372results about How to "Fast convergence" patented technology

The embodiment of the invention provides a fraud recognition method and a fraud recognition device. The method comprises steps: a transaction data sample set is acquired, wherein each piece of transaction data in the sample set at least comprises fund flow information, transaction party correlation information and transaction data fraud risk attributes; a fraud spreading relation graph is built according to the sample set, wherein each node in the fraud spreading relation graph is built according to the transaction party correlation information, and the directed edge between the nodes is builtaccording to the fund flow information and the transaction party correlation information; an initial node value is assigned to each node in the fraud spreading relation graph; according to a PageRankiterative updating algorithm, the node value of each node is subjected to iterative updating until iterative convergence to obtain a final node value; and if the final node value is higher than a preset threshold, the corresponding node is determined to be a fraud person node, and the transaction data correlated to the fraud person node are fraud risk data.

The invention relates to a multiscale fusion CNN-based super-resolution magnetic resonance image reconstruction method. The method comprises the following steps of: firstly preprocessing a low-resolution image and a corresponding high-resolution image, and constructing a training data set and a label data set; constructing a multiscale information-fusion full-convolutional neural network; inputting the training data set into the constructed multiscale information-fusion full-convolutional neural network to carry out training so as to obtain a learnt convolutional neural network model; inputting the test low-resolution image into the learnt convolutional neural network to obtain a reconstructed high-resolution image. Through a multiscale fusion unit, features of different convolutional layers are mapped and fused, so that flat structures stacked by multiple convolutional layers of traditional convolutional neural networks are overcome, the convergence speeds of the networks can be improved, lost image details of low-resolution images can be reconstructed more rapidly, the reconstruction time can be shortened, the reconstruction efficiency can be improved and the resource waste can be avoided.

The present invention discloses a series-wound long short-term memory recurrent neural network-based heating load prediction method. The method comprises the steps of constructing a sample data set based on temperature, climate and heat supply data during a given period of time, and respectively subjecting the input data and the output data of the sample data set to standardized treatment; dividing the input data into two portions, respectively inputting the two portions into two independent long short-term memory recurrent neural networks to merge the two portions of the input data, inputting the output data to a long short-term memory recurrent neural network at a next layer, and finally inputting the data into two full connection layers; training a constructed series-wound long short-term memory recurrent neural network, and optimizing the network by adopting the parameter optimization-based adaptive torque estimation algorithm; inputting to-be-predicted data into the series-wound long short-term memory recurrent neural network, calculating and obtaining a heating load prediction result. The method of the invention can effectively discriminate input data, and accelerate the learning speed. Therefore, the learning efficiency is improved and the prediction accuracy is improved.

A cable damage localizing method for a cable-stayed bridge on the basis of improved back-propagation neural network is disclosed. The invention has the method that a BP neural network can be improved effectively based on how to effectively perform, thus a cable damage localizing of a cable-stayed bridge is carried out, and therefore, the cable damage localizing method for the cable-stayed bridge with the improvement is established. The invention proposes to improve traditional BP algorithm jointly with Bold Driver technology, momentum item increase, simulated annealing algorithm and random hill-climbing algorithm. The cable damage localizing method has the advantages of accelerating convergence, improving learning-efficiency improvement and preventing the learning process from falling into local minimum and, in addition, the invention can reasonably determine network topologies.

The invention discloses a user comment emotion analysis system and a method based on an attention convolution neural network. The invention is a modular design and mainly comprises four modules, namely, a word embedding module, a convolution module, an attention module and a classifier module. The word embedding module uses the low-dimensional vector to represent the comment text, the convolutionmodule extracts the local features of the comment by convolution operation, the attention module determines the weight of the local features by comparing the similarity, and calculates the final feature expression of the comment by weighting, and the classifier module classifies the emotion according to the final feature expression. The invention overcomes the shortcomings of the traditional neural network model feature extraction method by adding the attention mechanism into the neural network model. After a large number of data training, the attention mechanism can judge the importance of different words in comments, so that the model can notice the part of comments that has the greatest impact on emotion, and improve the accuracy of emotion classification.

A spacecraft robust finite time saturation attitude tracking control method is disclosed. The present invention relates to the spacecraft robust finite time saturation attitude tracking control method so as to solve a rigid spacecraft attitude tracking control problem in the conditions of model uncertainties, external disturbance torques and actuator saturation and is aimed at the problems of controller chattering, a complex controller structure, many setting parameters and a limited application range of a control algorithm in an existing method. The method of the invention comprises the steps of (1) establishing a rigid spacecraft attitude kinematics and dynamics model which is an attitude tracking system, (2) defining a rapid nonsingular terminal sliding mode surface and auxiliary system according to the step (1), and (3) carrying out robust finite time saturation attitude tracking controller design, carrying out controller designing when the synthetical uncertainty delta upper bound of the attitude tracking system is an unknown constant, and carrying out adaptive controller designing when the synthetical uncertainty delta upper bound is an unknown function. The spacecraft robust finite time saturation attitude tracking control method is used for the aerospace field.

The invention discloses a multi-UAV (Unmanned Aerial Vehicle) cooperation path planning method based on an ant colony algorithm. The method comprises the following steps of (1) analyzing a UAV flight environment, and building environment modeling based on a Voronoi diagram; (2) calculating consideration of an edge in the environment modeling based on the Voronoi diagram; (3) utilizing the ant colony algorithm to plan an initial path for a UAV; (4) smoothing the initial path of each UAV to judge whether cooperation is achieved or not, and carrying out corresponding operation according to the results. According to the multi-UAV cooperation path planning method based on the ant colony algorithm provided by the invention, multiple different kinds of unmanned aerial vehicles are mutually cooperated, intelligently and autonomously adapt to complicated and changeable war environment factors, and dynamically adjust self-strategies so as to cooperatively accomplish the combat mission.

A large-scale scene three-dimensional reconstruction method for fusion of additional information includes: extracting SIFT (scale invariant feature transform) points of all images, performing image matching, and structuring external-pole geometric graphs to obtain trajectories corresponding to all three-dimensional spots; according to inertial measurement unit information or compass angles, obtaining initial camera rotation matrixes of all images, iteratively searching currently reliable connecting edges from the external-pole geometric graphs and performing global optimization by the aid of the edges; initializing the center of a camera to be a GPS (global position system) corresponding to the images to obtain initial projection matrixes of the images according to image initializing focus information, the rotation matrixes and the center of a camera, and iteratively triangulating and adjusting in bundle according to the projection matrixes and the trajectories of the three-dimensional spots. The large-scale scene three-dimensional reconstruction method is rapid in calculation, the obtained three-dimensional spots are reasonable and reliable, image mismatching sensitiveness is low, generalization performance is high, and the method is applicable to both orderly and disorderly image sets.

The invention discloses a method for evaluating the electric power system risk based on fault pre-scanning. The method comprises the steps that firstly, according to the current structure of an electric power system and a planning scheme, a corresponding planning scheme model is established; secondly, based on device data and a power grid structure, predicted faults and the possibility of the predicated faults of a power grid are analyzed, and then a pre-scanning fault set is formed; thirdly, through the risk state based on direct-current power flow, screening and ordering are conducted on the predicated fault set, so that a fault state list is formed; fourthly, all power flow corresponding to the faults in the list is worked out, and for the fault with the power flow out of range, the consequence of the fault is load loss; fifthly, according to the load loss of the fault states and the possibility of the fault states, the system risk indicator is worked out; sixthly, according to the system risk indicator, the system weak link is found out. By the adoption of the method for the evaluating electric power system risk based on fault pre-scanning, analysis of all the fault sates in the fault set through alternating-current power flow is avoided, so that the calculation amount of follow-up alternating-current power flow analysis and risk indicator processing is greatly reduced, and a large mount of calculation time is saved.

Power flow models of Interline Power Flow Controllers (IPFC) for large-scale power systems are studied, in details. Mathematical models of the IPFC, using the d-q axis decompositions of control parameters are derived. In this framework, for each IPFC, only two control parameters are added to the unknown vector in the iteration formula and the quadratic convergence characteristic is preserved. Simulations results from several practical large-scale power systems embedded with multiple Convertible Static Compensators (CSCs) demonstrate the effectiveness of the proposed models. Comparisons with existing models are made to elucidate the performance of the convergence.

The present invention discloses a medical image segmentation method based on a dual-path U-shaped convolutional neural network. A dual-path module and a U-shaped network are combined, original features are reused through stacking operation, and features are added to generate new features so as to improve the network training precision; the U-shaped convolutional neural network configuration is employed to perform transverse feature stacking between blocks so as to greatly improve the network training convergence speed and more rapidly obtain an ideal training model. The medical image segmentation method can meet the demands of data size of medical images requiring being processed is large and high processing efficiency and high precision requirements and can achieve a very good performancein the biomedicine detection application.

The invention discloses a partial discharge pattern recognition method based on a mixed neural network algorithm, the method comprises a judge process including extracting ultrahigh frequency partial discharge signal characteristic and establishing a value set library and discharge types; the method specifically comprises: establishing a typical partial discharge model firstly, acquiring a discharging ultrahigh frequency signal, and performing mixing frequency and reducing frequency process; then generating 5 kinds of two-dimension spectra according to the discharge signal, and extracting 37 kinds of statistical characteristic quantity (value set) to form the value set library; finally comparing the value set library and a value set corresponding to the statistical characteristic quantity calculated by a fault signal through the mixed neural network algorithm, and recognizing the partial discharge type of a transformer. According to the invention, the signal of partial discharge ultra wide band is fully ultilized, the single neural network pattern recognition problem is overcome, the partial discharge value set library of the transformer is established to perform pattern recognition for different kinds of discharging, the valuable data is provided for transformer partial discharge on-line monitor, and the method has good practical engineering value.

The invention belongs to the cross technical field of the computer vision and natural language processing, and provides an image title automatic generation method based on multi-modal attention. The alignment problem of vision feature and language feature and the sentence feature ignoring problem in the word prediction process in the traditional method based on neural network are solved, the convergence speed and the quality of the image title are improved. The method comprises the following steps: automatically performing feature extraction on an image region by utilizing a convolutional neural network; and then realizing sentence feature extraction by utilizing LSTM with vision attention; and finally producing a final image title by designing the LSTM with the multi-modal attention (vision attention and hidden variable attention). The experiment proves that the proposed method has excellent result on the MS COCO and like standard data set.

An object of the invention is to provides a particle swarm optimization method based on a multi-strategy synergistic function. The technical scheme of the method comprises a first step of initializing a particle swarm, wherein NP particles are initialized, a second step of calculating fitness values of the NP particles, a third step of determining particle speeds and position change modes, a fourth step of executing Cauchy variation on positions of the particles, and a fifth step of determining a particle execution stop condition. The method is suitable for optimization solving of a function, takes the full advantage of elite reverse learning, and improves a function optimization speed and function optimization precision; and the particles are prevented from falling into a local optimal value by utilizing gauss variation, and variation is performed on the positions of the particles by utilizing the provided Cauchy variation a Cauchy distribution proportion parameter of which decreases linearly and progressively, so more excellent particles are generated to guide other particles to move toward a better solution direction, the function optimization precision is improved, and function optimization stability are also improved.

The invention discloses a polarization code decoding algorithm based on deep learning. A multi-dimension zooming Min-sum belief propagation decoding algorithm is provided to accelerate the convergence speed of the decoding algorithm; and then, a polarization code decoder based on a deep neural network is realized according to the similarity between the factor graph of the BP algorithm and the deep neural network, a deep neural network decoder is trained by using the deep learning technology, compared with the original BP decoding algorithm, the decoding iterations of 90% is reduced, and meanwhile better decoding performance is obtained; and finally the invention provides hardware implementation of a fundamental operation module of the deep neural network polarization code decoder, and 50% of hardware consumption is reduced by using the hardware folding technology.

The invention relates to an adaptive amplifying digital baseband pre-distortion method. The method comprises the following detailed steps of: step A, judging whether an input signal and an output signal of a pre-distortion system are equal or not, if so, disconnecting with a pre-distortion training module, otherwise, performing step B; and step B, using a training method of a memory multinomial coefficient of Kalman filter to set the to-be-estimated memory multinomial coefficient of the pre-distortion training module to be a state variable, establishing a state space model, and using an estimation value of the prior time and a sampling value of the present time to update the state variable to acquire the estimation value of the present time. Compared with the traditional pre-distortion scheme, the technical scheme provided by the invention has the advantages of high robustness and convergence rate; and the requirement of updating parameters in real time can be met through the technical scheme.

The invention provides an intelligent vehicle dynamic obstacle avoidance method and an intelligent vehicle dynamic obstacle avoidance system based on deep reinforcement learning. The intelligent vehicle dynamic obstacle avoidance method comprises the steps of: S1, acquiring an image of an intelligent vehicle at a moment t; S2, inputting the image into a neural network model, and outputting a probability corresponding to execution of each action by the intelligent vehicle; S3, selecting an executed action of the intelligent vehicle at the moment t; S4, recording simulation data of the intelligent vehicle at the moment t; S5, setting t to be equal to t+1, repeating the S1-S4 until the simulation ends, and archiving the simulation data; S6, and transferring simulation data from a positive sample experience pool or a negative sample experience pool for training the neural network model while circulating the S1-S6 continuously until a dynamic obstacle avoidance strategy of the intelligent vehicle can perform dynamic obstacle avoidance completely in the simulation process. The trained dynamic obstacle avoidance strategy is applied to the dynamic obstacle avoidance under a man-machine cooperation mechanism, so as to complement the advantages of a human driver and an intelligent machine on behavior decisions of emergency obstacle avoidance of the intelligent vehicle, thereby achievinga unified and superior decision-making method. The man-machine cooperative dynamic obstacle avoidance method is applied to the field of intelligent decision making of intelligent vehicles.

The invention provides a method of enhanced satellite-based navigation using low earth orbit (LEO) satellites and comprises the following steps. Step 1: obtaining through the Internet the IGS precisesatellite ephemeris and satellite clock data, the GNSS system integrity warning data, and the corrected ionosphere parameters, etc., and uploading the data from ground station to LEO navigation satellite through the upload channel. Step 2: completing the real-time, high-precision positioning and the high-precision timing of the satellite in orbit. Step 3: the LEO navigation satellite compiles itsorbital parameters, the precise IGS data, and the corrected ionosphere parameters for transmission. Step 4: the high-precision receiver on ground receives the GNSS signal and the LEO satellite navigation signal and outputs the high-precision navigation and positioning. The invention makes real-time and high-precision orbit determination and provides enhanced navigation for users on the ground based on the uploaded precise IGS satellite ephemeris data and the high-precision dynamic model.

The invention discloses a self adaptive flow control method used for the data transmission of wireless video supervisory system, which comprises following steps, (1) initializing the video compression chip;(2)activating data frame protecting thread, and estimating the situation of physical channels according to the overflow condition of buffer area so as to adjust the frame loss value; (3)activating users terminal feedback information receiving process thread and monitoring and receiving the network situation information sent back from the user terminal so as to adjust the frame loss value; (4) activating self adaptive flow control thread, and adjusting the terminal object code rate in a self adaptive way according to the frame loss value in a certain period; (5) activating the sending thread. Through the detection of the buffer utilization factor of a control chip on the monitoring terminal and the packet loss information feedback generated in network transmission of user terminal, the invention can calculate the frame loss value, measure the necessity of code rate changes and adjusting step length of code rate, and choose the biggest data transmission speed in a self adaptive way so as to get the best video quality.

The invention discloses a repeater with internet connection sharing (ICS) function and an implementation method thereof. The device is used for generating a feedback echo signal in a numeric field and cancelling partial echo signal interference in the received signal. The echo cancellation method comprises three parts, namely pseudo code insertion, channel initial estimation and adaptive tracking, wherein the inserted pseudo code is a constant amplitude zero auto correlation (CAZAC) code; in the channel initial estimation part, by utilizing the characteristic that the code is strongly auto correlated and weakly mutual correlated, the impact response and delay information of a channel are estimated through related operation; and an adaptive variable step-size algorithm is adopted in the adaptive tracking part, and the channel impact response obtained through initial estimation is used as an initial value of operating the algorithm, so that the convergence rate of the algorithm is improved. The invention has the advantages that: the convergence rate of an adaptive filter can be improved, and the performance of an echo canceller is improved; meanwhile, the algorithm complexity is not high, the calculated amount is small and the algorithm has implementability.

A method for raising convergence speed of route information protocol route includes detecting whether neighbor router can be always reached or not by router using route information protocol according to quick detection session set up with neighbor router and deleting route item correlated with said neighbor router when neighbor router not being able to be reached is detected. The device for realizing said method is also disclosed.

The invention discloses an apparatus and a method for detecting harmonic current. Voltage and current signals measured by a sensor enters a sampling maintaining A/D converting circuit through an anti-aliasing filtering circuit; a lock phase loop synchronization pulse generating circuit provides a conversion signal for the sampling maintaining A/D converting circuit; a digital signal processor uses an LMS/LMF self-adaptive harmonic current detection algorithm to detect a harmonic current value in load current; a keyboard circuit and a colorful liquid crystal display realize a man-machine conversation; and a ferroelectric storage stores data. The invention uses the improved LMS/LMF algorithm to update weight and uses the self-correlation estimation of the normalized prior error signal and the last error signal during a step-size iteration, thereby effectively avoiding the influence caused by a lower signal-to-noise ratio and obtaining the better stable effect. The invention uses a TMS320F2812 chip to realize on the specific algorithm; a DSP program is easy to operate to continuously detect signals to be detected for a long time; and the hardware circuit of the whole apparatus is simple, small and convenient to carry and store data.

The invention discloses a method for predicting the available parking space occupancy of a parking lot in a short time. The method comprises the following steps of: performing multi-resolution and N-scale decomposition on the time sequence c0 of the available parking space occupancy by using a wavelet function so as to obtain a low-frequency coefficient vector cN of which the scale is N and N high-frequency coefficient vectors d1, d2, ..., dN with different scales; reconstructing the low-frequency coefficient vector cN and the high-frequency coefficient vectors d1, d2, ..., dN so as to obtain N+1 time sequences CN, D1, D2, ..., DN; and establishing a weighted Markov chain prediction model for predicting to obtain N+1 prediction results and accumulating the N+1 prediction results so as to obtain a prediction result which corresponds to the original available parking space occupancy. A wavelet analysis-weighted Markov combined prediction model is provided according to the short-time change characteristic of the available parking space occupancy of the parking lot, so that convergence rate is increased, and a network parameter and a training parameter are easy to determine.

Techniques for an adaptive synchronous switch in switching regulators are described, one aspect of which is to achieve a more optimal on/off timing of a synchronous switch that is controlled by the comparator in a feedback control loop and thereby improves power conversion efficiency and system performance; One approach samples a node in the output of the switching regulator and generates a sampled error signal that is analyzed to determine if the current comparator offset is too high or too low relative to a target switching regulator output value at least in part based on the sampled error signal value, and accordingly generates a compensated feedback error signal and applied to the compensated feedback error signal to an input of the comparator to have the effect of a comparator offset adjustment signal.

The invention brings forward a linear Kalman filter algorithm of a butterfly structure, and is used for carrying out dynamic balancing on reception signals after dispersion recovery in a coherent light communication system receiver, ie, de-multiplexing of the signals and compensation for inter-symbol crosstalk due to residual dispersion, bandwidth restriction, polarization mode dispersion, etc. are simultaneously realized. The realization steps are as follows: a tap matrix used for carrying out de-multiplexing on the signals and balancing the inter-symbol crosstalk is firstly predicted; output signals are calculated according to the predicted tap matrix and input signals; an auxiliary constraint is determined through utilization of a radius; updating of tap coefficients is realized through the linear Kalman filter; and the updated values are regarded as predicted values of iteration at the next moment; and the output signals are enabled to reach an optimal dynamic balancing effect. The algorithm brought forward by the invention has advantages of simultaneously realizing de-multiplexing and inter-symbol crosstalk compensation, suppressing singularity, being independent of phases and symbol types, having a fast convergence speed and a fast trackable polarization rotation rate.

The invention discloses an improved local information-based model image segmentation method, and mainly solves the problem of false segmentation caused by non-ideal non-uniform grayscale image segmentation effect of an improved CV model and relatively low robustness of a local information-based model to an initial contour at present. The method is implemented by comprising the steps of inputting an original image and setting the initial contour; setting default parameters and important parameters; combining a global grayscale fitting value and a local grayscale fitting value of an improved kernel function into a new grayscale fitting value of a weighted target and a background; obtaining gradient descent flow by utilizing an energy functional of a CV model in which penalty terms are introduced; and evolving a level set function according to a level set iteration formula, and through iteration, outputting a segmentation result. According to the method, a non-uniform grayscale image is effectively segmented; the robustness to the initial contour is enhanced; the initial contour is converged to a target contour more quickly; compared with other related models, the method has higher segmentation precision and efficiency; and the method is used for segmentation of artificially synthesized images, non-uniform grayscale images and infrared images.

The invention provides a method and a device for echo cancellation, wherein a signal strength value of a distal end signal is used as one of parameters for adjusting a simulation echo signal transfer function; nonlinear factors are led in when the signal strength of the distal end signal exceeds the strength threshold of an abnormal signal, so that the convergence speed of the echo simulation algorithm is increased, and the problem that non-convergence is possible to be caused by too high signal strength of the distal end signal is avoided; in addition, the adaptive capacity of the method forecho cancellation on the environment is improved, so that the effect of echo cancellation can be improved.

The invention discloses a maximal power point tracking (MPPT) control method for a single-phase single-stage photovoltaic inverter. The method comprises the steps of: sampling output current and output voltage of a photovoltaic cell, calculating output power of the photovoltaic cell; respectively filtering direct current components of the output voltage and the output power through high-pass filtering to obtain alternating current components of the output voltage and the output power; regarding the alternating current component of the output voltage as disturbance quantity of optimizing control of an extreme value, multiplying the disturbance quantity by the alternating current component of the output power, and then obtaining estimated work voltage of the maximal power output point by low-pass filtering and integration, superposing the disturbance quantity on the estimated work voltage, and repeatedly carrying out the steps to ensure that the photovoltaic cell works near the maximal power output point all the time. According to the maximal power point tracking control method for the single-phase single-stage photovoltaic inverter disclosed by the invention, the inherent quality of the system is sufficiently utilized without additionally injecting a disturbance signal, and the effect of the external disturbance signal on the system is eliminated. In addition, the maximal power point tracking control method for the single-phase single-stage photovoltaic inverter can ensure less calculation of the MPPT control method, thereby facilitating implementation.

A method for estimating load inertia and system for controlling motor speed using inverse model are provided, suitable for an alternating current (AC) servo module. An inverse system is connected to an actual system to inversely deduce a substantial torque output, wherein when the value of the speed signal generated by the actual system gradually increases, a net inertia ratio is accumulated; whereas when the value of the speed signal generated by the actual system increases to a fixed value, the net inertia ratio is updated.

The invention discloses an unmanned aerial vehicle route planning method based on an improved Q-learning algorithm. The method comprises the steps of: training an unmanned aerial vehicle in differentsimulation environments to obtain a priori knowledge list, and guiding the unmanned aerial vehicle to explore by utilizing the priori knowledge obtained by training in an unknown environment, so thatthe exploring step number of the unmanned aerial vehicle in the unknown environment is reduced; and introducing a single-position action value function convergence criterion, so that the traditional Q-learning based on a Markov process chain convergence principle is changed, and the convergence speed of an action value function is increased.