78results about How to "Avoid local minima" patented technology

The invention relates to a method for carrying out face three-dimensional reconstruction at any viewing angle on the basis of a self-adaptive deformable model. The method includes the steps of (1) obtaining face image data and screening a face image with high definition as original data, (2) positioning feature points, (3) coarsely estimating the angle of a face according to the positioning result of the feature points, (4) building a face three-dimensional deformable model, adjusting the feature points of the face to be at the same dimension as the face three-dimensional deformable model through translation and scaling and extracting coordinate information of the points corresponding to the feature points of the face to form a sparse face three-dimensional deformable model, (5) iterating face three-dimensional reconstruction by means of the particle swarm optimization algorithm according to the coarsely estimation value of the angle of the face and the sparse face three-dimensional deformable model to obtain a face three-dimensional geometric model, (6)mapping input face texture information in a two-dimensional image to the face three-dimensional geometric model in a texture pasting method after the face three-dimensional geometric model is obtained, so that a complete face three-dimensional model is obtained. The method can be widely used in the field of identity identification.

The invention discloses an unmanned aerial vehicle (UAV)-unmanned ground vehicle (UGV) combined formation cooperative control method, comprising the following steps: step 1, establishing nonlinear dynamical models of unmanned vehicles in UAV-UGV combined formation; step 2, processing the nonlinear dynamical models of a UAV and a UGV via equivalent transformation, and taking acceleration as a common control target quantity, obtaining a unified control model taking acceleration as a control input in the combined formation; step 3, establishing a ground-air combined formation structure based on a virtual pilot to obtain a stable control signal for the UAV-UGV combined formation and obtain an error model of the combined formation, wherein the control signal is acceleration obtained in step 2 as a common control target quantity; and step 4, designing a UAV-UGV combined formation controller by adopting a RBF (Radial Basis Function) network algorithm according to the control model, the error model and the acceleration serving as a control signal and a control target quantity at the same time, so that the combined formation is stable and reliable.

The invention discloses a method for evaluating the degree of a mechanical fault of a frame-type circuit breaker based on a vibration signal. The vibration signal in the method is a mechanical vibration signal collected by a frame-type circuit breaker mechanical fault detection system in the switching process of the frame-type circuit breaker. The method comprises the steps: employing a wavelet packet to carry out the denoising preprocessing of the vibration signal; carrying out the adaptive decomposition of a denoised vibration signal through employing a local mean decomposition algorithm; screening out the former d PF components with the maximum correlation with an original vibration signal; carrying out the improved multiscale arrangement entropy analysis of al PF components, and carrying out the dimension reduction of a feature vector formed by the above improved multiscale arrangement entropy values through the PCA method; building a fault feature vector; constructing a multi-class supporting vector machine, and carrying out the pattern recognition; carrying out the quantitative evaluation of the severity of the mechanical fault happening in the switching process of the circuit breaker through referring to the fault degree characteristic curves in different fault modes. The method is stable, is reliable and is effective.

The invention discloses an adaptive control system based on a radial basis function (RBF) neural network sliding mode control for a micro-electromechanical system (MEMS) gyroscope, and the system comprises a gyroscope and a control circuit, wherein the control circuit comprises a sliding mode controller and an RBF neural network; the difference of displacement of the three-axis gyroscope in the directions of three coordinate axes x, y and z and displacement of a reference model is taken as the input of the sliding mode controller. In the adaptive control system, an adaptive sliding mode control method is applied in controlling the gyroscope, so as to improve the stability and reliability of the system; and the RBF neural network is adopted to carry out adaptive learning on upper boundary of uncertain interference, thus reducing the influence of measurement error and external interference, effectively lowering the occurrence of buffeting, and achieving a better control effect.

The invention discloses a method for controlling a micro gyro based on a radial basis function (RBF) neural network sliding mode. Single-input single-output neural sliding mode control can be realized by using a switching function as the input of an RBF neural network, using a sliding mode controller as the output of the RBF network and using the learning function of the neural network; and a control effect can be achieved by integrating the advantages of sliding change structure control, an adaptive algorithm and the RBF neural network. The adaptive algorithm is used for adjusting the link weight of the RBF neural network in real time on line according to accessible conditions, so that a system finally achieves a sliding mode surface, completes tracking, and can adapt to sliding mode control strategies and timely correct and estimate all rigid errors, damping and the like; and the stability of a provided adaptive sliding mode controller exists according to the Lyapunov stability theorem, the system has good robustness, and digital simulation of the three-dimensional micro gyro proves that the method for controlling the micro gyro is valid.

An inventory forecasting model and its construction method based on chaotic neural network. The inventory of finished products is the key factor in precise distribution. If the inventory of finished products is sufficient, accurate delivery will be guaranteed, but the high inventory of finished products will bring a negative impact on the enterprise. The risk is high. On the one hand, it is difficult to process other materials after the original roll is processed into finished products. Once the user does not use it, it is likely to become a waste product. On the other hand, the finished product inventory takes up a large inventory space, which will make Limited storage capacity is getting tighter. The present invention divides the work into two phases. The first is the learning phase. The data of all the distribution users of the sample companies in the past three years are used as samples to establish a model, and these samples are used to learn and adjust the connection weight coefficients of the chaotic neural network, so that the network Realize the given input-output relationship; then the implementation stage, use the trained neural network to obtain the expected effect, establish a perfect calculation model, and realize the reasonable setting of the inventory.

The present invention discloses a GIS partial discharge type identification method based on GK fuzzy clustering. The method is characterized by comprising the steps of S01, constructing a GIS partial discharge gray scale image according to the acquired data; S02, extracting the fractal features of the GIS partial discharge gray scale image, namely a box dimension and an information dimension; S03, processing the fractal feature data further by a GK fuzzy clustering algorithm, isolating a GIS site interference signal; S04, designing a GIS partial discharge mode recognizer based on a least squares support vector machine classification algorithm; S05, identifying the GIS partial discharge type. By a GK fuzzy clustering method, the GIS site interference signal is isolated, thereby improving the accuracy of extracting the fractal features of the partial discharge signal. Meanwhile, by identifying the GIS partial discharge type by the least squares support vector machine classification algorithm, the correctness and rapidity of the discharge type identification are improved.

The invention discloses a thermal process model parameter identification method adopting an improved ant colony algorithm. The thermal process model parameter identification method comprises the steps of determining a system identification structure and parameters to be identified, determining algorithm path and initial pheromone distribution and completing the search through loop iteration. The thermal process model parameter identification method makes some corresponding improvements to a thermal process on the basis of the basic ant colony algorithm, and converts an identification problem into an optimization problem in a parameter space, so that the algorithm is more accurate and efficient. On the basis of known input and output data, the thermal process model parameter identification method adopts the improved ant colony algorithm on MATLAB software for carrying out efficient and parallel search on the entire parameter space, can identify model parameters quickly, and achieves the precise modeling purpose.

The invention relates to an independent method for controlling spacecraft attitude directing constraint attitude maneuver, belonging to the technical field of spacecraft attitude control. The method comprises the following steps of: constructing a navigation function V related to the motion of the tail end point of a current pointing vector r of a sensor on a unit spherical surface S by taking the tail end point of a target pointing vector rd of the sensor as a target point position, taking the tail end point of the current pointing vector r as a current position and taking a spherical doom formed by a pointing constraint as a barrier region; designing a control torque expression according to the navigation function, and regulating the amplitude of a control torque by changing control torque parameters to drive the spacecraft to make the sensor point to the target vector rd; and driving the spacecraft to rotate in the vector direction of the sensor by an angle theta after the sensor points to the target vector rd, so that a complete attitude maneuver process of the spacecraft is realized. According to the method, pointing avoidance of the barrier region can be processed definitely, a local minimum value can be avoided for a plurality of barrier constraints simultaneously, safe maneuver of the spacecraft to a target attitude is ensured, the requirement of boundedness on an executing mechanism is met, and independent control over spacecraft directing constraint attitude maneuver is realized.

The invention discloses a power amplifier behavioral modeling method based on a depth reconstruction model. The depth reconstruction model combines the advantages of a deep learning theory and an Elman neural network. A restricted Boltzmann machine is used to initialize the weight coefficient of the neural network. In a modeling process, the number of times of iteration is small, but faster convergence is acquired. Due to the fact that the output of the Elman neural network is related to immediate input and historical input, a power amplifier behavioral model is accurately reconstructed by describing the memory effect of a nonlinear system.

Present invention relates to video camera calibrating method only using one plane calibration object. It contains setting two sets quadrature parallel straight lines given that spacing as calibration object, shooting one image along video camera optical axis with plane calibration object oblique, to make two sets parallel lines approximately in vertical and horizontal two direction of image. Extracting intersecting point in image parallel line and proceeding straight line matching to parallel line, finding out one minimal residuum, using intersecting point of two minimal residuum fitting straight line as principal point initial value, finding out percentage distortion initial value according to residuum with relation, using all fitting straight line residuum sum as objective function, utilizing optimization method proceeding optimization to and percentage distortion, utilizing distortion model removing image distortion, finding out each intersecting point orthoscopic image co-ordinate, finding out video camera whole parameter by linear transformation utilizing intersecting point orthoscopic image co-ordinate and world co-ordinate in plane calibration object, finally to complete calibration.

The invention relates to a robot dual peg-in-hole assembling method utilizing a genetic evaluative algorithm based on learning and belongs to the technical field of robot automatic assembling. The method has the advantages that a regression prediction model is utilized to predict the fitness value of an actual robot assembling process, only a small number of optimal genes are subjected to actual tests every time, the assembling control algorithm in an actual environment is optimized on the basis of a small number of tests, and the assembling process is improved; the regression model of supportvector machines is utilized, the method has strong fitting ability to a complex nonlinear system containing noise, theoretical optical convergence can be achieved, the local minimum problem is avoided, the calculation complexity of the method depends on the number of the support vector machines instead of the dimensionality of sample spaces, and the dimensionality disaster problem can be avoidedto a certain degree.

The invention belongs to the field of prediction of amorphous alloy thermoplasticity forming performance, and discloses an artificial neural network predicting method of the amorphous alloy thermoplasticity forming performance. The predicting method comprises the following steps of a, selecting multiple performance parameters and collecting data of the performance parameters, dividing the data into a training sample, a verification sample and a to-be-predicted sample, and testing to obtain feature index test values corresponding to the training sample and the verification sample; b, selectingan artificial neural network model as an initial predicting model for the amorphous alloy thermoplasticity forming performance, adopting the training sample to train the artificial neural network model, and determining an improved predicting model; c, adopting the verification sample to verify the improved predicting model, finally obtaining a final predicting model, and adopting the final predicting model for prediction. By means of the artificial neural network predicting method of the amorphous alloy thermoplasticity forming performance, the amorphous alloy thermoplasticity forming performance is effectively predicted without experiments, guidance is provided for development of an amorphous alloy system suitable for thermoplasticity forming, the time for developing the new amorphous alloy system is greatly shortened, and the money cost for developing of the new amorphous alloy system is greatly reduced.

The invention discloses an X-ray machine movement trajectory planning method and device. The method comprises the steps that the target positions which an X-ray generating device and an X-ray receiving device of an X-ray machine are controlled to move to reach during positioning of a machine frame are set; in the movement process of the X-ray generating device and the X-ray receiving device, the current positions of the X-ray generating device and the X-ray receiving device are obtained in real time; if the current positions are not consistent with the target positions, the movement axis speeds of the X-ray generating device and the X-ray receiving device are calculated through an artificial potential field method; based on the calculated movement axis speeds, the X-ray generating device and the X-ray receiving device are controlled to move. According to the technical scheme, a movement path which is smooth, safe and suitable for simultaneous movement of multiple targets is easily and rapidly planned during positioning of the machine frame of the X-ray machine.

The invention discloses a mobile robot path planning method and system based on an improved RRT algorithm, which belongs to the field of robot moving path planning. In order to solve the technical problem of how to plan a better path of the mobile robot in various environments based on the RRT algorithm, the adopted technical scheme is as follows: a convergence factor is added on the basis of theRRT algorithm to improve the selection of growth points and exploration points of an extension tree, so that the aim of improving the convergence speed of the algorithm is fulfilled; then, dynamic step length adjustment is used for avoiding the situation that a local minimum value occurs when the extension tree carries out path planning; smoothing processing is carried out on the planned mobile robot path, so that the path length is shortened to tend to the optimal path; and the method specifically comprises the steps of S1 adding a convergence factor; S2 avoiding the local minimum value; andS3 carrying out smoothing processing. The system comprises a convergence factor adding unit, a local minimum value avoiding unit and a smoothing processing unit.

The invention discloses a neural-network-based rapid compensation method for photoelectric encoder. The method can increase compensation precision and compensation speed, and can simplify compensation procedures and reduce requirements to a needed instrument. The method includes: first, randomly rotating the photoelectric encoder in a circle, and acquiring a training sample; then adopting a single input and single output three-layer feed-forward Fourier neural network to establish a photoelectric encoder error compensation model, and estimating the number of nodes in hidden layers based on the characteristics of an object function; at last based on the actual result, correcting the number of the nodes so as to prevent the inefficiency due to repeated and varied attempts in a trial-and-error method. According to the invention, directed to the problem of slow convergence of the iteration training method, the method, through mapping the errors to nodes in the hidden layers of orthogonal triangle function basis, rapidly solves a minimum norm solution from the hidden layers to a weight of an output layer, which substantially reduces the time required by solving the weight of the neural network and greatly reduces time complexity and space complexity in calculation.

The invention discloses a measuring method of pulverized coal concentration. The measuring method comprises the steps that a wavelet neural network model is built, and training is conducted; wherein the wavelet neural network model comprises a cold primary air volume, a primary wind temperature, a coal feed quantity, a hot primary air volume, a coal mill inlet and outlet pressure differential, a coal mill outlet pulverized coal temperature, a separator outlet pressure and a total air volume which serve as wavelet neural network inputs and takes a concentration value of pulverized coal at the coal mill outlet as a wavelet neural network output; the trained wavelet neural network model is used for real-time online measuring of the pulverized coal concentration, newly sampled coal mill data serves as an input of the trained wavelet neural network model, and an output of the trained wavelet neural network model is the concentration value of the pulverized coal at the coal mill outlet. According to the measuring method of the pulverized coal concentration, dependence on a training sample set is low, the stability of the measuring method is high, the robustness is good, the method is not affected by field measurement environmental factors, and the error-tolerant rate is high; a wavelet neural network measuring system is simple in structure, convenient to install, free of interference of the field measurement environmental factors, high in sensitivity and low in maintenance cost.

The invention belongs to the technical field of neural network fault prediction, and discloses a fault prediction method based on a fast integrated convolutional neural network. The method comprises the following steps: (a) collecting an output signal of a to-be-predicted object, and carrying out time-frequency analysis on the output collected signal by adopting S transformation to obtain a time-frequency signal; (b) constructing a convolutional neural network structure based on LeNet-5, providing a maximum and minimum cosine cyclic learning rate scheduler based on a self-weight starting cosine cyclic learning rate, and setting a learning rate range of the convolutional neural network; and (c) training a time-frequency signal of the prediction data by using the convolutional neural network, and taking a prediction result of the SECNN as a final prediction result. According to the invention, fault prediction with high prediction speed and high prediction precision can be realized.

The invention discloses a parallel type genetic Elman neural network-based source driving 235U concentration recognition method, which mainly comprises the following steps: establishing a neural network model, wherein the neural network model is structurally divided into three layers which are a data allocation layer, a sub-network layer and a comprehensive decision-making layer respectively; preprocessing an acquired neutron pulse signal auto-correlation function; inputting a processed signal auto-correlation function sample into the data allocation layer of a parallel type genetic Elman network and adopting a cyclic random multi-point sampling method to allocate the data of the sample; inputting the allocated data into each genetic Elman sub-network in the sub-network layer respectivelyfor recognition and giving recognition results; and performing comprehensive processing on the recognition results of the plurality of sub-networks by the comprehensive decision-making layer to obtain a final 235U concentration recognition result. Through the method, relatively better 235U concentration recognition effect is obtained due to relatively higher data utilization ratio and a novel network structure.

The invention relates to a power or static power-based reinforced concrete simply supported beam fire model correction method, and belongs to the technical field of model correction methods. Accordingto the method, damage feature parameters are selected as input parameters for model correction; physical parameters under a corresponding damage state are selected as output parameters for model correction; the damage feature parameters are combined by utilizing frequency and vibration mode changes; and when to-be-corrected physical parameters are selected, structural boundary conditions and material performance parameters, which influence structural mode information and damage recognition, are comprehensively considered. According to the method, a support vector machine statistical learningalgorithm is applied and a distributed correction strategy is adopted, so that sample counting in once correction is decreased, training samples are greatly decreased, the model correction correctnessis improved, and corrected finite models can reflect structural vibration characteristics and other structural responses more really in fire processes or after fires.

The invention discloses a short-term load forecasting method based on improved HS-NARX neural network, the method comprises S1 collecting data and preprocessing; S2, establishing NARX neural network;the neural network is trained with the preprocessed data. 3, determining a fitness function of the HS algorithm; S4, setting parameters of the harmony search algorithm; S5 initialization parameters; S6 generates HMCR and PAR according to HMCRmean and PARmean, and the pitch adjustment bandwidth is (BWmax, BWmin); S7 generating (0, 1) random numbers, generating new harmonic vectors, and uses improved pitch adjustment rules and adaptive parameter tuning method to generate new harmonics; S8, comparing the generated new solution with the worst solution in the harmonic memory bank, if the new solution is better than the worst solution, replacing the worst solution, otherwise, not operating, recording HMCR and PAR again; S9 returning to S7 if the number of iterations is not reached, otherwise, the optimal solution is outputted; S10 mapping the optimal solution to the neural network, obtaining the weights W and the threshold theta of each layer of the network, and training the network and theload forecasting.

The invention discloses a traffic video vehicle identification method based on an SRC and support vector machine (SVM) combined classifier. The histogram of oriented gradient (HOG) features of the vehicle samples are trained by a dictionary training (K-SVD) algorithm to obtain a redundant dictionary, so that the sparse representation-based classifier (SRC) is constructed, and meanwhile, the HOG features of the vehicle positive and negative samples and samples to be classified are subjected to sparse reconstruction, and the SVM is trained by the reconstructed features. At the end, the combined classifier is formed by the linear weighting of the SRC and the SVM based on reconstruction to complete the comprehensive decision of vehicle identification. According to the invention, the identification rate and robustness of the whole system in the complex traffic scenes of adhesion, shielding, and target category diversity, and the training time is reduced.

The invention relates to an efficient evaluation method for the terminal entering state of Mars on the basis of intelligent learning, and belongs to the technical field of deep space exploration. A Mars entering trajectory optimization algorithm is used for optimizing maximum terminal height which can be reached by a spacecraft under different parameter combinations to provide the sample data of aprediction model based on Gaussian process regression. A genetic algorithm is used for the optimization solver of an entering trajectory under different scenes to avoid a local minimum value so as toguarantee the data quality of a Gaussian process training sample. A mean value function, a kernel function and a hyper-parameter are taken as the subject parameters of the Gaussian process and are selected as optimization parameters for describing correlation among samples so as to establish a Mars entering optimal terminal height prediction model based on the Gaussian process. By use of the method, the evaluation of the maximum terminal height under more than 3000 groups of different entering scenes can be finished in dozens of second orders, and an average relative error is within 4%.

The invention provides a hollow coil current transformer error compensation method based on an elastic network. The method comprises the steps that of collecting the influence quantity influencing three-phase error compensation of a hollow coil current transformer, wherein the influence quantity comprises environmental parameters and electrical parameters; collecting an error compensation quantity; normalizing the influence quantity and the error compensation quantity, calculating Pearson correlation coefficients of the influence quantity and the error compensation quantity, and performing feature selection on the main influence quantity by using a factor screening method based on an elastic network algorithm; taking dominant influence quantity of the hollow coil current transformer as aninput quantity, using an elastic network algorithm based on cross validation to carry out modeling prediction on error compensation, calculating a difference value between an actual compensation valueand a prediction compensation value, and taking an average absolute error and a root-mean-square error as prediction evaluation. According to the method, the error compensation trend of the hollow coil current transformer can be effectively predicted, and the measurement precision of the hollow coil current transformer can be effectively improved.

The invention discloses an automatic focusing method and device based on an RBF neural network, and the method comprises the steps: obtaining an image photographed at a position where an objective lens is located, and calculating a corresponding focusing evaluation value and an average gray value according to the image; inputting the focusing evaluation value and the average gray value into a preset RBF neural network model, wherein the RBF neural network model outputs the position of an objective lens at an optimal focusing point; and the objective lens is driven to be adjusted to the position of the optimal focus point by using a zoom motor. According to the application, the influence of a local peak value in an existing focusing search algorithm can be effectively avoided, the frequencyof back-and-forth movement of the zoom motor can be effectively reduced, and the focusing time is shortened.