37results about How to "Guaranteed training accuracy" patented technology

The invention provides a parameter synchronization optimization method and system suitable for distributed machine learning. A machine learning algorithm achieved in a parameter server distribution mode is used for overcoming bottlenecks, such as a large amount of parallel machine learning training time delay caused by insufficient network and parameter server resources, of an existing algorithm in the parameter synchronization process. The system comprises a resource monitoring and distributing module at the parameter server end, a parameter maintaining module at the parameter server end, server resource request modules of all working nodes, parameter synchronization time interval control modules of the working nodes, non-synchronization time accumulation modules of the working nodes, parameter calculation modules of the working nodes and parameter synchronization modules of the working node. According to the parameter synchronization optimization method and system, different synchronization time intervals are selected for the different working nodes to avoid request emergency situations by monitoring resource occupancy conditions of a parameter server; meanwhile, it is guaranteed that the selected time intervals can meet the requirements for communication frequency reducing and training accurate rate guaranteeing at the same time, and the bottlenecks of an existing distributed machine learning system in the parameter synchronization process are effectively avoided.

The invention discloses a federated learning method and system based on batch size and gradient compression ratio adjustment, which are used for improving model training performance. The method comprise the following steps: in a federated learning scene, enabling a plurality of terminals to share uplink wireless channel resources; completing the training of a neural network model together with anedge server based on training data of a local terminal; in the model training process, enabling the terminal to calculate the gradient by adopting a batch method in local calculation, and in the uplink transmission process, compressing the gradient before transmission; adjusting the batch size and the gradient compression rate according to the computing power of each terminal and the channel stateof each terminal, so as to improve the convergence rate of model training while ensuring the training time and not reducing the accuracy of the model.

The invention provides a small- and medium-sized reservoir dam safety evaluating method based on GRA-BP (grey relational analysis and back propagation) neural network in order to overcome the defects of existing small- and medium-sized reservoir dam safety evaluating technology and method. The method comprises the steps of screening main factors influencing small- and medium-sized reservoir dam safety based on grey relational analysis method, and constructing a small- and medium-sized reservoir dam safety evaluation index system; generating a network training inspection sample; determining topological structure of the BP neural network; setting main training parameters of the BP neural network, initializing network connection weights and thresholds, and setting network training end conditions, error precision and training frequency; correcting the network by using L-M algorithm accumulative network global errors; entering the training and inspection samples to train and inspect the safety evaluating BP neural network, and finally constructing an intelligent dam safety evaluating model based on BP neural network. The method provides the functions such as comprehensive safety evaluation and hidden peril analysis, and is highly scientific, effective and practical.

The invention discloses a method for screening the optimum indexes of a reservoir flood control dispatching scheme based on a BP neural network, comprising the following steps: generating network training samples; determining the topology of the BP neural network; training and inspecting the BP neural network; identifying the importance of each index in the BP neural network; calculating the degree to which the change of each index value affects the result, and analyzing the sensitivity of each index value; and determining the criterion and threshold of index screening. A lot of network training samples are generated through an index uniform discretization method, which effectively guarantees the training precision of the BP neural network. Based on an information storage and conversion mechanism of the BP neural network, the indexes and the influence of relative change of the indexes to the result of reservoir flood control dispatching scheme assessment are quantitatively evaluated according to the relative degree of importance and relative rate of contribution. By establishing a comprehensive judgment index, index screening is converted from a subjective process of analysis and judgment into a quantitative process of analysis and calculation.

The invention provides a method for constructing a risk evaluation model of green building material product authentication, comprising the steps of obtaining index parameters of green building material product authentication, wherein the index parameters are determined according to an authentication node and a risk point; acquiring sample values corresponding to the index parameters, and establishing a training sample and a test sample; inputting the index parameters into a neural network model and configuring the parameters of the neural network; training the neural network model by trainingsamples, and after the training, optimizing the neural network model by using test samples, wherein the optimized neural network model is a risk evaluation model of green building material product certification. The neural network model can extract the hidden knowledge and rules from the training data, and reduce the subjective factors and accidental factors in the evaluation process, which is more scientific, intelligent and expansible.

The embodiment of the invention discloses an image classification method and device, electronic equipment and a storage medium. The method comprises steps of acquiring a to-be-processed image, inputting the image to a pre-trained target image classification model, and the target image classification model is acquired through training in an unsupervised learning mode according to a first sample with a standard type label and a second sample without the standard type label; and determining an image type to which the to-be-processed image belongs according to an output result of the target imageclassification model. Through carrying out unsupervised iterative training on the image classification model through a large number of second samples without standard type labels and a small number offirst samples with standard type labels, manual classification and label setting do not need to be performed on each sample image so that the processing difficulty and the processing duration of thesample data can be reduced and the total training duration of the target image classification model can be further reduced.

The invention provides a method and a device for determining a learning rate. The method comprises the following steps: obtaining the number of intervals used for calculating the learning rate of a target layer in a deep learning model at intervals; and when the current iteration frequency is smaller than the interval frequency, in the process of carrying out learning training of current iterationon the target layer in the deep learning model, continuing to use the first learning rate in the previous iteration process to carry out learning training of the target layer. According to the technical scheme provided by the invention, the number of iterations for calculating the learning rate can be reduced, so that the long training iteration time caused by calculating the learning rate for the target layer in each iteration training in the prior art can be avoided.

The invention discloses a modulation mode recognition method based on spatial-temporal feature extraction deep learning. The method comprises the following steps: collecting a signal of a to-be-identified modulation mode; constructing an automatic modulation identification deep learning model comprising a parameter estimation module, a parameter change module and a spatial-temporal feature extraction module, and training the automatic modulation identification deep learning model; and performing modulation mode identification on the collected signals by adopting the trained automatic modulation identification deep learning model. In order to solve the problems that a modulation identification model in the prior art is relatively high in complexity and high identification accuracy is difficult to realize under the condition of low model complexity, the invention provides a spatial-temporal feature extraction automatic modulation mode identification deep learning model based on parameter estimation and transformation. The parameter quantity of modulation mode identification by using the model is less than that of an existing automatic modulation identification method based on deep learning, and the training overhead is lower than that of other methods with the same identification accuracy level.

The invention relates to a construction method and application of an industrial process fault diagnosis model, and the method comprises the steps: constructing a fault diagnosis framework which comprises a generator which carries out the coding-decoding-coding of each original sample generated by a feature extractor, and obtains a first hidden feature, a generated sample and a second hidden feature; training a generator by adopting the normal original sample set and a discriminator in the generative adversarial network and taking the discriminator to discriminate the generated sample as an original sample as a target; and using the fault score calculator is for generating a sample and a second hidden feature to perform fault diagnosis based on each original sample to be detected and the corresponding first hidden feature. The generator in the generative adversarial network is introduced into the fault diagnosis model. The generator has a coding-decoding-coding function, the discriminator in the generative adversarial network is adopted to train the generator only based on the normal original sample, and the problems that the fault diagnosis model is difficult to train, low in efficiency and poor in effect due to the fact that industrial fault samples are too few are solved.

The invention discloses a training method and device of a federated model based on a neural network, and a computer device. The method comprises the following steps: receiving a training request for carrying out iterative training on a preset federated model to obtain historical gradient information sent by each participant of the federated model to the federated model; screening the historical gradient information according to a preset LSTM model to obtain a plurality of participants which need to iteratively update the federal model at the current moment; and receiving gradient information generated after the local data of each participant in the plurality of participants train the corresponding local model so as to update the federal model. According to the invention, based on the neural network technology, the participants needing to participate in iteration training in each iteration process are obtained from all participants of the federated model by using the long-short-term memory artificial neural network, so that the calculation cost and the network communication transmission cost of the participants are reduced, the training speed of the federated model is accelerated, and meanwhile, the training precision of the federal model is improved.

The embodiment of the invention discloses an application program prediction model establishment method and device, an application program pre-loading method and device, a medium and a terminal. The application program prediction model establishing method comprises the following steps: obtaining a user behavior sample in a preset time period, wherein the user behavior sample comprises use time sequence association records of at least two application programs, grouping the use time sequence association records to obtain multiple groups of use time sequence association records, and training a preset SRU neural network model according to the multiple groups of use time sequence association records to generate an application program prediction model. According to the embodiment of the invention, the technical scheme is adopted; the application program use time sequence association record capable of truly reflecting the user behavior can be fully utilized; the application program pre-loadingmechanism is optimized, the accuracy of predicting the application program to be started is improved, on the premise that the training precision of the application program prediction model is guaranteed, the processing speed of training generation of the application program prediction model is greatly increased, and the time cost is saved.

The invention discloses a training data transmission method, which comprises the steps of: monitoring a current monitoring value used for reflecting a current training condition in a node, and judgingwhether the current monitoring value is greater than a preset value or not; if so, compressing obtained current training data into a preset type of compressed data; and establishing communication with the destination node, and sending the compressed data to the destination node. And after calculation is completed, decompression operation is carried out to continue the model training process. Therefore, by compressing the current training data into the compression data of the preset type, the data volume transmitted in the communication process is reduced while the model training precision isensured, the pressure of network resources is relieved, and the accuracy and high efficiency in the data transmission process are ensured. In addition, a training data transmission device, training data transmission equipment and a storage medium provided by the invention correspond to the method.

The invention provides a domain adaptive bearing voiceprint fault diagnosis method and system based on reinforcement learning, and the method comprises the steps: obtaining a fault diagnosis result according to an obtained rolling bearing voiceprint signal and a fault diagnosis model; wherein the loss function of the fault diagnosis model is the sum of the loss function of the domain adaptive network and the loss function of the classification network, the input of the classification network is the source domain label and the source domain output of the feature extraction network, and the input of the domain adaptive network is the source domain output and the target domain output of the feature extraction network; source domain output is obtained after the source domain data sequentially passes through the sparse auto-encoder and the feature extraction network, and target domain output is obtained after the target domain data passes through the feature extraction network; according to the method, the problem that a source domain data set and a target domain data set are inconsistent is considered, the sparse auto-encoder is combined, when data samples are insufficient, unknown fault diagnosis is subjected to high-precision classification through known feature signals, and safe operation of mechanical equipment is guaranteed.

The invention discloses a classification model training method and device, a classification method and device, a medium and equipment. The classification model training method comprises the steps of obtaining first sample data which is not provided with a classification label, inputting the first sample data into a pre-trained basic classification model, and determining the probability that the first sample data is provided with a preset classification label, wherein the basic classification model is obtained through training based on second sample data provided with a preset classification label; determining the weight of the first sample data based on the probability of setting a classification label by the first sample data; and training a to-be-trained classification model based on second sample data, the first sample data and the weight of the first sample data to obtain a target classification model. The label processing of the sample data without a label is realized, manual label setting of a sample is replaced, the time and labor cost of a sample data preprocessing process are reduced, and the weak supervision training of the classification model is further realized.

The invention discloses a welding seam identification method based on sub-regions and a BP neural network, and the method comprises the steps: S1, building a BP neural network model which uses each sub-region of a weld joint image as the input of a neural network, and determining the number of nodes of an input layer through the vector dimensions of the sub-regions; S2, training the BP neural network model; and S3, identifying a welding seam image by using the trained BP neural network model. The method can greatly reduce the training time of the BP neural network model on the premise of guaranteeing the training precision, can use the trained BP neural network model to recognize the sub-regions of the weld seam image when the weld seam image is recognized, and finally obtains a welding seam recognition detection effect picture. The invention has the advantages of high recognition efficiency, wide application range, high accuracy and the like.

The embodiment of the invention discloses a face recognition model training method, a face recognition model recognition method, a storage medium and related equipment. The training method comprises the steps of obtaining at least two groups of images under Nth iteration; according to the at least two groups of images, obtaining two groups of training data sets under the Nth iteration; processingthe first group of training data set by the auxiliary model under the Nth iteration to obtain a first feature group of the human face under the Nth iteration; processing the second group of training data set by the main model under the Nth iteration to obtain a second feature group of the human face under the Nth iteration; calculating a loss function value under the Nth iteration according to thefirst feature group and the second feature group; and at least determining whether to update the main model under the Nth iteration for the corresponding time according to the loss function value under the Nth iteration.

The invention discloses a method for predicting the performance of a centrifugal pump serving as a turbine based on an improved artificial neural network. The method comprises the following steps: firstly, calculating the lift HBEP, T and the flow QBEP, T of the optimal working condition point of the pump serving as the turbine at each specific speed in a segmented manner, and calculating the ratio a of the flow Qi to the flow QBEP, T at each working condition point in a turbine state and the square root b of the ratio of the Hi to the HBEP, T corresponding to the Qi; constructing a training set, wherein each training sample comprises geometric parameters, flow, a and b of the centrifugal pump and lift and efficiency corresponding to each flow in a turbine state; constructing an artificial neural network, and simultaneously performing L1 and L2 regularization on the artificial neural network; training the artificial neural network by adopting the training set; and finally, geometric parameters a and b of a to-be-predicted centrifugal pump in a turbine state are input into the trained artificial neural network, and lift and efficiency corresponding to each flow working condition are output. The method is wide in application range, high in prediction precision and short in calculation period.

The invention discloses a collaborative model training task configuration method for intelligent edge computing, the method is used for an edge computing node and comprises one or more training time slots, and each training time slot comprises the following steps: sending a model training request to one or more mobile devices; receiving an available state and a user data scale of the current timeslot reported by one or more mobile devices; based on the previously obtained task configuration result and the current available state of each mobile device, determining the number of small trainingwheels required by the mobile devices participating in training and the interactive model training; and performing interactive model training with the mobile devices participating in training until the determined number of small training rounds is reached, constructing an optimization problem with the purpose of minimizing the use of edge training resources according to the training effect and theuser data scale reported by each mobile device, and solving the optimization problem to obtain a new task configuration result. Compared with other methods, the training resource consumption is muchless, and the precision difference is small.