84 results about "Adaptive learning rate" patented technology

The invention belongs to the technical field of radars and particularly relates to a dual deep neural network-based radar range profile target identification method. The method comprises the steps of firstly performing preprocessing operations of random distance disturbance, sample expansion, noise addition and the like on range profile data of a target to enhance the robustness of an identification system; secondly in combination with a deep learning theory, proposing a dual deep neural network (DDNN) with an adaptive learning rate, and performing unsupervised pre-training and supervised fine adjustment on the DDNN to obtain DDNN model parameters; thirdly performing pre-identification on test samples by utilizing the DDNN to obtain pre-identification results of the samples in two sub-networks; and finally according to the pre-identification results, performing time-space multi-level decision fusion by utilizing an improved DS evidence theory to obtain a target identification result.

The invention discloses a core three-dimensional image super-resolution method, which comprises the following steps: (1) sending an image in a training set to a three-dimensional convolutional neuralnetwork proposed by the method, wherein the first layer of the network performs low-frequency feature extraction; (2) allowing the second to eleventh layers of the network to be responsible for learning a mapping relationship between low frequency and high frequency features; (3) allowing the twelfth layer of the network to use the learned mapping relationship to map the low frequency features into the high frequency features; (4) using a residual learning method to calculate a root mean square error, and accelerating the training by using the momentum gradient descent method; (5) using the adaptive learning rate and a gradient cutting method to optimize the training process during the process of training, and using the training configurations in (1) to (5) to perform continuous iterativetraining; and (6) using the trained network model to complete the reconstruction. The invention can improve the resolution of a rock CT three-dimensional image, restore more structure and details, andprovide clearer image samples for the next step of geology-petroleum research.

The invention discloses a fast self-learning improved ADRC control method for a non-linear system, which comprises the following steps: S1, creating an active disturbance rejection controller (ADRC) which includes a tracking differentiator (TD) part, an extended state observer (ESO) part, a non-linear state error feedback (NLSEF) part and a disturbance compensation part, wherein S1 includes S11, establishing a tracking differentiator (TD) control model, S12, establishing an extended state observer (ESO) control model, S13, establishing a non-linear state error feedback (NLSEF) control model, and S14, establishing a disturbance compensation control model; S2, creating a self-learning active disturbance rejection controller (SADRC): applying a self-learning method to the non-linear state error feedback (NLSEF) control model to establish a self-learning non-linear active disturbance rejection control system model; and S3, creating a fast self-learning active disturbance rejection controller (FSADRC): designing a learning rate oriented adaptive mechanism by using an additional momentum term method and establishing a fast self-learning model based on dynamic adaptive learning rate.

The invention discloses a fault diagnosis method and system for a heading machine hydraulic system. The method includes: building a fault diagnosis system framework through a fuzzy neural network method, building subsystems such as a parameter monitoring module, a fault knowledge base management maintenance module and an intelligent diagnosis reasoning module, taking a knowledge base of the heading machine hydraulic system as the basis and combining each module into a whole organically; building the overall structure of the knowledge base which comprises a fault type base, a fault knowledge base and a fault rule base according to the expert system design principle, introducing a relational data base into a knowledge base system, using the ACCESS as the database platform, building corresponding data sheets, realizing the functions of the expert system and managing and maintaining the knowledge base through fully utilization of the database technology; and building a fuzzy neural network fault diagnosis mode, reducing the network instability through the adaptive learning rate method and the additional momentum method, and training and simulating the fuzzy neural network model through actual data. Therefore, the fault diagnosis method and system for the heading machine hydraulic system can accurately reflect the faults of the heading machine hydraulic system.

An intelligent power generation control method based on multi-agent reinforcement learning having time tunnel thought includes the following steps: determining a state discrete set S; determining a combined action discrete set A; collecting real-time operating data of each power grid, calculating an instantaneous value of each area control error ACE(k) and an instantaneous value of a control performance standard CPS(k), and selecting search action a<k>; in the current state s, obtaining a short-term award function signal R(k) by a certain area power grid i; obtaining value function errors rho<k> and delta<k> through calculation and estimation; updating a Q function table and a time tunnel matrix e(s<k>, a<k>) corresponding to all states-actions (s, a); updating Q values and updating a mixed strategy pi(s<k>, a<k>) under the current state s; then updating a time tunnel element e (s<k>, a<k>); selecting a variable learning rate phi; and updating a decision change rate delta (s<k>, a<k>) and a decision space estimation slope delta<2>(s<k>, a<k>) according to a function. The intelligent power generation control method based on multi-agent reinforcement learning having time tunnel thought aims to solve the problem of equalization of multi-area intelligent power generation control, has a higher adaptive learning rate capability and a faster learning speed ratio, and has a faster convergence rate and higher robustness.

The invention discloses a moving target detection method based on a deep optical flow and a morphological method, which includes the following steps: (1) collecting video data, marking sample videos,randomly dividing the sample videos into a training set and a testing set, carrying out mean calculation on the processed training set and the processed testing set to form a training set mean file and a testing set mean file, and completing the preprocessing of the training set and the testing set; (2) constructing a fully convolutional neural network architecture composed of a coding part and adecoding part, and carrying out training by using the training set and the testing set through an adaptive learning rate adjustment algorithm to get trained model parameters; (3) inputting image dataneeding detection to a trained fully convolutional neural network to get a corresponding deep optical flow graph; (4) processing the deep optical flow graph through an Otsu threshold adaptive segmentation method; and (5) morphologically processing the data after threshold segmentation, removing outliers and slots, and finally obtaining a detected moving target area.

The invention discloses a road motor vehicle exhaust emission prediction method for improving an attention bidirectional long-short-term memory network. The method comprises the steps of 1, using PEMSand OBD detection equipment for jointly collecting motor vehicle exhaust emission data; 2, performing missing data compensation and normalization preprocessing on the tail gas emission data set; 3, establishing an improved Attention-Bi-LSTM attention bidirectional long and short term memory network model; 4, determining hyper-parameters of the model by adopting a pre-experiment; and 5, optimizingmodel parameters by adopting an adaptive learning rate algorithm to finish prediction model training. According to the method, all characteristic factors influencing road motor vehicle tail gas emission can be fully considered, the tail gas emission prediction precision is improved, and the method has a large application range, so that the PEMS tail gas emission test time can be effectively shortened, and the consumption of manpower, resources and time cost is reduced.

The invention relates to a mechanical property prediction method of a filling material for filling mining, especially to a mechanical property prediction method of a cemented filling material. On the basis of an improved BP neural network and an existing data sample, a mechanical property of a matching material that has not been tested can be predicted. The method comprises the following steps that: mechanical property influence factors of a cemented filling material are selected as input layer nodes, wherein the factors include slurry concentration and a dry material proportion and the like; mechanical property indexes needing prediction are selected as output layer nodes, wherein the mechanical property indexes contain a slump, a bleeding rate, and compressive strengths at different ages and the like; mechanical property values of different matching materials are tested and obtained data are used as training samples and test samples; input and output vectors are determined, a BP neural network that is improved by introducing an additional momentum method and an adaptive learning rate is constructed, and the network is trained and tested; and a condition needing prediction is inputted, thereby outputting a corresponding mechanical property value. According to the prediction method, the efficiency is high, the result is precise, and the cost is low; and demands of production and practice activities can be satisfied.

The invention relates to an electric energy quality prediction method based on similar days and improved LSTM. The method comprises the following steps: 1) collecting electric energy quality steady-state index data of a certain monitoring point in a certain region within a period of time and meteorological data of the region; 2) performing feature dimension reduction on the meteorological data byadopting a kernel principal component analysis method to obtain similar day feature vectors; 3) calculating meteorological factor matching coefficients of the historical day and the to-be-predicted day by adopting a grey correlation algorithm, and determining a similar day set; 4) selecting power quality historical data similar to the day to be predicted and similar day set data as a training sample set of the LSTM neural network, and optimizing LSTM neural network parameters by adopting a method of combining a Dropout algorithm and an Adam adaptive learning rate optimization algorithm, and taking the similar day feature vector as a model input variable to obtain a prediction result of the electric energy quality of each moment of the to-be-predicted day. Compared with the prior art, the method has the advantages of avoiding training overfitting, avoiding information interference and information repetition, improving prediction accuracy and the like.

An online Gaussian mixture learning model for dynamic data utilizes an adaptive learning rate schedule to achieve fast convergence while maintaining adaptability of the model after convergence. Experimental results show an unexpectedly dramatic improvement in modeling accuracy using an adaptive learning schedule.

The invention relates to a deep learning network intrusion detection method based on an improved learning rate. Determination method for learning rate and iteration frequency of improved deep belief network model in training is adopted, and the advantages of the improved deep belief network model and a softmax regression function are utilized to establish a deep belief network for network intrusion detection; and the softmax multi-classification combination model is used for training the model by using the public training data of the network intrusion detection data set and identifying and classifying the test data of the network intrusion detection data set by using the trained model. Rapid convergence of model parameters is realized by using an adaptive learning rate, and the optimized deep belief network-softmax multi-classification combination model is used for an intrusion detection system, the identification accuracy of attack behaviors can be effectively improved, and meanwhilethe detection efficiency can be improved.

The invention discloses an electric energy quality disturbance positioning and identifying method based on the lifting of the wavelet and the improvement of a BP neural network. The method comprises the following steps: using the Euclidean decomposition algorithm to obtain a db4 wavelet lifting scheme; performing lifting wavelet decomposition to the disturbance signal; in combination with the modulus maxima, performing the positioning detection to the disturbance mutation point peak; and utilizing the method of combining the self-adaptive learning rate with the incremental momentum term to improve and carry out disturbance identification to the BP neural network. The method of the invention can better obtain the disturbance period information, achieves fast positioning and high precision, and overcomes the shortcoming that a traditional BP neural network is prone to fall into local minimal points and slow convergence speed, therefore, achieving a high identifying rate for the power distribution network system's electric energy quality disturbance.

The invention relates to the technical field of wavelet neural network optimization, in particular to an adaptive learning rate wavelet neural network control method based on normalization lowest mean square adaptive filtering. The adaptive learning rate wavelet neural network control method comprises the steps that a control system model is built; unitization is conducted on all weight values of a wavelet network by layer; wavelet neural cell weight value optimization is carried out; an error signal and training cost are figured out; segment processing is conducted on derived functions of an activation function through a step function; fuzzy rules of fitting the derived functions are made; a membership function is determined; the proportion of each fuzzy rule in a derived function value is determined; a fuzzy system is output, and the activation function is displayed in a linearization mode; induction local areas of all neural cells are determined, and the neural cells are output; each local gradient function is solved; adjustment of the learning rate is conducted by an output layer in an adaptive mode; the range of the learning rate of the output layer is determined; the learning rate of a hidden layer is adjusted; neural cell synapse weight values are trained; a tracking control signal is output; closed-loop feedback control is completed. According to the adaptive learning rate wavelet neural network control method based on normalization lowest mean square adaptive filtering, the rate of convergence can be increased, and computation complexity can be reduced.

The invention discloses an adaptive learning rate BP neural network algorithm. The back propagation process of the BP neural network algorithm is optimized, and different learning rates are dynamically adapted to the connection weight of different neurons so that optimization of each direction is ensured to search the optimal solution of the direction, the adjustment efficiency of the weight can be enhanced to the greatest extent to accelerate the convergence speed of the whole training. The algorithm comprises the step one: network initialization; step two: sample inputting; step three: forward propagation; step four: result judgment; and step five: back propagation. The algorithm is simple in program and easy to implement and great in improvement effect so that the training time of the BP neural network can be greatly shortened, the problem of being prone to the local minimum of the present algorithm can be effectively overcome, the universality is high and thus the algorithm has great practical application value.

The invention discloses a decentralized federated learning method, device and system, and belongs to the field of federated learning, and the method comprises the steps: building a global communication network among a plurality of clients participating in federated learning, and enabling a communication path to exist between any two clients; each client respectively receiving the model parameters of other clients directly communicating with the client at the previous moment, respectively calculating the sum of the products of the obtained model parameters at the previous moment and the corresponding weight coefficients, and calculating the first product of the loss function gradient of the local model at the previous moment and a preset adaptive learning rate; updating a model parameter at the current moment to a difference value between the model parameter and the first product; and repeatedly executing the iteration updating operation until the loss function of the local model of each client is not higher than the corresponding threshold value, or until the number of times of repeated execution reaches the maximum number of iterations. While privacy and data security of each client are protected, each local model is globally trained.

The invention discloses an improved BP neural network-based coal pyrolysis product prediction method. A coal pyrolysis product is accurately predicted by adopting industrial analysis and element analysis data and pyrolysis temperature of coal; an initial weight and a threshold of a BP neural network are optimized by applying a PSO algorithm and GA combination method; and an adaptive learning rateis embedded in a calculation process of the BP neural network. The stability of the BP neural network is improved through the PSO algorithm and the GA; and the BP neural network is quickly converged by adding the adaptive learning rate, so that the calculation efficiency and the data prediction precision of the BP neural network are improved. The method has the advantages of simple model, easy operation, high prediction precision, favorability for promotion in the present stage and the like.

The invention discloses an image recognition method based on a self-adaptive full convolution attention network. The invention discloses a self-adaptive full convolution attention network-based imagerecognition method, belongs to the technical field of image recognition methods, and aims to provide a method for improving the image recognition speed and accuracy by using the self-adaptive full convolution attention network and further applying a trained model. According to the technical scheme, the image recognition method based on the self-adaptive full convolution attention network comprisesthe following steps of 1, training a neural network: a, selecting an activation function; b, a data preprocessing step; b, adaptive learning rate; d, defining a network structure: (1) training a fullconvolution attention network; (2) training an adaptive image; Secondly, the model trained in the first step acts on a target image set, the data obtained in the second step is compared with the target image set, and therefore images are recognized and classified; The method is applicable to the field of image recognition.

A seven-step hand washing method operation standardization automatic detection method comprises the following steps of S1, using a camera arranged above a wash basin for collecting a hand washing video; S2, performing screenshot on the hand-washing video and preprocessing the hand-washing video according to the actual condition of hardware and the requirement of the detection frame rate; S3, inputting each frame of preprocessed image into a clipped YOLOv3 gesture type detection network trained by adopting an adaptive learning rate algorithm to obtain a gesture type of each frame and a confidence coefficient thereof; S4, calculating the arrangement sequence, the effective time and the corrected average confidence coefficient of each gesture type based on all detected gesture information; and S5, based on the obtained characteristic value of each gesture type, scoring according to a normativity evaluation rule, and carrying out normativity evaluation on the hand washing operation according to the normativity evaluation rule. The objectivity of evaluation is improved, and implementation of effective supervision is facilitated; the gesture type detection speed and accuracy are improved; and evaluation rationality is improved.

The invention discloses an attention mechanism-based optical laryngoscope image lesion area labeling method, and mainly solves the problems that in the prior art, lesion area labeling is not performedon an optical laryngoscope image, and overfitting is easy to occur on a small data set. According to the implementation scheme, a laryngoscope image data set and an image data set label are obtained;scaling and centralizing the image data set, and marking the centralized image data set and the image data set label as a training data set; constructing an 18-layer network, taking the training dataset as training data of the network, and optimizing the network by using an adaptive learning rate optimization algorithm to obtain a trained network; and inputting a laryngoscope image into the trained network, obtaining a corresponding lesion area in the label data set according to the generated report, and labeling the lesion area. According to the method, overfitting of a small data set is avoided, the lesion area of the optical laryngoscope image can be obtained and marked, and a doctor can diagnose the optical laryngoscope image conveniently.

The invention discloses a CNN based fault diagnosis method of a DC / DC converter. The method comprises the following steps including 1) error data collection; 2) data preprocessing; 3) CNN deep training; and 4) diagnosis precision test and use. The CNN is used to train the data, deep learning training skills are combined, data reinforcement and adaptive learning rate are used to solve the fitting problem, and a fault feature is extracted; and the representative feature can be extracted from original data based on signal processing knowledge and engineering experience of the spectif equipment and fault type needless of establishing an accurate mathematical model, and compared with traditional artificial feature extraction and deep neural network methods, the method of the invention has a higher diagnosis performance and can achieve high diagnosis precision.

The invention discloses a multi-view language recognition method based on unidirectional self-tagging auxiliary information. The method comprises the following steps: firstly, implementing self-tagging on current words and word-level auxiliary information by virtue of a tagging model, so that probability distribution of self-tagging auxiliary characteristics of the current words is obtained; then, decoding the probability distribution of the self-tagging auxiliary characteristics by virtue of Viterbi, so that relatively accurate auxiliary characteristics are obtained, and bidirectional auxiliary information is converted into unidirectional auxiliary information; and inputting the unidirectional auxiliary information, together with the current words, into a multi-view language model for analysis, so that accurate semantics of the current words can be obtained. The multi-view language recognition method provided by the invention has the characteristics that on the basis of the word-level auxiliary characteristics in a multi-view neural network, adverse influence on post-text information is eliminated, the various word-level auxiliary information is adopted, the word-level auxiliary characteristics, which are represented as a tree structure, are introduced to the multi-view language model for training, in the tagging model and the language model, stable operators are adopted to regulate various adaptive learning rates and the like.

The invention discloses a dual-layer long-short term memory (LSTM) network-based early-stage state identification method of a 10kV single-core cable. The LSTM network-based early-stage state identification method is applicably used in the electrical field. The LSTM network-based early-stage state identification method comprises the steps of firstly, selecting from observable electrical quantity toobtain five types of current observable electrical quantity, and extracting a time sequence from the five types of current observable electrical quantity to construct cable early-stage state combination time sequence characteristic matrix; secondly, constructing a dual-layer LSTM network of an time sequence handling input according to the characteristic of identification matrix size; and thirdly,performing model training under supervised learning by a self-adaptive learning rate optimization algorithm to obtain a cable early-stage state identification model. By the LSTM network-based early-stage state identification method, big mass running by the cable can be fully utilized, the time sequence is extracted from five types of observable data to construct a combined time sequence characteristic matrix as an input of the dual-layer LSTM network under the condition that statistical characteristic is not used, a corresponding relation between an input and an output is determined by handling capability of the dual-layer LSTM on the time sequence input, and the cable early-stage state identification is further completed. By the LSTM network-based early-stage state identification method,the identification accuracy can reach 99.06%.

The invention discloses a visual statistical method for counting underground drill rods, which specifically comprises the following steps of: firstly, storing a drill rod unloading video, splitting frames, and normalizing the drill rod unloading video; establishing a binary classification data set, and dividing into corresponding categories of a database according to the content of the single-frame image; increasing the number of data sets through a preprocessing method combining rotation, overturning and brightness enhancement; training an improved self-adaptive learning rate ResNet-50 network; detecting the category of each frame of image in the video and outputting a confidence coefficient percentage; and when all the images pass through an adaptive learning rate ResNet-50 model, clearing confidence coefficients of all second-class non-rod-unloading results in the CSV file, filtering video output confidence coefficients by using an integral method, and finally, calculating the number of unloaded rods in the video through a falling edge. The method is high in detection precision, errors caused by model detection can be effectively reduced, the final detection precision is improved, and meanwhile the problem that the detection precision is reduced due to shielding is avoided.

The invention provides a motor fault diagnosis method based on GRU network stator current analysis. The method comprises the steps of collecting stator current variable data; wirelessly transmitting and storing the acquired stator current variable data to a cloud database; preprocessing collected stator current variable data sets, and dividing all the data sets into multiple groups of training, verification and test sets; in the working state of a three-phase induction motor, modeling the acquired stator current variable data by using a GRU neural network; updating GRU model parameters by adopting an Adam adaptive learning rate method; carrying out sequence division on training set data according to GRU input dimensions, sending data of each sampling point into a GRU unit, carrying out feature extraction by the GRU, starting to train the model, and testing a verification set to further debug GRU model hyper-parameters; taking the features extracted by the GRU as the input of a full connection layer; and performing final current signal fault diagnosis by using the data of the test set.

The invention discloses a parallel cold load prediction method based on a building space unit, and the method comprises the steps: completing the division of the building space unit according to a target building plane graph and a building space layout; designing a DCN topological structure of a distributed controller of the building space unit according to the division basis, and completing the installation of a DCN; establishing an improved adaptive learning rate deep belief network-partial least square method ADBN-PLSR cold load prediction model; enabling each DCN to download the predictionmodel; enabling a user to initiate a prediction instruction through any DCN, and enabling the DCN to execute an instruction task and transmit the instruction to the whole DCN network based on a spanning tree and a topological structure; enabling the DCNs to independently complete prediction of a controlled area in parallel, and finally acquiring a cold load prediction result of the whole buildingat an initiating node through a prediction result summation instruction. According to the space unit independent prediction method, the fluctuation characteristics of the building cold load are highly extracted, and the problems of low prediction precision and the like caused by the characteristics of high dimension, nonlinear dynamics and the like of cold load data are solved.

The invention discloses an urban traffic scene vehicle detection method based on a robust mixed Gaussian model, comprising the following steps: (1) collecting an urban traffic scene video in real time, and quickly getting an ideal urban traffic scene background model; (2) introducing an image counter, an image foreground detection counter, a background confidence image counter, an image update logo and the traffic status at each pixel of an image, and setting corresponding initial values; (3) judging the traffic status at each pixel of an image in the current scene; (4) judging whether the confidence of each pixel of a background model is updated; (5) judging whether the background model of each pixel is updated; (6) during background updating, updating the background model with an adaptive learning rate according to the traffic status in the current scene; and (7) carrying out urban traffic scene foreground detection. The method is used to realize vehicle counting, vehicle model classification, vehicle tracking and traffic parameter acquisition so as to realize intelligent management of tollgate video data.

The invention discloses a character-level language model prediction method based on the local perceptual recurrent neural network. A processing mode of the recurrent neural network is utilized, threelayers of networks are combined together in layers, the low layer obtains features among local characters, the high layer obtains semantic features of texts, so a new model is made to have the stronger information integration capability, and processing longer data sequences is easier. The method is advantaged in that the monitored BPTT-RNN (Recurrent Neural Network Reverse Propagation Algorithm) method is utilized for training; firstly, adadalta (Adaptive Learning Rate Adjustment) optimization training is utilized to a test set, BPC is lower than 1.45, rapid convergence is carried out, the SGD(Stochastic Gradient Descent) optimization method with the learning rate of 0.0001 and a momentum of 0.9 is utilized for training, and the better test result is obtained.

The invention discloses a foreground extraction method for highway video spilled object detection, and the method comprises the following steps: S1, building a Gaussian mixture model for a pavement ina video sequence, and initializing model parameters; S2, obtaining an initial background by using an adaptive learning rate method, and dividing the initial background to obtain an initial backgroundmodel group B; S3, matching video frame pixel points with the background model, judging whether the type of each pixel point belongs to a foreground or a background, and outputting a binarized foreground image; and S4, updating background model parameters for foreground extraction at the next moment according to a matching result, and judging whether the background model needs to be re-divided through a model weight attenuation strategy or not. The foreground extraction method provided by the invention has the characteristics of high real-time performance and strong adaptive capacity to environmental noise, and is high in accuracy and low in false alarm rate in actual detection.

The invention relates to an equipment fault diagnosis method with an adaptive learning rate based on deep learning. According to the method, a trained fault diagnosis model based on deep learning is used for processing to-be-diagnosed data collected in real time, wherein the fault diagnosis model adopts an adaptive learning rate to carry out iterative computation. The adaptive learning rate is specifically as follows: on the basis of the last round of learning rate, the current gradient value is used for adaptively adjusting the current round of learning rate. Compared with the prior art, themethod has the advantages of short model training time, high classification accuracy and the like.