40results about How to "Increase training rate" patented technology

The invention discloses a strip steel surface area type defect identification and classification method which comprises the following steps: extracting strip steel surface pictures in a training sample database, removing useless backgrounds and keeping the category of the pictures to a corresponding label matrix; carrying out bilinear interpolation algorithm zooming on the pictures; carrying out color space normalization on images of the zoomed pictures by adopting a Gamma correction method; carrying out direction gradient histogram feature extraction on the corrected pictures; carrying out textural feature extraction on the corrected pictures by using a gray-level co-occurrence matrix; combining direction gradient histogram features and textural features to form a feature set, which comprises two main kinds of features, as a training database; training the feature data with an improved random forest classification algorithm; carrying out bilinear interpolation algorithm zooming, Gamma correction, direction gradient histogram feature extraction and textural feature extraction on the strip steel defect pictures to be identified in sequence; and then, inputting the feature data into an improved random forest classifier to finish identification.

The invention provides a word vector training method and device. The method comprises the following steps: an internet webpage is captured, and training language materials are acquired and stored in a corpus; each training language material in the corpus is subjected to word segmentation, and an orderly word set corresponding to each training language material is obtained; a word list is established according to pre-collected user query logs; the training language materials stored in the corpus are distributed to nodes of a distributed word vector learning model; the distributed word vector learning model is configured to perform periodic word vector training on each word in the word list, and the word vector corresponding to the word in the word list is obtained. According to the word vector training method and device, the word vectors obtained through training can be well applied to search business, and fast iterative high-quality word vector training can be realized.

The invention discloses an intelligent rotary machine fault depth network feature identification method. A vibration sensor is arranged at a to-be-detected rotating mechanical part of a train rollingbearing; collecting an original vibration sequence when the rolling bearing works; decomposing and reconstructing the original vibration sequence through a singular spectrum analysis method; extracting a root-mean-square value of the reconstructed vibration sequence; standard deviation, skewness index and peak value; a fault position is judged by using a rotary machine fault position diagnosis model obtained by training of a support vector machine; and then, carrying out ensemble empirical mode decomposition on the reconstructed vibration sequence, calculating the permutation entropy values ofa group of decomposed intrinsic mode components, taking the permutation combination of the permutation entropy values as a detection characteristic, and judging the fault type by using a rotary machine fault type diagnosis model obtained by training of a support vector machine. The fault position and the fault type of the rotary machine can be detected more timely, and the fault diagnosis accuracy and reliability are improved.

The invention provides a neural network training method and device, an image classification method and device and a medium. The neural network training method comprises the steps that a source domainand a target domain are acquired, the source domain comprises a plurality of first samples, and the target domain comprises a plurality of second samples; respectively determining category weights ofvarious categories corresponding to the source domain and the target domain, and calculating the maximum mean value difference of soft weights of the source domain and the target domain based on the category weights; calculating a redundancy value for network compression based on the batch normalization layer coefficient of the neural network; and training the neural network based on the soft weight maximum mean difference and the redundant value.

The embodiment of the invention discloses a neural network training method, an image recognition method, a device, equipment and a medium. The neural network comprises a plurality of network layers and at least one enhancement layer, the at least one enhancement layer is arranged between the network layers, and the method comprises the following steps: acquiring an original image sample; inputtingthe original image sample into the neural network to obtain a prediction result; acquiring a real result corresponding to the original image sample; determining a loss value according to a loss function, the prediction result and the real result, and training parameters of the neural network based on the loss value; and when a training condition is satisfied, removing the enhancement layer in theneural network to obtain a trained neural network. Not only can the occupancy rate of the samples on the storage space be reduced, but also the generalization performance of the neural network can beimproved.

The invention relates to a semi-supervised extreme learning machine classification method based on graph balance regularization. According to the method, an adjacency graph based on label consistencyand an adjacency graph based on information structure consistency are balanced through a non-negative weight value, so that the graph balance is achieved, a Laplace regular term of an optimal graph can be obtained to constrain a model, it is considered that the weight of the label consistency graph is increased when the information consistency graph cannot well describe the structure information of the sample set, and otherwise, the corresponding proportion needs to be reduced. The method comprises the following steps of firstly, constructing an adjacent supervised graph between the training samples through the label consistency of the samples; and combining with a semi-supervised graph based on the sample information consistency to constrain the output of the model, changing the capability of describing data distribution by reasonably adjusting the proportion of the graph, and obtaining an optimal output weight vector after obtaining an optimal adjacent graph. The method has a wide application prospect in the electroencephalogram signal processing systems and the brain-computer interface systems.

The invention belongs to the technical field of mechanical vibration signal processing. In order to solve the problem of long dictionary training time in a K-SVD algorithm, the invention discloses a rapid dictionary learning algorithm for sparse representation of mechanical vibration signals, and the algorithm specifically comprises the following steps: S1, selecting a training sample to determinean initial dictionary and determine the number ml of atoms of multiple columns of samples adjacent to the optimal time sequence; S2, performing synchronous sparse coding on multiple columns of sampleatoms adjacent to a training sample time sequence by adopting a synchronous orthogonal matching pursuit (SOMP) method to obtain a sparse coefficient matrix A; s3, fixing the sparse coefficient matrixafter synchronous sparse coding, and updating the dictionary by adopting a least square method (SGK); and S4, repeating the step S2 and the step S3 until an iteration stop condition is met, and completing dictionary training to obtain a learning dictionary. By adopting the dictionary learning algorithm provided by the invention, the dictionary training rate can be greatly and effectively improvedunder the condition of ensuring the compression and reconstruction performance of the vibration signal.

The invention discloses a short-time traffic flow prediction method based on a gradient boosting decision tree, and the method for building a model comprises the steps: 1) carrying out the preprocessing of the traffic flow data of an original highway toll station; 2) performing data analysis, data aggregation (data slicing) and data feature extraction operation on the data; 3) constructing a gradient decision tree model, inputting data into the model and carrying out training work; 4) transplanting the model to a distributed platform, and performing segmentation point sampling statistical optimization and layer-by-layer training optimization adjustment; and 5) constructing a three-layer Stacking model, carrying out multi-model fusion, and then carrying out further training on the data. The related method designed by the invention has the function of rapidly extracting the vehicle flow characteristics, and can predict the vehicle flow in a short time.

The embodiment of the invention discloses a dialogue robot reply method combining RPA and AI, a retrieval model training method and device used for a dialogue robot combining RPA and AI, equipment anda medium. The dialogue robot reply method comprises the steps of receiving a questioning statement of a user; based on a preset retrieval model, determining a target similar problem meeting a similarity requirement with the questioning statement; and returning the answers of the knowledge points to which the target similar questions belong to the user. The method is advantaged in that the presetretrieval model is obtained by training similar questions in a preset knowledge base so as to establish an association relationship between the similar questions and the similarity. By adopting the method, question answering accuracy of the dialogue robot can be improved.

The invention discloses a traffic signal control model training method and a traffic signal control method, and the training method comprises the steps: obtaining the current traffic state data of a target intersection from a preset simulation environment, inputting the current traffic state data into an initial traffic signal control model, and obtaining a phase motion value space; selecting a pre-execution phase action by using the preset agent, and generating an experience sample corresponding to the pre-execution phase action based on the reward value; for each preset agent, determining a corresponding target experience sample from the shared experience pool by using the preset agent according to the corresponding exploration rate; and training the initial traffic signal control model by using the plurality of determined target experience samples to obtain a trained target traffic signal control model. According to the method, the traffic signal control model is trained by using the multiple agents in the training process, so that the training time is shortened, and the algorithm execution efficiency is improved.

The invention discloses a fractional order KVFD multi-parameter machine learning optimization method for viscoelasticity mechanical characterization of soft substances. The method comprises the steps:building corresponding K-tree dictionaries according to the conditions of solutions of three types of KVFD loading modes; judging the specific type of a to-be-tested curve, carrying out global search, obtaining the parameter [E0, alpha, tau] of the curve as a vector, zooming to a parameter interval, generating a preset number of curves according to the KVFD model corresponding to the to-be-testedcurve, adding random Gaussian noise, dividing into a training set and a test set, and transmitting the training set and the test set into a machine learning model for training; selecting the model with the minimum RMSE as a final model for training; performing parameter estimation on a to-be-measured curve by using the final model obtained in the step 3; and further learning a result obtained byparameter estimation through a Q-learning algorithm to obtain an optimization result. According to the method, the characteristics of parameter learning and a heuristic algorithm are combined, and theaccuracy and efficiency of parameter optimization can be greatly improved.

The embodiment of the invention discloses a distributed model training method and device, a storage medium and computer equipment. The invention is related to the field of machine learning of artificial intelligence; the method comprises the steps of obtaining a model structure and sample data of a target model; sending a model message to each graphics processor, and triggering each graphics processor to train a target model; based on the preset sample number of the sample labels and the number of the graphics processors, performing multiple times of data sampling on the sample data to obtain multiple groups of training sample data; sending the training sample data to a corresponding graphics processor, triggering the graphics processor to train a target model based on the training sample data, and updating training parameters of the target model among the graphics processors to obtain updated training parameters of the trained target model; and when the updated training parameters meet a preset training termination condition, outputting the trained target model. The speed of model training can be effectively improved.

The invention relates to a phase attention mechanism network model-based circuit breaker residual life prediction method, which comprises the following steps of: firstly, acquiring a vibration signal in an opening process, then optimizing a VMD algorithm, decomposing the vibration signal by using the optimized VMD algorithm, and selecting a modal component with relatively high kurtosis for reconstruction; then, according to the energy-entropy ratio, a contact breaking vibration segment is extracted from the reconstructed vibration signal; and finally, a prediction model fusing a stage attention mechanism is established, the prediction model takes a one-dimensional convolutional neural network and a GRU network as a trunk network, the stage attention mechanism is divided into two stages, the first stage is a distributed attention mechanism applied to the one-dimensional convolutional neural network, weighting is performed on an input sample in time and feature dimensions, and the second stage is a distributed attention mechanism applied to the GRU network. And in the second stage, weighting is carried out on the time dimension again by applying a time step attention mechanism of the GRU network. According to the method, the contribution degree of important information on the time dimension and the feature dimension to the prediction result is enhanced, and the prediction precision is improved.

The invention relates to a Faster R-CNN-based railway bullet train hood front opening and closing damage fault identification method. The invention relates to the technical field of headstock cover image recognition. The method comprises the steps of collecting an original image of a headstock cover; preprocessing the collected original image to reduce image noise; performing deep learning according to the preprocessed image to obtain a training sample; performing deep learning model training according to the obtained training sample to obtain a trained fault identification model; and according to the trained fault identification model, carrying out damage fault identification on the headstock cover. According to the method, the train hood front opening and closing damage fault is identificated and detected through a deep learning method, and identification errors caused by fatigue and personal judgment differences during manual detection are effectively avoided. Compared with manual detection, the deep learning method can improve the detection efficiency.

The embodiment of the invention discloses a federal model cooperative training method and device, computer equipment and a storage medium. The method comprises the following steps: constructing a globally shared federal model; issuing the federal model to a plurality of data servers agreeing to participate in model training, so that the data servers train the federal model according to local data to obtain training parameters; judging whether the precision of the federated model meets the requirement or not; if not, acquiring training parameters uploaded by a plurality of data servers; performing aggregation operation on the plurality of training parameters to obtain an operation result; and updating the federal model according to the operation result, and issuing the federal model to a plurality of data servers agreeing to participate in model training, so that the data servers perform training of the federal model according to local data to obtain training parameters. By implementing the method provided by the embodiment of the invention, high stability in the training process can be realized, and the training speed is high.

The invention discloses a handwritten form recognition method based on self-adaptive band differential gradient optimization. In the BP neural network parameter optimization algorithm for handwrittenform recognition, a common gradient descent algorithm is reintegrated and deformed by combining a traditional control theory thought; then, a differential link is added into a conventional gradient descent algorithm for advanced correction, and the future change trend of an error signal is forecasted through the change rate of errors, so that the precision is improved; and finally, adaptively adjusting the learning rate by using the average value of the stored exponentially attenuated past square gradients so as to accelerate the training rate. According to the method provided by the invention, a differential link is introduced, so that the training rate can be effectively improved, and the future change trend of error signals is forecasted through the change rate of errors. The learning rate can be adaptively adjusted, i.e., when the training is close to the optimal value, the learning rate is reduced due to the increase of the accumulated past square gradient, and the optimal point is prevented from being skipped due to the overlarge learning rate.

The invention discloses a CNN-ET model-based electrocardiogram monitoring method. The electrocardiogram monitoring method comprises the following steps: S1, collecting ECG signals of a subject to construct an original data set, and acquiring training data and test data; S2, performing normalization processing on the training data and the test data to obtain a training sample and a test sample; S3, constructing a CNN-ET hybrid model; S4, inputting the training sample into the CNN-ET hybrid model for training to obtain a trained CNN-ET hybrid model; and S5, inputting the test sample into the trained CNN-ET hybrid model to carry out classification and identification of the ECG signals. The CNN-ET model organically combines a convolutional neural network and extremely randomized trees, and integrates feature extraction of the ECG signals and a classification task of the ECG signals, so that the relevance and classification precision of the whole classification model are improved, and the classification and identification of the ECG signals can be effectively and accurately implemented.

The invention discloses a method and device for adversarial training by using Internet pictures based on active learning. The method comprises the steps of: selecting a part of picture samples for training through employing an active learning strategy, and employing a trained model as a discriminator to judge and select a to-be-selected picture sample; and selecting picture sample data from an original picture data set, generating adversarial samples by utilizing the selected picture sample data, and then performing adversarial training. The adversarial sample selection strategy problem is improved, the uncertainty of samples is measured based on the probability, representative samples are efficiently selected through the strategy, and then adversarial samples are generated for learning; and a classification maximum probability value is selected as an index for measuring the uncertainty of the samples. The device comprises a memory and a processor, and the processor implements the steps of the method when executing a program.