166results about How to "Solve vanishing gradient" patented technology

The invention discloses a human behavior recognition method based on an attention mechanism and a 3D convolutional neural network. According to the human behavior recognition method, a 3D convolutional neural network is constructed; and the input layer of the 3D convolutional neural network includes two channels: an original grayscale image and an attention matrix. A 3D CNN model for recognizing ahuman behavior in a video is constructed; an attention mechanism is introduced; a distance between two frames is calculated to form an attention matrix; the attention matrix and an original human behavior video sequence form double channels inputted into the constructed 3D CNN and convolution operation is carried out to carry out vital feature extraction on a visual focus area. Meanwhile, the 3DCNN structure is optimized; a Dropout layer is randomly added to the network to freeze some connection weights of the network; the ReLU activation function is employed, so that the network sparsity isimproved; problems that computing load leap and gradient disappearing due to the dimension increasing and the layer number increasing are solved; overfitting under a small data set is prevented; and the network recognition accuracy is improved and the time losses are reduced.

The invention provides a medical image segmentation method of a residual full convolutional neural network based on an attention mechanism. The medical image segmentation method comprises the steps: preprocessing a to-be-segmented medical image; constructing a residual full convolutional neural network based on the attention mechanism, wherein the residual full convolutional neural network comprises a feature map contraction network, an attention network and a feature map expansion network group; inputting the training set data into a residual error type full convolutional neural network for training to obtain a learned convolutional neural network model; and inputting the test set data into the learned convolutional neural network model, and performing image segmentation to obtain segmented images. According to the medical image segmentation method, an attention network is utilized to effectively transmit image features extracted from a feature map contraction network to a feature mapexpansion network; and the problem of lack of image spatial features in an image deconvolution process is solved while the attention network can also inhibit image regions irrelevant to a segmentation target in a low-layer feature image, so that the redundancy of the image is reduced, and meanwhile, the accuracy of image segmentation is also improved.

The invention discloses a P300 detection method based on a CNN-LSTM network which comprises the following steps: 1) designing a P300 character speller, collecting EEG data and determining a training set and a test set; 2) preprocess that training set and the t set; 3) Establish the algorithm model of combining CNN and LSTM; 4) train that model parameters by use a training set; 5) Testing model character recognition accuracy, P300 wave recognition accuracy, recall, accuracy, and F-measure. The invention combines CNN and LSTM neural networks to establish an algorithm model, which not only takesinto account the time and space characteristics, but also solves the problem of gradient disappearance or explosion in a simple RNN algorithm, and further improves the accuracy, and is a feasible method for detecting P300 signals.

The invention relates to a deep convolutional neural network based SAR image sea ice sorting method. The method includes steps of S01, segmenting a prior sea ice SAR image; S02, performing data pre-treatment; S03, performing model training and establishing a model; S04, performing treatment on a to-be-sorted sea ice SAR image; S05, combining sorting results. The method is advantaged in that the deep convolutional neural network model constructed in the method can realize characteristic automatic extraction based on images and excessive manual intervention is not needed; the method is an end-to-end SAR image sea ice sorting method and can reach sea ice monitoring service level and meet requirements on real time performance for offshore workers; image characteristics can be extracted automatically depending on a large amount of mark samples by the model and reliance on specialist knowledge is not needed; convergence is accelerated by utilizing a random gradient descending method and model training condition can be judged according to a loss function and accuracy; problems of gradient vanish or gradient diffusion in network parameter optimization counter-propagation are solved by utilizing an normative approach.

The invention discloses an underwater sound source positioning method based on deep learning, and the method comprises the steps: carrying out normalization operation of vector data simulated throughemploying a KRANEN program, carrying out the superposition of a 0-mean Gaussian random noise complex vector n, and obtaining the simulation sound field data p (f) at a frequency f; constructing a normalized covariance matrix H according to the analog sound field data p (f), performing Hermitian decomposition on the matrix H, and converting the complex matrix H into a real matrix capable of being processed by the convolutional neural network to obtain input data of the convolutional neural network; and training the convolutional neural network by using the input data to obtain an underwater sound source positioning prediction model, and predicting the distance and depth of the signal source according to the observed sound field data. LeNet-5 convolutional neural network and 56-layer deep residual network are used for underwater sound source positioning under single and multiple sound sources. An underwater sound source positioning algorithm with high precision and accuracy is acquired and the real-time performance of underwater sound source positioning is improved.

The invention provides a remote sensing image change detection method and system based on a U-net improved algorithm, and the method comprises: carrying out the preprocessing of a remote sensing image, and then integrating the preprocessed remote sensing images of different time phases to obtain an input image; performing down-sampling encoding and up-sampling decoding on the input image based onan improved U-net network; and finally, outputting a binary image whether the image is changed or not according to an analysis result. According to the invention, a network structure in a semantic segmentation direction is applied to the field of change detection, and a residual learning mechanism is introduced, so that the encoder can quickly converge and deepen the number of network layers; andmeanwhile, ASPP is used for enhancing the perception capability of the network to image features, so that the algorithm can be kept at a relatively high level in precision and efficiency, and has relatively high robustness. The method and the system are suitable for the field of change detection of remote sensing images, can be popularized to other fields, and have important significance.

The invention discloses a quick magnetic resonance imaging method based on a recursive residual U-type network. The method comprises the three steps of data processing, model training and image reconstruction. By adopting the method, a recursive residual module is added to the U-type network, the problems of gradient blast and gradient vanishing caused by back propagation of the network are solved, new network parameters are not added while the layer number of the network is increased, the quality of a reconstructed image is obviously improved, and more image details can be restored.

The invention discloses a method for predicting the residual service life of a rolling bearing based on EEMD-MCNN-GRU. The method comprises the following steps: preprocessing full-life-cycle vibrationdata of the rolling bearing; utilizing the ensemble empirical mode to construct degradation characteristics of the bearing vibration signals on different time scales so as to construct a degradationcharacteristic data set of each rolling bearing; integrating a multi-scale convolutional neural network feature extraction layer and a gate control cycle unit neural network time sequence prediction layer to build a residual life prediction model; normalizing the degradation characteristic data set input into the prediction model; carrying out segmentation processing on the normalized data; dividing the degradation feature data set into a training set and a test set, and taking the training data as the input of the whole prediction network model; and smoothing the prediction model output by using a moving average method, and outputting an optimal life prediction result. The prediction method which is more reliable and higher in robustness and generalization is provided for analysis of theresidual service life of the rolling bearing.

The invention discloses a construction method of a road blocking image semantic segmentation sample library for full convolutional neural network training. The construction method comprises the steps of performing vectorization, enhancement and standardization on samples; secondly, introducing a classical convolutional neural network type and a network structure improvement method, and explaining a network realization method and a training process; then, using the full convolutional neural network obtained through training for conducting remote sensing image road surface semantic segmentation, and on the basis that the road surface which is not damaged after disaster is extracted, judging the road integrity through the length proportion of the road which is not damaged before disaster and after disaster. The precision evaluation indexes of the improved full convolutional neural network model are superior to those of an original full convolutional neural network model. The improved model is more suitable for specific problems of post-disaster undamaged pavement detection and road integrity judgment, and the adverse effects of tree and shadow shielding on road blocking information extraction can be effectively overcome.

The invention discloses an image multi-subtitle automatic generation method based on a multiscale hierarchical residual network, and adopts an improved funnel network to capture multiscale target information. Firstly, when a funnel framework network is constructed, a densely connected polymerization residual block is put forward, and residual LSTM (Long Short Term Memory) is further put forward inorder to solve the problems of gradient vanishing and gradient explosion. By use of the method, high experiment performance is obtained, and the method has an obvious advantage on multi-subtitle taskacquisition.

The invention belongs to the technical field of CT imaging, and discloses a deep learning method for suppressing noise and artifacts in a low-dose CT image and facilitating later accurate medical diagnosis and analysis. The specific technical scheme is as follows: the double-attention generative adversarial network for noise reduction and artifact removal of the low-dose CT image is used for extracting non-uniform and irregular noise features and tissue texture features with complex distribution respectively; a Res2Net discriminator network for multi-scale feature extraction is designed, so that the discrimination capability of the discriminator is improved, and the stability and robustness of adversarial training are enhanced; a multi-description loss function combining artifact attentionloss, artifact consistency loss, structure constraint loss, adversarial loss and pixel-level L1 loss is designed to further improve the function of each sub-network. According to the method, the phenomenon of under-noise reduction or over-noise reduction caused by high similarity of noise artifacts and tissue structure distribution is avoided.

The invention discloses a gluing detection system and method based on deep learning, and the method comprises the following steps: constructing an initial deep residual error network model which is ofa five-layer structure and comprises an input layer, a convolution layer, a residual error layer, a full connection layer and a joint loss function layer; inputting the training sample into the deepresidual network model for training to obtain a gluing detection model; obtaining an original gluing image, and carrying out the preprocessing to obtain a to-be-detected gluing image; and inputting the to-be-detected gluing image into the gluing detection model, obtaining a score of gluing image detection, and judging whether the to-be-detected gluing image is qualified or not according to the score. According to the invention, the gluing of the vehicle window in the gluing workshop is automatically detected by constructing the deep residual network model, the gluing quality can be quickly andaccurately judged, the detection result is uploaded in real time, the timely processing is facilitated, and the automation degree and the production efficiency of the gluing workshop are improved.

The invention discloses a CNNLSTM fan fault prediction method based on an attention mechanism, and the method comprises the steps: firstly collecting the time sequence data of a fan through an SCADA system; secondly, a random forest algorithm is used for carrying out feature correlation degree analysis on the fan time series data, features with high fan fault correlation degree are extracted, a new feature data set is formed, the new feature data set is input into a CNNLSTM model to be trained, and an Attention mechanism is introduced into the model, so that the model pays more attention to the features related to fan faults; interference of non-related features is reduced, so that the accuracy and generalization ability of the fan fault prediction model are improved; and finally, inputting a to-be-predicted fan data set into the trained model to obtain a fault prediction result. The model provided by the invention is a fan fault prediction model which is suitable for processing time series data, high in accuracy, strong in generalization ability and outstanding in comprehensive performance.

The invention discloses a cement finished product specific surface area prediction method and system based on a long-term and short-term memory network. The method comprises the following steps: sorting training input data in a training input set according to a time sequence; inputting the sorted training input set into a pre-constructed long-term and short-term memory network model to obtain a cement finished product specific surface area prediction value at each moment; calculating a node error term of each neuron by adopting a time-based back propagation algorithm according to the trainingoutput set and the cement finished product specific surface area prediction value, wherein the node error term comprises a forgetting gate error term, an input gate error term and an output gate errorterm; training the to-be-trained parameters by adopting a random gradient descent method according to the node error term to obtain a trained long-term and short-term memory network model; and inputting the to-be-tested input set into the trained long-short-term memory network model to obtain a to-be-tested cement finished product specific surface area prediction value. The accuracy of cement finished product specific surface area prediction can be improved.

The invention discloses a deep learning-based method for predicting a lithologic sequence model through using seismic data. The method includes the following steps that: 1) the lithologic data of a target stratum section of a well in a work area and near-well trace seismic data are measured so as to be adopted as training data; 2) the near-well trace seismic data are normalized, so as to be converted to a range of -1 to 1; 3) a stacked recurrent neural network and a sequence-to-sequence recurrent neural network model are adopted to respectively train the processed near-well trace seismic dataand the well lithologic data, with the near-well trace seismic data adopted as observation data, and the well lithologic data adopted as target data, iterative computation is performed, so that a learning model reaches convergence; and 4) actual seismic data are inputted to the learning model which is obtained after calculation in the step 3), so that a predicted lithologic sequence is obtained. With the deep learning-based method for predicting the lithologic sequence model through using the seismic data of the invention adopted, a lithologic data body capable of effectively reflecting reservoir distribution can be generated under the control of a seismic data sequence, and an inter-well reservoir prediction problem can be solved, and a basis can be provided for exploration and development.

The invention discloses a crop disease identification method based on a deep fusion convolutional neural network. An IR _ CNN model provided by the method is formed by cascading effective modules in Inception v1 and ResNet50, and can be used for respectively extracting crop disease image diversity and deep features and fusing the crop disease image diversity and the deep features. The IR _ CNN model module is composed of neural networks with different branches, so that the width of the overall network is increased; full connection or even general convolution is converted into sparse connection, so that the calculated amount of the network is reduced. According to the method of the invention, the feature extraction capabilities of different network models are combined, so that x diversity features and deep features in crop disease images can be better obtained, the features are fused subsequently; various disease categories of different crops, especially complex crop diseases, can be better identified through training learning. The method has high identification precision.

The invention discloses a human face depth-prediction method based on convolution neural network, which comprises the following steps: generating a human face depth-prediction neural network in which a specially structured convolution neural network is created wherein the specially structured convolution neural network refers to the normalized operation and the excited operation in successive series-connection from the output end of the convolution layer, in series-connection with the aforementioned two operation convolution layers and in parallel to the KxK convolution; and initializing the convolution neural network in which the RGB human face images are inputted into the convolution neural network, and the minimized cost function is used for the iteration of the target so as to train and form the human face depth-prediction neural network.

The invention discloses a semantic similarity feature extraction method based on double selection gates, and relates to the field of natural language processing. The technical scheme is as follows: firstly, peforming word segmentation and vectorization representation on an input sentence pair to obtain a word vector; inputting the obtained word vector sequence into a bidirectional long-short-termmemory network; obtaining contextual information vectors of the two sentences; secondly, obtaining core feature vectors of sentence pairs through double selection gates respectively; obtaining vectorsof a sentence pair, then inputting the vectors into a multi-angle semantic feature matching network to obtain feature matching vectors of the sentence pair, finally, combining the two semantic feature matching vectors of the matching vectors through a bidirectional long-short-term memory network aggregation layer, and predicting the similarity of the sentence pair. The method effectively alleviates the problem of low matching efficiency caused by information redundancy, and avoids the cost problem of manual core information extraction.

The invention discloses a SAR image change detection method based on a complex neural network, which inputs two original SAR gray-scale images and uses a preliminary difference map obtained by a traditional method as a preliminary label. According to the preliminary tags, some pixel blocks in the two original images are selected as candidate training samples by confidence detection. A complex network is constructed, which consists of three layers of all connection layers and two layers of multiple batches of normalized layers. The complex training samples are selected from the candidate training samples according to the stochastic proportional method to train the complex network. The training complex network is used to test the test complex samples constructed directly from two original SAR images, and the final change detection results are obtained. The invention not only makes full use of the advantages of the traditional change detection result, but also displays the characteristicsof the original data, so that the neural network can better learn the relationship between the two images, thereby obtaining a better change detection result.

The invention provides a short-term water quality and water quantity prediction method based on deep learning, and the method comprises the following steps: A, preprocessing original water quality andwater quantity data, and dividing the processed data into a training set and a test set; B, inputting the training set into an LSTM network for training, and updating the weight by using an adam algorithm to obtain a prediction model; C, predicting a prediction value in the test set by using a prediction model based on the original water quality and water quantity data; D, inputting the prediction error into the ARMA model to obtain an error correction model of an error sequence; E, inputting to-be-predicted data into the prediction model and the error correction model, and geometrically adding settlement results to obtain a predicted value; the invention further provides a water quality and quantity prediction system. The invention has the advantages that the LSTM neural network and theARMA model are used for calculating the water quality and water quantity at the moment to be predicted and the prediction error respectively, higher universality and stability are achieved, and the water quality and water quantity prediction result is more stable.

The invention provides a wind power prediction method integrating a long-term and short-term memory network and an extreme learning machine, and the method comprises the steps of obtaining a wind power sequence and corresponding meteorological data, and carrying out the recombination of the feature data and meteorological feature data of the wind power sequence according to the frequency size, andforming a low-frequency combined input feature vector and a high-frequency combined input feature vector; inputting the low-frequency combined input feature vector into a trained long-term and short-term memory network prediction model to obtain a first prediction result, and inputting the high-frequency combined input feature vector into a trained extreme learning machine prediction model to obtain a second prediction result; and fusing prediction results of the long-term and short-term memory network prediction model and the extreme learning machine prediction model to obtain a final prediction result of the wind power. Different prediction models are set for components of different frequencies, and the prediction result of the prediction model is fused so that the wind power predictioneffect can be improved. Meanwhile, the strong coupling effect of the wind power meteorological information and the wind power is fully considered, and the accuracy of wind power prediction is improved.

The invention provides a meteorological data interpolation method and device, electronic equipment and a storage medium, and relates to the field of meteorological monitoring, and the method comprises the steps: obtaining a missing temperature segment based on obtained automatic observation data; and by taking the obtained manual observation data as a true value, carrying out interpolation on the missing temperature segment according to the trained meteorological data interpolation model so as to obtain complete high-frequency time sequence meteorological data. Complete high-precision frequency temperature observation data are obtained through low-frequency temperature observation data, data interpolation is performed on high-frequency observation temperature data with missing values, and the problem of low precision of current temperature observation data can be solved.

The invention discloses a rolling bearing fault classification method based on CNN and LSTM. The method comprises the steps: acquiring and preprocessing bearing vibration data, and constructing a dataset; denoising the processed vibration data by using a Butterworth Filter, and carrying out fast Fourier transform (FFT) to convert the preprocessed time domain signal into a frequency domain signal;employing a CNN network for learning to obtain image features of a time domain graph and a frequency domain graph, and carrying out image feature fusion through an Add layer; inputting image featuresobtained by CNN learning into an LSTM network, and realizing a classification function through a full connection layer and a Softmax function according to time sequence features contained in the LSTMnetwork learning features. The trained network is used for fault classification of test samples, and the method has important practical significance and practical value for early detection and analysis of weak faults of the rolling bearing.

The invention provides a rehabilitation detection method based on channel state information and BiLSTM-Attention, and the rehabilitation detection method comprises the implementation steps: collectingCSI data of a rehabilitation action in an indoor environment, and extracting amplitude information; performing preprocessing steps of low-pass filtering, normalization, principal component analysis and the like on the CSI data; segmenting the preprocessed signals, detecting starting and ending points of actions, and dividing segmented data segments into a training set and a test set; and inputting the training set into the deep neural network based on BiLSTM-Attention to train the action recognition model to obtain the rehabilitation action recognition model. The collected CSI test set data can be classified by adopting the model, and the purposes of rehabilitation action recognition and rehabilitation degree scoring are achieved. According to the rehabilitation detection method, the deepneural network based on BiLSTM-Attention is adopted for automatic learning and feature selection, and recognition of ten rehabilitation degrees of three different actions is achieved.

The invention discloses an electrocardiosignal atrial fibrillation detection method based on a one-dimensional dense connection convolutional network. The method comprises the following steps: step 1,acquiring a plurality of electrocardiosignal segments containing atrial fibrillation labels; 2, preprocessing the electrocardiosignal segments in the step 1 and taking the electrocardiosignal segments as training data for training a one-dimensional dense connection convolutional network model; 3, building the one-dimensional dense connection convolutional network model by utilizing a deep learning framework; 4, randomly selecting the size of an initial parameter, continuously sending the training data to the model in batches, and performing back propagation to update the network parameter toobtain an optimal parameter; 5, carrying out lightweight processing on the trained network, wherein the lightweight processing comprises parameter quantification and network pruning; and 6, collectingelectrocardiosignals of the patient, sending the signal waveform as input into the one-dimensional dense connection convolutional network model, outputting a result, and pre-judging whether the patient has atrial fibrillation or not.