44results about How to "Improve robust performance" patented technology

The present invention provides a particle filter-based gravity sampling vector matching positioning method. The vector matching algorithm takes into account the position correlation of the matching points, and solves the problem of low credibility of a traditional matching algorithm. The present invention utilizes vector consisting of gravity sampling points for matching; and the second estimation of the single point matching results is achieved by using Bayesian estimation based particle filter. After the single-point matching, Euclidean distance between adjacent points of the matching results is calculated to determine whether the distance meets the scope determined by the Euclidean distance between two adjacent sampling points in a non-error inertial navigation system. If the distance meets the scope, the matching result of the point is reliable; and if the distance does not meet the scope, single point matching is carried out again according to limitation conditions.

The invention discloses an integrated navigation method based on fault-tolerant Kalman filtering and belongs to the technical field of integrated navigation. The method comprises the following steps:selecting an integrated navigation state quantity to write a state equation, selecting a measuring quantity to write a measuring equation and carrying out discretization; constructing a conventional Kalman filterer for state estimation; constructing a fault detection function for detecting faults; constructing a fault-tolerant Kalman filer to estimate the state; feeding data of sub filters to a primary filter for information fusion to obtain global state estimation; and correcting a navigation parameter output by an inertial navigation system by means of an estimated value of global error of federated kalman. By adding a fault detection link and a fault-tolerant Kalman filter for an SINS / GPS sub system isolated as a result of faults, once the faults happen, fault points are returned to carry out fault-tolerant Kalman filtering again. Therefore, the utilization degree of sub system data is improved, so that the robust performance of the whole system is improved.

The invention discloses a depth self learning-based facial beauty predicting method. According to the depth self learning-based facial beauty predicting method of the invention, facial image local texture features extracted by an LBP operator have strong robustness to illumination and micro horizontal movement, so associative memorization of a network, for information representing facial beauty, can be facilitated when the facial image local texture features are adopted as the input of a CDBN, and therefore, possibility for the network to learn adverse feature description can be further decreased. A self-learning method can automatically improve the understanding of the CDBN for data feature distribution, and facial image beauty prediction can still obtain favorable facial image feature description under a condition that the type and number of training samples do not satisfy actual requirements. Through learning a deep nonlinear network structure, a network system, not relying on manual feature selection, combines low-level features such that more abstract and structural high-level distributed beauty feature representation can be formed, and therefore, one kind of automatic learning and feature extraction process can be expressed; and high consistency between facial image beauty mechanical scoring and manual scoring can be realized through using an SVM regression method.

The invention discloses a method for controlling the operation of a non-radial displacement sensor of a bearingless permanent magnetic synchronous motor. According to the method, firstly, a nonlinear mathematical model between radial displacement of a rotor and currents of a torque winding, voltage of a suspension winding, currents of the suspension winding and an angle of the rotor; then a multi-core least squares support vector machine is built by the adoption of linear combination of a polynomial kernel function, an index kernel function and a radial base kernel function; preprocessing and normalization are carried out on collected input data and collected output data, the multi-core least squares support vector machine is trained and adjusted, and a weight coefficient and a threshold value of the support vector machine are determined; a radial displacement prediction module of the multi-core least squares support vector machine is connected to the bearingless permanent magnetic synchronous motor in series; a displacement controller is designed, and the non-radial displacement sensor of the bearingless permanent magnetic synchronous motor can be controlled. A mechanical displacement sensor and an interface circuit of the mechanical displacement sensor of a prior system are removed, cost of the system is lowered, high-speed and ultra high-speed operation performance of the system is improved, and high adaptability, robustness and fault tolerance are achieved.

The invention relates to a battery management system based on a dynamic SOC estimation system. The system comprises a master control unit, a slave control unit and an upper computer, wherein the master control unit, the slave control unit and the upper computer are communicated through CAN buses, and the master control unit comprises a battery detection system, the SOC estimation system and a battery equilibrium strategy and control system. The SOC estimation system performs SOC estimation according to parameters of batteries in a battery pack, acquired by a current sensor and a voltage sensor. The dynamic SOC estimation is achieved by means of a dynamic SOC estimation method based on a model, an estimation result shows that the dynamic SOC estimation method has good inhibiting effect on noise of the system model and measurement noise, has good correction effect on an initial valve error of the system model and meanwhile has certain robustness on an identification result of a model parameter.

The invention discloses a hybrid electric vehicle belt-driven started generator (BSG) torque ripple compensating controller and a construction method of the hybrid electric vehicle BSG torque ripple compensating controller. The torque ripple compensating controller controls a composite controlled object. The torque ripple compensating controller comprises a PI adjuster, an extended inverter control module and a current compensating module. The composite controlled object comprises a hybrid electric vehicle BSG, a photoelectrical coded disk and a revolving speed and electrical degree calculation module. A neural network is adopted to build the current compensating module, the neural network is of a four-layer structure, and the neural network is trained by the adoption of the Powell algorithm. According to the hybrid electric vehicle BSG torque ripple compensating controller and the construction method of the hybrid electric vehicle BSG torque ripple compensating controller, the current compensating module is recognized through the neural network, the current harmonic wave of a BSG motor can be restrained effectively, and the torque ripple of the motor is reduced significantly, so that the hybrid electric vehicle BSG system has good dynamic and static characteristics and very strong robustness.

The invention relates to a plug-in hybrid power passenger vehicle energy management method based on a self-adaptive equivalent minimum consumption strategy, and belongs to the field of hybrid power passenger vehicle energy management. The plug-in hybrid power passenger vehicle energy management method comprises the steps that S1, a reference value of a power battery SOC is calculated according toa speed spectrum of a plug-in hybrid power passenger vehicle driving cycle; S2, a mathematical model, a longitudinal dynamic model, and a power balance equation between electronic devices of a power transmission system are established; S3,an objective function is established based on the self-adaptive equivalent minimum consumption strategy; S4, the Hamiltonian function is established by the objective function based on the Pontryagin minimum principle, and an optimal initial covariant variable value in the objective function is calculated by using the shooting method to obtain a power distribution strategy between an engine-generator unit and a battery unit. According to the plug-in hybrid power passenger vehicle energy management method based on the self-adaptive equivalent minimum consumption strategy, the process of working condition recognition prediction of the traveling speed is saved, operation is simpler and more convenient, calculation burden is small, the efficiency is high,and the result is strong in robustness.

The invention discloses a method for reconstructing an image based on a blind compressed sensing module and mainly aims at solving problems that only sparse signals can be monitored by traditional compressed sensing, and the quality of the reconstructed image is poor. The method is realized by the following steps: (1) carrying out redundance transformation on an input image to obtain a redundance matrix; (2) carrying out compressed observation on the redundance matrix under an observation matrix; (3) updating a sparse matrix by an OMP (Orthogonal Matching Pursuit) algorithm in an adaptive way according to a compressed observation result; (4) updating a sparse base by a singular value decomposition method according to the updated sparse matrix; (5) multiplying the updated sparse matrix and the updated sparse base to obtain a reconstructed image redundance matrix; and (6) carrying out reverse redundance transformation on the reconstructed image redundance matrix to obtain a reconstructed image; and evaluating the reconstructed image by the peak signal-to-noise ratio of the image. The method has the advantages of high reconstructed image quality and great noise inhibiting effect and can be applied in image denoising and image compression.

The invention provides a filtering backstepping ship movement control system based on a self-adaption fuzzy estimator. The filtering backstepping ship movement control system comprises a control system (2), a guide system (4), a diffeomorphism exchanger (6), a data processing system (7) and a sensor system (12). The guide system (4) obtains the expectation position, the expectation heading and the expectation speed of a ship at each moment. The sensor system (12) comprises a pose sensor (11) and a speed sensor (10). The data processing system (7) comprises a data fusion system (9) and a filtering system (8). The control system (2) comprises a filtering backstepping controller (3) and the self-adaption fuzzy estimator (5). The self-adaption fuzzy estimator (5) simultaneously receives data from the guide system (4) and the diffeomorphism exchanger (6). The filtering backstepping controller (3) simultaneously receives expectation information and differential coefficients of the expectation information which are provided by the guide system (4), new state variable information provided by the diffeomorphism exchanger (6), and estimated output, provided by the self-adaption fuzzy estimator (5), for unknown nonlinear functions.

A method and system for sliding mode control of permanent magnet synchronous motor based on model-free and non-singular terminal are disclosed in the invention. As compare with that traditional PI controller, the model-free and non-singular terminal sliding mode control method adopt by the invention can reduce the dependency of the controller on the system model, is more suitable for the nonlinearsystem such as the permanent magnet synchronous motor, and simultaneously adopts a sliding mode observer to estimate the unknown quantity, thereby enhancing the robust performance of the method of the invention. The control method of the invention has fast response speed and high control precision, and has certain fault-tolerant control function for permanent magnet loss-of-excitation fault, so that the permanent magnet synchronous motor can run efficiently and reliably under normal conditions or permanent magnet loss-of-excitation conditions.

The invention provides a finger vein identification method based on relative distance. The method comprises the steps of: carrying out partitioning and refining operations on a read-in finger vein image, extracting end points and intersections of the refined finger vein image to be used as a characteristic point set, defining the structural types of characteristic points, and finally carrying out matching of the finger vein image by calculating distances among the characteristic points for identity recognition. The invention is simple and easy to operate without positioning, reduces work load, improves recognition speed and accuracy, effectively overcomes influences of translation, rotation and the like on a recognition result, ensures that the system is improved in the aspect of recognition effect and has practical application value and development potential.

The invention discloses a star sensor star map identification method based on a convolutional neural network. The star sensor star map identification method comprises the following steps: carrying outstar filtration treatment on the original star catalog, establishing a navigation star catalog, carrying out statistic on constellations of the whole celestial sphere navigation stars, and numberingthe constellations, wherein a sample library is composed of a simulation star map and numbers of the constellations corresponding to the most stars; replacing the original star map with a sparse matrix, inputting a sample library star map into the convolutional neural network, and carrying out training; carrying out star image extraction on a star map obtained by shooting, converting the extractedstar image into a sparse matrix, then inputting into the convolutional neural network, carrying out coarse attitude star map identification, and obtaining a rough orientation; and identifying a fixedstar in a view field by applying a local sky area star map identification algorithm. The star sensor star map identification method disclosed by the invention has the advantages that a trained convolutional neural network is adopted for realizing coarse attitude whole celestial sphere star map identification, the navigation star catalog does not need to be searched, and the local sky area star map identification only needs to search a small part of a database; and the convolutional neural network has the capability of autonomously extracting characteristics of the original map and has stronganti-noise and anti-fake-star performance when being applied to star map identification.

The invention belongs to the technical field of computer vision, and particularly relates to an improved PSPNet-based semantic segmentation method, a lightweight MobileNetV2 network is used for learning feature information, a context semantic feature supplementing module is introduced to reserve more feature, a level set method is introduced as network post-processing after classification prediction is performed on each pixel point in an image in a network, so that a segmentation result is closer to a real contour of a target, and finally, the purpose of semantic segmentation is achieved. Thenetwork model is high in robustness in image semantic segmentation, and segmentation errors are reduced; the MobileNetV2 network with 17 anti-residual units is used as a front-end network of + PSPNet,so that the whole network tends to be lightweight; a context semantic feature supplementing module is introduced, and on the original basis, the problems of feature loss, fuzziness and the like in the sampling process are solved by means of an attention mechanism; and a level set method is introduced into the whole model as a post-processing mode, so that the segmentation precision of the whole model is improved.

The invention discloses a target tracking method based on HSV color covariance characteristics. The method mainly solves the problems that according to an existing tracking technology, the characteristics extracted from a color image target are large in information redundancy and poor in independence. The method comprises the implementation steps of firstly obtaining a candidate target under a particle filter framework through the target state predication, extracting hue, saturability, brightness and Laplacian response of a candidate target image to serve as the apparent characteristics, performing fusion to build a covariance operator, adjusting a tracking window and updating a characteristic template by calculating the similarity weight of the candidate target characteristics and the template, and finally updating the target state and effectively tracking the target according to the weight fusion candidate target. According to the method, the information redundancy of the target characteristics can be effectively reduced, the independence between the characteristics can be enhanced, the precision and robustness of target feature description can be improved, and the real-time precise tracking on a color image target can be achieved.

The invention discloses a defect detection algorithm based on a deep neural network Mask R-CNN, and belongs to the technical field of defect detection. The algorithm comprises the following specific steps of extracting features by using a feature pyramid network (FPN) based on ResNet50; extracting regions of interest (ROI) of a defect region by using a region proposal network (RPN) so as to obtaincorresponding anchor boxes; using a full convolutional neural network (FCN) to predict a pixel category in the ROI so as to realize defect segmentation; and finally, realizing prediction of the category to which each ROI belongs and corresponding anchor frame coordinates through a full connection layer of the network. Aiming at a magnetic shoe surface defect detection scene, the algorithm performs two improvements on a feature pyramid network (FPN) in MaskR-CNN: a C1 module is added in the FPN, and a pooling layer in a feature extraction layer of the C1 module is cancelled; and a CLAHE preprocessing module is added in front of a feature extraction layer of the FPN. Experimental results show that the algorithm of the invention has strong generalization ability and robustness, and can perform accurate defect segmentation on a magnetic shoe image.

The invention discloses a multi-time-phase remote sensing image building changing detection method based on an image block. The multi-time-phase remote sensing image building changing detection methodbased on the image block firstly inputs building detection rough results of different time phases, equally divides a length and a width of each detection result into N parts, obtaining a ratio of pixel quantities of a building in a same area of two time phases at any random time interval, divides a building change in this area into three changing modes which include 'dramatic increase', ' substantial unchanged' and 'dramatic decrease', marks the changing mode of each area on an image of a later time phase and outputs a detection result. Furthermore, the multi-time-phase remote sensing image building changing detection method based on the image block can perform quantitative analysis on the area change on the basis of ratios of building areas of the two time phases at the random time interval. The multi-time-phase remote sensing image building changing detection method based on image block can comprehensive consider information according to a context, overcomes noise in a high resolution image, effectively realizes fast automatic changing detection of the building, fully mines the changing mode of the building and reaches ideal accuracy when the building detection result is an average.

The invention relates to a fuzzy control method for an electric control system of a diesel locomotive, which is characterized in that after a fuzzy controller is used to fuzzily quantify the deviation between a power given value and a feedback value and the deviation change rate, the fuzzy quantity output is determined by a control rule table, After the fuzzy control quantity is defuzzified, the precise quantity is used for the input of the excitation current loop of the main generator. Improve the robustness of constant power control of electric-wheeled vehicles; reliably control complex systems that are difficult to establish accurate mathematical models; apply intelligent control theory to the transmission control of internal combustion locomotives, select more appropriate PI parameters, and load Or when the speed changes suddenly, the system response time is faster.

The invention discloses an intelligent robot grabbing method based on action demonstration teaching, and relates to the technical field of robot learning. The method comprises the following steps thatthe hardware environment building of an action demonstration teaching programming system is completed; a person demonstrates the grabbing operation to form a human teaching action video, and the person uses a demonstrator to control a robot to complete the demonstrating grabbing action to form a robot teaching action video; data of human and robot teaching action videos are gathered to be subjected to denoising and expanding operation; a meta-learning algorithm is adopted to directly and automatically learn priori knowledge from teaching actions of human and the robot so as to realize learning of new tasks. According to the meta-learning algorithm provided by the intelligent robot grabbing method based on the action demonstration teaching, one-eye learning of imitation learning can be realized in different background environments, different human demonstrators and different robots, and learning of an adaptive target loss function is realized by using time convolution, so that the network can capture multiple frames of human action image information at the same time; and the method has strong adaptability and robustness.

The invention discloses an action quality evaluation method based on uncertainty score distribution learning, and the method comprises the steps of taking a score label as a mean value, and generatingGaussian distribution as a supervision signal; sending the action video into a 3D neural network to predict a score label; optimizing the network by optimizing the KL divergence between the prediction score label and the supervision signal; and inputting the test video into the optimized prediction video score model, and selecting the score with the maximum probability value as the final prediction score. According to the method, the probability of the action quality score can be better described, and the ambiguity problem in the action quality evaluation score label is solved.

The invention discloses a tunnel cavity state radar spectrum image recognition model construction method and a tunnel cavity state radar spectrum image recognition method, and belongs to the technicalfield of image recognition. The method comprises the following steps: extracting time domain abstract semantic features from shallow to deep in a tunnel cavity state radar spectrum image through a convolutional neural network, and constructing a multi-source time-frequency domain complete global observation feature space of the tunnel cavity state radar spectrum image by fusing frequency domain gray statistical features extracted by Fourier transform; performing feature splicing on different feature vectors to obtain a final feature vector; and inputting the final feature vector into an SCN classifier to obtain a heterogeneous depth feature network model, so that the method is closer to a human vision model, the time domain and frequency domain features of the image are fully utilized, and the method has relatively high adaptability and robustness.

The invention provides an underwater acoustic leading signal detection method based on an accumulative correlation coefficient (ACC) under a sparse channel, which is applied to an underwater acoustic communication system under the sparse channel and can be used for reducing the false detection rate of leading signals, increasing the detection rate, improving the detection performance of the system, and increasing the communication efficiency of the system. The method comprises the following steps: performing sparse signal reconstruction and OMP algorithms, achieving the separation of main paths of the sparse channel by using a reconstruction process of the signals, calculating the relevancy between the signals of each path and the transmitted signals, accumulating to obtain a correlation coefficient, and performing leading signal detection based on the correlation coefficient. Compared with the prior art, the underwater acoustic leading signal detection method provided by the invention has the advantages as follows: 1, the method has strong robustness under additive white Gaussian noises and different types of interference; 2, compared with a detection technique based on a matching filter, the method has better detection performance under multi-path channels; and 3, compared with other prior arts, the method has more ideal detection performance under complicated and changeable actual underwater environments.

The invention relates to a sliding mode control-based MMC (Modular Multi-Level Converter) circulating current suppression method. With the method adopted, the quick suppression of the internal circulating current of a modular multi-level converter (MMC) device can be realized. According to the method, feedback linearization is carried out to decouple a system, and a variable-structure controller is designed to perform circulating current suppression. The method comprises the following main steps that: the current of upper arms and lower arms is measured, and internal circulating current is calculated; Park transformation is performed on the circulating current; the difference of a measured value and a reference value is calculated, and the difference is substituted into exponential reaching law calculation, an intermediate variable is calculated and is transmitted to a sliding mode controller, and the quadrature-axis component and direct-axis component of a circulating current calculation voltage value are obtained by means of the calculation of the controller; and the components of the circulating current calculation voltage value are substituted into a circulating current calculation formula, so that the reference voltage values of the upper arms and lower arms can be obtained; and the reference voltage values are transmitted to a modulation module, and trigger signals of modulated waves are generated, so that the on-off of sub-modules in the MMC can be controlled. When the sliding mode controller designed in the method performs circulating current suppression, the sliding mode controller has higher response speed and has higher stability after suppressing frequency-doubled circulating current.

The invention relates to an artificial intelligence technology, and discloses a disease analysis method, which comprises the steps: extracting feature data from case information of a preset type of disease, and performing vectorization processing to obtain a feature vector; training a to-be-trained disease analysis model according to the feature vector and an initial model gradient obtained from aserver to obtain an updated model gradient, and sending the updated model gradient to a server through encryption processing; when the model gradient in the convergence state transmitted by the server is received, updating the model gradient of the disease analysis model to be trained by using the model gradient in the convergence state to obtain a standard disease analysis model; and analyzing the case information of the target user by using the standard disease analysis model to obtain an analysis result of the preset type of disease. The invention further provides a disease analysis device, electronic equipment and a storage medium. The invention also relates to a blockchain technology, and the analysis result can be stored in a blockchain node. The accuracy of disease analysis can beimproved.

The invention discloses a multi-focusing image fusion method based on two-dimensional coupling convolution. The method includes the following steps: inputting an image to be fused, and dividing the image into small blocks with the size of M*N through a sliding window technique; and for a single-input and multi-output system model, solving an imaging system F<1><j> and an imaging system F<2><j> in the model through characteristic constant decomposition, and thus selecting clear image blocks. The method of the invention can accurately determine clear image blocks, so that a fused image exhibits great robustness. Moreover, significant clear details of the image can be well fused without artificial traces.

The invention discloses a depression degree evaluation method, system and device and a storage medium. The method comprises the following steps: acquiring a first video, wherein the first video is a video containing character expression change and character voice; randomly intercepting a first group of pictures, wherein the first group of pictures are continuous multi-frame pictures in the first video; inputting the first group of pictures into a facial expression recognition model for processing to obtain a plurality of expression feature vectors; inputting the plurality of expression feature vectors into a micro-expression depression recognition model for processing to obtain a first score; extracting a first audio, wherein the first audio is a voice segment corresponding to the first group of pictures; inputting the first audio into a voice depression recognition model for processing to obtain a second score; and evaluating the depression degree according to the first score and the second score. According to the method, the depression degree is evaluated by effectively combining the two indexes of voice and micro-expression, the evaluation precision is higher, and the robustness is higher. The method can be widely applied to the technical field of depression evaluation.

The invention relates to a Beidou No.3 short message channel scheduling method and system, and provides a terminal-satellite-ground station cooperative transmission mechanism based on an inter-satellite link and a satellite-ground link for a short message packet through operations of initiating a registration request, calculating, searching a table, broadcasting, forwarding, updating the table, returning and the like. By generating respective object service lists in a plurality of satellites and a ground central station, a transmission mechanism is shielded from opposite-end equipment, and the load of a terminal is reduced while the short message data packet is ensured to be accurately transmitted in a channel. The Beidou terminal card has the beneficial effects that single-transmitting and single-receiving, single-transmitting and multi-receiving, multi-transmitting and single-receiving and multi-transmitting and multi-receiving modes of the Beidou terminal card can be realized; all channel links are mutually redundant, and when some facilities break down or are damaged, the robustness is still high.

The invention discloses a visual positioning method of a mobile manipulator. The visual positioning method at least comprises the following steps: performing planar visual positioning by a camera, wherein the planar visual positioning comprises positioning of two translation degrees of freedom and one rotation degree of freedom; performing deep visual positioning compensation by a flexible mechanism, wherein the deep visual positioning comprises positioning of translation degree of freedom, preferably, the planar visual positioning comprises a step of obtaining an error of the translation degrees of freedom and an error of the rotation degree of freedom through a template matching mode. According to the visual positioning method, the technical problem of how to improve the robustness of the mobile manipulator is at least solved.

The invention discloses a structural damage assessment method based on distributed vibration data and convolutional self-encoding deep learning, and the method comprises the following steps: S1, selecting acceleration response monitoring points, and arranging an acceleration sensor at each monitoring point; s2, acquiring monitoring data of n acceleration sensors in a normal use state of the structure, and performing data preprocessing to form a data set for deep learning network training; s3, constructing a convolutional self-encoding deep learning network suitable for the data set in the step S2; s4, preprocessing massive structure monitoring data in a normal use state according to the step S2, and inputting the preprocessed data into a convolutional auto-encoder for training to obtain a deep learning network file; and S5, evaluating a structural damage state through a data reconstruction correlation function. According to the method, the data does not need to be pre-classified, the structure damage state is quantified in real time by using the real-time vibration monitoring data, and the score is given.

In order to solve the defects in the prior art, the invention provides a visual indoor positioning method and system based on architectural plane graph prior information, and the method comprises the steps: architectural plane graph preprocessing, data downsampling, input tensor construction, feature extraction and feature matching, registration and pose calculation, and model training. The method provided by the invention relates to an end-to-end deep learning model, the model innovatively introduces a K nearest neighbor distance histogram, a point cloud feature extraction network, a graph attention network, an optimal transmission algorithm and a self-supervised training method, so that the model has relatively strong robustness and adaptability, and the model has relatively small parameters and is easy to train and deploy; according to the method provided by the invention, the position of the equipment in the building body can be estimated only according to the building plane graph under the condition that the building body is not completely explored.

The invention discloses a wireless signal area intensity detection method based on deep learning. The method comprises the steps of carrying out short-time Fourier transform to obtain a two-dimensional signal speech graph; cutting, filling and deforming to obtain a speech graph sample; making a data set label; training the customized data set by using a deep learning one-dimensional compression model to obtain a model weight file; performing pretreatment; predicting the preprocessed picture to obtain regional signal intensity probability output; and obtaining position information of the signal and an image quality prediction value of the signal, and obtaining an optimal region. According to the invention, signal detection under complex conditions can be processed, and the robustness is very high; intensity probability output is carried out on each region of the signal, signal detection is realized, the pertinence is strong, and the false detection rate is low; and the position information of the signal can be obtained, and the quality evaluation of each region of the signal can also be obtained.