166 results about "Hypersphere" patented technology

A system and method of recognizing speech comprises an audio receiving element and a computer server. The audio receiving element and the computer server perform the process steps of the method. The method involves training a stored set of phonemes by converting them into n-dimensional space, where n is a relatively large number. Once the stored phonemes are converted, they are transformed using single value decomposition to conform the data generally into a hypersphere. The received phonemes from the audio-receiving element are also converted into n-dimensional space and transformed using single value decomposition to conform the data into a hypersphere. The method compares the transformed received phoneme to each transformed stored phoneme by comparing a first distance from a center of the hypersphere to a point associated with the transformed received phoneme and a second distance from the center of the hypersphere to a point associated with the respective transformed stored phoneme.

A system and method of recognizing speech comprises an audio receiving element and a computer server. The audio receiving element and the computer server perform the process steps of the method. The method involves training a stored set of phonemes by converting them into n-dimensional space, where n is a relatively large number. Once the stored phonemes are converted, they are transformed using single value decomposition to conform the data generally into a hypersphere. The received phonemes from the audio-receiving element are also converted into n-dimensional space and transformed using single value decomposition to conform the data into a hypersphere. The method compares the transformed received phoneme to each transformed stored phoneme by comparing a first distance from a center of the hypersphere to a point associated with the transformed received phoneme and a second distance from the center of the hypersphere to a point associated with the respective transformed stored phoneme.

A rolling bearing fault on-line detection and state assessment method is disclosed. The method comprises the following steps: twelve dimensional dimensionless parameters are extracted; the twelve dimensional dimensionless parameters comprise six dimensional time domain statistical parameters, three dimensional frequency domain statistical parameters and three dimensional dimensionless parameters in a small wave envelope spectrum; standardized reconstruction characteristic vectors can be obtained; whether a rolling bearing malfunctions is determined, and a state of the rolling bearing is assessed. Via the rolling bearing fault on-line detection and state assessment method, the twelve dimensional dimensionless parameters which can be used for effectively representing the state of the rolling bearing can be automatically extracted, the twelve dimensional dimensionless parameters are subjected to decorrelation and standardization operation, standardized reconstruction characteristic vectors that are distributed to form a hypersphere with an original point being a sphere center, and fault detection and state assessment of the rolling bearing can be realized via 2-norms of the standardized reconstruction characteristic vectors; difficult problems of long on line training time, low efficiency, and hard-to-obtain fault samples and the like of a rolling bearing state assessing model can be solved.

A hardware trojan identification method based on a single classification supporting vector machine belongs to the detection and identification field of hardware trojan chips. The invention solves the problems of huge error and low efficiency in the prior art because the technology of utilizing chip side channel information to identify the chip hardware trojan requires to observe an image manually. The method provided by the invention comprises the following steps: I, preprocessing, and acquiring a side channel information matrix; II, selecting a backward analysis chip to perform backward analysis, and determining whether a hardware trojan is contained; III, dividing the backward analysis chip not containing the hardware trojan into a train sample and a train optimizing sample; IV, utilizing the side channel information matrix of the train sample to establish the characteristic space of the chip; V, acquiring the side channel characteristic data matrix of a chip to be detected; VI, performing normalization processing; VII, picking up the normalized data of the train sample and the train optimizing sample; VIII, training the single classification supporting vector machine to constitute a minimum hypersphere; and IX, if beyond the minimum hypersphere, then determining that the chip to be detected is a hardware trojan chip.

The invention discloses a multi-dimensional inverted index and quick retrieval algorithm for large-scale image visual features. The algorithm includes: training a multi-layer codebook required for enhanced residual quantification by visual features of images and building a multi-dimensional inverted index with the trained codebook; according to the trained codebook, quantifying and encoding the image visual features, and according to calculated codes, inserting the codes into corresponding inverted lists in the inverted index; querying the image visual features to query the built multi-dimensional inverted index so as to obtain a query candidate set; optimizing the query candidate set by adaptive hypersphere filtering, and ranking filtered query results so as to finish retrieval of the image visual features. The algorithm has the advantages that as the image features are quantified and encoded, efficiency of quantifying the image features is improved; the multi-dimensional inverted index is built with the generated image codes, the inverted index can be built just by training few codebooks, and speed of building the index structure is increased.

The invention discloses an underdetermined blind source separation method based on density, mainly solving the problems that the computation complexity of the prior art is high, the prior art is easily influenced by an initial value and the number of source signals needs to be given. The method comprises the following steps: after the low energy sampling data of an observed signal is removed, projecting the observed signal onto a right half unit hypersphere; computing the density parameters of all projective points, and deleting the projective points with smaller density; utilizing an improved K-mean value clustering algorithm to cluster the surplus projective points and determining the optimal clustering number and clustering center; removing the cluster with less number of data objects, wherein the number of the surplus clusters is the estimated value of the number of source signals, and the corresponding clustering center is the estimated value of each column vector of a confusion matrix; and according to the observed signal and the estimated confusion matrix, adopting a linear programming method to recover the source signals. According to the method, the computation complexity is reduced, the influence of the initial value on the estimated performance is reduced, the confusion matrix can be estimated when the number of the source signals is unknown, and the estimating precision of the confusion matrix and the source signals can be increased.

A sphere decoder sets a Euclidean distance between a lattice vector obtained by using an MMSE or ZF estimate and a received signal as an initial radius, further reduces the initial radius, and searches lattices points included inside a hypersphere with the further reduced initial radius. In addition, one lattice vector having a minimum Euclidean distance is output. One dimension is selected to reduce an initial radius, and estimates in other dimensions are kept fixed, excluding the selected dimension. Then candidate lattice points are searched in the selected dimension, excluding a current estimate, such that a minimum Euclidean distance and a lattice point estimate corresponding to the minimum Euclidean distance are obtained. The initial radius is updated by the minimum Euclidean distance, and a final lattice vector is constructed by combining a lattice point estimate corresponding to the initial radius and the lattice point estimates in other dimensions.

The invention provides a finger vein image recognition method, which comprises the steps of: dividing finger vein ROI images into training and testing samples; normalizing image sizes to 46*102 pixels; utilizing a Gabor filter for enhancement; subjecting each pixel point to comparison and selection to obtain an ROI enhanced image; adopting a PCA for reducing dimensionality of the ROI enhanced image; subjecting the ROI enhanced image after dimensionality reduction to hyperspherical granulation to obtain a new granular set; and calculating distance between each large hypersphere in each finger vein ROI image of the testing samples and the new granular set for carrying out steps such as recognition. The finger vein image recognition method adopts the PCA and hyperspherical granulation combined method for granulating and fusing all finger vein ROI images belonging to the same individual into a large hypersphere, and better describes the generality of samples of the same individual acquired at different time. During recognition, the samples to be tested are processed into a hypersphere, and then subjected to distance comparison with the fused large hypersphere.

The invention provides a target re-identification method based on hypersphere embedding of a densely connected convolution network. At first, that features of underwater deformation object in the video sequence are extracted according to the secret-level connected convolution network DenseNet, so that that gradient disappear is greatly reduced, feature propagation is enhanced, feature reuse and parameter learning processes are supported; from the viewpoint of fine-grained classification, from local integration to global integration, the characteristics of underwater deformation targets are extracted by means of grouping average pooling, the more accurate feature expression ability of underwater deformation target is obtained; focusing on the inter-class difference of underwater deformation individual targets by using hypersphere loss, i.e. angular triple loss, the difference in regional classification are classified to avoid the direct measurement of the Euclidean distance between theindividual target coding features of underwater deformation, a complete and continuous underwater deformed individual target re-recognition model based on multi-point placement is constructed. The invention finally completes the close supervision and process tracking of the underwater deformation target individual in the close-range multi-field observation.

The invention relates to a calculation method for fast estimating the yield of an integrated circuit. The calculation method includes the steps that input original parameter variables are determined and orthogonalized according to elements or key elements in a circuit unit in the integrated circuit to be analyzed, and the number Y of the parameter variables is the number of technological parameters which are the most sensitive to the characteristics of the circuit; rmax is determined through original sampling points in normalized Gaussian distribution, and the cumulative distribution function value in chi distribution of the Y-dimension space is equal to a set sampling precision value; M evenly distributed sampling points are acquired in a hypersphere with the radius of rmax; the failure probability of the circuit unit is calculated on the basis of the M sampling points. The calculation method for fast estimating the yield of the integrated circuit is more precise and reliable in searching for failure areas; simulation research results show that the calculation method achieves a good compromise between efficiency, precision and searching for the failure areas, the yield of the digital circuit can be effectively estimated fast, and the efficiency is improved substantially.

A maximum hypersphere is created in the feature space according to support vectors, wherein the support vectors are one or more feature vectors in a feature space. A center of the created maximum hypersphere is calculated according to the support vector(s). A decision hyper sphere is created with the same center as the calculated center of the created maximum hypersphere. Feature vector(s) are classified within the decision hypersphere, as positive feature vector(s). False positive rate is kept to a predetermined level to provide effective object detection.

This invention publishes a secure group key management approach based upon N-dimensional hypersphere. After initialization, the GC admits the new members and assigns identifiers to them when there are new members joining the group, and deletes the leaving members' private information when there are members leaving the group. If a lot of members join and other members leave the group at the same time, the GC deletes the leaving members' private information, admits the new members, assigns indemnifiers to the new members, and then chooses mapping parameters, mapping each member's and its private information to the points in a multi-dimensional space. The GC calculates the central point of the hypersphere, and publishes the central point, the mapping parameter and the identifiers of leaving members if there are members leave. The group members calculate the mapping points, and then calculate the group keys. The invention can effectively reduce user storage, user computation, and amount of update information while re-keying. The independence of the group keys can be kept.

The invention discloses a grid curved surface reconstruction method for scene understanding, and belongs to the field of three-dimensional graphics. The method comprises the following steps of: automatically identifying abnormal data through a mathematical model based on an undirected graph network, and estimating and adjusting a point cloud normal vector in an internal curved surface mapping manner; solving low-latitude feature information of the point cloud model, extracting a skeleton center curve, and designing a block segmentation mode of a complex branch area; for the point cloud from which the branches are removed, adopting a depth map mapping mode to identify and construct a visual field visible area of the point cloud, and designing a visual clustering-based slice body segmentation mode; constructing a self-learning scene understanding model based on the segmentation result, and realizing automatic restoration of the missing region; finally, designing a point cloud reconstruction algorithm based on a hyperspherical mapping mechanism, and surface data reconstruction loyalty of the original point cloud is achieved, and the method can be applied to efficient and accurate reconstruction of a CAD model, a building, an organism and a point cloud model mixed with multiple scenes.

There are provided an image recognition apparatus and an image recognition method according to which, when registering a face model for a face image, face images are sequentially acquired from frames of a moving image, feature amounts are extracted from the sequentially acquired face images, and a hypersphere defined by Support Vector Data Description (SVDD) is obtained from the group of training samples. Then, training samples in a boundary region of the hypersphere are registered as boundary samples. When performing matching of a face image of a recognition target, the reliability is calculated using the boundary samples, and a person with the highest reliability is output as a recognition result.

The invention provides a Webshell detection method and apparatus based on deep learning and semi-supervised learning. The method comprises the following steps: obtaining original training samples, selecting labeled samples to perform word segmentation processing, analyzing the correlation between feature words and labels by chi-square test, and selecting the previous K feature words with the strongest correlation as screening feature words; performing feature word screening on unlabeled samples by using the screening feature words to serve as unlabeled sample features; training the obtained unlabeled sample features by using a neural network algorithm to obtain text vectors of the unlabeled samples; training a single-class SVDD model by using an unsupervised method, and optimizing a hypersphere radius to the minimum, wherein the maximum case comprises the unlabeled samples; and for a new labeled sample, performing incremental training on the SVDD model by using an online learning method to correct the single-class SVDD model; and applying the latest model to the prediction of new samples. By adoption of the Webshell detection method and apparatus provided by the invention, the missing report rate and the false reporting rate of the traditional webshell detection can be effectively improved.

The invention discloses a wireless sensor abnormal data detection method based on unsupervised learning. The method comprises the following steps of 1, acquiring m pieces of data continuously acquiredby a wireless sensor node to form a training sample set; 2, establishing a quarter hypersphere support vector machine model of which the sphere center is located at the origin of high-dimensional space coordinates, wherein the minimum support radius is R; 3, optimizing the parameters of the 1/4 hypersphere support vector machine model by applying a particle swarm algorithm and a training sample set to obtain an optimized model; 4, obtaining m + 1 pieces of data Tq continuously collected by the wireless sensor nodes, calculating the distance d (Tm + 1) of Tm + 1 in the mapping space of the optimization model to the sphere center, and if d (Tm + 1) is smaller than or equal to R, determining that Tm + 1 is normal data; if d (Tm + 1) is greater than R, using a set {T1, T2,... Tm} as a training sample set to retrain the model, and calculating the minimum support radius Rnew, and calculating the distance d (Tm + 1) new from the Tm + 1 in the updated mapping space of the model to the spherecenter, and if d (Tm + 1) new < = Rnew, determining that the data Tm + 1 is normal data, otherwise, determining that the Tm + 1 is abnormal data. According to the method, the unsupervised learning isadopted, and the samples do not need to be labeled, so that the detection accuracy is higher.

The invention provides an improved top speed learning model and a method for classifying modes of the top speed learning model, discloses an efficient neural network rapid learning method and a limited top speed learning machine and belongs to the field of mode identification and machine learning. The improved top speed learning model comprises (1) a concept of a limited parameter space, (2) a hypersphere limited condition in the limited parameter concept and (3) output weight learning. According to the concept of the limited parameter space, a lateral inhibition mechanism based on sample prior information is adopted and the content is particularly shown on the aspect of generation of a connection weight from an input layer to a hidden layer; according to the hypersphere limited condition in the limited parameter concept, selection of the connection weight from the input layer to the hidden layer is limited to a hypersphere; according to output weight learning, after a weight of the hidden layer is selected from the limited parameter space, learning is trained through a top speed learning machine model based on least squares, and finally an output weight of the model is obtained. According to the method, the classification and identification effects of the model can be greatly improved, and the training speed can be greatly increased.

The invention discloses a method for dynamically recognizing an overweight vehicle in a bridge monitoring system. The method comprises the steps of 1) ensuring a strain response time sequence to be astationary sequence and if not, transforming time sequence data into a stationary time sequence through step difference and order difference; 2) modeling the time sequence data using an SARIMA model,taking an AR coefficient of the model as a key feature to identify an abnormality, and then connecting the coefficients of different sensors in the same section in series to obtain a feature vector; 3) inputting the AR coefficient feature vector into a noise reduction automatic encoder for training, after the training is completed, obtaining a middle layer dimension in a network structure of the automatic encoder, that is, the required key feature, and taking the middle layer dimension as a final training feature; and 4) inputting the training feature into a one-class support vector machine with a kernel function for unsupervised training, with a training result being a hypersphere in a high-dimensional space, and then, using the hypersphere to determine whether the test data is overweightabnormal data. The method in the invention is simple and efficient in identification principle and has strong robustness.

The invention provides a big data sensitive characteristic optimization selecting based equipment failure early warning method and system. The method includes the following steps: collecting vibrationdata of equipment under a normal working condition, and performing time and frequency domain index feature extraction on the vibration data to form vibration-like feature vectors; applying a compensation distance assessment technique to perform optimization selecting on the vibration-like feature vectors, and commonly forming a sensitive vector set by the optimally selected vibration-like featurevectors and the process data of the equipment under the normal working condition to be a training sample which supports a vector data description model so that a SVDD hypersphere model of the equipment under the normal working condition can be formed through training; and processing testing vibration data by adopting the above same steps, and forming a testing sensitive vector set with the obtained optimized selection feature vectors and the process data under a testing working condition to input to the SVDD hypersphere model under the normal working condition, and performing early warning analysis on output results when the equipment is abnormal or is about to be abnormal. Thus, the intelligent maintenance of the equipment can be realized.

The invention relates to a method and a device of realizing lattice vector quantifies coding and decoding. Mainly in a method of the lattice vector quantifying coding and decoding, a new basic code book is selected out and directly or indirectly comprises all points whose value is less than or equal to 3 of any dimension in a RE<8> point set, and points on a fourth grade hypersphere in the basic code book are all odd points so that part of nearest adjoint point values can be comprised in the basic code book only via corresponding odd number processing. The capacity of the basic code book selected out in the invention embodiment is further enlarged, and at the same time, vector quantization Leader number in the selected basic code book is further reduced compared with an amount of code words of basic code book in existing technique, so that occupancy of memory space of a quanxtizer is reduced and a speed of searching for the basic code book is greatly improved.