93 results about "Sequence modeling" patented technology

The invention discloses a session recommendation method based on a space-time sequence diagram convolutional network. The method comprises the following steps: S1, modeling all session sequences intoa directed session graph; S2, constructing a global graph by taking common commodities in the session as links; S3, embedding an ARMA filter into a gated graph neural network, extracting a topologicalgraph signal which changes over time from the graph model, and obtaining a feature vector of each node involved in the session graph; S4, obtaining global preference information from historical sessions of the user by adopting an attention mechanism; S5, obtaining local preference information of the user from the last session clicked by the user, and obtaining final preference information of theuser in combination with the global preference information; S6, predicting the probability of possible occurrence of the next clicked commodity in each session, and giving a Top-K recommended commodity. According to the method, rich context relationships of clicked commodities can be captured from the global graph, global and local preferences of the user are accurately learned, the time attenuation effect of historical preferences of the user on current preferences is effectively evaluated, and accurate commodity prediction is provided.

The invention discloses a flood prediction method based on an attention model long-short time memory network, which comprises the following steps: collecting flood related data including flows and flood factors at different time points; introducing An attention model into the long-short term memory network, and constructing a long-short term memory network based on the attention model; Training the attention model-based long-term and short-term memory network by using the standardized flood related data; And carrying out flood prediction by using the trained long-short term memory network based on the attention model. According to the method, the sequence modeling capability of an original long-short-term memory network is reserved, an attention model is introduced, a circulation scheme isused for optimization, local situation information is described as a weight scheme, some flood factors which are not used for prediction are ignored, the prediction average time is greatly shortened,and meanwhile accuracy is improved.

The invention provides a multi-target voice enhancement method based on SCNN (Stacked Convolutional Neural Network) and TCNN (Temporal Convolutional Neural Network) joint estimation. On the basis of aSCNN and a TCNN, a new stacked and temporal convolutional neural network (STCNN) is provided; a log-power spectra (LPS) is used as the main characteristic, and input into the SCNN, so that high-levelabstract characteristics are extracted; then, a power function compression Mel-frequency cepstral coefficient (PC-MFCC) according with auditory characteristics of human ears better is provided; the TCNN takes the high-level abstract characteristics extracted by the SCNN and the PC-MFCC as the input; then, sequence modelling is carried out; furthermore, joint estimation on the clean LPS, PC-MFCC and an ideal ratio mask (IRM) is carried out; finally, in an enhancement stage, different voice characteristics have complementarity in a voice synthesis process; an IRM-based post-processing method isprovided; and enhancement voice is synthesized by adaptively adjusting the weight of the estimated LPS and IRM through voice presence information.

The invention discloses an industrial fault diagnosis method based on an improved KPCA (kernel principal component analysis) and hidden Markov model and belongs to the technical field of industrial process monitoring and diagnosing. Calculation efficiency of KPCA under the condition of large samples is greatly improved by a similarity analysis method, and industrial faults can be classified by means of high dynamic process time sequence modeling capability and time sequence model classifying capability of the hidden Markov model. Accordingly, compared with the existing methods of the prior art, the industrial fault diagnosis method has the advantages that complexity in calculation can be reduced, nonlinear characteristics can be more efficiently processed and nonlinear industrial fault diagnosis is high in accuracy since the nonlinear characteristics and massive data of industrial data are sufficiently considered.

The invention relates to a BERT embedding-based software programming field entity identification method, and belongs to the technical field of natural language processing, deep learning and software resource mining. The method comprises the following steps: firstly, carrying out text analysis and preprocessing on a data set of a software question and answer community StackOverflow by utilizing a natural language processing technology, determining a software programming domain entity category in combination with domain analysis, and carrying out manual annotation on sample data based on a Bartnatural language annotation tool to obtain a training set and a test set; secondly, obtaining semantic and vectorized representation of an input sequence through a BERT pre-training language model, and performing model training on the input sequence in combination with a BiGRU bidirectional recurrent neural network; and finally, modeling the input label sequence through a CRF conditional random field, thereby obtaining the label sequence with the maximum probability, and achieving entity identification in the field of software programming. Based on a deep learning training method, specific entities in the software programming field can be effectively identified under the condition of a small amount of labeled sample data.

The invention relates to a method for predicating long correlation sequences by utilizing a short correlation model. Aiming at self-similarity network flow, the invention provides an ARMA (autoregressive moving average model) self-similarity sequence predicating method based on EMD (empirical mode decomposition). The method comprises the following steps of: firstly decomposing the self-similarity network flow into a plurality of IMFs (Intrinsic Mode Functions) by adopting the EMD method, wherein due to the narrow-band characteristic of the IMF, the IMF is provided to be a short correlation sequence, so that the problem of modeling predication of the long correction sequences is converted into the modeling and predicating for the plurality of short correlation sequences, and the complexity of the model is effectively reduced; secondly predicating the decomposed IMF sequences by utilizing excellent short correlation modeling predication capacity of an ARMA model; and finally providing a method for improving the predication precision of the model, so as to effectively reducing the normalization error of mean square of the predication result. The method provided by the technical scheme of the invention has the advantages of high predication precision and low complexity, and the predication precision of self-similarity flow is higher than that of a neural network model.

The invention relates to the field of voice synthesis and voice conversion, and more specifically relates to a many-to-many voice conversion method based on double voiceprint feature vectors and sequence-to-sequence modeling. According to the method, a large number of parallel prediction is generated by employing the voice synthesis technology of a plurality of speakers, and great convenience is provided for model training; and then input source speaker features are modeled and mapped to target speaker features by employing a sequence-to-sequence neural network. In order to realize many-to-many voice conversion, voiceprint feature vectors representing the identities of the speakers are generated by employing a model of speaker verification tasks; the voiceprint feature vectors of the source speakers and the target speakers are added to the sequence-to-sequence model as auxiliary confidence; and through model training tests, good effects can be achieved.

The invention discloses a named entity recognition method for Chinese sentences, which comprises the following steps: inputting a Chinese character sequence into a recognition model, converting the Chinese character sequence into a character vector by the recognition model through a character embedding layer, and outputting the character vector to a convolutional network in the recognition model; the method also includes that the convolutional network performs convolution operation on each word vector to obtain a local semantic vector and outputs the local semantic vector to a self-adaptive combination layer in the recognition model; the self-adaptive combination layer performs attention calculation on the local semantic vector of the character and then splices the local semantic vector with the corresponding word vector to obtain a representation vector and outputs the representation vector to a sequence modeling network in the recognition model; and the sequence modeling network performs hidden layer modeling on the representation vector of the character and outputs the hidden layer vector obtained by modeling to a label reasoning layer in the recognition model to calculate a label corresponding to the hidden layer vector of the character. The local semantic information of the characters is extracted through the convolutional network and then is fused with the potential words based on the attention among the words, so that the utilization of the potential word information is realized, and the problem of wrong transmission of word boundaries is avoided.

The invention discloses a user behavior model training method, a recommending method, a model training device and equipment. The model training method comprises the steps of acquiring a user behaviorsequence; inputting the user behavior sequence into a user behavior sequence model under the current model parameter to obtain a current user expression; obtaining a first training sample according tothe current user expression and the user behavior sequence; determining a mutual information loss value by adopting a mutual information loss function according to the first training sample, and updating model parameters of the user behavior sequence model according to the mutual information loss value; and taking the updated model parameter as the current model parameter, and returning to execute the step of inputting the user behavior sequence into the user behavior sequence model under the current model parameter to obtain the current user expression until the current model parameter meetsthe preset condition. User behavior sequence modeling is realized through an unsupervised learning method based on mutual information maximization, the training time and cost of the user behavior sequence model are reduced, and the invention can be widely applied to the field of artificial intelligence.

The invention relates to a ViSAR-based anomalous change detection and tracking method. The ViSAR-based anomalous change detection and tracking method comprises the first step of target detection, wherein a foreground image and a background image are determined through the method of subtraction of a previous frame and a next frame of original three frames of an image of a video, a background region and a motion region are determined after setting of the foreground image and the background image of the image is completed, the motion region is demarcated as a target, position information and a statistical histogram of the demarcated target serve as the features of the demarcated target, the demarcated target and the position information and the statistical histogram which are carried by the demarcated target are brought into a next frame, and the second step is executed; the second step of target tracking and positioning; the third step of target anomalous change detection, namely dynamic sequence modeling. By the adoption of the ViSAR-based anomalous change detection and tracking method, monitoring equipment is not affected by weather conditions such as sunlight, cloud and mist, rainstorm and haze; the problems of defocusing and target blurring of SAR imaging are solved through a special algorithm.

The invention discloses an interest point prediction method based on a spatio-temporal point process. The interest point prediction method comprises the steps of S1, realizing spatio-temporal contextinformation integration and user sign-in sequence modeling based on the point process; S2, predicting user interests based on a spatio-temporal point process; and S3, realizing spatio-temporal contextand sequence perception prediction. According to the method, the behavior mode and interest of the user are extracted from the sign-in sequence of the user by using the spatio-temporal point process,the context interest of the user is predicted by combining the time-space context, and finally, the general interest and context interest of the user are comprehensively considered, so that the prediction effect is improved and the accuracy is improved.

The invention relates to a method for diagnosing the hub bearing fault of an automobile based on a Hankel matrix. The method comprises steps of firstly constructing the Hankel matrix to achieve the two-dimensional matrix representation of an original vibration signal and to improve signal correlation; secondly, establishing a convolutional neural network (CNN) of a known fault mode to enhance a hidden Markov model; and finally, diagnosing a fault mode of an unknown fault type by using the model. The CNN is a data-driven feature learning method, can perform convolution and sub-sampling operations on a two-dimensional signal represented by the Hankel matrix from the original data by using the CNN model, completely retains a part of representing the features of the signal, reduces the high-dimensional interference components in the signal, solves the feature automatic learning problem of the fault signals, and finally establishes the hidden Markov model of the known fault mode by using the self-learned feature. On the one hand, the method uses the CNN to automatically learn the features and reduce data dimension so as to obtain the distributed feature representation of the data; and on the other hand, determines the fault type by using the dynamic sequence modeling capability and the timing sequence mode classification ability of the Hidden Markov Model.

The invention provides an image character recognition method based on a space-time convolutional neural network. The method comprises the following steps: firstly, preprocessing an input image; extracting features of the input image by using a convolutional neural network; performing sequence modeling on the extracted features; and finally, predicting a final result according to the extracted feature sequence. The invention provides an image character recognition method based on a space-time convolutional neural network, and the method directly extracts text information in an image text, reduces the frequency of manual text input, saves the time of manual information input, reduces the workload, and improves the work efficiency.

The invention discloses a comment sentiment analysis method based on an improved neural network, and the method comprises the steps: constructing a comment representation matrix for inputted comment text data; calculating a comment representation matrix through a plurality of convolution kernels in sequence to obtain feature maps of different sizes; calculating each obtained feature map by using pyramid pooling to obtain a feature vector with a fixed length; splicing to obtain a feature vector of each feature map, and connecting the feature vectors with the full connection layer; mapping the output of the full connection layer into a probability distribution vector by using a Softmax function, wherein each dimension of the probability distribution vector corresponds to an emotion categoryin the emotion analysis task; and selecting the emotion category corresponding to the value with the maximum probability in the probability distribution vector as a comment emotion judgment result. According to the method, the text sequence can be effectively modeled, and the sequence characteristics of the text can be effectively reserved, so that the emotional attitude in the comment text content can be accurately and effectively recognized.

The invention relates to the technical field of communication signal processing, in particular to a Morse code automatic identification method based on a Bi-LSTM neural network. The method comprises the following steps: S1, constructing a convolutional neural network and a Bi-LSTM neural network, and performing sequence modeling in combination with the Bi-LSTM neural network and the convolutionalneural network to generate a multi-mode LSTM model; S2, training the multi-mode LSTM model in a joint training mode, and performing joint optimization on parameters of the Bi-LSTM neural network and the convolutional neural network; S3, acquiring a Morse code audio signal, and preprocessing the Morse code audio signal to obtain a preprocessed audio signal; S4, analyzing and converting the preprocessed audio signal to generate a frequency spectrum image of the audio signal; S5, inputting the frequency spectrum image into a multi-mode LSTM model, and outputting a probability vector result; and S6, judging the content of the Morse code by utilizing a probability vector result. According to the method, the deep neural network model based on Bi-LSTM can be utilized to efficiently and accuratelycomplete automatic identification of the Morse code.