36results about How to "Fast training" patented technology

The application relates to the technical field of artificial intelligence optical character recognition, in particular to an end-to-end communication big data journey card recognition method, and the steps of the method comprise the following steps: adjusting and correcting a text box according to size information of the text box, so that the size of the text box meets the required angle and proportion of adjustment; performing separated feature extraction on the corrected text box through a preset neural network to obtain feature data of a font; and obtaining text data corresponding to the feature data through a Bi-GRU recurrent neural network combined with an activation function.

The application discloses a data processing method, device and equipment and a computer readable storage medium. The method comprises the following steps: obtaining sample voice information and sample text information; processing the sample voice information through an acoustic model in an initial recognition model to obtain acoustic feature information; processing the acoustic feature information and the sample text information through a translation model in the initial recognition model to obtain a first predicted translation result and a second predicted translation result respectively; training the initial recognition model based on the first predicted translation result and the second predicted translation result to obtain a target recognition model; and the K vector and the V vector in the acoustic model are spliced with a prefix vector and / or the K vector and the V vector in the translation model are spliced with a prefix vector. The training efficiency of the target recognition model is improved.

The application provides an RGB-D image feature collaborative fusion method based on transfer learning and width learning, and comprises the following steps: obtaining an RGB-D data set, performing preliminary training through a neural network, and performing retraining in the data set after modifying the structure; after feature extraction, performing correlation analysis and fusion on RGB image features and depth image features; and using width learning to classify and identify the fused features. The application can reasonably fuse the features of RGB images and depth images, ensure that the feature information of color images and depth images can complement each other, improve the running speed of the system by using width learning, and finally make the classification result have higher accuracy and reliability.

The present application relates to the field of text semantic understanding, and more particularly to a text abnormal word recognition method and system based on a neural language model, which is used for recognizing network language, new words and wrong words, and the method comprises the following steps: step 1) collecting text to be recognized; step 2) determining the context word sequence of each word in the text to be recognized in a sliding window manner; step 3) inputting the context word sequence of each word into a pre-established and trained recognition model respectively to obtain the probability of occurrence of each word under the context word sequence; and step 4) comparing the probability of occurrence of each word with a set threshold value to determine whether the word is an abnormal word. The present application does not need to mark abnormal words and can be trained on a large amount of unsupervised data; the prediction probability threshold value of abnormal words and normal words determined by using a large number of positive and negative samples to predict probability statistics has universality; and the selection of negative samples is performed in a negative sampling manner, so that the training speed is accelerated.

An alloy performance prediction method fusing vision-language-process multi-modal data comprises the steps that a multi-modal data set of an alloy material is constructed, and the multi-modal data set comprises three kinds of modal input including process parameters such as temperature and time of solid solution and aging treatment, SEM image visual information and SEM image description text information and corresponding alloy mechanical property true values; training a visual encoder ResNet50 model through comparative learning and training a language encoder BERT model through mask language modeling to obtain an SEM image and vector codes of language description of the SEM image; splicing and fusing process parameters such as temperature and time of solid solution and aging treatment and the vector codes obtained in the second step, and training a random forest regression device according to corresponding alloy mechanical properties; and the random forest regression device obtained through training is used for alloy performance prediction. According to the method, the structured process data, the unstructured SEM image data and the derived text description data are subjected to collaborative fusion and joint modeling for the first time, complementarity among different modal data is fully utilized, and more comprehensive and more three-dimensional digital representation of the alloy state is constructed; the multi-modal fusion framework and the feature learning mechanism are suitable for wide material systems. Meanwhile, the adopted'depth representation + random forest 'hybrid model has the advantage of high training efficiency while ensuring high prediction precision.

The application provides a camera scheduling model training method, a violation event detection method and device. The camera scheduling model training method comprises: obtaining historical sample data of a camera; constructing a first sample probability distribution based on the historical sample data; sampling based on the first sample probability distribution to obtain first sample action data; and training a reinforcement learning model based on the first sample action data to obtain a camera scheduling model. The application uses historical sample data to construct a sample probability distribution, and then uses the sample probability distribution for sampling as the input of the reinforcement learning model to train the reinforcement learning model. Since the input of the reinforcement learning model is generated by the constructed sample probability distribution and does not truly respond to the environment, the problem of long time consumption for obtaining real samples is solved, a large number of samples can be quickly obtained for training, and the training efficiency of the camera scheduling model is improved.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training an action selection system using reinforcement learning techniques. In one aspect, a method includes, at each of a plurality of iterations: obtaining a batch of experiences, each experience tuple including: a first observation, an action, a second observation, and a reward; for each experience tuple, determining a state value for the second observation, including: processing the first observation using a policy neural network to generate an action score for each action in a set of possible actions; sampling a plurality of actions from the set of possible actions according to the action scores; processing the second observation using a Q neural network to generate a Q value for each sampled action; and determining the state value for the second observation; and using the state value to determine an update to current values of Q neural network parameters.

The invention discloses a base station energy-saving decision-making method for carrying out memory backtracking on DT based on state machine transfer, a state machine memory module with a backtracking function is designed to be used for a Decision Transformer to carry out long-term stable decision-making on base station energy saving, and the method comprises the following steps: recording past information such as states, actions and environment rewards, and arranging the information into a time sequence according to a specific sequence; accumulating the environment rewards in the trajectory, and calculating an accumulated expected return (return-to-go) in the trajectory; the method comprises the following steps: constructing a decision model, and taking a Decision Transform as a main body model; designing a state machine memory module and a state machine backtracking module; determining a loss function and training a decision model, obtaining a model meeting requirements as the decision model by adjusting a learning rate, an optimizer and regularization parameters, and storing the decision model; and inputting environment historical data, and outputting decision actions by the large model. According to the method, the problem that a traditional offline reinforcement learning method Decision Transform suddenly loses efficacy in a long-term decision that the reasoning track length exceeds the training sample track length is solved, and the stability of the energy-saving decision of the base station is improved.

The intelligent automobile lane-changing decision method and device, the electronic device and the storage medium provided by the embodiments of the present disclosure comprise: constructing a virtual lane-changing scene by using a simulation method; constructing a decision model for determining a mapping function between a state and a reward, obtaining a Q table for evaluating the lane-changing feasibility of an intelligent automobile based on the mapping function, and obtaining a lane-changing decision, i.e., "lane-changing" or "not lane-changing"; based on a greedy algorithm, making the intelligent automobile produce a random lane-changing behavior in the constructed virtual lane-changing scene, storing the state and the reward of the current lane-changing behavior in an experience replay pool, when the number of state-reward pairs in the experience replay pool reaches a minimum sample number, randomly sampling a plurality of state-reward pairs from the experience replay pool for single-step training of the decision model, storing a new state-reward pair generated in the training process in the experience replay pool, repeatedly performing the above training process until a maximum training number is reached, and obtaining a lane-changing decision model. The lane-changing decision generated by the present disclosure is high in safety and wide in applicability.

This invention discloses a substation acoustic signature fault diagnosis method based on LSTM-BHPSO. The method involves collecting substation acoustic signature signals and extracting features using MFCC to obtain a feature dataset, which is then divided into training and testing sets. An LSTM-based substation acoustic signature fault diagnosis model is constructed. The optimal initial parameters are obtained by iteratively optimizing the particle positions using the number of feature points extracted by MFCC, the number of batch samples, the number of network layers, and the number of hidden and fully connected layers. These optimal initial parameters are then input into the LSTM-based substation acoustic signature fault diagnosis model. Finally, the trained LSTM-based substation acoustic signature fault diagnosis model is used to diagnose substation acoustic signature faults. This invention offers good processing efficiency, avoids local optima, and facilitates accurate fault diagnosis results.

The embodiment of the invention provides a reliability detection and model automatic training method for a sealing structure of a wellhead. The method comprises the steps of obtaining a size data set of the sealing structure; the size data set comprises a plurality of size data; the size data represents information of a physical structure of the sealing structure; performing matching processing on the size data set and a data set in a preset model library to obtain multiple groups of matched data sets; determining an initial model corresponding to the matched data set based on a preset model library; based on a Bayesian optimization algorithm, processing the size data set and each initial model to obtain a reliability detection model; the reliability detection model is used for processing the size data of the sealing structure to obtain a reliability detection result of the sealing structure. The method is used for achieving the effect of accurately and effectively detecting the reliability of the sealing structure.

The invention discloses a retaining ring gluing image enhancement detection method based on a CycleGAN model. Comprising the steps of collecting an original gluing image; carrying out VGG16 neural network identification and classification; enhancing the image by using the improved CycleGAN network model; and in the gluing area, the gluing state is judged. According to the method, the low-contrast part of the glue area can be effectively enhanced, the detection precision is improved, the edge information of the glue image is enriched, and a more accurate gluing state image can be obtained. The final detection success rate of the image obtained through the method can reach 95%, more accurate information support can be provided for the gluing process of the industrial robot, the working stability between the retaining ring and the polishing head and between the retaining ring and the wafer is guaranteed, the yield in the chip manufacturing process is improved, and the method has actual production significance.

This invention discloses a method and apparatus for identifying road changes using remote sensing imagery. The method includes: acquiring new and old remote sensing images of unlabeled roads; cropping the remote sensing images; comparing the cropped road images; inputting the compared road change images into a deep learning model for self-supervised training to obtain a pre-trained model; training the pre-trained model using labeled road remote sensing images to obtain a classification and recognition model; and inputting the desired road change images into the classification and recognition model by transferring the parameters of the self-supervised model to obtain the road change results. The method provided by this invention can intelligently identify road changes by eliminating the need for manual labeling, achieving high efficiency, and reducing costs, thus enabling efficient monitoring and dynamic supervision of road changes. This provides technical and data support for promoting traffic road maintenance and optimization, and implementing road hazard detection and remediation.

This invention addresses the technical problems of existing unsupervised change detection algorithms, which mainly rely on algebraic transformations and spectral unmixing of spectra, resulting in poor performance and limited applicability in hyperspectral data with mixed pixels. It provides a hyperspectral image change detection method based on an unsupervised spectral unmixing neural network. The method includes the following steps: 1. Segmenting and mixing two hyperspectral images from different time phases to obtain hyperspectral image patch data, which is used as the training dataset; 2. Constructing a spectral unmixing network to generate endmember matrices and abundance matrices with image features; 3. Reconstructing the hyperspectral image patch data based on the endmember matrices and abundance matrices, and selecting reconstruction errors to train the spectral unmixing network; 4. Generating a change grayscale image based on the trained spectral unmixing network and the abundance matrix output by the abundance matrix generation module; 5. Binarizing the change grayscale image using a threshold segmentation algorithm to obtain the final hyperspectral image change detection result.

This invention discloses a salient object detection method based on a multi-scale dilated convolutional neural network. The method includes: extracting multi-scale features from the input image; inputting the multi-scale features into a dilated residual convolutional module to obtain fused features including contextual information of the multi-scale features; inputting the fused features into multiple channel attention modules to obtain multiple salient features; performing dimensionality reduction activation on each salient feature to generate a saliency map; and performing deep supervised training using a hybrid loss function that combines cross-entropy and cross-union loss. The method of this invention, based on a multi-scale dilated convolutional neural network, fully captures rich global and local semantic information in the image by using a dilated residual convolutional module, solving the problem of shallow encoder depth and insufficient information extraction. Simultaneously, the designed channel attention modules enable the network to focus on the target region, effectively improving the accuracy of object detection.

The invention provides a blasting vibration peak velocity prediction method fusing parameter uncertainty and data-driven optimization, which comprises the following steps of: firstly, acquiring data such as blasting parameters, lithologic indexes, geological conditions and actually measured vibration peak velocity (PPV), and establishing a basic database; a prior probability model is constructed, a joint uncertainty model is established in combination with probability disturbance and a fuzzy triangular number, a multi-dimensional disturbance sample is generated through a joint central value and a joint standard deviation and is fused with original data, and an extended database is formed. Feature analysis is carried out on the fused data, and input variables which have obvious influence on the vibration peak velocity (PPV) are screened; and constructing a BP neural network model based on the screened features, carrying out global optimization on a network weight and a threshold by adopting a PSO algorithm, and then carrying out local fine tuning by utilizing Adam. Finally, through training and verification, prediction of the vibration peak velocity (PPV) is realized, and model precision is evaluated through RMSE, MAE, MAPE, Rand other indexes. The influence of rock and soil parameter uncertainty on prediction precision can be effectively processed, and the reliability and applicability of blasting vibration prediction are improved.

This disclosure relates to batch normalization layers. The present invention provides methods, systems, and apparatus for processing input using a neural network system including a batch normalization layer, comprising a computer program encoded on a computer storage medium. One method includes: receiving a corresponding first layer output for each training example in the batch; calculating a plurality of normalization statistics for the batch based on the first layer outputs; normalizing each component of each first layer output using the normalization statistics to generate a corresponding normalized layer output for each training example in the batch; generating a corresponding batch normalized layer output for each training example from the normalized layer outputs; and providing the batch normalized layer outputs as input to a second neural network layer.

The invention discloses a fetal distress auxiliary diagnosis system based on continuous wavelet transform and a shallow convolutional neural network, and the system comprises a data obtaining and preprocessing module which is responsible for obtaining an original fetal heart rate signal and carrying out the data preprocessing of the original fetal heart rate signal; the time-frequency characteristic graph construction module is responsible for converting the signals processed by the data acquisition and preprocessing module from one-dimensional data into a two-dimensional time-frequency characteristic graph by using continuous wavelet transform; and the classification and recognition module is responsible for inputting the time-frequency characteristic pattern into a shallow convolutional neural network model and outputting a fetal distress state classification result. According to the method, a one-dimensional time sequence signal is converted into a two-dimensional time-frequency image, and joint distribution information of the signal in a time domain and a frequency domain is subjected to visual coding, so that a shallow convolutional neural network model can fully utilize the advantages of the shallow convolutional neural network model in image processing, deep pathological features hidden in an FHR signal are mined, and the deep pathological features hidden in the FHR signal are extracted. And the defect that the one-dimensional convolution is insensitive to the time sequence is overcome.

The application belongs to the technical field of neural network model training, and discloses a neural network model training method and system based on an improved random configuration algorithm, which comprises the following steps: under the condition of reducing the root mean square error of the neural network model output, screening suitable neurons as candidate hidden layer nodes through the inequality constraint condition of the random configuration algorithm; screening K neurons with the fastest training error reduction from the candidate hidden layer nodes, and selecting the neuron least related to the previous L-1 hidden layer nodes from the selected K neurons as the optimal hidden layer node; calculating the output weight through the least square method, and updating the structure of the random configuration network; and judging whether the network structure is completed by using the maximum allowable number of hidden layer nodes and the maximum allowable output error. The application can reduce the overall calculation amount of the algorithm while ensuring the prediction accuracy, and is suitable for application scenarios with high real-time requirements.

The application discloses a natural language to SQL conversion method based on a deep learning model, which is based on a deep learning model, wherein a natural language question of a user is predicted through a deep model, and a result of the prediction is converted into an SQL statement through a CYK algorithm, and the method comprises the following steps: step 1, preparing a corpus; step 2, establishing a table model; and step 3, performing SQL conversion through the CYK algorithm; the method is convenient and direct in converting human natural language into an SQL query statement and the like, and then the SQL statement can be directly used to obtain data that a user wants, the method simplifies a data query mode, the data that a user wants is inquired in a speaking mode, and the intelligent degree of a data acquisition mode is improved.

The application discloses an aircraft intelligent many-to-many game target allocation method, an electronic device and a storage medium, and belongs to the technical field of aircraft control. In order to realize the many-to-many fast game target allocation of an aircraft, an aircraft game confrontation relative motion model is established; a training environment of the intelligent many-to-many game confrontation of the aircraft, a state space of a plurality of intelligent agents and an action space of the intelligent agents are designed; observation values of the aircraft in the training environment are collected as the state space of the plurality of intelligent agents, input numbers of the aircraft are used as the action space of the intelligent agents, fuel consumption of the aircraft and reward values obtained by the aircraft are calculated; an OW-QMIX intelligent agent structure is constructed, then the state space of the plurality of intelligent agents, the action space of the intelligent agents and the reward values obtained are input into the OW-QMIX intelligent agent structure, and an aircraft intelligent many-to-many game target allocation strategy training result is output, and then simulation verification is carried out. The application realizes the many-to-many fast game target allocation of the aircraft.

The application discloses a target detection enhancement method and device based on a generative adversarial architecture and a storage medium, and relates to the technical field of target detection in computer vision. The application proposes a new target detection framework, which performs adversarial training on a target detection network that is difficult to further improve performance, can further improve the performance of the trained target detection network, does not increase parameters, can quickly and effectively improve the trained target detection network, has fast training speed, consumes few computing resources and is efficient, does not increase inference time or training difficulty as a cost, and is almost a plug-and-play training mode.

The present application relates to artificial intelligence technology, disclose a kind of voice conversion method, comprising: arbitrarily selecting one training data in training data set, obtain target training data;Respectively extract the voice feature of training voice and conversion speech label in target training data;Using the voice feature of the extracted training voice in conversion model is converted, and conversion voice feature is obtained;Conversion model is judged whether to converge based on the difference between conversion voice feature and the voice feature of extracted conversion speech label;When conversion model does not converge, the model parameters of conversion model are updated by second-order gradient self-adapting adjustment, and return arbitrarily selecting one training data step;When conversion model converges, the voice conversion of the voice to be converted is carried out using the conversion model at this time, and target conversion voice is obtained.The present application also proposes a kind of voice conversion device, equipment and medium, which can be used in financial field, and improves the efficiency of voice conversion of insurance notice matters for attention explanation voice.

This invention discloses an intelligent detection method for product surface anomalies based on machine vision, belonging to the fields of machine vision and anomaly detection technology. The method comprises the following steps: First, image data of the product surface is collected and filtered, comprehensively covering the surface of the product under test and dividing it into regions; the training set consists of normal surface images. Second, a suitable neural network structure is selected, and a loss function is constructed by combining MSE, Cosine, and SSIM; the model parameters are trained and saved. The anomaly heatmap of the verification image is compared with the heatmap of the normal image to determine the optimal threshold. Finally, preprocessing, network computation, and post-processing operations are performed on the test image to determine whether there are anomalies and their locations. This invention uses only surface images of normal products to train the neural network, eliminating the need to pre-collect images of various anomaly surfaces as training data, making it suitable for intelligent detection of various possible defects on the surface of products on assembly lines.

The invention discloses construction of a petroleum yield prediction model for domain knowledge migration in a multi-reservoir scene, and belongs to the technical field of petroleum engineering and intelligent prediction. According to the method, a domain knowledge migration mechanism is introduced, so that the model can share depth feature information among multiple oil reservoirs, the prediction precision and model robustness of small sample oil reservoirs are effectively improved, and the problems of poor model generalization, high training cost, unstable prediction and the like in a traditional method are solved. The method is clear in structure and moderate in calculation amount, has good interpretability and visualization ability, and can be widely applied to the fields of oil reservoir yield prediction, intelligent oil field management, production optimization decision making and the like.

The utility model discloses a quick mounting mechanism for an aero seat ground rail, which relates to the related field of aero seat ground rails, and comprises a bottom plate, and a nut is welded at the lower end of the bottom plate; the fixing sleeve is welded to the upper end of the bottom plate; the clamping blocks are movably arranged at the side opening of the fixing sleeve; the pressing shaft is in sliding fit with the guide hole of the fixing sleeve, and the outer conical surface of the pressing shaft is in contact with the side conical surface of the clamping block; the ground rail is arranged on the periphery of the main body of the fixing sleeve in a sleeving manner; the bolt penetrates through the pressing shaft, the fixing sleeve and the bottom plate and is in threaded connection with the nut; the bolt is screwed down, the bolt drives the pressing shaft to move downwards in the axial direction, the outer conical face pushes the side conical face to enable the clamping block to move outwards in the radial direction, the ground rail is clamped, and therefore the ground rail and the fixing sleeve are connected. The installation process is simple and fast, the installation time is greatly shortened, and the installation efficiency of the aero seat ground rail is effectively improved.

A system for selecting an action to be performed by an agent is provided, comprising: a worker neural network system configured to, at each of a plurality of time steps: receive a goal representation defining a goal to be accomplished as a result of an action performed by an agent in an environment, wherein the goal representation is based at least in part on an environment state of the time step and / or one or more previous time steps; and generate, based at least in part on the goal representation and the environment state of the time step, an action score for each action in a set of actions; and an action selection subsystem configured to, at each of the plurality of time steps: select, using the action scores, an action from the set of actions to be performed by the agent at the time step.

The invention is suitable for the technical field of power systems, and provides a power system fault classification method, device and equipment and a program product, and the method comprises the steps: obtaining electrical quantity time sequence data collected by power system monitoring equipment; performing multi-level signal decomposition on the electrical quantity time sequence data to obtain a plurality of signal components; extracting statistical features of the plurality of signal components to construct feature vectors to be classified; and inputting the feature vectors to be classified into a breadth feature mapping classification model trained in advance, and outputting a fault classification result, so that complex fault types can be effectively distinguished, and the classification accuracy and fineness are improved.

This utility model relates to an automatic cleaning device for large storage tanks, comprising: a base, a base support system, a pole system, a rotating seat, a rotating seat support system, an adjustment system, a power system, and a cleaning system. This device offers advantages such as convenient operation, high automation, strong safety, lightweight design, wide applicability, low operating costs and labor intensity, short construction period, high tank utilization rate, good cleaning effect, flexible cleaning effect acceptance methods, and no risk of operator poisoning. It is suitable for large storage tanks of various sizes and can effectively solve various problems in existing tank cleaning technologies.

The application discloses a micro-grid group scheduling optimization method and device based on a DDQN algorithm and an action mask, and the method comprises the following steps: regarding each micro-grid as an intelligent agent to construct a micro-grid group multi-agent model; the micro-grid group multi-agent model takes maximizing total operation income of the micro-grid group as a target, and considers power balance constraints, output constraints, ESS constraints and interactive power constraints; Markov modeling is performed according to the micro-grid group multi-agent model to obtain a Markov decision process; the micro-grid group is scheduled and optimized according to the Markov decision process in combination with a first algorithm to obtain a micro-grid group joint scheduling scheme; the first algorithm is based on the DDQN algorithm, and illegal actions are shielded through the action mask at each time step of the DDQN algorithm.