75results about How to "Reduce training time" patented technology

The invention discloses a dynamic space-time diagram flow prediction method and system based on course learning, and relates to the technical field of supply chain logistics data analysis, and the method comprises the steps: building a space-time matrix based on historical multi-source data, generating a dynamic adjacent matrix through learning, and carrying out the smooth fusion through combining a static diagram, and forming a dynamic diagram structure. And then space and time features are respectively extracted by using a graph convolutional network and a gating loop unit, and deep interaction and fusion are realized through a bidirectional cross attention mechanism. A multi-dimensional difficulty estimator is innovatively introduced, the prediction difficulty of each training sample is quantified from three dimensions of space, time and time-space coupling, the selection sequence of the training samples is dynamically adjusted based on an adaptive course scheduler, and progressive learning is realized. And finally, feature representation is obtained through global pooling, and multi-step traffic prediction is realized by adopting a parallel independent decoder, so that error accumulation is avoided. According to the invention, prediction precision and model training efficiency in a complex supply chain logistics scene are effectively improved.

The application discloses a continuous acquisition and data model construction system and method of an intelligent sensor. In order to overcome the problems that the training data acquisition of the chemical sensor in the prior art is inconsistent with the actual application, the data acquisition efficiency is low, and effective data is insufficient, in the training process, the sensor unit is placed in the test unit, and the concentration control unit is used to adjust the concentration of different samples, so that the data is continuously acquired. In the prediction process, the sensor unit is placed in the sample input unit, is input into the concentration prediction model after correction by the state calibration unit, and the sample concentration output is obtained. The continuous data sampling method is used, the recovery process is not performed, and the balance state is reached in a relatively short time, so that the training time is greatly reduced; the response and recovery stage data are continuously acquired, so that the training data is consistent with the signal and concentration change process in the actual application; in the prediction process, the state of the sensor is dynamically detected and different models are called, so that the model accuracy is improved.

The application provides a similarity feature constraint-based inverse synthesis prediction method, medium and equipment; wherein the method is: converting a to-be-predicted product into a corresponding product SMILES sequence; using a graph neural network to extract graph structure features corresponding to all characters in the product SMILES sequence; fusing the graph structure features and word vector representations to obtain a fusion vector; inputting the fusion vector matrix into a Transformer encoder and decoder to obtain a reactant set SMILES sequence, and converting the reactant set SMILES sequence to obtain corresponding reactants. The method integrates the graph structure features of the product molecule into the prediction features, improves the prediction accuracy effect, learns the commonality between the reactant and product sets through feature constraint, uses the commonality features to better help the Transformer model to perform inverse synthesis prediction, and further improves the prediction accuracy of the model.

This invention provides an energy efficiency optimization method for the large-scale deployment of ultra-reliable low-latency (URLLC) devices, comprising: Step 1: Calculating the short packet collision probability and short packet reception failure probability of a single device, and calculating the packet loss rate of the device based on the calculated short packet collision probability and short packet reception failure probability; Step 2: Calculating the packet loss rate of each device, and grouping devices with similar communication environments into one category based on the calculated packet loss rate; Step 3: Optimizing the energy efficiency of all devices categorized in Step 2 using a deep reinforcement learning method. The beneficial effects of this invention are: 1. The method can effectively reduce the transmission power of sensing devices sending short packets and the number of times the same short packets are repeatedly sent, significantly improving the communication energy efficiency of URLLC devices and extending their battery life; 2. The method of this invention uses a classification method, which significantly shortens the training time of deep reinforcement learning.

The invention relates to a circuit board defect detection method, device and equipment based on multiple sub-models and a storage medium, and is applied to the field of defect detection, and the method comprises the steps: obtaining an original image of a to-be-detected circuit board, aligning the original image according to a preset mother layout, and setting the aligned original image as a to-be-detected panoramic image; segmenting the panoramic image to be detected into a plurality of unit detection images according to a preset segmentation standard; matching a special sub-model from a preset detection model library according to the unit detection image, and performing defect detection on the unit detection image by adopting the special sub-model; and integrating the defect detection results of the unit detection images, and outputting the defect detection result of the circuit board to be detected. The circuit board defect detection method has the technical effect of improving the efficiency and accuracy of circuit board defect detection.

The application discloses a deep learning-based artillery barrel target image defogging method and system, which comprises the following steps: inputting an original artillery barrel target image into a trained defogging enhancement network model for processing, wherein the defogging enhancement network model comprises an improved U-Net module, an AOD-Net module and an atmospheric scattering model layer connected in sequence; the improved U-Net module is used to extract multi-scale features of the original artillery barrel target image, and output the multi-scale feature maps to the AOD-Net module; the AOD-Net module is used to extract features under different receptive fields of the multi-scale feature maps through different convolution layers, realize multi-scale feature fusion, and output input values of the atmospheric scattering model layer; and the atmospheric scattering model layer is used to obtain the defogged artillery barrel target image based on the input values. The model is constructed by connecting the U-Net and the AOD-Net in series, a new model optimizer is proposed, the algorithm of gradient descent is optimized, and a self-adaptive weight loss function is introduced, so that efficient defogging processing under complex fog and haze conditions is realized.

The application provides a photonic neural network system based on a multi-task neural network, which comprises: an optical computing module, taking a photonic artificial intelligence chip as a core, modulating an optical signal, and completing multiplication, summation and nonlinear operation of high-precision and high-speed optical analog quantities; a multi-task neural network module, having a main branch neural network and multiple secondary branch neural networks, and capable of simultaneously processing multiple classification tasks and regression tasks; and an optical computing communication and control module based on FPGA, connected with the optical computing module and the multi-task neural network module, and used for realizing real-time and high-speed communication of the optical computing module and the multi-task neural network module. The system has the characteristics of high bandwidth and low energy consumption, utilizes high-dimension parallel computing to greatly reduce the time of original neural network serial operation, has the advantages of large network scale and fast parallel operation speed, and can improve recognition accuracy and shorten training time on the basis of a single task.

The invention provides a system architecture for cross-data center large model training and a resource allocation method and system, and belongs to the technical field of cross-wide area distributed large model training. According to the method, large-scale model cooperative training across multiple data centers can be realized, and the bottleneck that the computing power of a single data center is limited is broken through. Through unified modeling and scheduling of calculation, memory and network resources, task loads can be intelligently allocated according to hardware performance and network bandwidth of different data centers, and efficient collaborative utilization of computing power resources is realized. The training task of the super-large-scale model can be rapidly completed in the heterogeneous computing power environment, and the training time is remarkably shortened. The provided flexible parallelism degree allocation method can be adaptive to different task and resource conditions, the proportion of data parallelism, model parallelism and pipeline parallelism is automatically adjusted, the parallelism efficiency is improved, and the communication overhead is reduced. A training time estimation function is integrated, the overall time delay and resource requirements can be predicted before task execution, and a basis is provided for scheduling decision making.

The application discloses a kind of multi-variable DC fault arc adaptive detection method based on similarity measure and transfer learning, to the system output current of different time period applies wavelet packet decomposition and coefficient reconstruction processing after extraction feature, and input to state identification model to carry out multi-period fault arc judgment, to current working condition through machine learning evaluation feature similarity measure value D and model output probability distribution P joint distribution mode, according to this, the extracted feature is distinguished into unidentifiable working condition type data and identifiable working condition type data, using field adaptive strategy to the simplest feature group of unidentifiable working condition type data carries out transfer learning, to update state identification model, to realize real-time accurate detection of DC fault arc under complex and variable working condition in this way.

This solution discloses a new method for identifying RNA pseudouridine sites. This method proposes to use a variety of feature representation techniques to extract sequence features, and then uses the SVM-RFE method for feature selection to compress the feature space and optimize the feature subset. The best feature set after feature selection is input into the kernel method KeMRF based on polynomial random forest to identify pseudouridine sites in the sequence. As a newly proposed classification method, compared with the traditional random forest, KeMRF not only optimizes the discriminant criterion for node splitting, but also combines with an easy-to-interpret kernel method, making the classification performance more superior. This method reduces the training time of the model, improves the classification performance of the model, and further enhances the accuracy of identifying pseudouridine sites.

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 application discloses a terminal equipment-based signed receipt information identification method and device, medium and equipment. The method comprises the following steps: obtaining a signed receipt image, identifying the current angle of the signed receipt image through an SVM network, and rotating the signed receipt to a standard angle according to the current angle; performing verification element detection on the signed receipt image by using a YOLO network, determining the bounding box position and category of the verification element, and identifying whether the bounding box position and category of each verification element in the signed receipt image match; if the bounding box position and category match, each verification element is cropped according to the bounding box position; the verification element is input into a text recognition model for text recognition; and if the text recognition result matches the signed receipt information recorded in advance, it is determined that the information identification of the signed receipt is completed. According to the technical scheme, main verification matching processing can be performed on the mobile equipment side, and in the case that the verification is unqualified, the mobile equipment needs to inform the business party of the unqualified verification element and the reason for the unqualification.

The present application belongs to the technical field of robot control, and particularly relates to a quadruped robot three-foot walking gait motion control method, steps of which comprise: in a single-leg damage scene, acquiring gait information of three-foot walking of the quadruped robot; designing a reference foot end position generator to generate a reference foot end position, and performing real-time optimization on the reference foot end position through a teacher strategy network; controlling the quadruped robot based on the optimized foot end reference position, and performing interactive training on the teacher strategy network by using motion data of the quadruped robot under different three-foot walking gaits; training a student strategy network by using a supervised learning method to approximate the teacher strategy network; and loading the student strategy network on a physical robot to output corresponding joint control commands of the quadruped robot. Compared with the prior art, the present application can realize stable three-foot motion of the robot under any single-leg damage condition, and provides key technical support for three-foot motion control of the quadruped robot under special working conditions.

The application provides an antenna array surface fiber grating layout method based on a genetic algorithm and a neural network, and relates to the technical field of deformation sensing of an antenna array surface structure and optimization of fiber grating layout. The method analyzes strain sensitive positions of the array surface structure through the genetic algorithm and the neural network, arranges fiber grating sensors at optimal layout positions of the array surface structure, and realizes high-precision deformation sensing of the array surface structure under complex service loads. The method solves the problems of node layout redundancy, high data processing cost and insufficient displacement prediction accuracy in traditional antenna monitoring, and provides technical support for efficient sensing of antenna structure deformation and stable guarantee of electromagnetic performance.

The application discloses a power transformer fault sound print detection method based on an extreme learning machine, first, audio signals of transformer faults are collected for framing, and preprocessed through fast Fourier transform; then, sound print characteristic parameters of the audio signals are extracted through an improved Mel filter, a fault sound print detection model based on the extreme learning machine is constructed, and model training is conducted; finally, a test set is established, real-time test samples are input into the trained fault sound print detection model, test identification is conducted, sound print identification results of transformer faults are obtained, and real-time fault sound print early warning is realized. The improved Mel filter can improve the spectral resolution of transformer sound print feature extraction, reduce the system complexity, the fault sound print detection model is constructed through the extreme learning machine algorithm, the transformer fault detection accuracy is effectively improved, and the method is easy to implement in engineering.

The invention discloses an elevator intelligent maintenance method and device based on DGConv and an improved CBAM attention mechanism, and the method comprises the steps: firstly obtaining fault-related continuous data and discrete data based on an elevator mainboard, carrying out the data length alignment of the continuous data through dynamic time warping, and combining with the discrete data to construct an elevator fault prediction data set; then, constructing a CNN-Transform hybrid neural network of cavity global convolution and an improved CBAM attention mechanism for predicting the occurrence probability of various faults at the next moment, and performing training based on an elevator fault prediction data set; and finally, fault maintenance priorities are allocated to different faults according to the fault probabilities, and a targeted intelligent maintenance strategy is generated. The method provided by the invention has a remarkable effect in intelligent maintenance of the elevator, can greatly reduce the manpower and material resource overhead of a traditional maintenance method, saves the maintenance time, and optimizes the maintenance strategy.

The application discloses a kind of online model parameter identification methods of servo driver based on AI algorithm, steps are as follows: S1. the parameter identification network of being served object is constructed, and current state and target value are input, and the parameter and the predicted state of next time of identification are output;S2. sampling strategy and parameter identification network iterative training are carried out;S3. parameter identification is carried out after the data of true machine sampling is optimized to network.The identification method uses AI algorithm for the system identification of servo motor can solve the problems such as low identification precision and low efficiency, to promote the application in industrial field.

The application discloses a kind of associated multi-intersection traffic flow prediction methods.The steps of the present application are as follows:1, collect the traffic flow data of associated intersections, and divide the intersection traffic data into training set, validation set and test set after preprocessing;2, use CNN to extract spatial features from the input intersection traffic data;3, use Transformer to extract time features by inputting spatial features;4, after the Decoder layer is fully executed, the three time window data are finally input into three vectors, and the three vectors are stacked and input into the average pooling layer;5, set the model parameters;6, train the model until the maximum training period, and use the final model to predict the traffic flow of associated multi-intersections.The application uses CNN and Transformer to extract the spatial and temporal features of associated multi-intersections.Learning time encoding is used to embed the position encoding of Transformer, which injects position information and time information into the model together, helping the model to better learn the time features of traffic volume.

The application discloses a visual rich-text page restoration method based on an attention network, which comprises the following steps: inputting a document image into a trained document image rotation angle classification network to obtain the angle at which the document image needs to be rotated and correct the document image, and then using a projection correction algorithm to perform secondary correction on the document image; performing element detection on the document image through a page analysis network and performing corresponding processing to obtain element content corresponding to each element; using a plurality of deep learning algorithm network models connected in series to restore the document image into an editable document; and adding a contrast denoising training strategy in the page analysis network training to improve the accuracy of page analysis, because the plurality of network models are connected in series, according to the rule of the weakest link, the effect of page restoration will be affected by the page analysis model, therefore, the addition of the contrast denoising strategy can improve the accuracy of the network, promote the convergence of the network and shorten the training time.

The application relates to a video instance segmentation method based on a dynamic convolution decomposition of a lightweight attention mechanism; the method first inputs a video frame, and a backbone network independently extracts a feature map of each frame in the video; then the feature map extracted by the backbone network enters an HQT encoding and decoding module, the position change of an instance in each frame is accurately positioned by combining an output head, and three prediction branches are used to supervise model training; in the application, a convolution layer of an encoding network adopts dynamic convolution decomposition, a more compact model is obtained, model training is easier, the required parameter quantity is greatly reduced, the training speed is improved, and the training time is shortened; a lightweight HQT encoding and decoding module is provided, which helps to reduce model parameters and improve efficiency; the loss function is improved, model training is more stable, the convergence speed and convergence precision are improved, and the problems of imbalance between foreground and background in samples and imbalance between foreground categories under a long tail condition are relieved.

The invention discloses a small-diameter tube welding defect detection method based on space guide feature selection. A used model comprises a backbone network, a pixel decoder and a query decoder. A multi-scale feature map is extracted from an input image through a backbone network, the multi-scale feature map is fused from top to bottom and from bottom to top in a pixel decoder, and a large-scale pixel decoding feature map, a secondary large-scale pixel decoding feature map, a secondary small-scale pixel decoding feature map and a small-scale pixel decoding feature map are obtained; the object query vector interacts with the small-scale pixel decoding feature map, the secondary small-scale pixel decoding feature map and the secondary large-scale pixel decoding feature map in sequence in a query decoder to obtain a prototype query vector; and interacting the prototype query vector with the foreground-enhanced large-scale pixel decoding feature map to obtain a welding defect detection result. Aiming at the characteristics of non-uniform gray distribution, smooth edge and the like of an X-ray small-diameter tube welding image, the backbone network extracts features from multiple angles, and the detection precision is improved.

This application discloses a method for detecting lateral impacts in robots based on an impact-sensing neural network, comprising: Step A, controlling the robot's sensing device to collect raw pose data and extracting a sequence of measurement values ​​to be sensed from the raw pose data; Step B, classifying the sequence of measurement values ​​to be sensed using a trained impact-sensing neural network to obtain a first target prediction probability and a second target prediction probability; Step C, determining whether the first target prediction probability is greater than the second target prediction probability. If so, it is determined that the robot has detected a lateral impact; otherwise, it is determined that the robot has detected no lateral impact. This method avoids stopping the use of the neural network for classification and recognition after validating the raw pose data collected in real time by the sensing device, and also avoids starting the neural network for classification and recognition only after performing Kalman filtering on the raw pose data collected in real time by the sensing device, thus improving the efficiency of the robot in detecting lateral impacts.

The invention relates to the field of satellite frequency orbit resources, in particular to a frequency orbit declaration coordinated question-answering system and method based on a large model. The system comprises a data processing module used for obtaining and preprocessing original data and constructing a pre-training data set and a question and answer pair data set; the model training optimization module is used for taking the pre-training data set as an original data set, taking the question and answer pair data set as a target data set, and generating a training data set through topic distribution modeling and text importance evaluation in combination with a sampling screening strategy; the big model pre-training module is also used for optimizing the training process of the big model based on data screening of training loss and a dynamic batch updating strategy so as to complete big model pre-training; and the model fine tuning deployment module is used for performing fine tuning on the pre-trained large model by adopting the question and answer pair data set, and deploying the fine-tuned large model to realize a frequency orbit declaration coordinated question and answer function. According to the invention, the efficiency and accuracy of frequency track declaration coordination can be improved.

This invention proposes a high- and low-voltage line strong discharge monitoring system and method based on video image recognition, relating to the field of strong discharge monitoring. It includes a camera for acquiring video streams of high- and low-voltage lines and sending them to an image analyzer; the image analyzer extracts keyframe images from the video streams, extracts static and dynamic spark features based on the keyframe images, and inputs a feature vector composed of these features into an extreme learning machine for spark recognition, generating a discharge warning signal, which is then sent to a remote server; the remote server receives and displays the discharge warning signal. This invention improves the accuracy of video spark image recognition and can quickly and accurately determine the discharge location, ensuring the safe and stable operation of the power grid and reducing economic losses caused by transmission line discharge faults, thus having significant social and economic implications.

This invention relates to the field of scene text segmentation, specifically a scene text segmentation method based on an improved SAM visual segmentation large model. Based on the SAM visual large model, this invention extracts text content perception features through an image content perception module and text edge perception features through a text edge perception module. Furthermore, the text feature fusion module extracts and calculates text edge perception feature maps, which are then added to the vectors requiring attention calculation before each self-attention calculation in the SAM encoder. This improves the accuracy of SAM in text segmentation and shortens the model training time while maintaining generalization.

The application discloses a CNN-LSTM-based phase calibration method for mode division multiplexing optical fiber communication and belongs to the technical field of optical fiber communication. By constructing a double-input CNN-LSTM hybrid neural network architecture, MIMO input signals and original target signals of mode division multiplexing optical fiber communication are acquired, data preprocessing is carried out, the spatial feature extraction capability of a convolutional neural network and the time series modeling capability of a long short-term memory network are combined, effective compensation of phase noise in long-distance strong coupling mode division multiplexing optical fiber communication is realized, complex MIMO signals are separated into real and imaginary parts and sliding window training samples are generated, and a double-branch network containing an image input layer, a convolutional layer, an LSTM layer, a deep connection layer and a self-defined phase calibration layer is constructed. The problems that the phase noise compensation capability of a traditional algorithm is limited and the convergence performance is poor under a strong coupling condition are effectively solved, and the equalization performance and transmission quality of a mode division multiplexing optical fiber communication system are significantly improved.

This utility model discloses a drilling device for automotive stabilizer bars, relating to the technical field of automotive stabilizer bar processing equipment. It includes a drilling device, a first support block, and a second support block. A fixed frame slides on the top of the first support block, and a guide block is fixed at the center of the top of the fixed frame. Horizontally mounted top plates are fixed to the side walls of the top of the fixed frame at both ends, close to each other. An inclined surface is provided on the top of the guide block near each of the two top plates. A slider slides vertically on one side wall of the second support block. In this utility model, a flipping mechanism automatically flips the stabilizer bar after drilling at one end, allowing drilling at the other end. This eliminates the need for manual flipping of the stabilizer bar, avoiding manual alignment of the drilling position. Furthermore, during drilling, the device can press down and fix the stabilizer bar, thereby reducing deviation or vibration during drilling, ensuring drilling accuracy and stability, and improving product quality.

The application relates to the field of wind power prediction, and discloses a wind power prediction method and system based on DIRMO and a differentiated target function. In order to solve the defects that a single model is used for prediction in the existing multi-step power prediction, time series modeling and feature processing cannot be simultaneously considered, the prediction accuracy is reduced, the model robustness is insufficient, and the search efficiency is low, the power and wind speed data after preprocessing are denoised, normalized and feature-extracted to obtain multi-scale features; the normalized power and wind speed data are subjected to time series modeling to generate a preliminary prediction result in the future; the multi-step prediction task is divided into several groups, and a multi-output problem is converted into a single-output problem for training and prediction; the hyperparameters of a LightGBM model are automatically optimized, the optimized hyperparameters are used for training the LightGBM models of the groups, and the final power prediction is carried out based on the preliminary prediction result of the GRU and the multi-scale features. The application is mainly used for predicting wind power.