242 results about "Shortterm Memory" patented technology

The invention discloses a mechanical equipment residual service life prediction method and system. The method comprises the steps that a time convolution network serves as a feature extraction algorithm, a long-term and short-term memory network serves as a regression prediction algorithm, a deep neural network life prediction model is constructed, and the deep neural network life prediction modelis trained; according to the model of the tested equipment and the data acquisition time sequence, constructing the acquired real-time operation data of the tested equipment into a service life prediction data set with time sequence characteristics; and carrying out prediction processing on the life prediction data set by using the deep neural network life prediction model to obtain the residualservice life of the tested equipment. A state monitoring signal output by a sensor for monitoring mechanical equipment has the characteristics of a time sequence; a time convolution network and a longshort-term memory network are combined, a deep neural network life prediction model is established for RUL prediction of mechanical equipment, the problems of over-fitting and gradient disappearanceexisting in a common deep neural network model are solved, and the prediction accuracy is improved.

The invention discloses an air quality space-time prediction method based on a long-term and short-term memory neural network. Particulate matter concentration data of an experiment station and a nearest adjacent station, meteorological data and gaseous pollutant data in the same period are integrated and converted into a supervised learning data format, normalization processing is carried out onthe data, and a prediction sequence of the air mass concentration is obtained by training the data through the long-term and short-term memory network. The method comprises the following steps: S1, acquiring historical air quality data and meteorological data; S2, performing data preprocessing, including abnormal value elimination, missing value interpolation processing, extraction of particulatematter concentration data of adjacent stations and data normalization, on the historical air quality; S3, converting a data format from a sequence to input and output sequence pairs; S4, dividing thedata set into a training set and a test set, and initializing various hyper-parameters of the long-term and short-term memory network; and S5, testing the model effect through prediction on the test set. According to the method, the prediction precision of the air quality data can be improved.

Numerous embodiments are disclosed for an analog neuromorphic memory system for use in a deep learning neural network. The analog neuromorphic memory system comprises a plurality of vector-by-matrix multiplication arrays and various components shared by those arrays. The shared components include high voltage generation blocks, verify blocks, and testing blocks. The analog neuromorphic memory system optionally is used within a long short term memory system or a gated recurrent unit system.

Disclosed is a semiconductor device used to embody a neuromorphic computation system and operation method thereof. By comprising a floating body as a short-term memory means electrically isolated from the surroundings and a long-term memory means formed at one side of the floating body not formed of a source, a drain and a gate, a low power synaptic semiconductor device is provided, which can be mimic not only the short-term memory in a nervous system of a living body by an impact ionization, but also the short- and long-term memory transition property and the causal inference property of a living body due to the time difference of signals of the pre- and post-synaptic neurons.

The invention discloses a face-changing video detection method based on a long short-term memory network. The detection method comprises the steps of video frame extraction, image feature extraction and long short-term memory network training test. The video frame extraction is responsible for extracting a key frame in a video clip and continuous multi-frame images after the key frame, cutting a face area in the image, processing the face image by using high-pass filtering, and extracting detail information in the face image; performing feature extraction on an image by using an Xception convolutional neural network trained in an Image Net image classification data set; and taking the output of the convolutional neural network as the features of the images, splicing the features extractedfrom each frame of image into a feature sequence, and inputting the feature sequence into a long short-term memory network for training to finally obtain a high-precision face-changing video classifier. According to the method, the inter-frame inconsistency existing in the forged video is fully utilized, the detection precision of the forged video is greatly improved, and a very good classification effect is achieved.

The invention discloses a time sequence remote sensing image-oriented land utilization classification method, and the method comprises the following steps: carrying out principal component analysis onmultispectral images forming time sequence remote sensing data to obtain images of three principal components; pre-training each three-waveband image, and extracting a feature image; sequentially inputting the feature images into a semi-supervised convolution long-term and short-term memory network model for training according to a time sequence; and utilizing the trained model to predict and classify the image of the last time phase to obtain a classification result. According to the invention, time context information of a time sequence and space and spectral characteristic information of aremote sensing image are comprehensively considered; the problems existing in the prior art are well solved by utilizing the pre-training model and the semi-supervised classification learning mode, so the land utilization classification method is more suitable for classification scenes with less training sample data and missing or partially missing remote sensing image data, and a better land utilization classification result is obtained.

The invention discloses a modulation signal identification method based on wavelet transform and a convolutional long short-term memory neural network, and the method comprises the steps: firstly obtaining a wireless continuous time signal in advance through a wireless communication system, and forming a data set; secondly, filtering the noisy signal by selecting a reasonable threshold value, andthen reconstructing a wavelet coefficient obtained after processing by utilizing inverse wavelet transform to recover an effective signal; finally, executing the signal feature extraction capability of the convolutional neural network and combining with the memorability of the long short-term memory network, fully learning global features and effectively classifying signal samples with time sequence. A wavelet denoising preprocessing technology is used for suppressing high-frequency noise of an input signal, a convolutional long-term and short-term memory neural network is constructed, globalfeatures are fully learned, and then signal samples with time sequence are more effectively classified; recognition accuracy under a complex environment is improved. therefore, the invention is a modulation identification method suitable for a real channel environment.

The invention discloses a medical text named entity recognition method and device. The method comprises the following steps of: respectively inputting a medical text into a forward long and short-termmemory network and a backward long and short-term memory network so as to obtain a first output result and a second output result; respectively mapping the first output result and the second output result by utilizing a first activation function, and combining the mapped results to obtain a third output result; calculating the third output result by utilizing a second activation function so as toobtain an n*r-dimensional matrix P; and substituting the matrix P into a conditional random field transfer matrix, and calculating and obtaining a global optimum label sequence corresponding to a named entity. According to the method, the medical term recognition correctness and recall rate are high, the calculation speed is high, and medical term recognition can be rapidly carried out, so that model calculation and prediction can be carried out.

The invention discloses a traffic flow data prediction method. The prediction method comprises the steps of obtaining historical traffic flow data related to a to-be-predicted moment; extracting spatial features of the historical traffic flow data by using the trained spatial feature extraction network; inputting the extracted spatial features into a trained time sequence feature extraction network, wherein the time sequence feature extraction network comprises a first long-short-term memory neural network and a second long-short-term memory neural network; wherein the first long-short-term memory neural network outputs long-term time sequence characteristics, and the second long-short-term memory neural network outputs short-term time sequence characteristics and congestion characteristics; and outputting traffic flow data at the moment to be predicted according to the long-term time sequence characteristics, the short-term time sequence characteristics and the congestion characteristics. Compared with a traditional long-term and short-term memory neural network, the long-term and short-term memory neural network based on the ordered neurons is used, influences brought by emergencies can be better captured, and the method is more suitable for variable traffic environments.

The invention discloses an electric quantity sale prediction method based on a long and short term memory network. The method comprises the following steps of S1, determining an influence factor influencing electric quantity sale data; S2, calculating a Pearson correlation coefficient r of electric quantity sale data of an industry to be analyzed and data of each influence factor; S3, using a k-means cluster algorithm, clustering the Pearson correlation coefficient r of each industry and acquiring several clusters after clustering; S4, carrying out normalization processing on daily total electricity consumption data of each cluster and carrying out normalization processing on the data of each influence factor; and S5, based on an electric quantity sale prediction model of a long and shortterm memory network LSTM, acquiring a total electric quantity sale prediction result. In the invention, electric quantity sale data and data characteristics of correlation influence factors can be automatically learned, and based on the long and short term memory network, multi-condition electric quantity sale data is modeled so as to realize accurate prediction of electric quantity sale.

A method for learning a recurrent neural network to check an autonomous driving safety to be used for switching a driving mode of an autonomous vehicle is provided. The method includes steps of: a learning device (a) if training images corresponding to a front and a rear cameras of the autonomous vehicle are acquired, inputting each pair of the training images into corresponding CNNs, to concatenate the training images and generate feature maps for training, (b) inputting the feature maps for training into long short-term memory models corresponding to sequences of a forward RNN, and into those corresponding to the sequences of a backward RNN, to generate updated feature maps for training and inputting feature vectors for training into an attention layer, to generate an autonomous-driving mode value for training, and (c) allowing a loss layer to calculate losses and to learn the long short-term memory models.

The invention relates to an emerging hot topic detection system on the basis of multiclass feature fusion. The emerging hot topic detection system comprises a data preprocessing module, a hierarchicalsequence model, a word encoder layer, a sentence level feature solving layer, a topic level feature solving layer and a topic predicting module. The data preprocessing module is used for preprocessing microblog texts; the hierarchical sequence model is used for training bidirectional recurrent neural network models and training inputted microblog texts by the aid of bidirectional LSTM (long shortterm memory) networks; the word encoder layer is used for vectorizing various words in sentences and forming preliminary vector representation; the sentence level feature solving layer is used for constructing static feature vectors for microblog sentences, linking the static feature vectors of the microblog sentences with neural network dynamic features of the sentence level feature solving layer and forming microblog sentence vector representation; the topic level feature solving layer is used for constructing static feature vectors for topics, linking the static feature vectors of the topics with neural network dynamic features of the topic level feature solving layer and forming topic vector representation; the topic predicting module is used for predicting the topics. The emerging hot topic detection system has the advantages that the emerging hot topic detection system is based on bidirectional long short term memory network architecture, the corresponding dynamic features and the corresponding static features are added, and accordingly the emerging hot topic detection capability can be improved.

The invention discloses a convolutional long-term and short-term memory network space-time sequence prediction method improved by utilizing an attention mechanism. The method relates to the field of computer prediction, and specifically comprises the following steps: (1) extracting spatial features by an asymmetric convolution block high-dimensional feature extractor; (2) a ConvLSTM encoder decoder architecture embedded with an attention module predicting extrapolation features; (3) reversely reconstructing a feature result; (4) L1 and L2 regularization optimization algorithms being carried out; and (5) predicting a space-time sequence image. According to the method, high-dimensional features of spatio-temporal sequence data can be well extracted through the multi-layer convolutional neural network, and the high-dimensional features are used as input of the model, so that the problem that high dimensions cannot be calculated is solved, and spatial key information is emphasized; according to the improved ConvLSTM, the spatial and temporal features can be better learned to realize more accurate extrapolation; the method is suitable for all sequential images.

The invention relates to the field of electric power systems, and discloses a new energy consumption electric quantity prediction method based on a long-term and short-term memory neural network, andthe method comprises the steps: A), collecting historical statistical data related to new energy consumption electric quantity in an operation electric power system of a to-be-predicted region; b) analyzing the bad data, and carrying out data preprocessing to obtain sample data; c) constructing a long-term and short-term memory neural network model; and D) predicting the new energy consumed electric quantity by using the long-term and short-term memory neural network model to obtain a predicted value of the new energy consumed electric quantity. According to the method, effective historical characteristics are mined under the condition of completely utilizing historical data, and parameter optimization is carried out on the long-short-term memory neural network model by utilizing a long-short-term memory neural network method, so that accurate prediction of the new energy consumption electric quantity in the operation electric power system of the region to be predicted is realized.

The invention discloses a knowledge tracking system based on a hierarchical memory network. The system comprises a controller assembly, a hierarchical memory assembly, a read head and write head assembly, the read head and write head assembly is arranged between the controller assembly and the hierarchical memory assembly, and the read head and write head assembly is used for writing input information processed by the controller assembly into the hierarchical memory assembly to be stored and updated; and the hierarchical memory matrix assembly comprises a working storage unit, a long-term storage unit, a segmentation module and an attenuation module, wherein the segmentation module is used for dividing input information into working memory information and long-term memory information and storing the working memory information and the long-term memory information into the working storage unit and the long-term storage unit respectively, and the attenuation module is used for attenuatingthe long-term memory information stored in the long-term storage unit and then storing the attenuated long-term memory information into the long-term storage unit. According to the knowledge trackingsystem based on the hierarchical memory network, the modes of long-term memory and short-term memory of human beings are simulated, input knowledge information is classified, attenuated and stored, and prediction is more accurate.

The invention relates to a construction method of a converter device IGBT residual service life prediction model, in particular to a construction method of a rail transit converter device IGBT residual service life prediction model based on a long-short-term memory network. The problem that the LSTM is used for constructing the remaining service life prediction model of the IGBT of the converter device is solved. The construction method is realized by the following steps: 1, collecting IGBT accelerated aging data; 2, performing data normalization processing; 3, constructing and training a long-term and short-term memory network; and 4, verifying the prediction model. By determining the characteristic parameters, constructing the LSTM deep network architecture and specifying the network training parameters, the construction method obtains an IGBT residual service life prediction model of the converter device which reaches the required prediction error index. The prediction model constructed by the construction method is based on a long-short-term memory network and is applied to prediction of the residual service life of an IGBT of a converter device, especially a traction converterof rail transit.

The invention discloses a method for detecting character interaction in an image based on multi-feature fusion, and the method comprises the steps: detecting all instance information, including human body position information, object position and category information and the like, in a picture through a target detection algorithm, then inputting the information into a trained character interaction behavior recognition network, and detecting an interaction behavior between a character pairs in a to-be-detected picture. According to the method, on the basis of global spatial configuration of a pose capture interaction relationship, effective information provided by a human and object intersection area is focused, finer local features are learned, the matching probability of correct human interaction pairs is increased, people, objects and background region information are effectively screened and utilized by means of a short-term memory selection module, and the precision of character interaction detection is improved through fusion of various features.

The invention relates to a dangerous condition early warning and forecasting method for pipe jacking downward penetrating process of an existing box culvert. The method comprises the steps that 1, multiple characteristic parameter data of the existing box culvert and the surrounding soil bodies of the existing box culvert are collected through a monitoring device; 2, format processing is conductedon the collected data; 3, the soil strength in a disturbance zone is obtained through inversion by using a dichotomous displacement ratio method; 4, multiple characteristic parameters, the weight ofthe soil strength of the disturbance zone affecting culvert safety coefficients and gray relational grade are obtained through a hierarchical analysis gray relational grade method; 5, a long and shortterm memory circulation neural network model is established to predict the destruction occurring time of the existing box culvert; 6, a Kalman filtering method is adopted to predict the destruction occurring time of the existing box culvert; 7, prediction results of step 5 and step 6 are combined to conduct early warning and forecasting on dangerous conditions before the critical destruction occurring prediction time. Compared with the prior art, the method can more reasonably perform quantification early warning and forecasting on the dangerous condition for pipe jacking downward penetratingof the existing box culvert in a manner which is more closer to actual conditions.

The invention discloses an image text recognition method, device and equipment and a computer storage medium. The image text recognition method comprises the following steps: extracting spatial features of a target image by using a convolutional neural network; utilizing a long-term and short-term memory network to extract time sequence features of the target image according to the spatial features; determining at least one text region in the target image according to the spatial features and the time sequence features; and identifying text information in the text region. According to the embodiment of the invention, the irregular image text in the target image can be quickly and accurately recognized, and the performance of image text recognition is improved.

The embodiment of the invention provides a network service identification method and device. The method comprises the following steps: inputting the feature data of a service data flow into a trainedlong-short-term memory network model, and outputting a service type label corresponding to the service data flow, the long-short-term memory network model being obtained by training based on the feature data of the sample data flow and a predetermined service type label of the sample data flow; and obtaining the service type of the service data flow according to the service type label output by the long-short-term memory network model. The LSTM network model is obtained by training based on the feature data of the sample data flow and the predetermined data flow type label, so that the LSTM network model has characteristic long-term memory and short-term memory, and can fully consider the characteristics of each data packet in the data stream in time dimension and space dimension, and accordingly the identification of the service type of the data stream is more accurate.

The invention discloses a power load prediction method based on EEMD secondary decomposition. The method comprises the following steps: constructing a load time sequence; preprocessing the data; carrying out primary signal decomposition; carrying out secondary signal decomposition on the high-frequency signal; performing time sequence combination prediction; and outputting a load prediction result. According to the invention, the power load time series data is mined by using the multi-layer long short-term memory network. The non-stationary nonlinear original time sequence is converted into aplurality of sub-sequences through a signal decomposition mode, and the decomposed high-frequency sub-sequences are subjected to secondary decomposition, thereby obtaining the implicit deep features of the data, and effectively improving the load prediction precision.

The invention relates to the fields of air combat environments, data processing, deep learning and the like, and provides a method for realizing aircraft trajectory prediction by using a long short-term memory (LSTM) network under an uncertain sensing condition. Therefore, the technical scheme adopted by the invention is as follows: the method for predicting the trajectory of the aircraft based on the long-short-term memory network comprises the following steps of: eliminating noise interference carried by a sensor feature vector by using Kalman filtering; data preprocessing including downsampling, invalid value elimination and missing value complementation is carried out on the directly obtained state parameters, in addition, in order to improve the calculation stability, data is subjected to normalization processing, and the value range of input data is included in the interval of [0, 1]; and an LSTM-based trajectory prediction model is created, input and output of the network is defined, and the network is supervised and trained. The invention is mainly applied to the prediction occasion of the flight path of the unmanned aerial vehicle.

There is provided a neural network system for detection of malicious code, the neural network system comprising: an input receiver configured for receiving input text from one or more code input sources; a convolutional neural network unit including one or more convolutional layers, the convolutional unit configured for receiving the input text and processing the input text through the one or more convolutional layers; a recurrent neural network unit including one or more long short term memory layers, the recurrent neural network unit configured to process the output from the convolutional neural network unit to perform pattern recognition; and a classification unit including one or more classification layers, the classification unit configured to receive output data from the recurrent neural network unit to perform a determination of whether the input text or portions of the input text are malicious code or benign code.

The invention discloses a weak light color imaging method based on deep learning. The method comprises the following specific steps: (1) constructing a convolution long-term and short-term memory network, and denoising an original image acquired by a sensor in combination with spatial and temporal information of multiple channels; (2) constructing a residual network, learning a mapping relationship from the denoised original image in the step (1) to an RGB image, and performing demosaicing processing on the image; and (3) constructing a cyclic generative adversarial network, learning conversion from night illumination to daytime illumination, and performing reillumination conversion on the image obtained in the step (2). The method aims at weak light imaging application. In order to solvethe problem that a clear and high-contrast color image is difficult to recover due to an extremely low signal-to-noise ratio of an acquired signal, high-quality, color and real-time imaging of a low-illumination scene is realized by utilizing deep learning in combination with denoising, demosaicing and heavy illumination processing methods, and the method is widely applied to the fields of military affairs, security and protection monitoring, scientific research and the like.

The invention relates to artificial intelligence, and provides a service system anomaly detection method and device, computer equipment and a storage medium, and the method comprises the steps: constructing a multi-scale signature matrix according to the multivariate time sequence data of each index generated by a service system; inputting the multi-scale signature matrix into a convolution layerto encode a spatial mode of the multi-scale signature matrix, and outputting a spatial feature map; inputting the spatial feature map into an attention-based convolutional long-short-term memory network layer, and updating the hidden state of the spatial feature map through the attention-based convolutional long-short-term memory network layer to obtain an updated spatial feature map; inputting the updated spatial feature map into a deconvolution layer to decode and reconstruct the updated spatial feature map to obtain a reconstructed signature matrix; comparing the reconstructed signature matrix with the multi-scale signature matrix, and determining an abnormal index of the service system. In addition, the invention also relates to a blockchain technology, and the multivariate time seriesdata can be stored in the blockchain. By adopting the method, the anomaly detection accuracy can be improved.

The invention discloses a power distribution system network loss prediction method based on a long-term and short-term memory network, and the method comprises the steps: obtaining a plurality of feature quantities of a prediction region at all moments, and constructing a time sequence of each feature quantity; obtaining a reference sequence and a comparison sequence from each characteristic quantity time sequence, calculating the correlation degree between the reference sequence and the comparison sequence by adopting a grey correlation analysis method, and selecting an influence factor corresponding to the optimal comparison sequence as an input influence factor; obtaining a training sample from each characteristic quantity time sequence according to the input influence factor; establishing a long-term and short-term memory network model, and training the long-term and short-term memory network model by using the training sample to obtain a network loss prediction model; and obtaining an input influence factor of the network loss value at the prediction moment, and performing prediction by using the network loss prediction model to obtain the network loss value at the predictionmoment. The method can improve the prediction precision and efficiency of the network loss of the power distribution system, so as to achieve the purposes of guiding the energy-saving work more efficiently and determining the energy-saving amount of a project.

The invention discloses a social network node classification method based on a dynamic graph. The relationship of different nodes before and after a time sequence is enhanced by utilizing a sparsemaxfunction; the sparse processing and the cell gating are enabled to act together by combining the long-term and short-term memory neural network, the front-back relationship and the dependency relationship of the time series data are better mined, and the change mode of the node state in the time series data is fully expressed, so that the social network node classification accuracy is improved. The invention provides a social network node classification method. For dynamically changing time series data in a social network, the problem that the mutual influence between nodes and the dependencyrelationship before and after different time cannot be effectively mined can be solved, and the method can be used for the classification problem of dynamic structure social nodes in the fields of social platforms, recommendation systems, information systems, medical health, movie and television entertainment and the like.

The invention discloses a text generation method and device, relates to the technical field of data processing, and aims to solve the problem that a target text generated according to an existing model is inaccurate in the prior art. The method mainly comprises the steps of obtaining initial text data; calculating hidden space parameters of a variational auto-encoder of the initial text data according to a preset BERT language model; taking the initial text data, the hidden space parameters and an initial control condition as input data, taking a control statement corresponding to the initialtext data under the initial control condition as output data, and correcting the weight of a training LSTM (Long Short Term Memory) decoder by adopting a sequential reverse transmission algorithm so as to train the LSTM decoder; and generating a target statement of a to-be-tested statement by taking the to-be-tested statement and the target control condition as input data of the LSTM decoder. Themethod and the device are mainly applied to a similar text extension process.

The invention belongs to the technical field of information security, and discloses a password guessing set generation system and method. The method comprises the steps of: generating a probabilisticcontext-independent grammar based on personal information and a password database, dividing character strings in the probabilistic context-independent grammar into character strings suitable for or not suitable for training a long-term and short-term memory neural network according to classification rules, training a convergent long-term and short-term memory neural network model, generating a password segment and a probability corresponding to the password segment by using the convergent long-term and short-term memory neural network model, mapping the probabilities corresponding to the password segments into new probabilities, and sorting the password segments in a descending order according to the new probabilities, and generating passwords sorted in probability descending order. According to the method, the defects that a long-term and short-term memory neural network cannot identify the composition structure and semantic information in the password and is poor in interpretabilityare overcome; the defect of poor generalization capability of probabilistic context-independent grammar is overcome; and the problem that the probability of generating the password segment by the long-term and short-term memory neural network is low is solved.