34 results about "Gaussian density" patented technology

The invention discloses a method utilizing the multi-scale multi-task convolutional neural network for population counting of stationary images. According to the method, firstly, an inverse Gaussian density map is combined with an original Gaussian density map to form a combined density map; secondly, non-overlapping sampling of an input image is carried out to acquire several image sub-blocks, and the network is trained based on the image sub-blocks and corresponding true combined density maps thereof; overlapping sampling of the input image is carried out in the same pace, combined density maps of each image sub-block predicted by the MMCNN are superposed to reconstruct the combined density map of a complete crowd image, and population counting is realized. For the problem of populationscale difference, learning characteristics of different scale networks are measured through a fractional loss function, moreover, the population combined density map, the density level and foreground / background classification are simultaneously predicted in a multi-task mode, estimation accuracy of the combined density map is improved, and thereby an uneven population density problem is ameliorated.

A method for adapting a codebook for speech recognition, wherein the codebook is from a set of codebooks comprising a speaker-independent codebook and at least one speaker-dependent codebook is disclosed. A speech input is received and a feature vector based on the received speech input is determined. For each of the Gaussian densities, a first mean vector is estimated using an expectation process and taking into account the determined feature vector. For each of the Gaussian densities, a second mean vector using an Eigenvoice adaptation is determined taking into account the determined feature vector. For each of the Gaussian densities, the mean vector is set to a convex combination of the first and the second mean vector. Thus, this process allows for adaptation during operation and does not require a lengthy training phase.

A system and method for detecting erosion in turbine engine blades is provided. The blade erosion detection system includes a sensor data processor and a cluster analysis mechanism. The sensor data processor receives engine sensor data, including exhaust gas temperature (EGT) data, and augments the sensor data to determine sensor data residual values and the rate of change of the sensor data residual values. The augmented sensor data is passed to the cluster analysis mechanism. The cluster analysis mechanism analyzes the augmented sensor data to determine the likelihood that compressor blade erosion has occurred. Specifically, the cluster analysis mechanism performs a 2-tuple cluster feature analysis using Gaussian density functions that provide approximations of normal and eroded blades in a turbine engine. The 2-tuple cluster feature analysis thus provides the probability that the sensor data indicates erosion has occurred in the turbine engine.

The invention relates to an improved particle filtering method based on the niche genetic algorithm. The method comprises the following steps of: (1) sampling based on the initial probability distribution to obtain initial particles and setting the initial weight; (2) based on the filtering estimations of M particles at (k-1)th moment, carrying out EKF or UKF on each sampled particle to obtain the mean value and the covariance matrix corresponding to the kth moment, and respectively sampling n particles from each disposal distribution by using Gaussian density as the proposal probability density and using the mean value and the covariance matrix of each particle as the mean value and the covariance matrix of the distribution to obtain a set formed by nM particles; wherein n and M are natural numbers; (3) respectively updating the weights of the Nm particles to obtain the weight of each particle; and (4) when the obtained particle set has particles are less than the effective sample capacity, resampling with the niche genetic algorithm. The invention improves the particle filtering, inhibits the degeneracy phenomena and the particle-lack problem caused by simple random resampling, and improves the diversity and the adaptability of the particles, thereby improving the performance accuracy of the particle filtering.

The invention discloses a reservoir classification method based on nuclear magnetic resonance logging, comprising the steps of: obtaining a nuclear magnetic resonance transverse relaxation time T2 spectrum of a depth point to be classified of a reservoir to be classified; calculating the nuclear magnetic resonance porosity of the depth point to be classified of a reservoir to be classified in dependence on the nuclear magnetic resonance T2 spectrum; employing a doublet Gaussian density function to fit the nuclear magnetic resonance T2 spectrum to obtain parameters representing pore structure characteristics of the depth point to be classified of a reservoir to be classified; employing a cluster analysis method to clarify the depth point to be classified of a reservoir to be classified in dependence on the nuclear magnetic resonance porosity of the depth point to be classified of a reservoir to be classified and the parameters representing pore structure characteristics of the depth point to be classified of a reservoir to be classified; and determining the reservoir type of the reservoir to be classified in dependence on the classification result of the depth point to be classified of a reservoir to be classified. The technical scheme provides strong technical support for accurate divide and reasonable development of a reservoir type.

The invention relates to the field of intelligent charging, and particularly discloses an intelligent charging system and a working method thereof, which extract correlation characteristics among charging data such as voltage, current, temperature, pressure and the like of a battery monomer which are acquired in real time through a characteristic encoder, and meanwhile, store the correlation characteristics among the charging data, such as the voltage, the current, the temperature, the pressure and the like, of the battery monomer. In addition, a convolutional neural network model based on a deep learning technology is adopted to mine high-dimensional implicit features of the electrolyte surface image, and a Gaussian density map is applied to fuse the charging feature vector and the surface feature vector, so that the classification result is more accurate. Therefore, a more accurate control type of the charging current can be obtained to judge whether the charging current needs to be increased, decreased or maintained, so that the charging effect can be better.

The invention discloses an objective non-reference image quality evaluation method based on independent subspace analysis. The method comprises steps that, 1, independent subspace analysis on mass original images in a widely-known database is carried out to obtain a series of relatively independent image characteristics, histogram distribution statistics is carried out, and an edge distribution statistics curve is obtained by employing generalized Gaussian density (GGD) model and is taken as a benchmark reference; 2, image characteristics of to-be-measured distorted images are extracted based on independent subspace analysis, and the GGD model is employed to obtain statistics distribution of the characteristic information; and 3, the distorted image characteristic information statistics distribution obtained through processing and benchmark reference statistics distribution are contrasted, and European style distances corresponding to all integrated characteristic information are measured and accumulated and are taken as quality tolerance of the to-be-measured distorted images. Through the method, objective image quality evaluation and subjective evaluation have good consistency, and the method has better performance compared with traditional image quality evaluation methods.

The invention relates to the field of intelligent manufacturing design, and particularly discloses a technological parameter optimization method and system for an aluminum-zirconium-carbon sliding plate and electronic equipment. According to the method, a context-based encoder model is adopted to carry out context encoding on physical performance parameters of the aluminum-zirconium-carbon slide plate prepared after a certain component in a preparation formula of the aluminum-zirconium-carbon slide plate is changed so as to obtain global physical parameter associated information, and distribution of data in a high-dimensional feature space is further corrected by utilizing a Gaussian density map; in this way, the parametric feature vectors are properly fused, so that the feature distributions can converge on the profile relative to each other. Meanwhile, on this basis, an image of a microstructure is introduced as a medium to improve the incidence relation among the parameters, and the accuracy of a decoding regression result is higher based on responsiveness estimation of a Gaussian density map, so that the optimal granularity and the optimal content of the silicon powder are determined through an artificial intelligence algorithm, and the accuracy of the decoding regression result is improved. Therefore, the finally sintered aluminum-zirconium-carbon material has better physical and chemical properties.

A method for characterizing jitter of an internal clock signal of a circuit may include generating a series of samples of the internal clock signal by a reference clock signal, comparing the word formed by the N most recent samples of the series to an N-bit pattern, where N is an integer greater than, or equal to 2, and incrementing a first counter if the word complies with the pattern. The method may also include incrementing a second counter when the count of the first counter reaches a first threshold X1, and incrementing a third counter when the count of the first counter reaches a second threshold different from the first. The method may include calculating an average p and a standard deviation σ of a Gaussian density curve as a function of the counts reached in the second and third counters.

The invention discloses a dense crowd counting method and system based on hidden density distribution and a terminal. The method comprises the steps: obtaining a self-adaptive hidden Gaussian densitygraph through a Gaussian network according to a crowd point graph; according to the counting loss item, the smooth item and the Bayesian item, guiding the hidden Gaussian density graph optimization, so as to enable the generation quality to be higher; taking the hidden Gaussian density map as a training target, combining an adversarial loss function and a Bayesian loss function, and outputting thedense crowd image as a predicted density distribution map; summing the predicted density distribution maps to obtain the predicted density people number. The density predictor, the hidden Gaussian density generator and the discriminator are alternately trained and cooperatively optimized. According to the method, the precision is improved to a large extent, the robustness is good, and the application value is high due to the fact that the parameter quantity and the operand of the inference stage are not increased.

The invention relates to the field of intelligent charging of electric vehicles, and particularly discloses a charging alarm system for an electric vehicle and a working method thereof, and the system employs a convolutional neural network model of a deep learning technology to extract high-dimensional implicit features of a magnetic field distribution diagram based on the magnetic field distribution of a transmitting coil and a receiving coil. And high-dimensional correlation feature information between the magnetic field distribution of the transmitting coil and the magnetic field distribution of the receiving coil is fused by using a Gaussian density map, so that the classification result is more accurate. In this way, the type of the foreign matter can be accurately recognized, and therefore different alarm levels are set based on different types.

The method comprises: aligning the length of phoneme correlated to the input words to multi model states; for each of the multi model states, selecting a mixing from the polyhybrid Gaussian density model state to generate a state sequence; using the state input sequence as the input words to generate voice parameter sequence.

The invention relates to the field of lead frame plastic package integrated circuits, and particularly discloses a method and a system for detecting punching forming of a lead frame plastic package integrated circuit. According to the method, a convolutional neural network model is adopted to extract high-dimensional features of an induction matrix from an array transmission type photoelectric sensor and a regression reflection type photoelectric sensor, and meanwhile, a Gaussian density map is adopted to fuse a first feature map and a second feature map; and further, gradually simplifying the three-dimensional Gaussian density map by using a Gaussian mixture model, and comprehensively training a convolutional neural network by using a density map simplification loss function value and a classification loss function value, thereby helping the convolutional neural model to learn consistent feature representation in high-dimensional features. Therefore, the abnormity can be better detected, so that the equipment is protected, and the yield of products is ensured.

The invention discloses a real-time track obstacle detection method based on a three-dimensional point cloud, and the method comprises the steps: carrying out the processing of three-dimensional point cloud sequence data collected by a laser radar, firstly carrying out the coordinate transformation of the point cloud, converting the coordinates in a European coordinate system into the coordinates in a spherical coordinate system, putting each point in the point cloud into a certain voxel of a cone by using a cone voxelization down-sampling method so as to reduce the calculation amount of subsequent steps; inputting the downsampled points into a local feature coding module, searching local point clouds by using K-nearest neighbor (KNN), aggregating geometric features of the local point clouds, and connecting the centroid, neighbor point coordinates, relative coordinates and Gaussian density features of the local point clouds into a vector; connecting all local point cloud information into a matrix through traversal, and obtaining high-dimensional local feature information of each local point cloud through MLP and maximum pooling; and finally, utilizing multi-scale three-dimensional sparse convolution to realize track real-time identification of a single-frame image through a plurality of down-sampling and up-sampling modules.

The invention provides an electric power system noise self-adaptive robust state estimation method. The method comprises the following steps that 1, L measuring fracture surfaces are obtained, wherein the L is an integer; 2, measuring error estimation is conducted on each measuring fracture surface to obtain the error vector; 3, the statistical learning method is used for carrying out estimation to obtain the parameters of a general Gaussian density model GGD on the basis of the error vector obtained through the estimation of the L measuring fracture surfaces to obtain the distribution type of measured noise, and the best corresponding robust state estimation model is selected according to the distribution type of the measured noise. According to the electric power system noise self-adaptive robust state estimation method, due to the fact that the distribution rules of the noise can be obtained through the statistical learning process, and the distribution rules of the noise are matched with the robust state estimation method in an on-line mode, so that the self-adaptive effect of various kinds of noise types is achieved, namely, the best estimation result which is closer to the state variable true value can be obtained in any noise distribution type.

The invention discloses a visual vibration amplification method, a visual vibration detection method and a visual vibration detection system based on morphological component analysis, and belongs to the technical field of computer vision. The method comprises: acquiring a video file comprising a target object, and determining a region of interest of the target object in each frame of image in the video file; adopting a quality factor adjustable wavelet dictionary and a discrete sine and cosine dictionary to respectively represent a structure component and a texture component of each frame of video; adopting a threshold selection algorithm based on generalized Gaussian density distribution to determine thresholds of the structural components; separating an image structure component and a texture component by adopting a self-adaptive MCA; according to the separated structure components, in combination with the Euler visual angle principle, achieving amplification of micro-vibration signals, reconstructing an amplified video, meanwhile, extracting visual vibration signals for the structure components, and achieving measurement of multiple vibration frequencies.

The invention discloses an interactive yield optimization method and system based on a recurrent neural network, belongs to the technical field of chemical reaction yield optimization, and aims to solve the technical problem of how to obtain a relatively high chemical reaction yield on the premise of reducing the experiment cost. Comprising the following steps: acquiring various experimental condition parameters; simulating a current chemical reaction through a mixed Gaussian density function to obtain a reaction yield; constructing a chemical reaction model based on the recurrent neural network model; and initializing a historical data set, training the chemical reaction model by taking the encoded experimental condition parameters and the corresponding chemical reaction yield as current experimental condition parameters and corresponding reaction yield, outputting experimental condition parameters of the next round as target experimental parameters, and performing a chemical experiment based on the target experimental condition parameters. The obtained reaction yield is used as a target reaction yield; and under the condition that the target reaction yield reaches a threshold value, carrying out multiple rounds of training on the chemical reaction model.

The invention relates to the field of medical data processing, and particularly discloses a data processing system for early warning of osteofascial compartment syndrome through forearm skin data. A convolutional neural network model and an encoder model comprising a one-dimensional convolutional layer and a full-connection layer are adopted to carry out implicit feature extraction on a surface image, skin tension data and skin hardness data of the skin of a forearm; and the tension feature vector, the hardness feature vector and the surface feature map are fused by using a Gaussian density map and Gaussian responsiveness distribution so as to effectively unify the tension feature and the hardness feature on the basis of high-dimensional feature distribution of the surface image feature. Therefore, whether the front arm of the patient has the osteofascia space syndrome or not can be early warned in time by processing the medical data.

The embodiment of the invention provides a lithium ion battery SOC estimation method and device. The method comprises the following steps: determining a mapping relationship between each model parameter and an SOC value in a battery equivalent circuit model; running at least two Kalman filters in parallel, and establishing a battery discrete state space model in combination with the battery equivalent circuit model and the mapping relationship between each model parameter and the SOC value; and estimating the SOC value of the battery discrete state space model through Gaussian and Kalman filtering algorithms. According to the lithium ion battery SOC estimation method provided by the embodiment of the invention, a plurality of Kalman filters can be combined in the form of different weight coefficient proportions, and finally, the SOC value of the battery discrete state space model is estimated through Gaussian and Kalman filter algorithms to obtain the optimal SOC estimation value, so that the SOC estimation accuracy of the battery discrete state space model is improved. Equivalently, a plurality of Gaussian density functions are adopted to accurately describe the process noise andthe measurement noise in the whole process according to a certain weight coefficient, so that the accuracy of the SOC estimation result is greatly improved, and the universality is high.

The invention discloses a livestock inventory method, device and equipment, and a storage medium. The livestock inventory method comprises the steps of obtaining a image for inventory of a place wherelivestock is located; marking coordinate points at the geometric center point and a plurality of feature points of each target in the image for inventory; based on the coordinate points, establishingan elliptical Gaussian model taking the geometric center point as a Gaussian kernel, and obtaining a Gaussian density map; and inputting the for inventory image and the corresponding Gaussian densitymap into a deep learning neural network model, and performing regression training and integral processing to obtain the number of livestock. The embodiment of the invention can effectively improve the accuracy of the statistical result of the live pig number.

The invention relates to the field of wearable device interaction, and particularly discloses an interaction control system of a wearable device and a working method of the interaction control system. A convolutional neural network model based on a deep learning technology is adopted to mine correlation features between data items and between time points, wherein the data items and the time points are collected by a plurality of sensors pasted on an arm. Meanwhile, local correlation features in the signal oscillogram of the bio-electricity signals captured by the myoelectricity sensor within the preset time period are extracted, and a Gaussian density map based on likelihood maximization is further adopted to improve the fusion degree in a high-dimensional space, so that the classification result is more accurate. In this way, the willingness expressed by the human body action can be more accurately obtained from the obtained data, and then the experience of a user when the user controls the wearable device is improved.

The invention relates to the field of intelligent wheelchair interaction, and particularly discloses a finger gesture interaction system of an intelligent wheelchair and an interaction method of the finger gesture interaction system. The method comprises the following steps: performing high-dimensional feature extraction on an acceleration signal, an angle signal and an electromyographic signal by adopting an encoder comprising a one-dimensional convolutional layer and a full-connection layer and a convolutional neural network, and fusing an acceleration feature vector, an angle feature vector and an electromyographic feature map based on a Gaussian density map and Gaussian responsiveness distribution, on the basis of feature distribution of the myoelectricity feature map, classification vectors of actual high-dimensional data distribution of the response acceleration features and the angle features are obtained, feature fusion is effectively achieved, and classification accuracy is improved. In this way, the interaction instruction of the gesture action type can be recognized and judged more accurately, and then the control effect of the intelligent wheelchair is better.

The invention provides an electric power system noise self-adaptive robust state estimation method. The method comprises the following steps that 1, L measuring fracture surfaces are obtained, wherein the L is an integer; 2, measuring error estimation is conducted on each measuring fracture surface to obtain the error vector; 3, the statistical learning method is used for carrying out estimation to obtain the parameters of a general Gaussian density model GGD on the basis of the error vector obtained through the estimation of the L measuring fracture surfaces to obtain the distribution type of measured noise, and the best corresponding robust state estimation model is selected according to the distribution type of the measured noise. According to the electric power system noise self-adaptive robust state estimation method, due to the fact that the distribution rules of the noise can be obtained through the statistical learning process, and the distribution rules of the noise are matched with the robust state estimation method in an on-line mode, so that the self-adaptive effect of various kinds of noise types is achieved, namely, the best estimation result which is closer to the state variable true value can be obtained in any noise distribution type.

The invention relates to the field of instruction interaction of wearable devices, and particularly discloses an arm gesture interaction method and system of a wearable device and an electronic device. According to the method, a sequence of an arm image and a signal oscillogram of a bio-electricity signal are processed through a convolutional neural network model, so that motion information in the image sequence and local correlation features in the signal oscillogram are extracted; and the feature vector and the feature matrix are effectively fused based on a likelihood maximized Gaussian density map, so that the subsequent classification result is more accurate. Thus, the gesture arm type can be recognized more accurately, the willingness expressed by the behaviors of the human body can be understood more accurately, and then more satisfactory experience feeling is brought to people.

The invention relates to the field of power supply control, and particularly discloses a power supply control system of terminal equipment with an NFC function and a working method of the power supply control system, high-dimensional implicit mode information between working states of an SIM card and working states of NFC is mined through a deep learning technology based on a neural network model, and in the process, the working states of the SIM card and the working states of the NFC card are subjected to the deep learning technology. And a Gaussian density map is adopted to accurately express respective information and classification characteristics of associated information of the two, so that whether the power supply power of the power supply needs to be adjusted or not is judged. Through the mode, the SIM card and the NFC card can be ensured to work normally under the condition of no unnecessary power fluctuation.