19 results about "Linear embedding" patented technology

The invention discloses a transformer abnormity identification method based on voiceprint feature analysis, and belongs to the field of power equipment state monitoring and intelligent diagnosis. The method comprises the following steps: firstly, analyzing an iron core acoustic mechanism based on a magnetostrictive effect, and establishing a three-dimensional model through finite element simulation to obtain vibration and sound field characteristics; in a complex substation environment, a hybrid noise reduction method combining density peak clustering and a CEEMDAN-wavelet threshold is provided, and the signal-to-noise ratio is effectively improved. Then extracting Mel-frequency cepstrum coefficients (MFCC) and spectrum features, and performing local linear embedding (LLE) dimension reduction to form a compact feature set; in the recognition stage, a convolutional neural network framework is designed, specifically, a spectrogram and an energy spectrum are modeled through a two-dimensional CNN, an MFCC tensor obtained after dimensionality reduction is modeled through a three-dimensional CNN, and accurate diagnosis of mechanical faults such as core looseness is achieved. The method has the advantages of being non-contact, anti-noise and high in recognition precision, and real-time diagnosis and early warning of mechanical abnormity of the transformer can be achieved under complex working conditions.

A method and an apparatus for forecasting a future time target variate, a computer device, and a storage medium are disclosed. Specifically, a method for forecasting a future time target variate is disclosed. The method includes: dividing a history target variate and covariates of target variates into a plurality of patches channel-wise, and performing linear embedding, wherein the covariates of target variates comprise a covariate of the history target variate and a covariate of the future time target variate; performing position embedding on a plurality of tokens and a learnable forecasting token; wherein the plurality of tokens are linear embedded from the plurality of patches; inputting the plurality of tokens and the learnable forecasting token into an encoder; and forecasting and outputting the future time target variate based on the encoder and the learnable forecasting token. According to the foregoing manner, in a case of zero-shot training, the future time target variate can be accurately forecast through limited history target variates and covariate information of the target variates, thereby having a large inspiration effect on the field of basic models in time series forecasting.

The invention discloses an automatic driving vehicle dynamics modeling method based on a time-varying parameter neural network, and belongs to the field of vehicle dynamics. Comprising the steps of obtaining operation data of a vehicle under different working conditions, performing preprocessing, and constructing a sample set according to a time sequence slicing mode; constructing a neural network model based on the physical constraints and the time-varying parameters; training and verifying the neural network model by adopting the sample set to obtain a neural network model after parameter optimization for state prediction; the neural network model comprises an input layer, a GRU network, a physical constraint layer and an output layer; input data enters the neural network model through the input layer, and dimension transformation and time sequence packaging are carried out through the linear embedding layer; then time sequence features are extracted through the GRU network, bounded mapping is carried out on output vectors of the GRU network through a physical constraint layer, a time-varying parameter estimation value is obtained, the time-varying parameter estimation value is substituted into a vehicle kinetic equation for calculation, and a physical prediction state is obtained.

The invention discloses a time series prediction method based on a multi-modal enhanced large language model, which comprises the following steps: acquiring historical time series data, constructing a standardized input matrix and setting core task parameters; dividing data blocks through a sliding window mechanism, and converting the data blocks into uniform dimension features through linear embedding; a semantic prototype is generated based on a large language model vocabulary, and time sequence features and semantic features are fused through multi-head cross-attention; alternately splicing the data blocks and the corresponding semantic information, and constructing a self-multi-modal input sequence; designing a three-level structured prompt including context, task target and modal guidance, and fusing the three-level structured prompt with a multi-modal sequence; and training the lightweight model by adopting a frozen training strategy, and outputting a prediction result of a specified time step in the future. According to the method, through single-source data enhancement and prompt guidance, the time sequence reasoning capability of a small-parameter large language model is activated, high-precision prediction is guaranteed, efficient deployment is achieved, and the method adapts to long-term and short-term prediction tasks in the fields of electric power, traffic, meteorology and the like.

The invention relates to a river flow prediction method, which comprises the following steps of: 1, dividing a daily runoff data set, and performing multi-stage decomposition on a training set by using Haar wavelets to obtain de-noising parameters; 2, multiplexing the verification set and the test set to obtain data; step 3, carrying out first-order differential transformation, and zooming to an interval of [0, 1] by using MinMaxScaler to obtain a normalized value of a differential sequence; 4, mapping the data to a high-dimensional space through a linear embedding layer, and then carrying out position coding to obtain data; step 5, inputting the data obtained in the step 4 into a DTransformer encoder, so that a catastrophe point obtains a higher attention weight so as to strengthen difference characteristics of adjacent time steps and obtain encoder output; 6, inputting the data obtained in the step 5 into an LSTM decoder, and performing feature recombination to obtain encoder output; 7, mapping the data obtained in the step 6 to 7-dimensional output through a full connection layer; and 8, performing inverse normalization and inverse difference transformation on the data obtained in the step 7 to obtain final output.

The application belongs to the technical field of image processing, and discloses a chest and abdominal cavity bleeding automatic segmentation method and device based on deep learning, which comprises the following steps: obtaining and normalizing chest cavity or abdominal cavity CT volume data, and then performing linear embedding to generate initial features; inputting the initial features into an encoder for multi-stage encoding, in which the encoding features of different levels are respectively enhanced in the frequency domain according to different semantic levels to generate level-aware frequency domain enhanced features; the input features are modeled by a deformable mixed window multi-head self-attention mechanism in at least one stage of the encoder, and the attention calculation results are weighted and fused by using the frequency domain guide weight generated based on the frequency domain enhancement results; the encoding features are input into a decoder for multi-stage decoding and fusion with the corresponding level features of the encoder to gradually restore the spatial resolution; and the bleeding area segmentation results corresponding to the CT volume data are generated according to the final output features of the decoder. The application can efficiently and accurately segment the bleeding area and assist clinical decision-making.

The invention discloses a dual-path Transform-based sky and ground wave identification method in a Rowland system. The method comprises the following steps: performing learnable discrete wavelet decomposition on a preprocessed signal received by a Rowland receiver to obtain a low-frequency component and a high-frequency component; projecting the components to a potential embedding space through independent linear embedding layers, and obtaining an embedding vector with position information in combination with position coding; inputting the embedded vector into a double-path Transform encoder, and respectively extracting low-frequency and high-frequency characteristics; the encoded features are reconstructed and fused through learnable inverse wavelet transform, and unified global representation is obtained; inputting the global representation into two layers of stacked Encoder for further optimization to obtain deep features; and finally, inputting the deep features into a multi-layer perceptron to realize sky and ground wave recognition. According to the invention, by introducing learnable wavelet decomposition and reconstruction, a dual-path encoder structure and a sparse attention mechanism, the sky and ground wave signals are effectively separated and fused, so that the success rate of ground wave recognition is remarkably improved.

The present disclosure provides a training method of an image extraction model and an image extraction method, which can be applied to the fields of image processing and pattern recognition. The image extraction model comprises a linear embedding network, an encoding network and a decoding network. The training method comprises: processing a sample image obtained by using the linear embedding network to obtain an embedding feature map; processing the embedding feature map by using the encoding network to obtain i encoding feature maps; processing the i encoding feature maps by using the decoding network to obtain a weight fusion feature map, wherein the decoding sub-network is constructed based on a local-global attention layer and a weight feature fusion layer; performing segmentation head mapping processing on the weight fusion feature map to obtain an image segmentation result, wherein the image segmentation result represents a geological change attribute of a target geographical environment region; and training the image extraction model according to the image segmentation result and label data corresponding to the image segmentation result to obtain a trained image extraction model.

The present invention relates to the technical field of artificial intelligence, and disclosed are a CA-SwinT-based hierarchical image classification method and system. The method comprises: S1, acquiring an original image, and preprocessing the original image; S2, performing feature extraction on the image by means of a SwinT module; S3, using a dual cross-attention mechanism to fully fuse extracted image features and classification label information so as to construct fine mapping between the image and a hierarchical label; and S4, on the basis of the fine mapping in S3, generating a corresponding label by means of a label generation module, and updating the classification label information. S2 comprises: patch-based processing, linear embedding processing, two-stage Swin module processing, patch merging processing, two-stage Swin processing, and linear embedding processing. In the present invention, rich feature information of an image can be effectively extracted from spaces of different dimensions, thereby identifying a multi-level label of the image more efficiently.

A maskless Transform framework-based intelligent identification method for operation events of a small nuclear power system comprises the following steps of: 1, acquiring multi-source heterogeneous time sequence data related to an operation state by using a sensor system or a simulation platform carried by the small nuclear power system to form an original data set; 2, preprocessing the original data set; 3, creating a linear embedded layer network, and mapping a multi-dimensional parameter vector in the sample sequence to a high-dimensional feature space by using the network to form an event data embedded vector; 4, performing position coding on the event data embedding vector, and adding time sequence information in the sample sequence into the event data embedding vector; (5) a maskless Transform encoder network formed by stacking N identical encoder layers is built, and N is a positive integer greater than or equal to 1, and N is a positive integer greater than or equal to 1; 6, the event data embedding vector is input into a maskless Transform encoder network, and a feature sequence is output; and 7, calculating the probability that the sample sequence belongs to each operation event category through a Softmax function, and judging the category with the maximum probability as a final recognition result. The method is high in recognition accuracy and strong in time sequence characteristic capturing capability.

This invention provides a method, apparatus, and device for identifying grain origins based on hyperspectral images, relating to the field of image recognition technology. The method includes: acquiring hyperspectral sample images of grains from different origins using a hyperspectral imaging device; constructing a pseudo-RGB image of each hyperspectral sample image by selecting three representative bands; performing grain instance-level segmentation on each pseudo-RGB image to obtain a corresponding segmentation mask; extracting the average spectral information of each grain using the segmentation mask and constructing a spectral information matrix; sequentially performing principal component analysis, local linear embedding, and sparse subspace clustering on the spectral information matrix to obtain a dimensionality reduction matrix; training a fully connected neural network model based on the dimensionality reduction matrix to obtain a grain origin traceability model, and then performing grain origin traceability based on the grain origin traceability model. This invention can significantly improve the accuracy and processing efficiency of grain origin identification.

The invention discloses a partial discharge fault diagnosis method based on multi-scale convolution, and aims to improve the diagnosis precision of a partial discharge fault of power equipment. The method comprises the following steps: reading original binary partial discharge time sequence PRPS signal data, and carrying out preprocessing; the method comprises the following steps: extracting local features based on multi-scale one-dimensional convolution, generating sequence input by combining multi-scale Patch division and linear embedding, carrying out sequence modeling by utilizing a multi-layer Transform encoder, and realizing fault classification through attention pooling and a multi-layer full-connection classifier; for a single normalized input signal, convolution feature extraction, multi-scale embedding, Transform coding, attention aggregation and classification prediction, and fault category and corresponding probability output are performed in sequence. According to the scheme, the accuracy and the automation level of partial discharge fault diagnosis can be improved, and the method is suitable for scenes such as state evaluation and fault early warning of power equipment.

The application belongs to the field of magnetic resonance sounding signal noise filtering, and is a kind of magnetic resonance sounding signal extraction method based on intelligent optimization manifold learning, the parameter group of the local linear embedding manifold learning method is initialized, the genetic algorithm in the intelligent optimization algorithm is used, the signal-to-noise ratio is taken as the fitness function, and the parameter group in the local linear embedding manifold learning method is optimized, the local linear embedding manifold learning method uses the optimized parameter group to sequentially perform first processing and second processing on the magnetic resonance sounding signal, removes random noise, and obtains the final denoised magnetic resonance sounding signal. The application effectively retains the signal characteristics through the nonlinear dimension reduction of manifold learning, avoids the information loss caused by frequency band selection, maintains the local relationship between data points in the dimension reduction process, ensures that adjacent points in m-dimensional space remain adjacent in d-dimensional space, only compresses and filters noise, and realizes signal extraction.

The present application relates to the technical field of power grid super-harmonic, and specifically discloses a power grid super-harmonic online monitoring and identification method and system, which utilizes a preset learning type measurement matrix to perform non-uniform compression sampling on original signals to be measured of a power grid, so as to obtain a low-dimensional observation vector sequence; the sequence is subjected to linear embedding processing and superposition position coding, so as to generate a feature vector containing time sequence information; the feature vector is input into a pre-trained Transformer reconstruction model, a high-dimensional time domain waveform after restoration is output through a decoding layer of the Transformer reconstruction model, so as to realize reconstruction of a super-harmonic signal; the reconstructed waveform is subjected to feature analysis, and monitoring parameters of the power grid super-harmonic or a harmonic source identification result is output. The present application can obtain a high-frequency signal under the condition of lower sampling rate of hardware, can reduce data transmission amount, and thus effectively reduces system cost; compared with a reconstruction algorithm of traditional compression sensing which depends on iterative solution, the present application only needs one-time forward calculation to complete signal reconstruction, and response speed is improved.

The invention discloses an unmanned aerial vehicle detection and classification method based on deep learning. A received unmanned aerial vehicle radio frequency signal is processed to generate sparse enhanced MVDR spectrum features; based on an improved self-attention learning feature extraction and coding module and a convolutional coding module, a SunnyitUAVCNN neural network is constructed; sequentially carrying out sequence segmentation, linear embedding and spectrum physical position coding on the sparse enhanced MVDR spectrum features, carrying out feature hierarchical extraction and enhancement through the constructed neural network, and constructing a residual structure to enhance feature representation; and finally, introducing a classification confidence weighting mechanism, performing dimension compression and probability mapping on the feature matrix processed by the neural network, calculating classification confidence, and comparing the classification confidence with a preset threshold to complete unmanned aerial vehicle detection classification. According to the invention, the problem that unmanned aerial vehicles with unknown models or signal disguise cannot be effectively identified in the prior art is solved.

A CT apparatus in which the x-ray source is coupled to a source robotic arm and the detector is coupled to a detector robotic arm. A motion correction module utilizes a locally linear embedding motion correction algorithm to estimate the geometry-describing parameters associated with the positions of the source and detector and the angle of the detector. These estimates are used to reconstruct the image data and produce corrected images with fewer errors resulting from patient movement, misalignments, and coordination issues in the system.

The application relates to a kind of asymmetric neighborhood aggregation residual graph convolution collaborative filtering recommendation method and equipment.The method constructs the two-part graph of user and item according to user feature information, item feature information and its interaction data, the feature information of user and item is one-hot encoding, then using linear embedding layer is mapped to initial node embedding vector;Node embedding vector is carried out feature propagation in improved graph convolution layer, the weighted average of node embedding vector of all layers is obtained, and the final embedding vector of user and item is obtained;Inner product operation is carried out on the final embedding vector in the prediction layer, and the score of target user to target item is obtained.The application uses a new feature propagation mechanism to update the node embedding vector of user and item, fully captures the high-order interaction relationship between user and item, not only relieves the over-smoothing effect, but also fully utilizes the low-order node features of user and item, thereby effectively improving the accuracy of recommendation.

The invention discloses a video-oriented intelligent animation generation method, and the model of the intelligent animation generation method employs an encoder and decoder architecture, and comprises a shallow feature extraction module which is used for carrying out the feature extraction of an input video sequence and a style reference image, and obtaining a video shallow feature and a style shallow feature; and the feature embedding and coding module is used for carrying out block linear embedding on the video shallow layer features, adding space-time position codes and coding the style shallow layer features to obtain style feature vectors. According to the invention, through the space-time dynamic attention path in the STSFA module, the combined modeling of the space structure and the time dynamic is realized, the long-time-history space-time dependency relationship in the video can be accurately captured, and for the shielding scene, the dynamic state of the shielded area can be deduced through the motion trail of the adjacent frames.