70results about How to "Increase the learning rate" patented technology

The invention discloses a gesture recognition method based on 3D-CNN and convolution LSTM. The method comprises the steps that the length of a video input into 3D-CNN is normalized through a time jitter policy; the normalized video is used as input to be fed to 3D-CNN to study the short-term temporal-spatial features of a gesture; based on the short-term temporal-spatial features extracted by 3D-CNN, the long-term temporal-spatial features of the gesture are studied through a two-layer convolutional LSTM network to eliminate the influence of complex backgrounds on gesture recognition; the dimension of the extracted long-term temporal-spatial features are reduced through a spatial pyramid pooling layer (SPP layer), and at the same time the extracted multi-scale features are fed into the full-connection layer of the network; and finally, after a latter multi-modal fusion method, forecast results without the network are averaged and fused to acquire a final forecast score. According to the invention, by learning the temporal-spatial features of the gesture simultaneously, the short-term temporal-spatial features and the long-term temporal-spatial features are combined through different networks; the network is trained through a batch normalization method; and the efficiency and accuracy of gesture recognition are improved.

The invention relates to a learning process that is personalized to the learner. Student performance and learning behavior is tracked and analyzed, feedback and reinforcement is given, and educational or training stakeholders are engaged. Continuous improvement of the student, instructor, curriculum, and learning process is envisioned to accelerate learning.

The invention discloses a dialog strategy online realization method based on multi-task learning. According to the method, corpus information of a man-machine dialog is acquired in real time, current user state features and user action features are extracted, and construction is performed to obtain training input; then a single accumulated reward value in a dialog strategy learning process is split into a dialog round number reward value and a dialog success reward value to serve as training annotations, and two different value models are optimized at the same time through the multi-task learning technology in an online training process; and finally the two reward values are merged, and a dialog strategy is updated. Through the method, a learning reinforcement framework is adopted, dialog strategy optimization is performed through online learning, it is not needed to manually design rules and strategies according to domains, and the method can adapt to domain information structures with different degrees of complexity and data of different scales; and an original optimal single accumulated reward value task is split, simultaneous optimization is performed by use of multi-task learning, therefore, a better network structure is learned, and the variance in the training process is lowered.

The invention discloses a fault location method based on a residual and a double-stage Elman neural network for a hydraulic servo system, comprising the following steps of: obtaining the input / output signals of the hydraulic servo system in a normal working state, an electronic amplifier fault state and a leakage fault state, training a fault observer by virtue of the input / output signal in the normal state, and obtaining a real-time residual signal by the fault observer at first, and then training a state follower in real time and on line to obtain a network connection weight corresponding to the real-time signal, and training an RBF (radial basis function) fault locator by using the time-domain characteristic value of the residual signal and the network connection weight as the training input samples of the RBF fault locator. Both of the fault observer and the state follower are realized by the improved Elman network. Whether the system has a fault or not at present can be judged by comparing the time-domain characteristic value with a fault threshold, and the type of the fault can be obtained by the fault locator. The fault location method disclosed by the invention realizes fault location for the hydraulic servo system, and has high location accuracy and engineering applicability.

The invention belongs to the image processing technology field and particularly relates to a breast ultrasound image tumor segmentation method based on the full convolution network. The method comprises steps that the full convolutional neural network based on cavity convolution is constructed and is for rough segmentation of an ultrasound image to obtain a breast tumor; in the constructed DFCN network, cavity convolution is utilized, so the network is made to maintain the relatively deep-level feature map resolution to ensure that the tumor is well segmented in the presence of a large numberof shaded areas; in addition, the batch normalization technology is utilized in the DFCN network, the network is made to have the higher learning rate, and the training process is accelerated; a dynamic contour PBAC model based on the phase information is utilized to optimize the rough segmentation result to obtain the final fine segmentation result. The experimental result shows that the tumor can be precisely segmented, and the good segmentation result is achieved especially for ultrasound images with blurred boundaries and many shadows.

The invention discloses a neural-network self-correcting control method of a permanent magnet synchronous motor speed loop. The method is characterized by: taking a current loop and a motor as generalized objects; firstly, collecting information, such as a rotating speed, a current and the like; using an adaptive linear time-delay neural network to carry out off-line parameter identification to the motor; then, taking a weight obtained through off-line learning as an initial value of on-line learning; finally, carrying out on-line parameter identification to the system, calculating a load torque of the motor according to the identified parameter; designing a neural-network self-correcting control law according to the obtained parameter value and a load disturbance value, adjusting the network weight on line according to an error between a controlled object and an identification model, and then setting the parameter of the neural-network self-correcting controller on line so as to realize online adjustment of the controller parameter. Uncertainty of the system and influence brought by the external disturbance can be eliminated. Dynamic performance and an anti-disturbance ability of a servo system can be improved.

The invention discloses a media access control (MAC) address hardware learning method and system based on a hash table and a ternary content addressable memory (TCAM) table, and relates to the MAC address learning field. According to the MAC address hardware learning method and system, when no MAC address conflict happens, the hash table is used for storing learned MAC addresses, and a static random access memory (SRAM) or a dynamic random access memory (DRAM) is applied to the hardware of the hash table; and when MAC address conflict happens, the TCAM table is used for caching conflicted MAC addresses, a TCAM storage is applied to the hardware of the TCAM, idle table items can be positioned through one-time searching due to the fact that parallel seeking is performed on the hardware of the TCAM, and the number of the table items of the TCAM is the number of the conflicted MAC addressed which can be practically cached. The MAC address hardware learning method is achieved on a general programmable exchange chip, does not need support of hardware circuits, is high in learning efficiency and small in occupied internal memory resources, is flexible in application due to the fact that the general algorithm is adopted, and can achieve complete control on conflict probability.

The invention discloses a WSN (Wireless Sensor Network) anomaly detection method based on an MEA-BP neural network. The method comprises the following steps: initializing various distributed sensor nodes, and starting to acquire data by various sensor nodes; using a K-means algorithm to perform space clustering on the various sensor nodes to obtain a plurality of cluster structures; using a mind evolutionary algorithm to perform parameter optimization on a BP neural network, optimizing the weight and threshold of the BP neural network through a convergence and dissimilation operation, obtaining optimal weight and threshold, inputting the optimal weight and threshold, and establishing an MEA-BP neural network model; and adopting a distributed algorithm to execute anomaly detection on the sensor nodes in each group of clusters independently, after anomaly detection is finished, transferring a detection result to cluster head nodes of the group of clusters for further verification by the sensor nodes. The WSN anomaly detection method based on the MEA-BP neural network provided by the invention improves the algorithm performance of the BP neural network, accelerates the learning rate of the BP neural network, effectively improves the accuracy of the abnormal data detection and reduces the false positive rate.

The invention discloses a method for removing station captions and subtitles in an image based on a deep neural network, and relates to the technical field of image restoration, and the method comprises the following steps: S1, building an image restoration model; S2, preprocessing images of the training set; S3, processing training data: taking the training image as a real image Pt; Setting a pixel point RGB value in a Mask1 region in the training image as 0, and taking the pixel point RGB value as a training image P1; Setting a pixel point RGB value in a Mask2 region in the training image as0, and taking the pixel point RGB value as a training image P2; S4, training the image restoration model to obtain a trained image restoration model; S5, image restoration; The method comprises the following steps of: preprocessing an image or a video needing to remove station captions and subtitles; According to the image restoration method, based on the deep learning idea, station captions andsubtitles in the image are automatically and rapidly removed, the processing process is clear and clear, restoration real-time performance is high, and the application range is wide.

Artificial neural networks (ANNs) are a distributed computing model in which computation is accomplished using many simple processing units (called neurons) and the data embodied by the connections between neurons (called synapses) and the strength of these connections (called synaptic weights). An attractive implementation of ANNs uses the conductance of non-volatile memory (NVM) elements to code the synaptic weight. In this application, the non-idealities in the response of the NVM (such as nonlinearity, saturation, stochasticity and asymmetry in response to programming pulses) lead to reduced network performance compared to an ideal network implementation. Disclosed is a method that improves performance by implementing a learning rate parameter that is local to each synaptic connection, a method for tuning this local learning rate, and an implementation that does not compromise the ability to train many synaptic weights in parallel during learning.

The invention discloses a vehicle type recognition method based on a convolutional neural network under the vehicle-mounted environment. A semantic compact bilinear pooling method for vehicle type recognition is put forward, which combines the hierarchical tag tree and compact bilinear pooling and demonstrates superior performance on the CompCars data set and the Stanford auto data set. With application of the method, the compact bilinear pooling method uses semantic connection between different levels of semantics of the vehicle and makes them mutually enhanced during training. The softmax loss function is promoted to the loss avoidance function aiming at making full use of prior knowledge. Experiments show that the invention improves the accuracy of the vehicle type recognition task on the CompCars data set and the Stanford auto data set.

The invention discloses an infrared image heterogeneity correction method based on trilateral filtering and a neural network. The original infrared image of the nth frame in the original infrared image sequence is sequentially loaded as the current frame image, and the corrected gray value of the i-th row and the j-th column of the current frame image is determined, and the n-th frame of infraredimage processingis original through the fast three-edge filtering algorithm; the expected value qn(x) of the pixel x is obtained according to the deviation between the expected value of the i-th row j-th column pixel x of the current frame image and the corrected gray value of the i-th row j-th column pixel, the adaptive iterative step size is updated to obtain the n+1th frame of the original infrared image, the pixel gain parameter and the pixel offset parameter of the corresponding position of the pixel in the i-th row and the j-th column, the original infrared image of the n+1th frame is corrected by the pixel gain parameter and the pixel offset parameter. According to the method, the learning rate of parameters can be improved, and the non-uniformity correction effect of the image canbe improved.

The invention relates to a small sample target detection method based on attention and contrast learning, and belongs to the field of artificial intelligence and image processing. The invention relates to the small sample target detection method which combines data enhancement, an attention region suggestion network (Attention RPN) and comparative learning. The method is based on a Faster R-CNN (Convolutional Neural Network) network, and comprises the following steps of: adopting a Few-shot Mosaic data enhancement module for enriching the comparison between a small sample background instance and a Novelclass instance and a Base class instance, enhancing the attention of a model on a foreground by an Attention RPN (Regression Coordinate Compensation) module based on regression coordinate compensation, and improving the expression of instance-level features by a contrast learning module. According to the method, the new class detection precision of the Faster R-CNN on a small sample is improved, and meanwhile, relatively high base class detection precision is kept; the dependency of Faster R-CNN on the new class training sample size is reduced, the new class migration ability is improved, and the effectiveness of the method is verified on COCO and VOC data sets.

The invention discloses a unit dynamic fire risk assessment method based on a neural network, and the method comprises the following steps: S1, recognizing basic elements of building fire risk assessment according to the type of a building. According to the invention, the analytic hierarchy process AHP and the BP neural network are combined, and the invention can be applied to real-time evaluationof fire risks of various types of social units at the same time; the method not only keeps experts to quantify quantitative and non-quantitative evaluation factors in subjective understanding and judgment, but also keeps the nonlinear mapping capability of the BPNN neural network. By automatically determining the input dimension of the complex fire risk evaluation system, the learning rate of theneural network is improved, the convergence rate of the neural network is increased, the evaluation result is more accurate and credible, automatic fire risk prediction becomes possible, then a scientific basis is provided for preventing and controlling fire, and subjectivity and blindness during fire prevention are reduced.

The invention relates to a method for improving the pattern recognition precision trough combining with data representation and a pseudo-inverse learning auto-encoder. Based on the pattern recognitiontheory where a sample is linearly inseparable in a low-dimensional space and may be separable in a high-dimensional space, the method employs the advantages of quick learning based on the pseudo-inverse learning auto-encoder, and can achieve the quick and accurate training of a stacked auto-encoder deep neural network. The method comprises the steps: increasing the dimensions of data through receptive fields: employing four specific receptive field functions for increasing the dimensions of original data, wherein the four receptive fields are a receptive field based on a kernel function, a receptive field based on function connection, a receptive field based on nonlinear transformation, and a receptive field based on random mapping; employing the data transformed through the receptive fields as the input of the auto-encoder, and employing a pseudo-inverse learning method for quickly obtaining a weight matrix of a neural network. The method has remarkable advantages in improving the precision of pattern recognition, is suitable for most of regression and classification problems, does not need complex counterpropagation calculation and time-consuming super-parameter optimization, and facilitates the hardware implementation at a mobile terminal.

The invention relates to a wireless communication technology, in particular to an anti-interference wireless communication method based on deep reinforcement learning. The method comprises the following steps: using two convolutional neural networks: one convolutional neural network calculates a value function, and the other convolutional neural network performs action selection based on a calculation result of the value function; adopting priority experience sampling in an experience playback stage, so that experience samples with higher priorities are sampled preferentially, updating parameters of the convolutional neural network based on the experience samples, and updating the priorities of all the experience samples through calculation of the updated convolutional neural network; adopting a forward action reservation strategy, designing a Gaussian-like function to judge the value of the current action, and dynamically adjusting and controlling the probability that the current action is continuously executed. According to the method, the optimal sending power and the optimal communication frequency band can be intelligently selected, the learning speed of the whole system is improved, and the optimal sending mode can be learned under the condition that a third-party attacker model is unknown.

Aiming at the defects in the existing sewage treatment control technology, the invention discloses a prediction control method based on an extreme learning machine (ELM). The method provided by the invention comprises the following stepsthatsewage process data are collected, an extreme learning machine is used for establishing a system model containing dissolved oxygen and nitrate nitrogen in thesewage process, the real-time state of the system is accurately described, a predictive control algorithm is adopted for rolling optimization, a control target and various constraints are embodied inan optimization performance index, and the model is updated on line according to real-time data. The flow optimization control of the sewage treatment process is realized, the control quantity can beadjusted in time according to the control condition, the stability of the control process is ensured, and the self-adaptive optimization control can be carried out according to the change condition ofthe process, so that the energy consumption of the sewage treatment process is reduced. The extreme learning machine is used as a prediction model of prediction control, so that the generalization ofthe system is improved, a local optimal solution is avoided, the model prediction speed is increased, and the calculation time is shorter when relatively high precision is obtained.

The invention relates to an efficient asynchronous federated learning method for reducing communication times. The method comprises the following steps: firstly, designing a hyper-parameter r which adaptively changes along with version obsolescence, reducing errors brought to asynchronous federal learning by the version obsolescence, and guiding model convergence; and to solve the large federated learning communication traffic, increasing the learning rate and the decreasing the local round number in the early stage, and then gradually reducing the learning rate to increase the local round number, so that the performance of the model can be basically unchanged under the condition of effectively reducing the total communication round number of model training, and the system can better carry out asynchronous federated learning.

The invention discloses an SE-FPN-based target detection model training method and device and a target detection method and device, and the training method comprises the steps: zooming a plurality oftraining pictures according to different zooming coefficients, and splicing the training pictures into a new picture which comprises a plurality of targets of different sizes; distributing the plurality of targets with different sizes to different pyramid feature layers of an SE-FPN target detection model according to a predetermined distribution strategy; in each pyramid feature layer, finding mpositions closest to a central point according to true values of training samples of the pyramid feature layer, calculating DIoUDg of all anchors of the m positions and the true values, calculating aDg mean value mg and a standard deviation vg, obtaining a threshold tg, and selecting the central position which is larger than the tg and located in a target frame and anchor output; and calculatinga classification loss function and a position regression function, and training a model through a back propagation algorithm. The SEFPN-based target detection network model is constructed, an image preprocessing mode and a sample selection strategy are improved, the model is trained, the model is applied to target detection, and the target detection efficiency is improved.

A system and method of real-time optimization of drilling performance metrics during a well drilling operation, for oil and gas as well as geothermal wells, or wells drilled for any other purpose. In a preferred form, the system receives information about allowable drilling metrics and real-time information of performance indicators. The drilling performance metrics and performance indicators are used to build a model to predict drilling parameters likely to optimize one or more drilling performance metrics.

The invention discloses an industrial big data mining-based state predicting method. The method comprises the steps of step 1, data acquisition, i.e., taking a sample reflecting a system history operating state as a training set, wherein xi is a system state variate, i.e., the input of a model, and ti is a concerned predictive index, i.e., the output of the model; step 2, building OS-ELM (online sequential extreme learning machine) models, i.e., using a training sample of step 1 to build a plurality of OS-ELM models, and calculating to obtain a plurality of predictive values; step 3, building EOS-ELM (enhanced online sequential extreme learning machine) models, i.e., averaging predicting results of the OS-ELM models to obtain predicting results of the EOS-ELM models. The problem that the system state in the current industrial system is difficultly predicted is solved, and stability and reliability of prediction are improved.

The invention discloses a composite material defect detection method and system based on infrared and ultrasonic signal fusion, and the method comprises the steps: collecting a data set containing a composite material infrared signal and an ultrasonic signal, and dividing the data set into a training data set and a verification data set; constructing a signal feature learning and fusion classification model based on deep learning, and inputting the training data set into the signal feature learning and fusion classification model for training; and inputting the verification data set into the trained signal feature learning and fusion classification model to obtain a composite material defect detection result. The method and system effectively solves the problems of great influence of human factors on defect type judgment and difficulty in defect qualitative analysis in ultrasonic detection and the problems of low defect type classification accuracy and incapability of reflecting defect positions well in infrared thermal imaging detection, and realizes objective judgment on the defect types and positions of the composite material; and the accuracy of defect type classification is improved.

The invention discloses a system and a method for providing example sentences according to input types. By the technical means of judging the input type of a queried material, then, searching the example sentences with the input type corresponding to the queried material in an example sentence library, and displaying the example sentences and the sentence pattern of the example sentences, the system and the method can directly output the example sentences suitable for the current situation, and realize the technical effect of improving the learning rate of the users.

The invention discloses a bearing fault diagnosis method and system based on Transform and data enhancement. The method comprises the following steps: acquiring data in a bearing operation process; performing data enhancement processing on the acquired data, and dividing the enhanced data into a training data set X and a verification data set Y; constructing a feature extraction network based on a Transform; establishing a network training and verification framework, selecting data of the training data set X and the verification data set Y in batches, and sending the data into the established feature extraction network to obtain a feature extraction network output result; and analyzing an output result of the feature extraction network, obtaining a fault diagnosis type of the bearing, and completing bearing fault diagnosis. According to the method, the network performance is generalized by using a data resampling enhancement method, and meanwhile, the Transform is used as a feature extractor, so that the fault diagnosis verification accuracy of the method reaches 95%.

The invention discloses a multi-target rice mill unit scheduling optimization system based on ACO-BP. A 4*4 rice milling unit control system is designed, multi-target processing can be achieved among units through scheduling optimization, processing parameters of each rice milling are optimized, a rice breaking rate of a rice mill can be reduced, and processing efficiency of the units can be improved. The unit control system optimizes an internal utilization algorithm of the units, can regulate and control the processing parameters of each rice milling according to a single processing target, realizes an online whitening precision intelligent control of the units and reduces a operation cost of the rice mill. The rice milling unit control system is built in a manner of optimizing a BP neural network by using an ant colony optimization algorithm (ACO), the ant colony optimization algorithm can accelerate a learning rate of a neural network, convergence to optimal parameters is faster, and a neural network model for optimally regulating and controlling the rice milling units is built. A built rice milling unit database can perform iterative optimization on processing parameters and processing schemes through evaluation of products and learning ability of the neural network, and self-learning of the database is realized.

The invention discloses an irrigation time calculation method based on APSO-ELM and fuzzy logic. The method is specifically implemented according to the following steps that 1, an irrigation online monitoring system is constructed, crop growth environment parameters are monitored in real time, an APSO-ELM model is established, growth environment data obtained through monitoring serve as input through the APSO-ELM network prediction model, and the evaporation capacity of a prediction reference crop is output; a fuzzy control irrigation time forecasting model is established; 2, the crop evaporation transpiration and soil humidity drop rate data obtained in the step 1 are taken as input of fuzzy control, the irrigation time required by the crops is input into the fuzzy control irrigation timeforecasting model, and finally the irrigation time required by the crops is output, so that irrigation time calculation is completed. The problem that the network forecasting precision is unstable due to the randomly generated weight and threshold of the input layer and the hidden layer of the ELM algorithm in the prior art is solved.

The invention discloses a semi-supervised building change detection method and system based on CutMix-ResNet. A remote sensing data set is divided into a labeled training data set and an unlabeled test data set; performing data preprocessing on the training data set; training a region-level teacher model and a sample-level teacher model by using an enhanced data set obtained by performing CutMix on the training data set; and jointly training a student model by using the two teacher models and a small amount of labeled data sets. According to the method, the original label data set is subjected to data enhancement by adopting the CutMix technology, and the respective self-supervision loss and the mutual cross pseudo-supervision loss of the two teacher models are optimized, so that the co-trained student models are more robust and complete, the generalization performance and the accuracy of the change detection model are improved, and the method can be used for change detection of remote sensing data.

The invention discloses a neural-network self-correcting control method of a permanent magnet synchronous motor speed loop. The method is characterized by: taking a current loop and a motor as generalized objects; firstly, collecting information, such as a rotating speed, a current and the like; using an adaptive linear time-delay neural network to carry out off-line parameter identification to the motor; then, taking a weight obtained through off-line learning as an initial value of on-line learning; finally, carrying out on-line parameter identification to the system, calculating a load torque of the motor according to the identified parameter; designing a neural-network self-correcting control law according to the obtained parameter value and a load disturbance value, adjusting the network weight on line according to an error between a controlled object and an identification model, and then setting the parameter of the neural-network self-correcting controller on line so as to realize online adjustment of the controller parameter. Uncertainty of the system and influence brought by the external disturbance can be eliminated. Dynamic performance and an anti-disturbance ability of a servo system can be improved.

The invention discloses a hydroelectric generation prediction method based on an extreme learning machine. The method comprises the following steps: obtaining parameter data information from a hydroelectric generation system and preprocessing data; dividing the data into two mutually exclusive parts, performing data training on one part, and performing data testing on the other part; acquiring training data, and establishing a model by adopting the training data; performing module training by adopting methods of cross validation, grid search and model evaluation and obtaining an optimal model;using the trained optimal ELM model to predict test data, acquiring and outputting a prediction result, and ELM being an extreme learning machine model. Through the method, higher learning speed andbetter generalization ability are displayed; the hydroelectric generation is more accurately and effectively predicted, the cost is reduced, and the learning rate is improved.

The invention relates to the technical field of air conditioner energy consumption prediction, and particularly discloses a hybrid optimization algorithm-based energy consumption prediction method. The method includes comprises the following steps: acquiring air conditioner historical energy consumption preprocessing data of a data acquisition and processing device; according to an optimized extreme learning machine algorithm, processing and analyzing training set data in the air conditioner historical energy consumption preprocessing data, and then constructing an energy consumption prediction model; and inputting prediction set data in the historical energy consumption preprocessing data of the air conditioner into the energy consumption prediction model to obtain a prediction result. The invention further discloses a cloud computing platform and an air conditioner energy consumption prediction system based on the hybrid optimization algorithm. According to the air conditioner energy consumption prediction method based on the hybrid optimization algorithm, prediction of air conditioner energy consumption can be effectively and accurately achieved.