77 results about "Batch training" patented technology

The invention provides a multiple fault diagnosis method of complex equipment based on DNN, which comprises the following steps: 1, collecting multiple fault logs of the equipment, and counting and summarizing fault information; 2. relying on expert knowledge to locate the root cause of multiple faults; 3, integrating the log information according to the time characteristic to obtain a time sequence data set; 4, preprocessing the time series data set; 5, numerically quantizing the data set; using word embedding as the first layer network of DNN; adding Dropout after the network layer; 6, establishing an LSTM network layer; 7, establishing an MLP network layer; 8, setting the learning parameters of DNN; 9, dividing the data set; 10,learning and testing DNN by using partitioned datasets. Theinvention processes the multiple fault logs to obtain the sequential data set, and establishes a DNN model including a word embedded network layer, an LSTM network layer and an MLP network layer; After the data sets are partitioned, the DNN is learned by batch training data sets and verification data sets, and the accuracy of DNN in identifying root causes of failure is evaluated by test data sets.

The invention discloses a transformer fault diagnosis method based on Bi-LSTM and analysis of dissolved gas in oil. The method comprises: collecting fault DGA monitoring data of each substation, carrying out normalization, sequence expansion, noise superimposing and the like on the data, and extracting fault feature information based on a non-coding ratio method; carrying out length ranking on a DGA sequence, carrying out grouping and filling, and classifying groups into a training set and a verification set; constructing a deep learning frame based on Bi-LSTM, inputting data, and carrying outtraining; and then carrying out diagnosis and network updating by combining actual test data to obtain a fault diagnosis model with the high diagnosis accuracy and portability. According to the invention, the influence of the noise and error on the diagnosis during the DGA data monitoring process is reduced effectively; and the Bi-LSTM-based transformer fault diagnosis model is constructed by considering the complex correlation between different sequences. With introduction of links of sequence sorting, grouping, filling and the like, a problem of different sampling lengths of different transformers in the actual engineering is solved by using the batch training strategy.

The invention discloses an inter-carrier interference resistant OFDM detection method based on deep learning. The method can be applied to a high-speed mobile OFDM communication system and an OFDM system with relatively large millimetric wave band carrier phase noise and can effectively resist against the inter-carrier interference brought by Doppler frequency offset and the phase noise. Accordingto the inter-carrier interference resistant OFDM detection method disclosed by the invention, a deep network structure is designed for an approximation ML detector by using a deep expansion mode on the basis of a projection gradient descent method, the training algorithm is the Adam algorithm, a small batch training mode is adopted, each batch contains multiple input and output OFDM symbols andcorresponding channel matrixes H, that is, each batch reflects the changes of the channels within a period of time. Different types of channel information are retrieved at first during the training, and then deep learning is performed by using the channel information to converge a loss function to a small value. An OFDM signal is demodulated by using a trained deep detection network to effectivelyimprove the performance of the OFDM system that is affected by the inter-carrier interference generated by greater Doppler frequency offset or phase noise.

The invention discloses a BP neural-network classification method based on a Hadoop. The BP neural-network classification method based on the Hadoop comprises the following steps that data are preprocessed, Map tasks are started at Mapper ends of all nodes on a Hadoop platform, a training datum is obtained at each Mapper end and the training data serve as weight calculation modification values of the current network and the modification valves are transmitted to Reducer ends; Reduce tasks are started on the Reducer ends of all nodes on the Hadoop platform, all modification values of one weight are obtained by each Reducer end and the average value of the modification values is calculated out and serves as the output; the manner of batch training is adopted and the weight values of all layers are modified; the steps are repeated until the error reaches the preset precision or the frequency of study is larger than the preset maximum frequency and a BP neural-network model is obtained. Otherwise, iteration continues. Parallel computing can be achieved according to the BP neural-network classification method based on the Hadoop.

The invention relates to a low-rank constraint online self-supervised learning scene classification method. The method comprises the following steps: performing training and feature extraction on off-line image data; carrying out small-batch training to obtain an initial metric learner; inputting online data images sequentially and extracting image features; judging whether each image feature has a label; if the image feature has the label, updating the metric learner; if the image feature has no label, measuring the similarity between the image feature and each training sample, and utilizing a generated bidirectional linear graph to transmit the label; judging feature vector similarity scores of the sample; if the scores are high, updating the metric learner; and otherwise, inputting online data images. According to the scene classification method, self-updating can be realized gradually and useful information obtained from marked samples and unmarked samples can be combined; and the framework of a unified on-line self-updating model is utilized to process online scene classification, so that the on-line automatic scene classification can be achieved, the accuracy of classification is ensured, and work efficiency is improved.

An intrusion detection method and system for Internet of Vehicles based on Spark and combined deep learning are provided. The method includes the following steps: S1: setting up Spark distributed cluster; S2: initializing the Spark distributed cluster, constructing a convolutional neural network (CNN) and long short-term memory (LSTM) combined deep learning algorithm model, initializing parameters, and uploading collected data to a Hadoop distributed file system (HDFS); S3: reading the data from the HDFS for processing, and inputting the data to the CNN-LSTM combined deep learning algorithm model, for recognizing the data; and S4: dividing the data into multiple resilient distributed datasets (RDDs) for batch training with a preset number of iterations.

The invention provides a multi-level progressive classification method and system based on a neural network and a Bayesian model, and the method comprises the steps: carrying out the data preprocessing of the neural network, and preparing a feature vector and a prediction result of a neural network training model for a first model; training a neural network: training a neural network model on theprepared data, and constructing a large-class hierarchical classifier model; training a Bayesian model, and establishing a Bayesian network model of the category under each large category; and predicting the samples to be classified. According to the method, advantages and disadvantages of different models are fully utilized, batch training is carried out on mass data according to requirements oflevels, labels of samples to be classified are determined through a series of models, and corresponding solutions are provided for training and prediction of multiple models.

The invention relates to an image-text cross-modal hash retrieval method based on large-batch training, and belongs to the field of cross-modal retrieval, and is used for solving the problems that theexisting deep learning-based cross-modal hash retrieval method, particularly a triple-based deep cross-modal hash method, is long in small-batch training time, limited in obtained sample number and not good enough in gradient, so that the retrieval performance is influenced. The method comprises the following steps: preprocessing an image and text data; carrying out hash code mapping; establishing a target loss function L; inputting the triple data training model in a large-batch manner; and performing cross-modal hash retrieval by using the trained model. According to the scheme provided bythe invention, the triple data is input in a large-batch manner for training, so that the time of each round of training is shortened; because more training samples exist when the parameters are updated each time, a better gradient can be obtained, orthogonal regularization is used for the weight, the gradient can be kept during gradient transmission, model training is more stable, and the retrieval accuracy is improved.

The invention discloses a flower recognition method based on a convolutional neural network (CNN) with a ReLU activation function, which belongs to the technical field of image recognition. The method comprises steps: basic parameters of the CNN are set; the weight and the bias term are initialized, and convolutional down sampling layers are designed layer by layer; a random sequence is generated, 50 samples are selected at each time for batch training, a forward process, error conduction and a gradient calculation process are completed, and a gradient sum is updated to a weight model for next weight updating; and a well-set training function and an updating function are called for training, and the accuracy of a sample is tested. Flower recognition can be carries out effectively at a high speed under influences of lighting, rotation and occlusion conditions.

The invention discloses a multi-class energy consumption prediction method based on a circulating neural network. A circulating neural network model is formed by pre-training, and the method comprisesthe following steps: step S1, loading original energy consumption data on the basis of the circulating neural network model, judging missing value and abnormal value from the original energy consumption data, and detecting and processing the missing value and abnormal value; step S2, extracting the time series characteristic data from the original energy consumption data on the basi of the original energy consumption data, establishing a characteristic set of the circulating neural network model, and normalizing the characteristic set; step S3, batch training being carried out on the featureset after the normalization processing, a multi-class output neural network being established by combining the non-time series feature data, and the multi-class energy consumption data output by the multi-class output neural network being predicted. The utility model has the advantages that a multi-class output neural network is established by using the time series characteristic data of the original energy consumption data and the non-time series characteristic data to predict the energy consumption.

The invention discloses a distributed training method and device for a deep learning model and computing equipment. The method comprises the following steps: in each training step, acquiring a predetermined number of training data from a training data set as batch training data; calculating a gradient of a model parameter of the deep learning model on the batch training data, and taking the gradient as a local gradient; calculating an accumulated value of the local gradients of the preset number of training steps as an accumulated gradient; communicating with other computing nodes, and exchanging accumulated gradients of each other; and calculating a gradient average value of the accumulated gradients of all the computing nodes, and updating the model parameters based on the gradient average value.

The embodiment of the invention provides a deep learning model training method, a working node and a parameter server. The deep learning model training method applied to the working node comprises thesteps that a first statistical parameter sent by a parameter server is received, and the first statistical parameter is determined by the parameter server according to historical training data of a target layer of a target model; when the target layer is trained based on target batch training samples, target statistical parameters of the target layer are obtained, and the target statistical parameters are statistical parameters of the target batch training samples; and an actual statistical parameter of the target layer is determined based on the first statistical parameter and the target statistical parameter, batch standardization is performed on the target batch training samples based on the actual statistical parameter, and the target statistical parameter is sent to the parameter server. According to the embodiment of the invention, the training efficiency of the deep learning model can be improved.

The invention provides an implementation method of a gradient descent width learning system. The gradient descent width learning system is composed of three dense layers, namely a feature layer, an enhancement layer and an output layer, and a merging layer. The feature layer maps the input data into mapping features by using random mapping to form feature nodes; the enhancement layer enhances themapping features output by the feature layer to form enhancement nodes; the merging layer merges the outputs of the feature nodes and the enhancement nodes and inputs the merged outputs into the output layer as a whole; and the output layer maps the output of the merging layer into the final output of the network, a small batch of training samples are continuously input into the gradient descent width learning system during training, and the gradient descent width learning system adopts a gradient descent method to update the weight of the network, so that the loss of a mean square error MSE loss function is gradually reduced. When the gradient descent width learning system implemented by the method continuously performs small-batch training sample training, the regression performance during batch training can be obviously improved, and the method can be applied to regression tasks.

The invention discloses a noise elimination method based on VAD and ANN, and the method comprises the steps: firstly employing the VAD technology for dividing voice with noise into a voice frame and anoise frame; secondly carrying out the different noise reduction of the voice frame and the noise frame through ANN; thirdly employing a sequential training of the reference noise corresponding to the voice frame, and employing batch training for the reference noise corresponding to the noise frame. According to the invention, the voice with the noise is divided into the voice frame and the noiseframe through the VAD technology. Through the quick convergence and high convergence precision of the batch training of the neural networks, the method achieves the processing of the noise frame. Onethe one hand, the method enables adaptive filtering to be convergent in one piece of noise frame data; on the other hand, the method also improves the convergence precision.