34results about How to "Solve long-term dependence" patented technology

The invention discloses a multi-dimension labelling and model optimization method for audio and video. The method specifically comprises the following steps: first, carrying out sample management andsorting, carrying out de-duplication aiming at sample data of an input system, carrying out numbering, and establishing a sample labelling task library; at the preprocessing stage of audio data, carrying out audio extraction on video data of the task library, and completing the preprocessing operation for the audio data; at the audio content analysis and feature extraction stage, after the audio preprocessing is completed, carrying out deep analysis according to a labelling standardized system configured at the background, and outputting label data; S304, at the video content analysis and feature extraction stage, carrying out image analysis on the video content, and carrying out deep analysis according to the labelling standardized system configured at the background, and outputting the label data; S305, carrying out feature fusion and label generation, namely, fusing the recognition features and label information, and outputting a label result of the sample; carrying out manual rechecking and model optimization, wherein the label result data generated by the system can be subjected to artificial re-check conformation.

The invention provides an end-to-end classification method of a large-scale news text based on Bi-GRU and a word vector. The end-to-end classification method comprises the following steps: S1. word Embedding word-level semantic feature representation is performed; S2. the attention weight Bi-GRU word level sentence feature coding model is constructed; S3. the Bi-GRU sentence level feature coding model based on the attention weight is established; S4. hierarchical Softmax is applied to realize end-to-end classification implementation. According to the method, the dimension of the vector can bereduced and the problem that the features are too sparse can be effectively prevented. The final output vector is optimized and the effectiveness of model feature coding is enhanced. The problem thatthe model is difficult to train because of the high dimension can be avoided and the additional semantic information can also be provided. The feature extraction model and various common classifiers can be flexibly combined so as to facilitate replacement and debugging of the classifiers. The computational complexity is reduced from | K | to log | K | in comparison with that of Softmax.

The invention provides a rolling bearing remaining life prediction method based on a long-short term memory network. The rolling bearing remaining life prediction method comprises the steps that characteristics of abrasion signals of the rolling bearing is extracted; principle component analysis is conducted on the extracted characteristics to obtain fusion characteristics; normalization processing is conducted on the fusion characteristics; cyclic overlapped interception is conducted on fusion characteristic data in a set step-size to obtain short sequences; the short sequences are divided into a training set and a prediction set; an LSTM deep learning network is constructed; the LSTM deep learning network is trained through the training set; the LSTM deep learning network is verified through the prediction set; and training results and prediction results are subjected to inverse normalization processing and output. According to the rolling bearing remaining life prediction method based on the long-short term memory network, an LSTM prediction model is provided based on the field of deep learning and has high applicability and accuracy in fault time sequence analysis, and the problem of long-term dependence in time sequence is solved.

The invention discloses a residual network and attention mechanism-based drug relationship extraction method. The method comprises the following steps of: S1, carrying out vector representation on words in a drug entity relationship data set; S2, carrying out time-series modeling on a drug relationship statement by utilizing a two-layer bidirectional long-short term memory model neural network; S3, importing residual connection into the constructed two-layer bidirectional long-short term memory model neural network; S4, decomposing a deep semantic meaning automatically obtained by the two-layer bidirectional long-short term memory model neural network into a memory space and an attention space, fusing memory information and attention information, and inputting the fused information into aSoftmax classifier to extract a drug relationship. According to the drug relationship extraction method disclosed by the invention, dependency relationships between long-distance words are effectivelysolved, gradient diffusion is overcome, model overfitting is prevented, the model robustness is good and the classification effect is good.

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 provides a manufacturing method for continuously rolling a seamless steel pipe by using a hollow mandril. The method comprises the following steps of: designing the steel grade and the specification of the hollow mandril; performing continuous casting, cogging, and forging; rolling a mandril material, performing tempering heat treatment; and detecting the performance of the mandril material, and processing and connecting the mandril material. The method has the advantages that: the problem that the retained mandril depends on import for a long time is solved, so that the production localization of the retained madril is realized; the service life of the mandril is close to or the same as that of the solid mandril of the same specification which is imported from foreign countries or bought in China; and the quality of the inner and outer surfaces of a seamless steel pipe product in the production process is good, 1.5 to 1.7kg of steel is consumed for one ton of mandrils, the cost of one ton of steel is 30 yuan which is only one third the original cost, and the production cost is greatly reduced. Due to the development and use of the hollow mandril, the consumption cost of a tool is reduced, and the market competitiveness of a steel pipe product is improved; the technology promotion of the mandril manufacturing industry in China is driven, and the method has a profound and lasting significance for the top potential and consumption reduction of metallurgical industry.

The invention discloses an emotion recognition method based on a bidirectional gating circulation unit network and novel network initialization. The emotion recognition method includes the steps: extracting high-dimensional features of three modes of text, vision and audio, and aligning according to the word level; performing normalization processing, inputting the data into a bidirectional gatingcirculation unit network for training; adopting a network initialization method to initialize the weights of the bidirectional gating circulation unit network and the full connection network at the initial training stage of each modal network; performing feature extraction on the state information output by the bidirectional gating circulation unit network by adopting a maximum pooling layer andan average pooling layer; and splicing the two pooled feature vectors to serve as input features of the full connection network, and inputting a to-be-identified text, vision and audio into the trained bidirectional gating circulation unit network of each mode to obtain emotion intensity output of each mode. According to the emotion recognition method, the problem of long-term dependence can be solved; the robustness of the bidirectional gating circulation unit network in training is improved; and the emotion recognition accuracy based on the emotion time context information is improved.

The present invention discloses a PMU primary frequency modulation load forecasting method based on the associated fully connected neural network and LSTM, in particular comprising the following steps: selecting training data, verifying the data, establishing a joint neural network model, training the joint neural network model, and inputting the forecasting sample set into the joint neural network model trained. The method of the invention considers the correlation between the historical data of the load and the power in the ultra-short-term load forecasting, adopts the structure of the LSTMneural network and the fully connected neural network, and effectively solves the problem of long-term dependency. The invention also has the advantages of simple algorithm, short running time and high prediction accuracy, and provides a solid guarantee for the stable operation of the power network.

The invention relates to an air quality prediction method in gridding monitoring. The method comprises the following steps: firstly, performing data cleaning on position information and historical air pollutant concentration information of each monitoring station in gridding monitoring input by a user, then inputting the processed data into a GCN to extract space correlation information among the monitoring stations, and inputting the data with the space information into an LSTM to extract time features; and finally, integrating the features extracted by the GCN and the LSTM through a linear regression layer, generating a prediction result, and returning the prediction result to a user. The effectiveness of the method is verified through related experiments.

The invention provides a remote sensing image semantic description method based on multistage feature fusion, and belongs to the field of remote sensing image processing and computer vision, and the method comprises the following steps: obtaining a high-resolution remote sensing image, and constructing a remote sensing image semantic description data set; training a semantic classification model of the image by using the semantic annotation data set, extracting word description from the image and encoding to obtain semantic features; training a target detection model by using the target detection data set, extracting region-level features of the image and encoding the region-level features to obtain visual features; aggregating the obtained semantic and visual features, namely splicing the two groups of features together; and taking the aggregated multi-level features as the input of Transform, and training an image natural language generation model. The semantic and visual features of the image are utilized, the extracted information comprises the scene information, the regional visual information and the semantic relation of the object, and the generated image semantic description is high in readability and high in accuracy.

The invention discloses a news website general crawler design method based on a GRU neural network. HTML data preprocessing is carried out on HTML page content, target data is constructed, charactersare marked, a character dictionary is constructed, the HTML content is converted into a digital vector, and finally batch filling is carried out; establishing a GRU neural network, using a Cross Entropy as a loss function, and using a pre-trained character vector to train and predict the GRU neural network by an Embedding layer; and based on the Scrapy crawler framework, constructing a whole-station crawling crawler. According to the method, after a crawler crawls HTML content of any news page, the HTML content is transmitted into the model trained by using the neural network algorithm designed by the invention, so that news texts can be automatically extracted, and customized time and manpower are saved.

The invention belongs to the technical field of shield tunneling, and particularly relates to a shield attitude position deviation prediction method. According to the method, various parameters of a completed shield construction project serve as source domain data and are trained in a pre-training model, relevant parameters of a feature extraction layer in the pre-training model are extracted, two new full connection layers are overlaid behind the feature extraction layer, and a shield attitude deviation prediction model is formed; taking each parameter in the current shield construction project as target data, and training on the shield attitude deviation prediction model by using the target data so as to obtain prediction of shield tunneling deviation; and various parameters of the completed shield construction project are used as source domain data, so that enough training data for training in the initial stage of shield construction is ensured, and the prediction effect of the shield attitude deviation prediction model is further ensured. The situation that the prediction effect is inaccurate due to the fact that only a small amount of data is used for training in the initial stage of the shield is avoided.

The invention discloses a single-mode label generation and multi-mode emotion discrimination method based on a Transform algorithm, and the method comprises the steps: 1, obtaining a multi-mode non-aligned data set, and carrying out the preprocessing of the multi-mode non-aligned data set, and obtaining an embedded expression feature of a corresponding mode; 2, establishing an ITE network module, and extracting intra-modal features; 3, single-mode label prediction and multi-mode emotion decision discrimination label fusion generation are carried out; 4, establishing an inter-modal BTE network module and a modal enhancement MTE network module, and obtaining inter-modal features and modal enhancement features through a global self-attention STE network module; and 5, obtaining a label of multi-modal emotion deep prediction. According to the method, for the condition that a current multi-modal data set only has one multi-modal label, decision fusion is carried out through a self-supervised weighted voting mechanism to generate a single-modal label, and based on the use of various cross-modal TE, data between modals are fully interacted, so that the precision of multi-modal emotion discrimination can be improved.

The invention provides a child pneumonia auxiliary diagnosis model and a training method thereof; the training method comprises the steps: obtaining a medical image of a child pneumonia patient and a corresponding medical diagnosis statement, enabling the medical image to serve as a training image set, and enabling the medical diagnosis statement to serve as a training statement; extracting an image depth feature vector from the image training set data through a CNN neural network to obtain a depth feature image set, and performing word vector training on the training statement through a word2vec model to obtain a depth feature vector word set; and carrying out feature fusion on the deep feature image set and the deep feature vector word set, and then carrying out training through an LSTM neural network so as to obtain a trained child pneumonia auxiliary diagnosis model. According to the method, the existing medical image of the child pneumonia patient and the corresponding medical diagnosis statement are trained, and the model obtained through training serves as a tool for a doctor to learn diagnosis or provides effective reference opinions for clinical diagnosis of the doctor.

The invention relates to deep learning in the field of artificial intelligence, in particular to learning of a system call sequence. The method comprises: data collection and processing, model construction, model training, model evaluation and sequence generation. The data collection and processing comprises the following steps: firstly, collecting a system call sequence with parameters and a sequence in a trace format, and then coding the sequences into input data suitable for model training. The model construction comprises: selecting RNN and LSTM neural network models, and determining a network structure as an input layer, a hidden layer and an output layer. Model training includes batching input data, initializing network parameters, calculating a value of a loss function to adjust the network parameters. Model evaluation includes calculating a normalized edit distance between test sequence data and a prediction sequence. The sequence generation comprises the following steps: randomly selecting initial system call and sequence length, generating an integer sequence according to a model obtained by training, and decoding the integer sequence into a system call sequence. And the generated sequence is used as the input of the kernel fuzzy test, so that the vulnerability mining efficiency is improved. The process is shown in Figure 1.

The invention relates to an LSTM-based optical cable manufacturing equipment fault remote prediction system. The system comprises a detection node and a data processing node, wherein the detection node comprises a microprocessor, a data acquisition module, a communication module, an analog-to-digital conversion module and a power supply module, the data processing node comprises an upper computerand a display module. When the system works, the microprocessor in the detection node controls the sensor of the data acquisition module to acquire and detect typical fault process parameter data of an optical cable production line, the microprocessor processes the acquired and detected data and wirelessly transmits the processed data to the data processing node through the communication module ofthe microprocessor, and the upper computer receives a data signal and sends the data signal to the communication module of the microprocessor; the trained LSTM network is called to analyze and calculate the data, and finally the equipment operation state model is outputted to a display screen to complete fault prediction. The system can prevent problems of the optical cable production line in thecase of sudden faults, reduce operation and maintenance cost and improve the coping capability of the production line for the sudden faults.

The invention relates to a lower limb motion intention prediction method based on an attention mechanism, which comprises the following steps of: 1, acquiring a gait signal in a lower limb motion process, performing normalization processing on the gait signal, and dividing training set data and test set data; step 2, constructing a prediction model, wherein the prediction model comprises an input module, a convolutional neural network module, a long-short term memory neural network module and an attention mechanism, and when the input module, the convolutional neural network module, the long-short term memory neural network module and the attention mechanism are connected in sequence, an output result of the attention mechanism passes through a full connection layer; 3, pre-training the prediction model by using the training set data, and determining a time step length; then training the pre-trained prediction model; and 4, applying the trained prediction model to prediction of the motion intention of the lower limbs. According to the method, the position with obvious change of the joint angle is amplified through the attention mechanism, and the prediction error of the joint angle can be effectively reduced.

The invention relates to a haze prediction method based on a global attention mechanism, and belongs to the technical field of artificial intelligence information prediction. The method comprises thefollowing steps: firstly, acquiring haze data of an environment monitoring point, processing the acquired haze data, training a haze prediction model based on a global attention mechanism, and outputting a final prediction result by using the haze prediction model. In a haze prediction task, a global attention mechanism is introduced, different influence factors are endowed with different weights,and the problem that the information transmission distance is too long is effectively solved. A bidirectional gating recurrent neural network is introduced, the influence of previous moment data in training data on subsequent moment data is introduced, the association of the subsequent moment data and the previous moment data is analyzed, the problem of long-term dependence in haze prediction data is solved, and the haze data at the future moment can be accurately predicted. The method has good expansibility, the network structure can be dynamically changed according to the data characteristics of different regions, and the haze prediction method suitable for the local region is obtained.

The present invention claims to protect a character recognition system and method based on the combination of neural network and attention mechanism, which specifically includes: a convolutional neural network feature extraction module for spatial features of text images; inputting the spatial features extracted by the convolutional neural network To the two-way long-short-term memory network module, the two-way long-short-term memory network can extract the sequence features of the text; the extracted feature vectors are semantically encoded, and then the attention weights of the feature vectors are assigned through the attention mechanism, so that attention can be focused on the weights Higher feature vector; the decoding part of the model is realized by nesting long-term short-term memory network, and the features extracted by attention and the prediction information at the previous moment are used as the input of nested long-term short-term memory network, and long-term short-term memory is used before and after The purpose of the network is to maintain the time characteristics of the feature vector, so that the model pays attention to the continuous change of the position point with time; the invention can more accurately detect the text area in the natural scene, and has a good effect on the small target text and the text with a small inclination angle. Good detection effect.

The invention provides a method for manufacturing a hollow mandril of a continuously rolling seamless steel pipe. The method comprises the following steps of: designing the steel grade and the specification of the hollow mandril; performing continuous casting, cogging, and forging; rolling a mandril material, performing tempering heat treatment; and detecting the performance of the mandril material, and processing and connecting the mandril material. The method has the advantages that: the problem that the retained mandril depends on import for a long time is solved, so that the production localization of the retained madril is realized; the service life of the mandril is close to or the same as that of the solid mandril of the same specification which is imported from foreign countries or bought in China; and the quality of the inner and outer surfaces of a seamless steel pipe product in the production process is good, 1.5 to 1.7kg of steel is consumed for one ton of mandrils, the cost of one ton of steel is 30 yuan which is only one third the original cost, and the production cost is greatly reduced. Due to the development and use of the hollow mandril, the consumption cost of a tool is reduced, and the market competitiveness of a steel pipe product is improved; the technology promotion of the mandril manufacturing industry in China is driven, and the method has a profound and lasting significance for the top potential and consumption reduction of metallurgical industry.

The invention provides a short-time logistics demand prediction method, device and equipment and a readable storage medium, and relates to the technical field of logistics demand prediction, and the method comprises the steps: obtaining first information, wherein the first information comprises feature information obtained after decomposition of historical logistics data in a first time period through EEMD and LMD, establishing an LSTM prediction model for logistics demand prediction, and correcting logistics demand prediction through an error correction mathematical model. According to the method, the EEMD-LMD-LSTM-LEC model is provided from two perspectives of feature decomposition and feature extraction aiming at the characteristics of non-stationarity, strong randomness, local mutation, nonlinearity and the like of short-time logistics data, so that the problems of relatively large prediction error and prediction hysteresis of direct prediction caused by non-linear unsteady original requirements are solved. And the prediction precision is improved.

The invention provides a modulation signal identification method for an evolutionary long-short term memory network. The method comprises the following steps: constructing a data set; constructing a target function; initializing parameters of a cheongfish predation searching mechanism; calculating a fitness value, and determining the position of the elite semaphorus and the position of the injured sardine; selecting a sailfish attack selection strategy, and updating the position of the sailfish; pursuing the preys, and updating the position of the sardine; calculating a fitness value, determining the sardines caught by the cheongfish, and determining the positions of the elite cheongfish and the injured sardines; judging whether an iteration termination condition is met, namely, the maximum number of iterations is reached or all the sardine is captured by the cheongfish, if the iteration termination condition is met, continuing to run downwards, otherwise, enabling g to be equal to g + 1, and returning to continue; and training the digital communication signal modulation identification LSTM network with the optimal hyper-parameter by using the training set. According to the method, the optimal LSTM network model parameters are obtained by designing a culture cheongfish predation search mechanism.

The invention discloses a Chinese statement simplification method and device. The method comprises the steps of carrying out the word segmentation processing on a received text sequence to obtain a word segmentation sequence; performing encoding processing on the word segmentation sequence through a trained encoder to generate a semantic vector; and decoding the semantic vector through the traineddecoder to generate a simplified output sequence, wherein the encoder and the decoder are both LSTM models. According to the scheme provided by the embodiment of the invention, the combination of LSTM is selected for encoding and decoding in the model training process because the LSTM can effectively solve the problem of long-term dependence in sequence prediction.

The invention discloses a general crawler design method for news websites based on a GRU neural network. HTML data preprocessing is performed on HTML page content, target data is constructed, characters are marked, a character dictionary is constructed, the HTML content is converted into a digital vector, and a batch is finally filled; Build a GRU neural network, use Cross Entropy as the loss function, and the Embedding layer uses pre-trained character vectors to train and predict the GRU neural network; build a full-site crawler based on the Scrapy crawler framework. After the crawler crawls the HTML content of any news page, the present invention transfers it into the model trained by the neural network algorithm designed by the present invention, and can automatically extract the news text, saving time and manpower for customization.

The invention discloses a PMU primary frequency modulation load prediction method based on an associated fully connected neural network and LSTM, which specifically includes selecting training data, verifying data, establishing a joint neural network model, training the joint neural network model, and inputting the prediction sample set into the trained joint neural network. Network model; the method of the present invention considers the correlation between the historical data of load and power in the ultra-short-term electric load forecasting, adopts the structure associated with the LSTM neural network and the fully connected neural network, and effectively solves the long-term dependence problem ; The present invention also has the advantages of relatively simple algorithm, short running time and high prediction accuracy, which provides a solid guarantee for the stable operation of the power grid.

A complex human behavior recognition method based on multi-feature fusion CNN-BLSTM includes the following steps: segment continuous sensor data through a sliding window, extract features from the segmented sensor data, and use a feature selection algorithm to classify the series of artificial Extract features for screening and retain dominant features; input the segmented behavior data into the deep learning model for training, first perform one-dimensional convolution pooling processing through the convolutional neural network, and then use the average pooling process through the bidirectional long-term and short-term memory neural network The layer extracts the salient features of the state output by the bidirectional long-short-term memory neural network, and finally fuses the pooled feature vector with the previously extracted advantage feature vector as the input feature of the fully connected layer to obtain the output of complex behavior recognition. The invention can fully excavate the characteristics of sensor data and improve the recognition accuracy of complex human behaviors.

The invention discloses an electric vehicle speed reducer remaining service life prediction method based on LSTM. The method comprises the steps of: preprocessing full-life-cycle vibration data of an electric vehicle speed reducer; constructing degradation characteristics of reducer vibration signals on a time domain, a frequency domain and a time-frequency domain, thereby constructing a degradation characteristic data set of each electric vehicle reducer, and constructing a residual life prediction model by using an LSTM neural network; normalizing the degradation characteristic data set input into the prediction model; carrying out segmentation processing on the normalized data; dividing the degradation characteristic data set into a training set and a test set, and taking training data as the input of the whole prediction network model; and performing network parameter adjustment by using a single variable method, and outputting an optimal life prediction result. The prediction method which is more reliable and higher in generalization is provided for analysis of the remaining service life of the speed reducer of the electric vehicle.

The present invention provides an auxiliary diagnosis model for childhood pneumonia and a training method thereof. The training method includes acquiring medical images of children with pneumonia and corresponding medical diagnosis sentences, the medical images being used as a training image set, and the medical diagnosis sentences being used as training sentences ; Extract the image depth feature vector from the image training set data through CNN neural network to obtain a depth feature atlas, and perform word vector training on the training sentence through the word2vec model to obtain a depth feature vector word set; The feature fusion of the deep feature vector word set and the deep feature vector word set is carried out, and then the LSTM neural network is used for training, that is, a trained child pneumonia auxiliary diagnosis model can be obtained. The present invention trains the existing medical images of children with pneumonia and the corresponding medical diagnosis sentences, so that the model obtained by training can be used as a tool for doctors to learn diagnosis or provide effective reference opinions for doctors' clinical diagnosis.

The invention discloses a microgrid load prediction method based on a long and short term neural network model, and relates to the technical field of microgrid load prediction, and the method comprises the steps: obtaining the weight of each influence factor when the clustering interval is maximum, and extracting the feature vector of each influence factor in a current group through a double-layer convolutional neural network; optimizing the long-short-term neural network through an optimizer, and carrying out feature vector screening through the optimized long-short-term neural network; weighting processing is carried out on the screened feature vectors through attention optimization, and optimized feature vectors are obtained; and connecting the optimized feature vectors through a full connection layer and obtaining the predicted load of the current group. According to the invention, the optimized feature vectors are connected through the full connection layer, the predicted load of the current group is obtained, and the long-term and short-term neural network is adopted to solve the problem of long-term dependence in the training process.