34 results about "Momentum factor" patented technology

The invention discloses an image significance detection method which uses confrontation training to generate a convolution neural network model, which belongs to the field of computer vision and image processing. The method comprises the steps of data preprocessing, network structure, suitable parameter selecting, and training with a random gradient descending method and an impulse unit. According to data preprocessing, a large amount of collected data and labels are preprocessed. According to network structure, a network structure and a specific kernel function are designed. Suitable parameters including learning rate, a momentum factor and the number of images inserted into the network each time are selected. The random gradient descending method and the impulse unit are used for training to reduce the possibility of network over-fitting. According to the invention, a significance map can be accurately acquired.

The invention relates to a wind power variable-pitch multi-variable fuzzy neural network PID control method. The control method includes the following steps that a fuzzy parameter setting module is used for presetting the weight of a PID neural network module; the error between a rotating speed reference value and the actual rotating speed output of a wind driven generator is calculated through a PID calculation module to obtain a reference output quantity of torque of the wind driven generator; the error between the power output value and the power reference value of the wind driven generator and the error change rate are set through the fuzzy parameter setting module to obtain a presetting parameter of the weight of the PID neural network module; through a negative gradient algorithm with a momentum factor, the weight of the PID neural network module is trained, and the reference value output of torque and the reference value output of the pitch angle of the wind driven generator are adjusted. The output power of the wind driven generator can be stabilized nearby a rated valve, and safety of a fan is ensured.

The present invention relates to a rockburst dynamic prediction method based on BP neural network modeling. The method comprises the steps of: determining and acquiring rockburst influence factors; performing quantification processing on qualitative description parts in influence factor indexes, and obtaining an initial population; performing BP neural network training on the eight acquired influence factors separately; optimizing a number of neurons, an algorithm learning rate and momentum factors by using a genetic algorithm, and obtaining an optimal hidden layer node number; and performing prediction on rockburst of a mine by using a BP neural network algorithm model obtained through training, and obtaining a risk level of the rockburst of the mine. The method provided by the present invention has relatively high reliability, overcomes the defect of no association between the rockburst and the influence factors of the rockburst in the current rockburst prediction process, implements middle and short term dynamic prediction on the rockburst, and can be widely applied to the field of mine rockburst prediction.

The invention relates to a network security situation evaluation method based on CS and improved BP neural network. The method comprises four steps of S1. acquiring network security situation elements, forming a training sample set and a test sample set, and determining a BP neural network structure; S2. seeking an optimal initial weight and a threshold by using a CS algorithm; S3. introducing a momentum factor and a gradient factor to improve the BP neural network; S4. training the improved BP neural network, finally, using the trained network in network security situation evaluation so as toobtain a final situation value and a security level. Network security situation is evaluated precisely and quantitatively by using the improved BP neural network, so that subjective effects of expertopinions in traditional evaluation methods are lowered, and overall network security situation is reflected objectively and comprehensively; and the network security situation is improved by combining the CS algorithm and introducing the momentum factor and gradient factor, the convergence speed is improved, time and space overheads are reduced, and accuracy and practicability of network securitysituation evaluation are improved.

The invention discloses a transformer noise prediction method based on a wavelet neural network and the wavelet technology. A neuronal hyperbolic tangent S-type excitation function of a hidden layer in the traditional BP (back propagation) neural network is replaced with a wavelet-based function, momentum factors are introduced when parameters of the neural system are adjusted, and accordingly, a prediction model is higher in convergence speed and higher in error precision. Vibration and noise digital signals are decomposed by means of the wavelet decomposition technology, wavelet low-frequency coefficients obtained are used as input-output pairs for the prediction model, the wavelet low-frequency coefficients obtained by prediction are reconstructed by means of the wavelet reconstruction technology after modeling, and predicted noise digital signals are obtained. The transformer noise prediction method based on the wavelet neural network and the wavelet technology has the advantages that fewer training samples are required, time of training neurons in the neural network is shortened, and the problem that poor prediction effect is caused by ambient high-frequency interference noise contained in actually-measured transformer noise data is further avoided.

The invention discloses a multiple BP neural network load prediction method based on grey correlation degree. The method comprises the following steps: 1) carrying out load sequence correlation analysis based on grey correlation degree; 2) carrying out clustering based on a shortest distance method to determine a member set of a multiple BP neural network; 3) determining multiple number of the multiple BP neural network based on effectiveness index; 4) improving the BP neural network by introducing a momentum factor and adopting a multiple-calculation averaging method to solve the problem of easy local convergence of the BP neural network, and improve anti-vibration capability thereof; and 5) carrying out short-term power load prediction through an established multiple BP neural network prediction model. The method solves the problem of easy local convergence of the BP neural network and improves anti-vibration capability thereof; and compared with a conventional BP neural network prediction model, the multiple BP neural network has a better prediction effect.

The invention provides a BP (Back Propagation) neural network algorithm based method for analyzing coating aging. The method has the advantages of higher flexibility and forecast precision and better hereditability and comprises the processes of signal forward propagation and error backward propagation, wherein in the forward propagation process, an input sample is imported from an input layer and then transmitted to an output layer after the sample is processed through various buried layers layer by layer; if the actual output of an output layer does not accord with an expected output, the process is turned to an error backward propagation stage; in the error backward propagation process, an output error is backwards transmitted to an input layer through the buried layers in a certain form layer by layer, and the error is shared by all units of all the layers so as to obtain error signals of all the units of all the layers, wherein the error signals are used as references for correcting the weight values of all the units; and the weight value adjustment process of all layers of signal forward propagation and error backward propagation is carried out in cycles until network output errors are reduced to an acceptable degree or a preset number of times is finished. The method is characterized in that a momentum item delta W(t)=eta delta X+alpha delta W(t-1) is added, wherein alpha is a momentum factor alpha belonging to the set of (0, 1); the learning rate is adaptively regulated, if a total error E rises after the adjustment of a batch of weight values, eta is equal to beta eta (theta>0), and if the total error E drops after the adjustment of a batch of weight values, eta is equal to theta eta (theta>0); and a steepness factor is introduced, and when an error curve plane enters a flat area, a changed output quantity is set, wherein lambada is the steepness factor, in the flat area, lambada is larger than 1, and after quitting the flat area, lambada is equal to 1.

The invention relates to a power cable state evaluation method based on momentum BP neural network, which comprises the following steps: momentum BP neural network with momentum factor is established;the neural network is trained by collecting multi-group data of power cable state index, the weight of the neural network is adjusted gradually by output value, target value, weight change and momentum factor of the neural network, so that the error between output value and target value is minimized globally, after the training, the power cable state index data to be evaluated is inputted to themomentum BP neural network, and the running state of the power cable is judged. Momentum BP neural network has fast convergence and minimal global error. The method can quickly and accurately establish the mapping relationship between power cable status index data and evaluation conclusion, evaluate power cable status comprehensively and accurately, and provide scientific basis for cable operationand condition-based maintenance.

The invention discloses a slope deformation and soft soil foundation settlement prediction method based on a GABP neural network, and the method comprises the steps: carrying out the optimization design of the structure, initial connection weight, initial threshold value, learning rate and momentum factor of the neural network through employing a genetic algorithm, and positioning a better searchspace in a solution space; and then optimizing the connection weight and the threshold of the network again in the small solution spaces by using a BP algorithm, and searching an optimal solution, sothat the optimized BP neural network can better predict the output of the function. According to the method, the BP neural network method and the genetic algorithm are combined, and the advantages ofthe two methods are fully utilized, so that the improved method not only has strong learning ability and robustness of the BP neural network, but also has global optimization ability of the genetic algorithm, and has the advantages of high prediction precision, high network convergence speed and the like; and a good effect is achieved on slope deformation and soft soil foundation settlement prediction.

The invention discloses a quasi-PR grid-connected inverter optimization control method based on self-adaptive harmonic elimination. For the shortcomings of high parameter precision requirement, poor anti-disturbance capability, difficult realization and the like of an ideal PR controller in an actual system, a quasi-PR controller is applied to a single-phase photovoltaic inverter system; in view of factors such as the bandwidth limit of an inverter, the precision of a detection system, and the like, characteristic harmonic detection is adopted for selectively compensating 3rd, 5th and 7th harmonics with relatively high content and relatively low orders; and a self-adaptive learning rate and a steepest descent method of additional momentum factors are adopted for improving a BP neural network, so that while the training time is shortened, the shortcoming that the BP neural network easily obtains a local optimum solution is overcome, and the problems of inaccurate harmonic compensation,overlong system response time and the like are solved.

The invention provides an UPFC control method based on nerve network sliding mode control, which utilizes a converter targeting a parallel connection side and a serial connection side, and, on the basis of the vector control, constructs the state space of the parallel connection side and the serial connection side of the converter, adopts the radial basic function (RBF) nerve network algorithm and the momentum factor based on the classic gradient decedent algorithm to perform regulation on the hidden layer node center, the node width and the network weight value. The UPFC control method disclosed by the invention can realize the whole process adaptive control of the sliding mode state, removes the sensitivity of the sliding mode control to the external parameters, realizes the decoupling control of the active power and the reactive power, inhibits the oscillation during the system interference, and fast approaches the object value of the system operation. The structure is simple and the operation is reliable, and has good adaptability and robustness. The invention makes up the blanket in the field and provides necessary technical support and the beneficial reference to the improvement of unifying power flow controller stable control system.

The invention discloses a self-adaptive acoustic feedback suppression method. On the basis of an LMS algorithm, by introducing a momentum factor, namely a prior error, and simultaneously combining with an NLMS algorithm, an update step of a self-adaptive algorithm is optimized and a convergence speed is accelerated, thereby adjusting a central frequency of a wave trap more quickly and reducing thedistortion of a finally obtained output signal. According to the self-adaptive acoustic feedback suppression method, through a self-adaptive wave trapping method, a change of a howling frequency canbe tracked and the central frequency of the wave trap is automatically adjusted. When sampling data are limited, relatively accurate frequency estimation precision can be obtained, and thus the distortion degree of the signal after wave trapping can be reduced. A recursive process is used by the self-adaptive algorithm and only time domain processing is involved, so an FFT operation does not needto be performed, and the calculation complexity is greatly reduced. By introducing the prior error and a changeable convergence factor to perform step adjustment, the convergence speed of the self-adaptive algorithm is further accelerated; and while the algorithm has relatively good acoustic feedback suppression effect, the timeliness on signal processing can be met.

The invention discloses a hybrid neural network (HNN) algorithm-based performance assessment method used for a complex industrial product. The method comprises the steps of firstly determining HNN structure parameters; secondly initializing a connection weight of a neuron of each layer and a characteristic point of a membership function, and setting an error limit value, an iterative frequency, a learning rate and a momentum factor; thirdly performing quantification and normalization on input fuzzy sample data, and performing normalization on a quantitative numerical value; fourthly calculating a learning error derivative of the neuron of each layer in repeated iteration, correcting the connection weight, and adjusting the characteristic point of the membership function by adopting a gradient descent method; and finally performing repeated iteration until a set error is reached, and giving out a quantitative performance prediction result of the complex industrial product through an HNN algorithm. The HNN algorithm provides a solution with high prediction accuracy for the problem in complex industrial performance assessment based on qualitative, quantitative and qualitative-quantitative combined data input and possibly different dimension numbers of input data items in actual conditions.

The invention discloses an intelligent real-time prediction method for a high-speed large-range maneuvering target track. The method comprises the following steps of firstly, proposing a learning sample establishment method; constructing a target motion law learning and training mechanism based on an improved BP neural network; and finally, through a single-step prediction and rolling prediction method, realizing the intelligent, rapid and accurate prediction of the high-speed large-range maneuvering trajectory of the aerospace moving target. According to the invention, only the history of theaerospace moving target and the position data at the current moment need to be known, the motion model of the target is not needed, meanwhile, by designing a momentum factor and adopting a variable step size iterative strategy, the convergence speed of the traditional BP neural network is increased, and oscillation during the convergence process is reduced, and the precision of trajectory prediction is greatly improved. The method can be directly applied to the trajectory prediction problems of various high-speed and high-maneuverability targets, has the higher applicability, and provides thetheoretical basis and the technical reserves for the subsequent tasks, such as monitoring, tracking and intercepting the hypersonic aircrafts, such as X-37B, etc., and the like.

The invention provides a back propagation method for an out-of-order data stream in big data. An improved back propagation algorithm based on dynamical adjustment (IBPDA) is provided to solve the problem that association rules are difficult to obtain from the out-of-order data stream in the big data. A dynamic self-adaptation structure adjustment mechanism is used, a network training structure is adjusted in a self-adaptation mode according to environment requirements, invalid training nodes are automatically deleted, and optimized iteration of the training process is achieved; three factors of a neural network, namely a learning index, a momentum factor and a scale factor, are dynamically adjusted in the web-based learning process to achieve the aims of increasing learning response speed and enhancing network stability. As is shown in a simulation result, by means of the dynamic self-adaptation structure adjustment mechanism and dynamic adjustment of the three factors of the neural network, the method can obtain more convergence times, effectively improve the convergence rate and improve whole network performance.

The invention discloses a neural network laser cutting quality prediction method. The method comprises the following steps: taking a to-be-processed original part as a sample to carry out a cutting experiment according to different laser cutting parameter sets; acquiring experimental data of the cutting experiment; constructing a neural network, and setting a transmission function, a learning rate, a training frequency, a training target, a momentum factor and a training parameter of the neural network; preprocessing the experimental data, and training and verifying the neural network by usingthe experimental data; and simulating a cutting process by utilizing the neural network model to search an optimal target laser cutting parameter set meeting the quality requirement. According to themethod, experimental data is acquired by using fewer cutting experiments to perform training modeling on the three-layer reverse transmission neural network, all laser cutting parameters are predicted by using the three-layer reverse transmission neural network, and a target prediction result meeting customer quality requirements is selected; the method has the advantages of being short in period, low in cost, high in efficiency, capable of obtaining the optimal laser cutting parameters more quickly and the like.

A multi-factor risk analysis system may include a resource data store defining, for each of a plurality of resources: a resource identifier, a current resource value, and a resource type. A back-end application computer server may establish a set of risk parameter diversification objectives and determine a set of eligible resources by accessing the resource data store. The server may then calculate a value factor characteristic and a momentum factor characteristic for each of the eligible resources. A set of index resources may be constructed from the set of eligible resources based on the risk parameter diversification objective and a risk optimization process utilizing a weighted value factor characteristic and momentum factor characteristic of each resource. An interactive user interface may then be rendered including indications associated with the constructed set of index resources, the value factor characteristics, and the momentum factor characteristics.

The invention proposes an LCL filtering-based RBFNN segmentation online optimization passive control system and method. Grid-side three-phase voltage and current signals are acquired through three-phase voltage and current signal sensors respectively, and coordinate transformation is performed; a passive control Hamiltonian model based on an IDA-PBC algorithm is built according to coordinate axisvoltage and current, and improved d,q-axis switching function is built; particles containing parameters such as an RBFNN learning rate and a momentum factor under different load resistances are subjected to offline optimization through PSO to obtain an optimal particle set; the load resistances calculated by DC-side voltage and current sensor signals are used as segmentation triggering conditions,an RBF-PID model is built through the optimal particle, and a controller model is used to realize segmentation optimization control; the RBF-PID after parameter optimization is used for optimal solution for stably-operating Im; and according to the optimized Im and in combination of the d,q-axis switching function, control is carried out, IGBT control signals are generated by SVPWM, and rectification control is realized. The control precision is higher, and the robustness is better.

The invention discloses an improved GA-based (genetic algorithm-based) method. The invention discloses a gas emission amount prediction method of a BP network model, and relates to the technical fieldof coal mine underground stope face gas prevention and control. GA-BP network model is combined with a genetic algorithm and a BP algorithm, and on the basis of keeping the original adaptivity and fault tolerance, the optimal initial weight value and threshold value are selected through global search. Therefore, the learning speed of the network is accelerated, and the global optimization capability is improved to a certain extent. GA-BP network model is combined with a main factor analysis method, main factors are extracted through main factor analysis to replace original input variables, the network structure is simplified, and variable redundancy information is eliminated. Meanwhile, a genetic algorithm (GA) is adopted to optimize the initial weight value and the threshold value of thenetwork, and a momentum factor is added to optimize the updating mode of the weight value, so that the search is prevented from falling into a local minimum value, and the prediction accuracy is improved. And finally, selecting actual gas emission monitoring data as label data and input data, and carrying out simulation and analysis on different network models.

The invention provides a blind source separation method under a time-varying channel based on an adaptive momentum factor, belonging to the signal processing field, aiming at ensuring the separation convergence speed while effectively reducing the separation steady-state error, and realizing the following steps: constructing a separation index and obtaining an adaptive update formula of the separation index; obtaining the iterative step length of the separation matrix; acquiring adaptive momentum factor; obtaining the iterative formula of the separation matrix; getting an estimate of the source signal. The invention introduces an adaptive momentum factor in the iteration of the separation matrix, under the condition that the separation convergence speed is not reduced, the steady-state error of the separation is reduced, and the accuracy of the source signal estimation is improved.

The invention relates to a BP (back propagation) neural network-based momentum face detection method. According to the method, Gabor features and a momentum factor back propagation algorithm are combined. The method includes the following steps that: the Gabor features of training sets are extracted; the Gabor features are inputted into a momentum factor back propagation neural network so as to perform training; and the trained system is adopted to detect whether a face exists in an inputted image, and a face exists in the inputted image, the face is marked by a rectangle. In order to improve the training effect of a traditional back propagation algorithm, the momentum factor is added to the algorithm, and therefore, the vibration trend of the neural network in training can be effectively slowed down, and the algorithm can be prevented from falling into the local minimum; and the added momentum factor can adaptively adjust the weight of each layer of the back propagation neural network. Numerous experimental results show that, compared with classical or the most advanced face detection models, the BP (back propagation) neural network-based momentum face detection method of the invention is effective and competitive.

The invention discloses a depolarization multiplexing method based on a momentum factor, and belongs to the technical field of coherent light communication. The method comprises the following steps: firstly, initializing a demultiplexing matrix A (n), wherein the demultiplexing matrix is a 2 * 2 matrix; inputting the two paths of polarization state signals X '(n) and Y' (n) into A (n) to obtain ademultiplexed signal; finding out constellation points X # (n) and Y # (n) closest to X # '(n) and Y #' (n), and calculating an initial error epsilon1 (n); inputting the constellation points X # (n) and Y # (n) into an inverse matrix A-1 (n) to obtain pseudo observation signals X ''(n) and Y ''(n), and calculating a reverse observation error epsilon2 (n); then constructing a momentum factor omega(n) by the gradient information of the matrix A (n); updating the multiplexing matrix A (n) by using a gradient descent method; and finally repeating the above steps, solving to obtain a demultiplexingsignal, and updating the demultiplexing matrix A (n). The invention also discloses a depolarization multiplexing system based on the momentum factor, and the convergence speed of the depolarization multiplexing algorithm is further improved by introducing the new dimension of the momentum factor. Therefore, the technical problem of polarization state tracking in a lightning environment is solved.

The invention provides a neural network weight training method and application thereof. The training method for the weight of the neural network comprises the following steps: firstly, taking a predetermined number of short-term power load data as an initial training set, and taking an initial weight as a memory weight; obtaining a momentum factor, a resistance factor and an estimation factor of the weight training model by adopting a data linear analysis method; acquiring a weight updating speed according to the momentum factor, the resistance factor and the estimation factor, acquiring a weight variation according to the weight updating speed and the memory weight, taking an intermediate weight as a new memory weight, repeating the process for each piece of training data in an initial training set, and taking a final intermediate weight as an optimal weight.

The invention provides modeling of a Hammerstein nonlinear dynamic system and concentration control of a continuous stirring reactor, and the method comprises the steps: building a continuous stirring reactor model by using the Hammerstein nonlinear dynamic system, approaching a static nonlinear module by using a three-layer BP neural network, and building a linear dynamic module model by using a transfer function model. In the method, firstly, according to input and output data of a binary signal, a least square method is adopted to estimate parameters of a dynamic linear module in a model; secondly, on the basis of input and output data of random signals, parameters of a static nonlinear module in the model are estimated through a stochastic gradient descent optimization algorithm containing momentum factors; and finally, a reactant concentration control problem of the continuous stirring reactor system is converted into a linear system control problem by utilizing a special structure of a Hammerstein nonlinear system, so that the design of the control system is simplified, and a better control effect is achieved.

The invention discloses a depolarization multiplexing method based on a momentum factor, and belongs to the technical field of coherent light communication. The method comprises the following steps: firstly, initializing a demultiplexing matrix A (n), wherein the demultiplexing matrix is a 2 * 2 matrix; inputting the two paths of polarization state signals X '(n) and Y' (n) into A (n) to obtain ademultiplexed signal; finding out constellation points X # (n) and Y # (n) closest to X # '(n) and Y #' (n), and calculating an initial error epsilon1 (n); inputting the constellation points X # (n) and Y # (n) into an inverse matrix A-1 (n) to obtain pseudo observation signals X ''(n) and Y ''(n), and calculating a reverse observation error epsilon2 (n); then constructing a momentum factor omega(n) by the gradient information of the matrix A (n); updating the multiplexing matrix A (n) by using a gradient descent method; and finally repeating the above steps, solving to obtain a demultiplexingsignal, and updating the demultiplexing matrix A (n). The invention also discloses a depolarization multiplexing system based on the momentum factor, and the convergence speed of the depolarization multiplexing algorithm is further improved by introducing the new dimension of the momentum factor. Therefore, the technical problem of polarization state tracking in a lightning environment is solved.

The invention discloses an imitating method of a power amplifier and an imitating device of the power amplifier. The method adopts an improved back propagation (BP) neural network to imitate the power amplifier. The improved BP neural network adopts Levenberg-marquardt (LM) learning algorithm and training algorithm with momentum factors added to update weight values and threshold values of a model, and therefore model building speed is effectively increased and the model is prevented from being sunk into a local minimum. Disclosed is the imitating device of the power amplifier further. The improved BP neural network is adopted. The imitating device of the power amplifier is further used for building a power amplifier imitator with a parallel structure. Imitation and training are free of mutual interference. Working efficiency of the power amplifier imitator is improved. Compared with the prior art, the imitating method of the power amplifier and the imitating device of the power amplifier are quick in training convergence speed and good in instantaneity.