11results about How to "Avoid Gradient Explosion" patented technology

The application discloses a kind of physical constraint under aero-magnetic data deep learning forward and inversion method and system. Including: constructing the forward network based on deep operator network DeepONet, the network adopts logarithm-linear dual space mechanism, to log resistivity profile and flight height as input, predicts log quadratic field;Inversion network is constructed, with the input of observed transient electromagnetic response and flight height, and the output of resistivity profile;When training, the forward network is pre-trained to convergence under data-driven loss and Maxwell equation physical constraint loss;Inversion network is trained under the joint supervision of its model space loss and physical data space loss, wherein the physical data space loss is obtained by comparing the re-synthesized response with the original observation after inputting the resistivity profile predicted by the inversion network into the frozen forward network.The application deeply integrates physical laws into data-driven models, and realizes the integrated efficient solution of aero-magnetic forward and inversion.

This design is a multi-task automatic EEG detection and prediction system for epilepsy. The system first receives the EEG signals from epilepsy patients to be identified or the collected patient EEG signals used for training the module through a preprocessing module. The preprocessing module performs noise reduction, filtering, segmentation, and normalization on the signals, and outputs the processed patient EEG signals to the training module. In the training module, the collected patient EEG signals are used for training and testing. First, feature extraction is performed using a multi-scale convolutional network. Then, the detection and prediction branches incorporating a dual attention mechanism are trained, and the trained model parameters are saved for use in the detection and prediction module. Finally, the detection and prediction module provides detection and prediction results for the EEG signals from epilepsy patients to be identified. This invention employs a modular design, achieving real-time detection and prediction of noisy, unbalanced EEG signals, providing a reliable system for practical applications.

The application discloses a secure communication trajectory planning method for unmanned aerial vehicles with integrated control of an adaptive antenna, constructs an air-ground secure communication model containing rigid body dynamics constraints, and determines that the attitude of the unmanned aerial vehicle body is uniquely determined by the translational acceleration vector and the gravity acceleration; for horizontal-horizontal (HH) and vertical-vertical (VV) two kinds of airborne antenna configurations, an explicit coupling model of the antenna gain with respect to the acceleration is established; a non-convex trajectory optimization problem is constructed to maximize the average secrecy capacity; a sequential convex approximation combined with a convex difference programming strategy is adopted for solving, wherein the implicit function differentiation method is adopted to solve the gradient singularity problem of the VV antenna null point; and finally, a flight strategy containing an optimal acceleration sequence is output. The application utilizes the active change of the flight attitude of the unmanned aerial vehicle to realize mechanical beamforming, significantly improves the secrecy capacity through "active drift" or "null alignment" maneuvering, and the generated trajectory strictly satisfies the physical feasibility.

This invention discloses a method for constructing a road cavity detection model based on 3D ground-penetrating radar (GPR). The method includes constructing a hybrid spatiotemporal detection model, which comprises a feature preprocessing module, a soft clipping module, a spatiotemporal feature extraction module, and an output module. The feature preprocessing module performs feature pre-extraction on preprocessed and labeled data, outputting a pre-extracted feature map. The soft clipping module performs spatiotemporal perception on the pre-extracted feature map, outputting a clipped feature map. The spatiotemporal feature extraction module analyzes the clipped feature map, outputting feature information with global spatiotemporal information. The output module integrates the feature information and predicts the existence of cavities and the range of their occurrence channels. By combining the correlations between data in the spatiotemporal dimension, it accurately detects underground cavities, solving the problems of high computational load and difficulty in adjusting model parameters in existing 3D GPR high-resolution data interpretation. This provides an efficient and reliable technical solution for intelligent road cavity detection and infrastructure safety maintenance.

The invention provides a multi-beam synthesis method and device based on a collaborative null generation strategy, and belongs to the technical field of antenna design. The method provided by the invention comprises the following steps: randomly generating a design index in a beam direction range, and constructing a mask matrix based on the design index; constructing a multi-beam comprehensive model, inputting the mask matrix into the multi-beam comprehensive model, and training the multi-beam comprehensive model to output amplitude excitation and phase excitation of each beam; constructing a target mask matrix according to beam performance indexes in an actual application scene; and inputting the target mask matrix into a trained multi-beam comprehensive model, outputting amplitude excitation and phase excitation of each array unit for different beams, loading the amplitude excitation and the phase excitation to a multi-beam phased array, and generating a multi-beam far-field directional diagram. The invention provides a multi-beam synthesis method and device based on a collaborative null generation strategy. The method and the device are used for reducing the loss of equivalent omnidirectional radiation power on the premise of effectively suppressing multi-beam interference.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training diffusion neural networks using differentiatable reward functions.

This invention discloses a multi-level land resource segmentation migration fine-tuning method with input alignment, comprising the following steps: acquisition and processing of remote sensing image data and label data related to the land resource segmentation task; selection and slicing of sampling areas based on prior knowledge of the study area; data augmentation to expand the dataset and pixel value normalization; partitioning the model dataset using stratified sampling techniques; determining the optimal initial learning rate through hyperparameter optimization, entering input feature alignment, and establishing a quick connection between the input alignment module and the model output; employing three Transformer architectures and their pre-trained models for land resource segmentation, and optimizing the model training process using gradient-value-constrained gradient clipping and two-stage training strategies; step six: segmentation prediction and visualization. This invention improves the model's training efficiency and segmentation performance, achieving high-precision multi-level land resource segmentation and recognition.

This invention discloses a recurrent neural network training method and system based on Ising machine optimization, belonging to the field of neural network training technology. The method transforms the weight training task of a recurrent neural network into a quadratic unconstrained binary optimization problem, mapped to an Ising model; it utilizes the physical evolution or annealing process of the Ising machine to obtain the spin configuration of the minimum energy state; finally, it decodes and restores the optimized weights and loads them into the recurrent neural network to complete the training. This invention achieves accurate representation of continuous weights through binary discrete encoding and offset matrices, supporting accelerated solutions using various Ising machine hardware (such as optical Ising machines, quantum annealing machines, etc.). Compared with existing gradient-based training methods, this invention completely avoids the gradient vanishing and gradient exploding problems, significantly improving the stability and accuracy of long sequence predictions. Simultaneously, based on advanced Ising machine computing equipment, the core training time can be shortened from seconds to milliseconds, providing a new technical path for efficient, low-power AI systems.

The application discloses a kind of based on sub-domain self-adaptive multi-channel synthetic aperture radar moving target detection method, obtains the simulation and measured data of multi-channel synthetic aperture radar, makes network training and verification dataset, adds label for clutter and moving target;Subdomain self-adaptive residual network for ground moving target detection is built;The network training and verification data are preprocessed, input into subdomain self-adaptive residual network, save the optimal network after training ends;The radar data to be measured is slid window, and network test dataset is made;The preprocessed network test data are input into the saved network, and the predicted label is obtained;All prediction results are windowed, and the ground moving target detection result is obtained.The application trains network by using labeled simulation data and unlabeled measured data, reduces the difference between simulation and measured data by subdomain self-adaptive, solves the problem of lack of labeled data in the field of radar, and effectively improves the detection performance of ground moving target.

The invention belongs to the technical field of casting production workshop scheduling, and particularly discloses a casting workshop multi-target full-process scheduling method and device, equipment and a storage medium. According to the invention, the deep reinforcement learning environment is established according to the production data and the workshop scheduling task; performing agent training based on the target neural network framework to obtain a process scheduling agent model of the casting enterprise; wherein the target neural network framework comprises an action network and a plurality of evaluation networks; and carrying out multi-target full-process scheduling on the production tasks of the workshops of the casting enterprises according to the scheduling targets of the workshop scheduling tasks based on the process scheduling agent model. Through the above mode, the feature vector generalization deep reinforcement learning algorithm capable of describing the global state of the casting workshop and the casting task is utilized, that is, agent training is performed in a deep reinforcement learning environment, and multi-target full-process scheduling is performed based on the process scheduling agent model, so that the efficiency and accuracy of scheduling the casting workshop can be effectively improved.

The invention relates to the technical field of image processing, in particular to a tunnel lining defect ground penetrating radar image enhancement method, medium and equipment, and the method comprises the following steps: constructing a data set; constructing a generative adversarial network model based on improved deep convolution; the obtained tunnel lining structure defect ground penetrating radar image data set is used for training in the constructed improved deep convolution-based generative adversarial network model, and hyper-parameters of a generator and a discriminator are continuously adjusted through a test result, so that the adversarial network model is globally optimal; and generating a tunnel lining structure defect ground penetrating radar image by using the improved deep convolution generative adversarial network model. According to the invention, the technical problem that the difference between the image generated by the existing image processing method and the real ground penetrating radar image is large for the ground penetrating radar image of the internal defect of the tunnel lining is solved.