20results about How to "Mitigate Vanishing Gradients" patented technology

This invention belongs to the field of computer vision technology, specifically a tree-structured video description generation method based on perceptual grammar knowledge. This invention explicitly utilizes semantic information inherent in language, using dependency structure analysis tools to transform sequential sentences into a syntax tree structure. By analyzing the connections between parent and child nodes in the tree, the dependency grammar structure within the sentence is explicitly modeled. A perceptual context attention network is used to model the contextual information generated along different paths during the generation process. Simultaneously, reinforcement learning and iterative generation training methods are introduced during the training phase to further improve model performance. Qualitative and quantitative experiments demonstrate that the model has the ability to generate more accurate and semantically richer descriptions.

The application discloses a kind of ball grid array package chip solder ball bubble defect automatic detection method and system, it is related to chip intelligent detection technical field, the method is guided by designing semantic cross-layer fusion framework, realizes multi-scale feature balanced fusion in neck network by introducing feature alignment and redistribution module, strengthens small target details and semantic information using local semantic enhancement module, and the positioning accuracy of bubble is optimized in combination with smooth geometry positioning loss function.The application significantly improves the detection capability of small, fuzzy and occluded bubbles, has strong robustness and generalization, and can be widely used in solder ball bubble, material porosity and surface defect industrial vision detection scene.

The invention provides a multispectral image classification method based on an adaptive feature fusion residual network, and belongs to the technical field of computer vision and remote sensing image processing. The method comprises the following steps: firstly, preprocessing an original multispectral image; then, a novel adaptive feature fusion residual network model is constructed, the core of the model comprises a dual-path feature extraction backbone network, and spectral features and spatial context features are extracted respectively; a self-adaptive feature fusion module is introduced, multi-level features from double paths are dynamically learned and fused through a channel attention mechanism, and optimal integration of spectral information and spatial information is achieved; and finally, completing pixel-level or image-level classification by using the fused advanced features. According to the method, the problems of insufficient utilization of spectrum-space joint features of multispectral data and weak model generalization ability of a traditional method are effectively solved, and the precision and robustness of multispectral image classification are remarkably improved.

This invention discloses a method for monitoring the state trends of hydropower units based on two-stage signal decomposition and the IBiLSTM model. The method involves collecting vibration signals from the hydropower units for preprocessing; constructing the ITGCOA optimization algorithm and designing a fitness function; using the ITGCOA optimization algorithm to adaptively optimize the SVMD and BAACMD models, achieving initial decomposition of the preprocessed signal and secondary decomposition of the sub-mode component with the highest center frequency obtained from the initial decomposition; calculating the fuzzy entropy values ​​of the remaining sub-mode components for reconstruction; and fusing the high-frequency feature sub-sequences obtained from the secondary decomposition with the reconstructed feature sub-sequences to construct the input sequence of the prediction model. This input sequence is then used to construct the IBiLSTM prediction model for monitoring the state trends of hydropower units, achieving high-precision prediction. Compared with existing technologies, this invention improves the efficiency and accuracy of hydropower unit state trend prediction, accurately warns of abnormal unit operating conditions, ensures the safe and stable operation of the units, and improves the overall efficiency of the power plant.

The application belongs to the field of traffic flow prediction, and provides a city traffic flow prediction method, system and product under complex road network conditions, and the technical scheme is as follows: time mixing features are extracted based on city road network traffic flow time series data, intersection information between different traffic features is captured based on the time mixing features, mixed feature representation is obtained by fusing the time mixing features and the intersection information between different traffic features; global dependence in a long sequence is captured by selective state space modeling based on the mixed feature representation, and dynamic time series features are output; after the dynamic time series features are enhanced, a trailing time dimension is projected to obtain prediction values of each feature variable of the traffic time series data. In city traffic flow prediction, data can be dynamically processed according to real-time traffic conditions, multivariate information can be effectively utilized, and efficient long sequence modeling can be performed, and the prediction accuracy of time series data is significantly improved.

This invention belongs to the field of materials science, specifically relating to a crystal structure generation method based on variational autoencoders and Cartesian coordinate derivation. The aim is to construct a crystal structure design method driven by target properties. The method includes: acquiring real crystal structure data and performing data augmentation processing; training a variational autoencoder using the crystal structure dataset as the real sample and the target material properties as the input conditions; converting the real crystal structure data in the crystal structure dataset into crystal diagrams, performing feature derivation on the edges of the crystal diagrams based on the geometric information of atoms in the Cartesian coordinate system, and training a graph neural network based on the edge-derived features; inputting the target material properties into the fully trained variational autoencoder to generate candidate crystal structures; inputting the candidate crystal structures into the fully trained graph neural network for property prediction; comparing the predicted property values ​​with the target material properties within a preset target tolerance range and selecting the desired crystal structure.

This invention provides a method and system for identifying the lubrication state of self-lubricating spherical friction pairs, belonging to the field of mechanical equipment condition monitoring. The method collects acoustic emission signals during the operation of the friction pair using an acoustic emission sensor. After normalization, overlapping sliding window sampling, and data augmentation preprocessing, the signals are input into a dual-stream time-frequency fusion deep learning model. The model extracts complementary features through parallel time-domain and frequency-domain feature extraction modules, performs spatiotemporal modeling using a bidirectional GRU and Transformer encoder layer, and optimizes the model by combining a focus loss function and the AdamW optimization algorithm. Finally, the lubrication state prediction result is output. This invention fully utilizes the time-frequency characteristics of acoustic emission signals, achieving high accuracy, strong anti-interference capability, and good real-time performance. It can effectively identify the lubrication break-in period, stable period, and rapid deterioration period, providing a guarantee for the reliable operation of self-lubricating spherical friction pairs.

This invention proposes a target detection method for low-light environments, comprising the following steps: acquiring road traffic image data under nighttime or low-light conditions; annotating the road traffic image data with targets; constructing an enhanced detection network LLE-YOLO based on an improved version of YOLOv11n, where the C3k2 module in the backbone network is replaced with a CSP-PMSFA module, and the Conv module is replaced with a ContextGuidedBlock module; the C3k2 module in the neck network is replaced with a CSP-PMSFA module, and an SPDConv module is introduced; a DetectAux module is introduced in the head network; the low-light traffic perception dataset is divided into a training set, and the enhanced detection network LLE-YOLO is trained; after the enhanced detection network LLE-YOLO is trained, the auxiliary detection head and its related branches are removed, and end-to-end inference is performed on targets in low-light environments, outputting target category and location information. This method achieves lossless downsampling, adaptive enhancement of multi-scale features, effective suppression of background noise, and guidance of hard sample learning, thus solving the problem of target detection in low-light environments.

This application discloses a training method and related apparatus for a quantum generative adversarial network (GAN), belonging to the field of quantum computing technology. The GAN includes a generator and a discriminator. The method includes: using the generator to obtain generated samples against random noise; using the discriminator to distinguish between real samples and the generated samples, obtaining a discrimination result; updating the parameters of the generator and the discriminator based on the discrimination result, the loss function of the generator, and the loss function of the discriminator, to obtain a trained GAN. At least one of the generator and the discriminator includes a quantum convolutional layer and a quantum residual neural module connected sequentially. The quantum residual neural module includes a first qubit and a second qubit for encoding and evolving each feature data. Two qubits; a first quantum logic gate and a second quantum logic gate acting on the first qubit for performing an identity mapping operation on feature data; a third quantum logic gate acting on the second qubit and entangled with the first qubit between the first and second quantum logic gates, wherein the third quantum logic gate includes an encoding module for encoding feature data and parameterized training logic gates located before and after the encoding module for implementing residuals, and the encoding module includes an encoding logic gate and entanglement gates located before and after the encoding logic gate, the entanglement gates realizing the entanglement of the first and second qubits; the second qubit corresponding to one feature data at an adjacent position is the first qubit corresponding to another feature data, the feature data being obtained based on a quantum convolutional layer. Applying this application can effectively suppress the gradient vanishing problem and reduce the waste of training resources caused by training failures.

The invention discloses an integrated circuit time sequence prediction method, device, equipment and product. The method comprises the steps of obtaining a GBA time sequence path report of a target integrated circuit; constructing a triple of each time sequence arc in the target integrated circuit based on the GBA time sequence path report; the time sequence arc comprises nodes and edges between the nodes, the nodes are pins of a logic unit in the target integrated circuit, and the triple comprises node features, edge features and a node time sequence diagram; inputting the triple of each time sequence arc into a pre-trained time sequence prediction model for time sequence prediction to obtain prediction delay and prediction conversion time of each time sequence arc; the time sequence prediction model comprises multiple layers of residual error edge networks connected through a residual error connection mechanism, each layer of residual error edge network aggregates and encodes node features and edge features of each time sequence arc in the target integrated circuit, and the obtained encoding node features, edge features and node time sequence diagrams are transmitted to the next layer of residual error edge network as input. The technical problem that the time sequence prediction efficiency of a time sequence calculation method is low is solved.

The invention relates to the technical field of geotechnical engineering and artificial intelligence crossing, and discloses a grouting data enhancement method and system for an improved table generative adversarial network, and the method comprises the steps: obtaining grouting table data, carrying out the preprocessing, obtaining the preprocessed grouting table data, and generating a condition vector; processing the preprocessed grouting table data and the condition vector by using a data enhancement model; outputting the synthesized grouting table data and the authenticity judgment result; training a generator module and a discriminator module by using a loss function according to the authenticity discrimination result, respectively configuring learning rate adaptive schedulers for the generator module and the discriminator module, and adjusting the learning rate according to the loss change; and based on the trained data enhancement model, generating enhanced grouting table data and performing similarity evaluation. A small sample grouting data set can be effectively expanded, and sufficient training samples are provided for a data-driven grouting parameter prediction model.

The application discloses a discharge detection device and method for a general high-low voltage switch cabinet for mines, and belongs to the technical field of switch cabinet detection. The device comprises a modularized sensor assembly, a signal acquisition unit, an edge computing node unit and an upper monitoring platform, and adopts a UHF sensor, an ultrasonic sensor and a transient ground voltage sensor for cooperative detection. The method comprises multi-source signal synchronous acquisition, adaptive denoising based on a sparrow search algorithm optimization, double-domain feature extraction, discharge type identification based on an attention-enhanced convolutional neural network, and multi-sensor information fusion and hierarchical early warning based on evidence theory. The application compensates for the blind area of a single detection method through cooperative work of multiple types of sensors, improves the denoising effect through adaptive parameter optimization, enhances the recognition accuracy through double-domain features and an attention mechanism, and improves the diagnosis reliability through multi-sensor fusion, so that accurate detection and intelligent diagnosis of local discharge of a mine switch cabinet are realized.

The present application relates to the field of artificial intelligence and data processing technology, specifically to a temperature control sensor data completion method based on a generative adversarial network, comprising obtaining a time series data matrix of a temperature control sensor containing missing data; constructing an adversarial model containing a generator and a discriminator; inputting the data matrix into a forward topological hard embedding layer, calculating the time and space partial differential physical residual and embedding the neuron transmission equation to output a latent feature matrix; inputting the latent matrix into an active feature selection module, eliminating redundant vectors and generating a synthetic sequence matrix; inputting the synthetic sequence and the real sequence into the discriminator to obtain the confidence, calculating the composite loss of the adversarial, reconstruction and residual constraints; using a gradient penalty mechanism to iterate the network parameters until convergence, and finally outputting the target temperature time series to complete the matrix. The present application effectively solves the problem of missing sensor data by introducing partial differential physical residual constraints and active feature selection, improving the adversarial network interpolation accuracy and the reliability of the time series sequence.

The application discloses a soft-hard interbedded rock mechanical parameter prediction method and system based on a residual attention network, relates to the technical field of rock mechanical parameter prediction, and has the advantages that the traditional neural network is prone to gradient disappearance when processing a deep network, which influences the training effect; the existing method lacks an attention mechanism and cannot effectively identify and strengthen key features; and the modeling capability for interlayer interaction is insufficient; the application provides a soft-hard interbedded rock mechanical parameter prediction method based on a residual attention network, which comprises the following steps: obtaining structure parameters and target mechanical parameters of a soft-hard interbedded rock sample; converting the structure parameters into an enhanced feature vector; constructing a residual attention network model; inputting the enhanced feature vector into the residual attention network model; and outputting a mechanical parameter prediction result and reliability evaluation information.

The invention discloses a communication adaptive adjustment method and system based on an underwater acoustic communication frame blank time slot, and the method comprises the steps: obtaining an underwater acoustic communication frame, and collecting environment noise data based on the blank time slot of the underwater acoustic communication frame; performing time-frequency transformation on the environmental noise data to generate a time-frequency characteristic graph; the time-frequency feature map is input into a residual hybrid expert network, and the residual hybrid expert network is obtained by adopting joint loss function training and comprises a shared backbone network, a gating network and a plurality of expert networks; extracting a high-dimensional feature vector from the time-frequency feature map through a shared backbone network; inputting the high-dimensional feature vector into a gating network, and outputting a hard routing selection result by the gating network, which is used for indicating to activate a certain expert network; according to a hard routing selection result, inputting the high-dimensional feature vector into a corresponding expert network for processing to obtain an environment state classification result; and acquiring an underwater acoustic communication physical layer processing strategy matched with the communication environment according to an environment state sensing classification result, and acting on the current underwater acoustic communication frame.

This invention belongs to the field of character recognition technology and provides a handwritten letter recognition method based on piezoresistive signal detection, including the following steps: collecting the temporal piezoresistive signal during the letter writing process as raw data; preprocessing the raw data to obtain training data; constructing a recognition model, inputting the training data into the recognition model to extract temporal feature vectors, and performing classification training; using the trained recognition model for forward inference to confirm the letter category; this invention accurately captures changes in writing force and temporal logic based on piezoresistive electrical signals, significantly improving the accuracy of handwritten letter recognition.

This invention discloses a pseudo-anomaly enhancement method and related apparatus based on the SegAug visual segmentation model. The method segments a training set of normal images using the SegAug visual segmentation model to obtain component masks and background masks. Using the component and background masks of each training set sample, three types of pseudo-anomalies are generated through pasting, filling, and introducing scars. A multi-class classifier is constructed to promote self-supervised learning. Normal samples and pseudo-anomaly samples are mixed and fed into the training to extract deep feature information containing local texture features and global high-level semantic features. A local feature distribution for detecting local anomalies to identify structural anomalies is constructed, and a global feature distribution for capturing global information to identify logical anomalies is constructed. Local anomaly scores and global anomaly scores are calculated. Based on the local and global anomaly scores, an anomaly score is defined for the test image, completing the pseudo-anomaly enhancement. This effectively enhances the model's ability to detect global high-level semantic and logical anomalies.

The present application relates to the technical field of three-dimensional point cloud data processing and graph convolutional neural network, and particularly to a multi-domain adaptive graph convolution method and device based on visual selectivity. The present application takes the primate visual selectivity theory as the core basis, first constructs a multi-scale visual hypersphere neighborhood space, adaptively completes the polarization gridding of the neighborhood through clustering analysis, then designs a visual selectivity adaptive graph convolution operator to complete single-neighborhood feature extraction, constructs a deep network based on the visual hierarchical serial processing logic to realize high-level feature abstraction, and finally completes the point cloud classification task through multi-domain feature fusion. The present application solves the problems of information loss, topological structure damage, insufficient feature expression capability, and poor multi-scale feature adaptability in the existing 3D point cloud processing technology, and can be widely applied to 3D point cloud processing scenes in the fields of autonomous driving, augmented reality, robot perception, etc.

The invention provides a three-dimensional frequency spectrum prediction method, system and device based on a multi-scale map convolutional network and a medium, and belongs to the technical field of frequency spectrum prediction.The method comprises the steps that receiving power data of multiple frequency points are continuously collected with the resolution of preset duration, three-dimensional tensor data are formed, and a feature matrix is constructed; a learnable parameter matrix of a preset dimension is adopted to transform the feature matrix, a self-adaptive threshold value is set according to the median of scores of all nodes, and a weighted adjacency matrix is generated; performing first-order graph convolution operation and second-order graph convolution operation on the weighted adjacency matrix, and performing feature transformation by adopting a multi-dimensional learnable weight matrix to form multi-dimensional fusion features; and expanding the multi-dimensional fusion features according to a time sequence, constructing a gating mechanism through element multiplication and residual connection, and outputting a prediction result through a linear transformation layer. The accuracy of spectrum prediction is improved, and the model can output a more accurate prediction result when processing spectrum data of a long time sequence.

The invention provides an anion exchange membrane water electrolysis hydrogen production control method and system based on data-driven model predictive control, and the method comprises the steps: collecting operation data, and constructing a system non-mechanism model through a BP neural network algorithm; predicting output parameters of the AEM water electrolysis hydrogen production system according to the system non-mechanism model; inputting the output reference value y (t) into a system non-mechanism model and a controller; the x (t) is optimized through an optimization model constructed through an NSGA-II algorithm; the delta x (t) is applied to an AEM water electrolysis hydrogen production system, and actual output and predicted output are obtained; comparing to obtain an error value, and performing feedback correction on the controller; and updating the system non-mechanism model and the prediction model. According to the method, the non-mechanism model and the prediction model of the system are dynamically updated, the dynamic change of the AEM water electrolysis hydrogen production system can be reflected in real time, the control model is dynamically updated, and the precision of the control model is effectively improved.