65 results about "Neural network topology" patented technology

The invention discloses a dissolved oxygen control method based on a dynamic radial basis function neural network, which adopts the following steps of: determining a control object; designing a dynamic RBF neural network topology structure used for a dissolved oxygen DO controller during sewage treatment; correcting sample data; training a neural network by using part of the corrected data, controlling dissolved oxygen (DO) by using the trained RBF neural network, and taking an error between anticipant DO concentration and actually output DO concentration and an error change rate as the input of the RBF neural network, wherein the output of the RBF neural network is the input of a frequency transformator, and the frequency transformator achieves the purpose of controlling a blower by adjusting the rotating speed of an electromotor so as to finally control aeration rate. The output of the whole control system is the actual DO concentration; the control effect of the controller is improved; the dissolved oxygen meets the anticipant requirements quickly and accurately; and the problem of poor self-adaptive capability based on switch control and PID control is solved.

The invention discloses a mini-inverter fault detecting method based on a neural network expert system. The mini-inverter fault detecting method includes: step 1, building an initial knowledge base, step 2, confirming a neural network topology structure and network parameters, and building an expert system based on the neural network comprising an input layer, a middle layer and an output layer, and step 3, sending data detected in real time to the expert system through a data processing module of the neural network expert system, calling data from the knowledge base by the expert system to compare the data with the data detected in real time, judging whether the system is faulted or not by the aid of speculation and analysis of an inference engine, outputting faulted information, timely detecting fault, providing related handing information and adopting multi-layer inputting topological structure. The mini-inverter fault detecting method is simple, efficient and high in detecting accuracy, so that maintenance staff can fix the fault as soon as possible, and loss and hazard caused by the fault can be prevented.

The invention relates to a photovoltaic failure detection method based on an improved particle swarm optimization Elman network, which is characterized by comprising the following steps: (1) initializing particle swarm algorithm; (2) constructing an Elman neural network topology structure; (3) determining the particle evaluation function and calculating the particle fitness value; (4) updating theparticles and introducing the mutation operator to obtain new population particles: re-determining the individual extreme value and the global extreme value, and obtaining the optimal particle when reaching the set precision or the maximum number of iterations; (5) obtaining the optimal weight values according to the optimal particles obtained in the step (4) to carry out network training and result prediction. The method obtains the optimal weight value of the neural network through the improved particle swarm algorithm, overcoming the defect of the Elman neural network trapped in local optimal solution, greatly improving the prediction efficiency and speed, and facilitating the maintenance and management of the photovoltaic power generation system.

The invention discloses a hyperspectral image waveband selecting method applying a neural network to carry out sensitivity analysis. The method comprises the steps that firstly, a subspace dividing method is used for predicting some waveband combinations with the poor relevancy, a training sample and a testing sample are determined according to a pre-selected surface feature type and original surface feature information, and a BP neural network topology structure is determined; secondly, the BP neural network is optimized through a differential evolution algorithm; finally, the Ruck sensitivity analysis is executed through the optimized BP neural network, sensitivity analysis results of all testing sampling points are integrated through a comprehensive judgment function, and the waveband having the greatest effect on the classification result is finally screened out.

The invention provides a deformation prediction method and apparatus based on Kalman filtering and a BP neural network, wherein the deformation prediction method based on the Kalman filtering and the BP neural network comprises: obtaining the deformation monitoring data of a monitored object in a project as a training sample; establishing a BP neural network topology model according to the training sample; learning the BP neural network topology model by using a Kalman filtering algorithm according to preset training parameters to adjust the weights of the neurons in the BP neural network topology model; and performing deformation prediction according to the BP neural network topology model with adjusted weights. The deformation prediction method based on Kalman filtering and a BP neural network can shorten the learning time of the BP neural network and improve the establishment efficiency of deformation prediction model in a deformation prediction process.

The invention discloses an industrial automation defect detection method based on deep learning, and relates to the technical field of product defect detection. Complex feature extraction work is avoided, the method has relatively high accuracy and relatively good generalization capability; under the conditions of small defect ratio and complex detection background, an applicable detection model can be quickly trained, a detection method with a better detection effect than that of a traditional detection method is obtained, the trained target detection model YOLO-V3 is used for processing a shot picture, and whether a detected object exists in the picture or not is judged; and the photographed pictures are recognized by utilizing a neural network topology Inception-V3 image recognition model obtained by migration learning of small sample data deployed on a server, and whether the target positions have defects or not is judged. The method is mainly applied to industrial production linedetection occasions.

The invention belongs to the field of industrial robots, and discloses an industrial robot constant-force grinding and polishing method based on big data. The method comprises the following steps of firstly, collecting robot running data, specifically, a six-component force sensor is connected with an industrial robot and a controller, and through continuous adjusting of the laminating degree of the same grinding track, a large amount of running data are collected for forming a training set; secondly, determining a BP neural network topological model; thirdly, according to the running data obtained in the first step, training the BP neural network topological model built in the second step; and fourthly, applying the trained BP neural network topological model to the grinding living example of the industrial robot without a sensor, obtaining a grinding force time domain curve in the running process of the industrial robot, and according to the preset grinding force threshold value, adjusting the track of the industrial robot, and obtaining the constant-force grinding effect. The path fine adjusting work is repeated, and the problems that grinding and polishing efficiency is low, and machining cost is high are solved.

The invention discloses an anti-interference method based on an intelligent antenna and a neural network algorithm. According to the method, a back frame structure of the antenna is used as a model; based on an artificial neural network, an intelligent antenna structure deformation estimator and a shape controller are designed; an improved error algorithm is adopted, so that the characteristics of high convergence rate and small calculation amount are realized; and a relatively high anti-interference characteristic is achieved. A typical applicable type is adopted by the conventional artificial neural network topology structure; the intelligent antenna is combined, and a bipolar intelligent antenna is used, so that the convenience in actual operation is achieved; the capacity of the 8-unit bipolar intelligent antenna can be met; the coverage aspect is within a normal cell coverage range; and obvious coverage loss does not exist. The transverse size of the bipolar intelligent antenna is reduced by over 50 percent compared with the conventional single-polar intelligent antenna and has obvious advantages of reducing the windward area, reducing the engineering mounting difficulty and reducing the electromagnetic radiation panic of common users.

The invention discloses a method and system for recognizing video images of motion states of vehicles in an expressway tunnel. The method comprises the following steps: carrying out partition on a video image of a vehicle moving object based on background modeling, and respectively carrying out extraction on multiple features of the vehicle image, wherein the features include a texture feature, a geometric feature and an edge feature; constructing a neural network topology architecture as a base classifier by using the extracted multiple features; carrying out integration on the base classifier by using an Adaboost method so as to form a strong classifier; carrying out vehicle image recognition through the strong classifier; and when the existence of a parking state is recognized, sending a signal to a video monitoring center, automatically switching the monitoring center to monitor the image, starting an intra-tunnel parking state alarm, and releasing a notice to the outside of a tunnel. According to the invention, the hazardous state of vehicle parking in the tunnel can be intelligently recognized by using images transmitted from a video surveillance camera back to the monitoring center, and an alarm is made in time, thereby improving the automation level of the management of the expressway tunnel.

The invention discloses a section operation performance detection method and system. The method comprises the following steps: step 1, determining a BP (Back Propagation) neural network topology structure, acquiring section performance detection index samples of different time periods, and establishing a sample set; step 2, optimizing a weight and a threshold of a BP neural network by use of a genetic algorithm, performing network training, and outputting an optimized BP neural network; and step 3, predicating section performance comprehensive indexes through the BP neural network optimized in the step 2 according to input section performance detection index real-time data. According to the method, various influence factors of controlling section operation performances are comprehensively contained, actual requirements of performing real-time detection and alarming on the section performance comprehensive indexes by an air traffic control unit can be met, and a data support effect on improving the control operation management level and optimizing a control airspace structure is achieved. The designed control section operation performance comprehensive detection system can be applied to the air traffic control unit and has high application engineering operability.

Methods and systems for neural network include configuring a physical network topology for a network that includes hardware nodes in accordance with a neural network topology, one of which is designated as a master node with any other nodes in the network being designated as slave nodes. One or more virtual neurons are configured at each of the hardware nodes by the master node to create a neural network having the neural network topology. Each virtual neuron has a neuron function and logical network connection information that establishes weighted connections between different virtual neurons. A neural network processing function is executed using the neural network.

The invention discloses an online biochemical oxygen demand (BOD) soft measurement method based on a dynamic feedforward neural network. The method comprises the following steps of: designing a dynamic feedforward neural network topological structure for BOD soft measurement of a sewage aeration tank, determining an input sample of the dynamic feedforward neural network, and performing online normalization processing on the input sample; calculating the variation condition of an ownership connection value connected with a hidden node in the neural network in each training process by employing a standardized data training neural network, judging the activeness of the hidden node, and splitting the hidden node with high activeness; judging the capacity of learning information of the hidden node by calculating the absolutely output variation condition of the hidden node in the training process, and deleting a hidden node without the learning capacity; adjusting parameters of the neural network; and determining the BOD of effluent of the aeration tank after the training process of the dynamic feedforward neural network is ended. The method has the advantages of high real-time property, high stability, high precision and high neural network generalization ability.

The invention provides a BP artificial neuron network's injector performance predicting method based on the use of deep Adaboost algorithm. The method comprises: collecting relevant parameters of a given injector; according to the neuron network's topological structure, determining the number of the neurons in the neuron network's input layer, the hidden layer and the output layer; inputting samples and starting to train the neuron network created from the step 2 and repeating the training several times; forming a week classifier after each training; recording the error of each training result; creating a strong classification function; synthesizing the weak classifiers into strong classifiers; creating a super-strong classification function according to the corresponding weights assigned according to the prediction result; synthesizing the strong classifiers into a super-strong classifier wherein the super-strong classifier is a deep BP- Adaboost neuron network; acquiring the real-time measurement data of the given injector; and inputting the data to the created and completed BP- Adaboost neuron network to obtain the output vector, or the prediction value. The method of the invention achieves high-precision predictions and consumes a shorter time to do so.

The invention relates to an engine ECU (electronic control unit) circuit fault diagnosis method, and belongs to the technical field of an engine. The method comprises the following steps that 1, a neural network topology structure is built, and a BP neural network input-output mode mapping relationship is built; 2, input samples are introduced from an input layer and are transferred to an output layer after the layer-by-layer processing of each hidden layer, wherein the initial weight value of a BP neural network is optimized through a genetic algorithm, and a better searching space is determined; the network is finely adjusted in a local solution space by an L-M method, and the optimal solution or the approximate optimal solution is searched out. The scheme has the advantages that the initial weight value of the BP neural network is determined by the genetic algorithm, the better searching space is used for replacing the random selection of the ordinary initial weight value, and the convergence speed is accelerated.

The invention discloses a risk rating method for optimizing a Hopfield neural network based on a firefly algorithm, and the method comprises the following steps: firstly, determining a performance period and a risk level, extracting a modeling sample customer, and obtaining customer data as a modeling index system, the customer data comprising the risk level and credit data affecting repayment performance; preprocessing the credit data, and randomly segmenting a training set and a test set; constructing a Hopfield neural network topological structure according to the data features of the modeling sample, determining the parameters of the network, and initializing the weight and threshold of the Hopfield neural network; and constructing a mapping relation between the weight and the threshold of the Hopfield neural network and a firefly algorithm, obtaining an optimal weight and an optimal threshold through the firefly algorithm, and training the Hopfield neural network by using the training set. According to the method, the optimal weight and threshold of the Hopfield neural network are determined by using the firefly algorithm, the convergence speed of the neural network is accelerated, the accuracy of the prediction model is improved, and the requirement of real-time evaluation of Internet financial credit can be met.

The present invention discloses a driving behavior scoring method based on the BP neural network. The method comprises the following steps: (1) collecting driving behavior data, and determining a driving behavior index system; (2) constructing a BP neural network topology model; (3) performing data normalization processing on the collected driving behavior data samples; (4) using the constructed BP neural network topology model for training learning; and (5) calculating the driver behavior score according to the scoring system formula. According to the driving behavior scoring method based onthe BP neural network disclosed by the present invention, the technical problems that the traditional scoring method has insufficient accuracy and one-sided evaluation are solved, and by constructingthe BP neural network, the driving behavior of the driver can be comprehensively, objectively, scientifically and normatively evaluated.

The invention discloses a nuclear accident source item inversion method, which comprises the following main steps: firstly, determining a target signal of a nuclear accident source item inversion andextracting a main influence factor in an environmental influence factor as an input variable by using a PCA method; then, determining a BP neural network topology and normalizing the input variables;after that, determining the number of hidden layer nodes of the BP neural network model and training the neural network model to debug an optimal parameter; finally, using a BP neural network weight and threshold value optimized through an MEA method to establish a PCA-MEA-BP neural network nuclear accident inversion model; and carrying out comparison and analysis on the PCA-MEA-BP neural networknuclear accident inversion model and an unoptimized BP neutral network model result. According to the nuclear accident source item inversion method in the invention, an effective characteristic quantity is extracted by using the PCA method, thereby improving the classification precision, reducing the training time of the neural network and simplifying the reversion model structure; moreover, the MEA algorithm optimizes the neural network weight and threshold value, thereby effectively reducing the network instability and improving the availability of the reversion model.

The invention discloses a prediction method for piston cutting deformation based on a BP neural network. The method comprises: piston cutting machining simulation is carried out through finite elementanalysis; taking the predicted deformation obtained by the simulation experiment as a training sample, establishing a BP neural network topology model according to the training sample, learning the BP neural network topology model according to preset training parameters, and finally predicting the piston cutting deformation according to the BP neural network topology model after the weights of the neurons are adjusted. Compared with a simulation process prediction method, the method has the advantages that the prediction time is greatly shortened, the establishment efficiency and prediction precision of the prediction model are improved, and a processing prediction guide with relatively high reliability can be quickly provided for production and processing.

The invention discloses a method for predicting blasting vibration characteristic parameters based on an SAGABP algorithm, and belongs to the technical field of blasting vibration. The method comprises the following steps: collecting blasting vibration influence factors in a blasting engineering field, and then determining hole depth, packing, chassis resistance line, elevation difference, blasting source distance, final assembly charge and maximum section charge as a training sample and a prediction sample according to a main analytic hierarchy process; determining a BP neural network topological structure, calculating an optimal weight value and a threshold value by applying a genetic simulated annealing algorithm (SAGA), and decoding and assigning the optimal weight value and the threshold value to a BP neural network system for training; preliminarily constructing a blasting vibration characteristic parameter prediction model; performing error analysis on a prediction result; and finally, carrying out field prediction on the blasting vibration characteristic parameters in the blasting engineering field. According to the method, the optimal solution can be searched only by optimizing a small number of samples through the improved hybrid intelligent algorithm. Meanwhile, the convergence speed is increased, and the situation of falling into the local optimal solution is avoided.

Techniques are disclosed relating to configuring a neural network based on information received via a dashboard user interface. In some embodiments, a computing system displays a dashboard that includes a set of plots for displaying data and user interface elements that may be used to configure the number and type of the plots. The plots may display information of various kinds, including raw or processed data, relationships between data, processes applied to data, etc. and may be different types, including, e.g., spark lines, scatter, or time series, etc. The dashboard module is operable to communicate the user input to a module operable to generate a neural network topology. User input to the dashboard may provide information regarding sources of data to be used for generating plots, or training or running the neural network. Results based on processing data using the trained neural network may be displayed on the dashboard.

The invention discloses a laser spectrum noise reduction method and a laser spectrum noise reduction device based on deep learning optimization S-G filtering, and belongs to the field of laser spectrums. The laser spectrum noise reduction method comprises the following steps of: collecting absorption spectrum data by adopting a gas laser absorption spectrum experimental device based on a QCL quantum cascade laser or other tunable laser sources, and acquiring the absorption spectrum data of a gas to be detected as an Adam algorithm neural network training sample; establishing an Adam algorithmneural network topology model according to the spectral data training sample, and selecting an optimal filtering parameter combination; adopting an S-G filtering algorithm to carry out adaptive filtering according to the optimized filtering parameters; correcting filtering values by using the constructed Adam algorithm neural network; and performing noise reduction processing on the spectral dataaccording to the optimized S-G filtering algorithm. According to the laser spectrum noise reduction method based on deep learning optimization S-G filtering provided by the invention, a signal-to-noise ratio of spectrum filtering can be improved, so that the absorption spectral line measurement of gas becomes more accurate.

The invention discloses a solder paste printing quality prediction method and system based on IGA-DNN, and the method comprises the steps: collecting solder paste printing key process parameters and solder paste printing relative volume original data, carrying out the preprocessing, determining a deep neural network topological structure according to the types of input and output variables, the solder paste printing relative volume predicted by the deep neural network is used as a fitness function of the genetic algorithm; finally, a genetic algorithm is used for optimizing and solving the weight and the threshold value of the deep neural network, algorithm model training is completed through a training sample. However, with a traditional neural network which adopts a gradient descent method to update hyper-parameters of the network, the network search speed is low, and local optimum is extremely prone to occurring. According to the method, the hyper-parameters of the deep neural network are initialized based on the genetic algorithm, the convergence speed and precision of the neural network are improved, high precision and stability of SMT solder paste printing quality prediction are achieved, and meanwhile corresponding technical method support is provided for SMT solder paste printing process parameter optimization.

The invention discloses a BP neural network microwave remote sensing soil moisture inversion method optimized by considering a firefly algorithm, and the method comprises the following steps: 1, obtaining a corresponding ALOS-2L waveband radar 1.1-level remote sensing image in a research region, carrying out the preprocessing of the image, obtaining a total backscattering coefficient, and simultaneously and synchronously acquiring CLDAS-V2.0 soil moisture data in the same time for model calculation and verification; 2, the research area being a vegetation coverage area, according to a water cloud (WCM) model, removing the influence of vegetation in the research area on the soil backscattering coefficient, obtaining the soil backscattering coefficient, meanwhile, obtaining Landsat-8 optical data in the same area within the same or similar time, calculating related vegetation indexes through band operation after preprocessing, and providing data support for the water cloud model; and 3, according to a BP neural network topological structure, establishing a corresponding data set for the soil backscattering coefficient obtained in the step 2 and CLDAS soil moisture data, and optimizing the BP neural network by using a firefly algorithm so as to perform soil moisture inversion.

The invention discloses an engine emission real-time prediction method, which comprises the following steps of: firstly, acquiring a plurality of known engine emission historical test data samples, dividing the samples into a training set and a test set to train a neural network, and calculating neural network output root-mean-square errors under different hidden layer nodes to determine a neural network topological structure; and then the initial weight and threshold of the neural network are optimized through a mind evolutionary algorithm, and finally an engine emission real-time prediction system is established by using an Adaboost algorithm. The problems that an existing engine emission data acquisition mode wastes time and labor, is limited by environmental factors, is high in instrument cost, is poor in transient emission measurement performance and the like are solved, and the transient emission data of the engine can be measured only by simply measuring the rotating speed, torque, power, track pressure, air-fuel ratio, oil consumption, EGR (exhaust gas recirculation) rate and SOI (oil injection time) in the operation process of the engine. Therefore, transient NOx emission, THC emission and CO emission of the engine can be accurately predicted in real time.

The embodiment of the invention discloses an advertisement recommendation method and system based on machine learning, and the method comprises the following steps: user behavior tracking: tracking the Internet behavior of each user, and mining and analyzing the real-time demands of each user; recommendation object classification: classifying recommendation objects into minimum units according toproject classification, and preliminarily establishing a mapping relationship between user real-time demands and the minimum units of the recommendation objects; recommendation object updating: updating the real-time demands of the user in real time according to the click rate of the minimum unit of the recommendation objects, and ranking the sequence list of the recommendation objects again; userdemand expansion: performing neural network topology on all real-time demands according to the click rate ranking of the minimum unit of the recommended objects, and updating the real-time demands ofeach user in real time; according to the scheme, different types of advertisement weights are distinguished in different time periods, so that the influence of search content types caused by work reasons on basic interest contents is avoided, and the accuracy and effectiveness of advertisement recommendation are improved.

The invention discloses a network topology inference method and system based on a convolutional neural network, and the method comprises the steps: judging three paths of sub-topology structures through the convolutional neural network, taking the information as an input, and inferring a network topology through employing a topology inference algorithm in the invention. The topology inference algorithm is the same as that of most traditional topology inference algorithms. According to the algorithm provided by the invention, the topology is constructed by determining the positions of the branch nodes and the nodes connected to the branch nodes; the difference is that three paths of sub-topological structure information are input into the algorithm, and compared with similar measurement type quantitative data input in a traditional method, the qualitative data is higher in error tolerance degree, and therefore better robustness is achieved.

The invention discloses an electronic expansion valve flow characteristic prediction method based on a particle swarm optimization BP neural network. The method comprises the following steps that an experimental data sample is obtained and preprocessed; constructing a BP neural network topological structure; optimizing an initial weight and a threshold value of the BP neural network by adopting a particle swarm algorithm; inputting the optimized initial weight and threshold into a BP neural network for sample training and testing; and performing flow characteristic prediction on the trained neural network under different working conditions. The method has the advantages that the flow characteristics of the electronic expansion valve are predicted by adopting a BP neural network method, the initial weight and the threshold value of the BP neural network are optimized by adopting the particle swarm optimization algorithm, the defect that the BP neural network is slow in convergence speed and even does not converge or falls into a local minimum value is overcome, the method is convenient and efficient, and flow prediction is accurate.

The invention provides a self-organizing neural network topology preservation reinforcing method based on deep learning. The method is used for settling a technical problem of requirement for improving self-organizing neural network topology preservation effect in prior art. The method comprises the steps of setting a network structure and a parameter of the self-organizing neural network, and normalizing input layer data; setting the number of input-layer neurons and the number of competition-layer neurons; performing rough adjustment on the weight vector of the competition-layer neurons of the self-organizing neural network, and obtaining a rough-adjusted competition-layer weight vector; by means of the rough-adjusted result, performing fine adjustment on the weight vector of the competition-layer neurons of the self-organizing neural network, and obtaining a fine-adjusted competition-layer weight vector; measuring the competition-layer weight vector, thereby obtaining a topology preservation reinforcing effect. The self-organizing neural network topology preservation reinforcing method has advantages of reducing difference between each competition-layer weight vector of the self-organizing neural network and an input sample, and improving topology preservation capability of the self-organizing neural network.

The invention discloses a probabilistic neuron circuit, and a probabilistic neural network topological structure and application thereof. The probabilistic neuron circuit comprises an integrating capacitor, a non-fixed threshold volatile device and a load resistor, wherein one end of the integrating capacitor is externally connected with a synaptic resistor and connected with one end of the non-fixed threshold volatile device, and the other end of the volatile device is connected with one end of the load resistor. The network topological structure comprises a plurality of input neuron circuits, a plurality of probabilistic neuron circuits and a lateral suppression neuron circuit, wherein each probabilistic neuron circuit is used for carrying out random excitation based on the non-fixed excitation threshold value of the probabilistic neuron circuit and an electric signal emitted by each input neuron circuit; and the suppression neuron circuit is used for suppressing the excitation of other subsequent probabilistic neuron circuits when receiving signals excited by the first n probability neuron circuits. According to the invention, the non-fixed threshold volatile device is introduced into the neuron circuit, so the application of the neuron circuit is greatly expanded, and the neuron circuit can be particularly used for solving non-deterministic problems and has a reliable solution result.

The invention discloses a BP neural network coal conveying fault prediction method based on genetic algorithm optimization. The method comprises the following steps: determining a BP neural network topological structure diagram; taking harbor office unit parameters as input variables in the BP neural network topological structure diagram, and performing data normalization processing to obtain an initial population; coding individual chromosomes in the initial population, and evaluating the fitness value of each chromosome; selecting excellent individuals according to the fitness values, and performing selection, crossover and mutation operation to obtain an optimal individual; establishing an improved BP network learning rate optimization model for optimally calculating a BP neural network weight and a threshold value; and decoding by using the optimal individual and assigning the optimal individual to the BP neural network as network initial weight and threshold input, and training the neural network to obtain an optimal prediction model. The coal conveying system fault can be predicted in advance, software and hardware are easy to implement, the cost is low, the convergence speed of a traditional neural network is increased, and the network training efficiency is improved.