426 results about "Descent algorithm" patented technology

The invention discloses an LSTM deep learning model-based hydropower unit fault diagnosis method and system. The method comprises the steps of obtaining sampling sequences of N different signal channels of a hydropower unit, and performing VMD on each time sequence to obtain K IMF components; constructing corresponding training sets and to-be-diagnosed sets; building an LSTM model for the trainingset of each IMF component, and performing feature extraction on each IMF component through two LSTM layers; connecting outputs of K LSTM layers of the same signal channel to a Dense layer; through aSoftmax layer, performing feature classification on outputs of multiple Dense layers; and training a deep learning neural network model through an RMSProp gradient descent algorithm, and diagnosing the to-be-diagnosed sets by the trained model. According to the method and the system, the relatively good signal-noise separation effect of the VMD is combined with the processing advantage of an LSTMto the time sequences, so that the hydropower unit fault diagnosis accuracy is effectively improved.

The invention discloses a radar one-dimensional range profile target recognition method based on a depth convolution neural network, includes the following steps: a data set is collected, the collected data is preprocessed, features are extracted from the preprocessed data, the HRRP signal is divided into two parts: low SNR and high SNR, A feature enhancement algorithm based on robust Boltzmann isconstructed, and a HRRP target recognition model based on convolution neural network and bidirectional loop neural network based on LSTM is constructed. The parameters of the network model are fine-tuned by using gradient descent algorithm, and an effective target recognition model is obtained. A radar HRRP automatic target recognition technology with small sample robustness and noise robustnessconstructed by the invention has strong engineering practicability, and a radar one-dimensional range profile target recognition model based on a convolution neural network and a cyclic neural networkis proposed from the aspects of feature extraction and the design of a classifier.

The invention discloses an image detection method. The method comprises the steps of acquiring training samples, which are obtained by expanding marked image data containing vehicles and pedestrians;classifying first features extracted from the training samples and obtaining the coordinates of a first detection frame; calculating total loss through a preset algorithm according to the classification result, the coordinates of the first detection frame and the real classification information, updating the parameters of a convolution neural network model according to the total loss through a gradient descent algorithm, performing iterative training on the training data to obtain a trained convolutional neural network model, acquiring a to-be-detected image including vehicles and pedestrians,and obtaining the corresponding vehicle and pedestrian information by utilizing a trained convolutional neural network model. The invention further discloses an image detection device and a computerreadable storage medium. Vehicle and pedestrian information in an image can be detected based on a convolutional neural network model, and the accuracy of image detection can be improved.

The invention provides an improved ceramic material member sequence image segmentation method of a fully convolutional neural network. The method comprises the steps of S10, performing manual marking on an acquired original image, classifying a target and a background by different kinds, obtaining a training label, and representing a label graph of a training sample in an index mode; S20, constructing an improved network model based on the fully convolutional neural network, and performing training; and S30, calculating a loss function and a reverse propagation calculation loss function according to a gradient reducing algorithm, and performing training learning on the network, wherein the learning rate is reduced to one tenth of the original learning rate when verification accuracy increase is stopped. The fully convolutional network is an improved structure based on a convolutional neural network. Based on keeping of good classification performance of the CNN, a spatial position relation between pixel matrixes in better kept, and global characteristic extraction is facilitated. The visual characteristic of an object can be comprehensively studied, and high interference resistance is realized. An objective target can be automatically divided from background, thereby realizing intelligent segmentation.

The invention discloses an MEC-based task cache allocation strategy applied in a mobile edge computing network. The MEC-based task cache allocation strategy comprises the steps: an edge cache system architecture of MEC servers is used; cache merging and cache segmenting of a task are conducted by means of a collaboration function of the MEC servers; access delay and energy consumption analysis areconducted under constraint conditions; and access delay and energy consumption are optimized by means of a coordination descent algorithm. Accordingly, more videos can be effectively cached on the MEC servers, the access delay is greatly shortened, and the energy consumption is greatly reduced.

The invention discloses a low-dose CT image denoising method, and mainly solves the problems of image edge blurring and detail loss caused by the fact of only considering restoring image grayscale information in the prior art. The implementation scheme is that 1) multiple low-dose and full-dose CT images are acquired from the same part of the body; 2) the acquired CT image data set is extended andnormalized and then block taking of each pair of CT images is performed so as to obtain a CT image block data set; 3) a twelve-layer full convolutional denoising neural network is established, the CTimage block data set is used as the network training data and the network is optimized by using a small batch gradient descent algorithm with the momentum so as to obtain the trained network; and 4)a complete low-dose CT image is inputted to the network so as to output the corresponding denoised CT image. The edge and the details of the image can be greatly kept in case of image denoising so that the method can be used for enhancement of the low-dose CT image and is convenient for the doctor to recognize the CT image case.

The invention relates to an electric power equipment detection algorithm based on a convolution neural network, comprising the following steps of (1) a labelImg tool is used for labeling the visible light image and infrared image of the electric power equipment, and obtaining the label about the electric power equipment in the image; (2) the images and device-level tags in the training set are inputted to the detection model of electric power equipment for training, and update the model parameters by using the gradient descent algorithm of driving quantity to back propagate; (3) in the testingphase, the prediction box is reduced by soft non-maximum suppression, and the prediction box whose confidence level is higher than the threshold value is outputted finally; (4) compared with the current power equipment detection methods, the algorithm can converge and obtain excellent detection performance without fine-tuning on the pre-training model, has excellent robustness for small power equipment, can effectively reduce the phenomenon of frame error and frame leakage, and has high detection accuracy and detection speed to achieve real-time processing.

The invention relates to a roadside air pollutant concentration prediction method based on reconstruction deep learning. On the basis of a reconstruction deep learning method, a depth reconstruction Elman model is provided according to the time-space distribution characteristics of roadside air pollutants; according to the characteristics of a restricted Boltzmann machine, initialization of the depth reconstruction Elman model is completed by utilization of part input data of a roadside air pollutant concentration data set; the depth reconstruction Elman model is trained by adoption of a gradient descent algorithm; and, due to the characteristic mapping function of the model, a real-time roadside air pollutant concentration prediction method based on the factors, such as road network information, weather information and traffic information, can be obtained.

The invention provides a measurement matrix optimization robustness method based on MSE. The method is based on a traditional model for optimizing a measurement matrix, a regular item is added and theregular item represents the mean square error of the original image and the reconstructed image. The method comprises the following steps of assuming that a mean square error obeys standard positivedistribution; carrying out the singular value decomposition by applying a center limit theorem and an equivalent dictionary, so that the optimization model of the measurement matrix is well simplified; finally using a gradient descent algorithm, so that the optimized measurement matrix is iteratively solved, the newly proposed image measurement matrix optimization model fully applies the information of the image, the cross correlation coefficient between the measurement matrix and the sparse base is reduced, the requirement on the sparsity is reduced, and the robustness of the image compressedsensing system is improved to a certain extent. Experiments show that independence between optimized measurement matrix columns is increased, and the reconstruction of high-quality image signals is facilitated.

The invention relates to a human face identification method. The method comprises the following steps of 1, reading a human face image sample data set; 2, establishing a deep convolutional neural network, wherein a residual error unit is introduced in the deep convolutional neural network; 3, updating parameters of the deep convolutional neural network by utilizing a gradient descend algorithm, and firstly mapping human face images of the sample data set into 512-dimension eigenvectors through the deep convolutional neural network; calculating a loss function and a gradient of the loss function, and updating the parameters of the deep convolutional neural network according to whether a gradient descend distance of the loss function is smaller than a preset threshold or not, wherein the loss function is formed by weighting a Softmax function and an A-softmax function; and 4, performing human face identification through the deep convolutional neural network subjected to the parameter updating. Not only the between-class distance but also the within-class distance are considered, so that the human face identification rate is increased.

The invention discloses an address similarity calculation method and device, equipment and a storage medium. Aiming at the problems that address matching has complex rules, an existing matching algorithm is not high in retrieval speed and accuracy, and the address matching efficiency is low, a solution is proposed. Input address information is expressed by a proper initial vector; an address similarity calculation model based on a twin neural network is used, a gradient descent algorithm of a ternary loss function is combined, thus obtaining a feature vector of the initial address vector; finally, the cosine distance or L2 distance between the feature vector and an address vector in a standard address data set is calculated, the known address vector closest to the input address vector is obtained, so that the address matching rule is simplified, the preferred accuracy of the same address is improved, and the retrieval speed and accuracy of the matching algorithm are further improved.

The invention provides a large-scale MIMO system CSI feedback method based on a long-short-term attention mechanism. The method comprises the following steps: firstly, generating sample data of a space-frequency domain channel matrix by using a COST 2100 channel model, and preprocessing the sample data through discrete Fourier transform and truncation operation to obtain a training set and a testset; secondly, establishing an ALSTM-CsiNet channel state information feedback reconstruction model by utilizing an LSTM and an attention mechanism; inputting the training set into an ALSTM-CsiNet channel state information feedback reconstruction model, and performing iterative training on the model by adopting a mean square error and an adaptive estimation gradient descent algorithm to obtain anoptimized ALSTM-CsiNet model; and finally, directly inputting the test set into the ALSTM-CsiNet model to carry out channel state information reconstruction. According to the invention, the LSTM is used to learn the time correlation of the channel, and the attention mechanism is used to perceive the local information and the automatic weighting feature information, so that the channel feedback precision is improved.

The invention discloses an SQL (Structured Query Language) conversion method and system based on language model coding and multi-task decoding. The method comprises the steps: combining a language model in combination with a field where a data set is located to carry out pre-training, and improving the feature extraction capability in the field; sequentially expanding the query database according to table names and column names, converting a two-dimensional table into a one-dimensional text sequence, and splicing the one-dimensional text sequence into an input sequence X in combination with user questions; inputting the sequence X into a pre-training language model, and outputting a coding result; a multi-task decoder composed of nine different neural networks is utilized to decode and restore the SQL fragments, and cross entropy loss is calculated; different weights are set for loss values of different neural networks, the sum is finally calculated as the total loss of the model, a gradient descent algorithm is utilized to optimize an objective function, and model training parameters are updated; after training is completed. Model parameters are stored, and a corresponding SQL sequence is automatically generated according to the user problem and the target database.

The invention discloses a super-resolution implementation method based on reconstruction optimization and a deep neural network. The method comprises the following specific steps of 1, establishing arelationship between a high-resolution image and a down-sampled image by means of a down-sampling method, and establishing a target function by means of a least square method; 2, utilizing a conjugategradient descent algorithm to conduct iterative optimization on the target function to obtain a high-resolution image based on a reconstruction optimization algorithm; 3, establishing and completinga three-layer full convolutional neural network through training; 4, inputting the high-resolution image obtained in the step 2 into the three-layer convolutional neural network to further improve theresolution of the image. By means of the method, good results can be achieved in both subjective evaluation and objective image quality evaluation.

A case reasoning method based on dynamic knowledge representation learning comprises the following steps: 1) obtaining all related data of a broken case, dividing the related data into five categoriesof entities including personnel, cases (events), articles, places and organizations, and extracting a corresponding relationship; 2) storing the extracted events as a quadruple format in the form oftime, entities and relationships, and storing the quadruple format into a graph database; 3) optimizing hyper-parameters of Gaussian process regression based on a gradient descent algorithm; 4) usinga recurrent neural network model to carry out recurrent event reasoning on the tetrad data; 5) performing first-degree and second-degree relationship search by using the graph database, and performinglink prediction based on a search result. According to the method, the entity and relationship embedding is carried out on the tetrad through a dynamic knowledge representation learning algorithm, the training and learning are carried out on the basis of a constructed knowledge graph, the possible criminal suspects are inferred, and the police affair work is simplified.

An image editing application (or a blur classification module thereof) may automatically estimate a coherent defocus blur map from a single input image. The application may represent the blur spectrum as a differentiable function of radius r, and the optimal radius may be estimated by optimizing the likelihood function through a gradient descent algorithm. The application may generate the spectrum function over r through polynomial-based fitting. After fitting, the application may generate look-up tables to store values for the spectrum and for its first and second order derivatives, respectively. The use of these tables in the likelihood optimization process may significantly reduce the computational costs of a given blur estimation exercise. The application may minimize an energy function that includes a data term, a smoothness term, and a smoothness parameter that is adaptive to local image content. The output blur map may be used for image object depth estimation.

The invention relates to a face recognition method based on a deep separable convolution model, the method comprising the following steps: 1, reading a face image sample data set; 2, establishing a deep separable convolution model, wherein the deep separable convolution model cascades a plurality of residual bottleneck module between two adjacent convolution modules; 3, updating the parameters ofthe deep separable convolution model by using a gradient descent algorithm; 4, carrying out face recognition through the deep separable convolution model after parameter updating. The method can improve the recognition speed on the basis of ensuring the face recognition accuracy. Experiments show that the method can ensure that the recognition accuracy is higher than 99%, so that the recognition speed on an ARMv8 mobile terminal reaches less than 300ms and the mobile terminal can have an accurate and fast face recognition function.

This application provides an airspace complexity evaluation method based on deep unsupervised learning for air traffic management, which includes the following parts. A stacked autoencoder is used to establish an airspace complexity evaluation model. Input the airspace complexity factors into the stacked autoencoder to obtain the low-dimensional embedded representations of the airspace complexity factors. Cluster the low-dimensional embedded points to capture the centroids of the airspace complexity data. The application utilizes the soft assignment distribution and real assignment distribution of the embedded representations to construct a training loss function which optimizes the airspace complexity evaluation model by gradient descent algorithm. The trained airspace complexity evaluation model and the three obtained cluster centroids describing the airspace complexity level are used to evaluate the current airspace complexity.

The invention provides a vegetable identification method and device. The method includes the steps of establishing the first convolution neural network model, collecting the images of each vegetable in different scenes, building the corresponding input dataset, iteratively training the first convolution neural network model according to a learning rate adaptively updated gradient descent algorithmand the input data set, obtaining the second convolution neural network model, carrying out the vegetable recognition according to the second convolution neural network model, and improving the efficiency of the vegetable recognition is improved. The first convolution neural network model is trained by a stochastic gradient descent algorithm, so that the second convolution neural network model obtained by the training can accurately identify the vegetable species under the complex background.

The invention discloses a multi-unmanned aerial vehicle task scheduling method and system and a storage medium, the first stage is a multi-unmanned aerial vehicle task allocation stage, a multi-unmanned aerial vehicle task scheduling problem is divided into a plurality of single-unmanned aerial vehicle scheduling sub-problems, and a simulated annealing algorithm embedded with a tabu table is proposed to realize multi-unmanned aerial vehicle task allocation; and the second stage is a single unmanned aerial vehicle task scheduling stage, and a variable neighborhood search descent algorithm is designed according to the task allocation scheme in the first stage by considering the observation capability of the unmanned aerial vehicle platform and the requirements of the tasks so as to provide an effective and feasible task scheduling scheme. And in the first stage, according to a feedback result in the second stage, combining the tabu factor, the transfer factor and the exchange factor to iteratively adjust and update the task allocation scheme until a stop criterion is met. In conclusion, a two-stage iterative optimization method is provided for solving the multi-unmanned aerial vehicle cooperative task scheduling problem. Simulation experiments verify the superiority and efficiency of the method.

The invention discloses a multi-unmanned-aerial-vehicle path planning method based on edge computing dynamic task arrival. The method comprises the steps that a system model of multi-unmanned-aerial-vehicle cooperative service users is established; constructing a multi-unmanned aerial vehicle path planning problem; simplifying the problem into an optimization problem in a single time slot; and decomposing the optimization problem into a user frequency optimization sub-problem and a joint optimization problem of multi-unmanned aerial vehicle path optimization and user association sub-problems.Compared with a single unmanned aerial vehicle scene, multiple tasks and energy queues of multiple unmanned aerial vehicles are added, and scheduling limitation conditions are added to guarantee cooperative communication among the multiple unmanned aerial vehicles. In addition, in order to solve the complex multi-unmanned aerial vehicle path planning and scheduling problem, the Lyapunov queue optimization theory and the block iterative descent algorithm are combined, and the problem complexity is further reduced through linear relaxation and continuous convex approximation. A simulation resultshows that compared with a single-unmanned-aerial-vehicle system, the multi-unmanned-aerial-vehicle service ground user system has the advantages that the energy consumption of the unmanned aerial vehicles is reduced, and the queue backlog task load is reduced.

The invention discloses a target radiation source positioning method based on TDOA. According to the invention, a corresponding non-line-of-sight sensor vector is defined and the non-line-of-sight sensor vector is initialized according to received time difference parameters so as to correct original time difference parameters. And an intermediate variable is selected, a target radiation source TDOA pseudo-linear estimation model is constructed, and weighted least square estimation is applied to obtain position estimation of the target radiation source. And a result is corrected iteratively byusing a gradient descent algorithm. And finally, final target position estimation is provided by utilizing a relationship between the intermediate variable and a target position estimation coordinateand applying weighted least square estimation again. After introduction of a proper intermediate variable, a target radiation source TDOA measurement equation which is not easy to solve directly is converted into a pseudo-linear estimation model; meanwhile, a gradient descent algorithm is combined, so that the influence on the positioning performance due to possible non-line-of-sight errors is reduced. And finally, the final solution of the position coordinates is further optimized.

The invention relates to a data classification system and method based on KL divergence optimizationmethod for classifying data based on KL divergence optimization. The method comprises the steps thatdata preprocessing is conducted on original images, t, texts and other data, and objects are modeled into multi-dimensional distribution; S; selecting a certain amount of triple from the tagged training data to carry out model training; T; the selected triple serves as training data, a linear mapping A is applied to all the mean vectors, t, the optimal linear mapping is learned through iterativeoptimization, and the learning process is based on the basic assumption of metric learning, t, that is, t, the distance between samples of the same kind becomes smaller, and the distance between samples of different kinds becomes larger; A; an intrinsic gradient descent algorithm is adopted for optimization, and after the gradient of an objective function is projected to the tangent space of the same manifold, Riemannian gradient descent is executed on the manifold of an SPD matrix with given affine invariant Riemannian metric; A; and calculating the KL divergence between the test set and thetraining set, and classifying the samples by adopting a K-nearest neighbor (KNN) classifier. The method can effectively improve the classification precision of the system, and has more stable performance.

The invention discloses a modeling solving method for a multi-level positioning inventory path problem. The method comprises the following steps: S1, acquiring basic logistics information of a relatedfactory, a distribution center and a point of sale; S2, establishing a three-level positioning inventory path problem model by using a fuel consumption model and the acquired basic logistics information, wherein the three-level positioning inventory path problem comprises distribution center site selection, a distribution vehicle path and replenishment quantity; S3, utilizing a greedy genetic algorithm to optimize the site selection of the distribution center and the path of the distribution vehicle; S4, optimizing the replenishment quantity by using a gradient descent algorithm. When an LIRPmodel is solved, a greedy and genetic hybrid heuristic algorithm (GGA) is designed to solve a site selection scheme, a path scheme and an inventory scheme in the logistics system, and the gradient descent algorithm (GD) is designed to realize secondary optimization of a replenishment quantity strategy, so that the inventory cost of the logistics system is reduced, and the total cost of the logistics system is further reduced.

The invention discloses a posture evaluation method based on dynamic step length gradient descent. The method comprises the steps that step A, an inertial navigation device IMU is provided with a triaxial accelerometer and a three-axis gyroscope, and data, output by the IMU, of the accelerometer and the gyroscope is collected; step B1, in order to correct motion noises of a motion carrier, data information of the accelerometer from the IMU is smoothed; then a quaternion set is obtained through a dynamic step length gradient descent algorithm; step B2, through a differential equation of quaternions, the other quaternion set is solved by utilizing data output by the gyroscope; step C, the sets of data obtained from the step B1 and the step B2 are fused, and the required posture information is obtained. The method adopts a dynamic step length gradient descent method for posture solution, and the minimum value of an error function is sought by utilizing the gradient descent method, so thatthe algorithm is simple, and the optimizing speed is higher; a one-dimensional optimizing method is adopted for conducting dynamic treatment on the gradient step length of the gradient descent method, and the dynamic performance of posture solution is strengthened.

The invention discloses a multi-objective optimized virtual machine deployment algorithm for a network range, called MVMDNR for short. The algorithm comprises: combining characteristics that network target range virtual machine deployment takes a scene as a unit and communication between scenes is isolated, solving the VMP problem by referring to the idea of the ant colony algorithm; when a virtual machine request is reached; initializing a control node cluster and an intimacy matrix, each control node selecting the deployment position of each virtual machine through probability, realizing positive feedback through intimacy enhancement and volatilization after each round of iterative search, and finally, screening a scheme that the load unbalance degree of a target range physical machine and the communication overhead of the virtual machines is minimum comprehensively. Compared with a random algorithm and an optimal adaptive descent algorithm, the multi-objective optimized virtual machine deployment algorithm provided by the invention has remarkable advantages in reducing the communication overhead of the virtual machine and the load unbalance degree of physical machine resources.

The invention discloses a resource optimization method in mobile edge computing task unloading and electronic equipment. The method comprises the following steps: constructing a queue stability indextaking minimization of the task queue length of a user terminal and the task queue length of an edge server as a target; constructing a network resource overhead index; constructing a random network resource optimization model by taking the queue stability index as a constraint condition; introducing a Lagrange multiplier to carry out variable relaxation on constraint conditions of the random network resource optimization model, and constructing a coupling model; and solving the coupling model based on a momentum stochastic gradient descent algorithm to obtain an optimal resource allocation decision of each time slot. According to the method, the coupling model of the random network resource optimization model is constructed based on the original coupling theory, the coupling problem is solved online based on the momentum stochastic gradient descent algorithm, and task queue overstock is reduced while algorithm convergence is accelerated. The task queue backlog can be further reduced under the condition that the network resource overhead is not increased.

The invention discloses a control method and system for realizing polarization stability, and belongs to the technical field of optical fiber communication and optical fiber sensing. According to thecontrol method and system of the invention, the output light polarization state of any input light polarization state after passing through a polarization controller is rapidly adjusted to be close toany specified target polarization state on a Poincare sphere through a rapid positioning algorithm, and then the output light polarization state is further stabilized to a target polarization state through a random gradient descent algorithm. Therefore, any polarization state can be stabilized to any set target polarization state. The specific implementation system is composed of an input end polarization analyzer, an output end polarization analyzer, a calibrated polarization controller, a single-chip microcomputer and the like; the input polarization analyzer and the polarization controllerwith the calibrated phase difference and voltage relation are used for rapid positioning, and meanwhile, the output polarization analyzer is combined with the stochastic gradient descent algorithm tofinally stabilize the output polarization state to a set value. According to the polarization stability control method, the searching speed and the local extreme value avoidance are balanced, and thepolarization state can be rapidly and stably changed to any designated target polarization state.

The invention discloses a multi-target urban logistics distribution path planning method. The method comprises the following steps: decomposing a three-target vehicle path problem with a time window into a plurality of single-target sub-problems through a group of uniformly distributed weight vectors; initializing the sub-problems by adopting a heuristic strategy; generating a filial generation for the sub-problem by using an evolutionary operator, and designing a target-oriented neighborhood operator to be combined with a variable neighborhood descent algorithm to serve as a local search strategy so as to improve the solving quality of the sub-problem; updating the solution of the sub-problem by adopting a Chebyshev aggregation function; optimizing a non-dominated solution in the archivesby adopting an external archive strategy based on a sorting and congestion degree mechanism; and S3, repeating the steps S3 to S4 until the set maximum number of iterations is reached, and providinga group of feasible vehicle distribution schemes for multi-target urban logistics distribution. Compared with single-target optimization, the method can provide richer decision information for a decision maker, and considers the quality of the solution on the premise of ensuring the convergence and diversity of the algorithm.

The invention provides a stable controllable image generation model training method based on a W distance, and the method comprises the following steps: a, preprocessing image data to acquire sample data of a training set; b, constructing a stable controllable image generation model based on a W distance; c, constructing an overall model according to a loss function of each network; d, alternatelyiterating model training by a gradient descent algorithm to ensure normal convergence of model parameters; and e, after the model parameters converge, respectively separating a generator network G, an encoder network E and a classifier network C as independent service products. The invention is an end-to-end network model, improves a model in the prior art from two aspects of a model structure and a distribution measurement standard, solves the problems of unstable model training gradient and unstable model gradient descending direction, can stably and directionally generate an image sample.Robustness in the model training process and controllability of the generated sample are improved.