characteristics so that the trained emotion classification model is obtained. The neural network model is adopted, the words are expressed by low-dimensional spatial vectors, the low-dimensional spatial word vectors are fused into the sentence vector characteristics according to the preset rules, and the emotion classification model is obtained by certain learning models through training so that word vector dimension can be effectively reduced, the dimensions disaster problem can be avoided, correlative attributes between the words can be mined and vector semantic accuracy can be enhanced.

The invention relates to an extraction method of EEG characteristics of imagination of single-side limb motion in a brain-computer interface device. The classification of the characteristics of imagination of single-side motion has stronger pertinence, which is decided by the nature of excluding same tasks and extracting different tasks of a CSP space filtering method; at the same time, due to the combination of CSP algorithm and FDA characteristics extraction, the dimension of input vectors is reduced, and the marketability of a classifier is enhanced; therefore, the classification accuracy rate is enhanced to a certain extent; adopting Fisher Differentiation and Analysis (FDA), ten-dimension input vector v1, v2(v1 is four-dimension, and v2 is six-dimension) are reduced into two one-dimension input vectors f1, f2, and then classification is carried out by a support vector machine, thereby not only improving classification accuracy rate, but also avoiding the problem of dimension disaster caused by too high dimension ,as well as being beneficial to the popularization of the classifier.

The invention provides a deep Q network based regional power-grid operation point dispatching optimization method. The method comprises the steps that a dispatching configuration of a regional power-grid operation point dispatching center is determined; a dynamic decision process for dispatching operation points of regional power grid is modeled into a corresponding Markov decision process model which comprises state, action, cost and optimization target function; the deep Q network is used to carry out strategy resolution on the Markov dynamic decision process; the operation point of the regional power grid comprises a routine thermal power set output of a next dispatching period, thermal power set output and the flexible load reduction amount; and dispatching of the operation point of the region power grid comprises, at least, that the power grid operation point in the next dispatching period is determined dynamically according to wind power and load power short-term prediction information and operation point information in the present period. Thus, effective response to randomness of the new energy output and load demands can be made, source-load interactive dispatching potential gives a full play, dynamic balance of the power of the regional power grid is maintained, and the operation efficiency of the power grid system is improved.

The invention discloses a stability evaluation method based on correlation between power grid operation and a transient stability margin index. The method comprises the following steps of establishing a power grid safety stability hierarchical correlation analysis model based on offline simulation data and real-time monitoring data of the power grid, and performing transient stability hierarchical correlation characteristic extraction according to the model; performing discrete semantic description of the power grid operation characteristic and the transient stability margin index, and performing correlation relation identification between the power grid operation characteristic and the transient stability margin index; and establishing a stability correlation characteristic rule database, and performing quick stability determination and reason analysis on power grid operation state change according to the rule database. The stability evaluation method has advantages of quickly evaluating the current stability level according to a system operation state, supplying quantitative information support for auxiliary decision of operation personnel, improving standardization, speed and self-adapting capability in power grid stability evaluation, and realizing wide application prospect.

A method for optimize that dynamic dispatch of power generation and transmission system include such steps as establishing a multi-area interconnected power network structure including conventional generator set, photovoltaic generator set, wind generator set, rigid load, flexible load and DC tie line, and establishing physical model of each unit. Then the dynamic dispatching problem of interconnected power network is established into the corresponding MDP mathematical model. Finally, the depth Q learning algorithm is used to solve the MDP mathematical model. Under the strategy, the dispatching mechanism can select a reasonable action plan according to the actual operation state of the power grid at the dispatching time, and realize the dynamic dispatching of the interconnected power gridgeneration and transmission system. The invention can effectively cope with the randomness of the new energy and load demand in the cross-region interconnected power network, promotes the consumptionof the new energy, is conducive to the safe and economic operation of the cross-region interconnected power network, and improves the stability of the operation of the power system.

The invention discloses an identifying method of brain cognitive states based on tensor locality preserving projection (Tensor Locality Preserving Projection, TLPP). The method comprises the following steps: 1) pretreating and grouping of fMRI (functional Magnetic Resonance Imaging) data of the brain cognitive states; 2) constructing a neighbor graph G and a corresponding incidence matrix S; 3) calculating characteristic decomposition of a training sample set, solving corresponding characteristic transformation matrix and calculating low dimensional imbedding of training samples; 4) classifying and identifying: calculating low dimensional imbedding of the training sample sets, and distinguishing and classifying the training sample sets by a tensor distance-based neighbor classifier. According to the method, dimensionality reduction and characteristic extraction are directly carried out on multidimensional tensors by TLPP algorithm, and characteristic dimensionality reduction is carried out on collected brain cognitive fMRI data, so that the brain cognitive states are effectively identified and classified. By combining the tensor distance-based neighbor classifier, the classifying accuracy is improved. The method not only inherits advantages of conventional methods, but also greatly reduces complex of time and space and overcomes curse of dimensionality. The method is less in calculated amount, less in memory consumption and shorter in time consumed.

The invention discloses a text clustering method based on random neighbor embedding. The text clustering method includes the following step that the text set is preprocessed and the text set is expressed as a normalized word-text co-occurrence matrix; high-dimensional text data is embedded into low-dimensional space by t-Distribution random neighbor embedding (t-SNE) to make the low-dimensional embedding points corresponding to the text with low similarity in high-dimensional space be far in distance and the low-dimensional embedding points corresponding to the text with high similarity be close in distance; and the plurality of low-dimensional embedding points are taken as the initial mass center of the K-means algorithm, and the K-means algorithm is adopted to perform clustering based on the low-dimensional space mapping point coordinates. According to the invention, the dimension disaster problem caused by text high-dimension sparse characteristics can be solved, the dimension of the text data is reduced, the operation time of the clustering algorithm is shortened, and the precision of the clustering algorithm is improved.

The invention relates to an HU invariant moment and support vector machine-based garment style identification method. The method comprises the steps of preprocessing a garment image to obtain an outer contour of a garment; extracting an HU invariant moment characteristic of the outer contour of the garment; and performing support vector machine (SVM)-based garment style identification. The preprocessing of the garment image refers to a process that the garment image is subjected to segmentation processing, a 8-adjacent connection region with a maximum area is found as a garment region, and internal pore filling is performed on the garment region; the obtaining of the outer contour of the garment refers to a process that the preprocessed garment image is subjected to external edge detection to obtain a contour image of the garment; the extraction of the HU invariant moment characteristic of the outer contour of the garment refers to a process that a 7-order HU invariant moment eigenvector of a contour shape characteristic of the garment is extracted; and the SVM-based garment style identification refers to garment style multi-classification identification performed by adopting an SVM multi-classifier. The method can achieve the identification accuracy of 83%, has a relatively good effect of identifying garment styles with similar contours, has the characteristics of quickness and accuracy, and can be suitable for identification of garment styles in garment images.

The invention discloses a multi-agent collaboration framework based on deep reinforcement learning, which comprises an agent, a billboard, and a an actor-critic-based deep reinforcement learning module, a next moment state calculation module. The agent is defined by a current state, a speed and an expected target; the billboard is responsible for storing, updating and transmitting information; anactor in the Actor-Critic-based deep reinforcement learning module selects a proper action according to the current environment state and the state of an agent, and training learning is continuously carried out through Critic in combination with evaluation given by the state sequence of each agent, so that an optimal control strategy is obtained; and the next moment state calculation module respectively calculates the state of each agent at the next moment according to the current state of each agent and the taken action, and interacts with the billboard. The multi-agent cooperation frameworkbased on deep reinforcement learning provided by the invention has relatively good expandability and relatively high universality, and can provide a technical scheme for realizing diverse multi-agentcooperation.

The invention discloses an optimized design method of an axial phase-separated magnetic suspension switched reluctance flywheel motor, comprising the following steps: an initial value of structural parameters of the axial phase-separated magnetic suspension switched reluctance flywheel motor is estimated; Based on the initial value, the three-dimensional finite element electromagnetic calculationmodel is constructed, and the key structural parameters are selected by sensitivity analysis. The key structure parameters are inputted into the three-dimensional finite element electromagnetic calculation module to obtain the output variables corresponding to the motor performance indicators to form the sample data set. The unified optimization objective function is constructed by taking the keystructural parameters as the structural parameters to be optimized, and the kernel limit learning machine algorithm is used to train the sample data set to construct the optimization model corresponding to the optimization objective function. the structure parameters to be optimized are taken as the input of the optimization model, the PSO algorithm with weight inertia adjustment is used to obtainthe optimization value of the key structure parameters, and the optimization operation of the motor is completed. The invention realizes the purpose of multi-objective synergistic optimization of theaxial split-phase magnetic suspension switched reluctance flywheel motor.

The invention discloses a CMR model for uniformly retrieving cross-media information, and aims to provide a cross-media retrieval model with accurate and rapid information. The method is realized through the following technical scheme: a multi-modal media information semantic feature unified expression and association module queries heterogeneous information input by the input and cross-media datamodule; the multi-modal semantic features are mapped into the same feature space and a multi-modal semantic association rule is constructed based on the multi-modal semantic feature extraction resultand the mapping from the bottom-layer features to the high-layer semantic features, thereby realizing the association between the bottom-layer features and the high-layer semantic features of the cross-media information and the association of the high-layer semantics of different modal information; the cross-media data index construction module is used for establishing a multi-dimensional retrieval index for the multi-modal data characteristics; the cross-media retrieval model construction module realizes unified retrieval of multi-modal information based on ontology, semantic network and knowledge graph technologies; and conflict detection and self-organization of retrieval results are realized through a multi-modal retrieval result association verification and organization module.

The invention discloses a video tracking method based on rank learning. The method comprises the steps of firstly compressing multi-scale image features by using a sparse measurement matrix based on a compressed sensing theory, secondly using a Median-Flow tracking algorithm as a predictor to obtain the rough position of a target and constructing a training data set for an RV-SVM algorithm, and finally sorting training samples and taking the RV-SVM algorithm as a binary classifier to separate the target and a background to achieve the purpose of video tracking. The training process of the RV-SVM algorithm is a linear programming problem, the training time of online learning is reduced, and the efficiency of a tracking system is improved. Through the combination of multi-scale image compression feature extraction, the Median-Flow tracking algorithm and the RV-SVM algorithm, problems of target scale change, partial occlusion, 3D rotation, posture change, target fast movement and the like in a video tracking process can be effectively processed.

The invention discloses an improved text classification characteristic selection method. The improved text classification characteristic selection method comprises the steps of obtaining a training set text; performing word segmentation and stop word removal on the obtained training set text; improving a characteristic selection method, performing division on all word sets by using text frequencies of characteristic words, performing characteristic selection on a low-frequency word set by using an information gain value, and performing characteristic selection on a high-frequency word set through an improved x2 statistic method; and combining characteristic words of two parts to form a final classification characteristic word set. With the adoption of the method, more representative classification characteristic words can be selected through carrying out a characteristic selection process twice, so that the classification efficiency and accuracy are improved.

The invention discloses a cascade reservoir ecological random optimization scheduling model and a solving method. The establishment of the optimization scheduling model comprises the following steps:firstly, establishing a cascade reservoir ecological random scheduling model target function with the minimum corrected annual flow deviation AAPFD index; secondly, describing constraint conditions ofthe cascade reservoir ecological random scheduling model, wherein the constraint conditions comprise a water level constraint, a discharge flow constraint, a reservoir capacity constraint and a guaranteed output constraint. The ecological random scheduling model is solved by utilizing an SARSA algorithm in reinforcement learning, so that the problem of dimensionality disasters caused by a randomdynamic programming algorithm can be solved. The method can improve the ecological demands on the premise of guaranteeing the output of the cascade reservoir hydropower station, and has higher guidingsignificance for the ecological utilization of reservoir water resources. Meanwhile, the proposed SARSA algorithm can well obtain the result of the cascade reservoir random ecological optimization scheduling model in a short time.

The invention provides a multi-agent adversarial decision-making method based on cooperative reinforcement learning and transfer learning, and the method is characterized in that the method comprisesthe following steps: defining the state space S = {s1, s2,..., sn} of an agent; setting an action space A to be equal to {a1, a2,..., an}; setting a value function matrix of the agent reinforcement learning model; calculating a value function sequence corresponding to the current state st by using an action evaluator, and selecting a corresponding action at through an action selector based on simulated annealing and softmax strategy; meanwhile, the state of the intelligent agent is changed, and the intelligent agent is transferred to the next state st + 1. After the action at is executed, theintelligent agent obtains a reward signal rt from the environment; the loss of experience storage can be reduced through a weight sharing mode, and the adversarial decision-making efficiency is improved. Through the migration learning method based on the attenuation function, the agent can reuse previous experience with a gradually decreasing probability, and the migration learning migrates the previously trained action evaluator weight to more adversarial decision scenes, thereby improving the generalization of the learning model.

The invention is applicable to the technical field of clouds, and provides a construction and evaluation method of a fault model based on an SVM (Support Vector Machine). The method comprises the following steps: selecting a radial basis function RBF to construct a cloud fault prediction model of the support vector machine SVM; training the support vector machine SVM based on a given cloud sampletraining set, wherein the training process is specifically as follows: based on the support vector machine SVM, converting cloud fault prediction into quadratic programming with a constraint condition; building a decision function based on the solution of the quadratic programming, wherein the decision function is a hyperplane; and performing fault evaluation on a test sample point based on the hyperplane. Compared with a BP model and an LVQ (Learning Vector Quantization) model, an SVM model can find a global optimal solution, avoid the curse of dimensionality, and improve the convergence speed rapidly at the same time; and in addition, the embodiment of the invention selects the radial basis function RBF to construct the cloud fault prediction model of the support vector machine SVM, therefore, the complexity of model construction can be reduced and the fault analysis efficiency can be improved while the accuracy is satisfied.

The invention discloses a principal component distribution function based software defect prediction imbalance data processing method and belongs to the technical field of software engineering application. The method comprises the steps that data acquired from a software data concentration is preprocessed to obtain an original sample set; dimension reduction processing is conducted on the original sample set by adopting a PCA algorithm to obtain a principal component data set including defect-free sample sets and defect sample sets; subsampling is conducted on the defect-free sample sets, and boundary samples and noise samples of the defect-free sample sets are removed; distribution fitting is conducted on principal component data corresponding to the defect sample sets to obtain new defect sample sets; new sample sets are obtained by screening the new defect sample sets; the Euclidean distances between the sample sets and the original sample set in the new sample sets are calculated to remove noise samples in the new sample sets. By the adoption of the imbalance data processing method, the software defect prediction accuracy can be effectively improved.

The invention discloses a method for extracting blood pressure data from a pulse wave signal and a device thereof. The method comprises: model training and blood pressure data extraction; wherein themodel training comprises: acquiring the pulse wave and corresponding blood pressure data, and performing preprocessing; performing complete waveform detection and separation of the pre-processed pulsewave; performing frequency domain transformation on each complete pulse wave waveform, and extracting the frequency domain features; taking the frequency domain feature as an input, taking the corresponding blood pressure value as an output, combining with the neural network for predictive model training, and obtaining a blood pressure data prediction model. The blood pressure data extraction includes the following steps: inputting the frequency domain feature to-be-predicted into a blood pressure data prediction model, and outputting a blood pressure value. The pulse wave is detected and separated, the frequency domain is transformed to extract features, which can effectively avoid the problem of difficult feature detection in a time domain feature extraction method, provides more accurate samples for subsequent model training with neural network, and improves the prediction accuracy of blood pressure measurement and reduces the prediction error.

The invention discloses a real-time video tracking method, comprising the following steps: firstly, compressing image features by dividing the tracking target into sub-blocks; secondly, adapting to the scale change and fast movement of the tracking target by constructing multi-scale candidate regions ; Finally, use the KCF (Kernelized Correlation Filter) algorithm to calculate the correlation between feature vectors to achieve the purpose of video tracking. The invention realizes accurate tracking of common targets in various scenarios, fully meets the real-time requirement, and has very high practical value.

The invention belongs to the field of power distribution network fault positioning, and provides a power distribution network single-phase earth fault positioning method and system. The method comprises the steps of selecting zero-sequence current and negative-sequence current information of a half cycle before and after a fault occurs as fault characteristic quantities; based on the topological structure of the power distribution network, constructing a single characteristic quantity differential node matrix by the fault characteristic quantity; fusing the fault characteristic quantity in a time domain-space domain to generate a high-dimensional fault information state monitoring matrix expanded by a single characteristic quantity differential node matrix; conducting dimensionality reduction on the high-dimensional state monitoring matrix to a two-dimensional space through a multi-dimensional scale dimensionality reduction method, solving eigenvalues and eigenvectors of the matrix obtained after dimensionality reduction, and obtaining eigenvectors of all sections; and quantizing the correlation of the feature vectors corresponding to the sections to obtain a correlation coefficient matrix, displaying fault nodes and non-fault nodes on a two-dimensional plane based on a DBSCAN clustering model, and positioning fault sections corresponding to the fault nodes.