3402 results about "Model parameter" patented technology

Alignment accuracy between two or more patterned layers is measured using a metrology target comprising substantially overlapping diffraction gratings formed in a test area of the layers being tested. An optical instrument illuminates all or part of the target area and measures the optical response. The instrument can measure transmission, reflectance, and/or ellipsometric parameters as a function of wavelength, polar angle of incidence, azimuthal angle of incidence, and/or polarization of the illumination and detected light. Overlay error or offset between those layers containing the test gratings is determined by a processor programmed to calculate an optical response for a set of parameters that include overlay error, using a model that accounts for diffraction by the gratings and interaction of the gratings with each others' diffracted field. The model parameters might also take account of manufactured asymmetries. The calculation may involve interpolation of pre-computed entries from a database accessible to the processor. The calculated and measured responses are iteratively compared and the model parameters changed to minimize the difference.

The invention discloses a convolution neural network parallel processing method based on a large-scale high-performance cluster. The method comprises the steps that (1) a plurality of copies are constructed for a network model to be trained, model parameters of all the copies are identical, the number of the copies is identical with the number of nodes of the high-performance cluster, each node is provided with one model copy, one node is selected to serve as a main node, and the main node is responsible for broadcasting and collecting the model parameters; (2) a training set is divided into a plurality of subsets, the training subsets are issued to the rest of sub nodes except the main mode each time to conduct parameter gradient calculation together, gradient values are accumulated, the accumulated value is used for updating the model parameters of the main node, and the updated model parameters are broadcast to all the sub nodes until model training is ended. The convolution neural network parallel processing method has the advantages of being capable of achieving parallelization, improving the efficiency of model training, shortening the training time and the like.

Apparatus and methods for universal node design implementing a universal learning rule in a mixed signal spiking neural network. In one implementation, at one instance, the node apparatus, operable according to the parameterized universal learning model, receives a mixture of analog and spiking inputs, and generates a spiking output based on the model parameter for that node that is selected by the parameterized model for that specific mix of inputs. At another instance, the same node receives a different mix of inputs, that also may comprise only analog or only spiking inputs and generates an analog output based on a different value of the node parameter that is selected by the model for the second mix of inputs. In another implementation, the node apparatus may change its output from analog to spiking responsive to a training input for the same inputs.

The invention belongs to the technical field of wireless communication networks, and discloses a federated learning information processing method and system, a storage medium, a program, and a terminal. A parameter serve confirms a training task and an initial parameter and initialize a global model. The parameter server randomly selects part of participants to issue model parameters, encrypts themodel parameters and forwards the model parameters through the proxy server; the participants receive part of parameters of the model and cover the local model, and the model is optimized by using local data; the participant calculates a model gradient according to an optimization result, selects a part of the model gradient for uploading, adds noise to the uploading gradient to realize differential privacy, encrypts the uploading gradient and forwards the uploading gradient through the proxy server; the parameter server receives the gradients of all participants, and integrates and updates the global model; and the issuing-training-updating process of the model is repeated until an expected loss function is achieved. According to the invention, data privacy protection is realized; the communication overhead of a parameter server is reduced, and anonymity of participants is realized.

The present invention discloses a deep learning-based insulator identification method. The insulator identification method comprises the steps of pre-processing an aerial image, and secondly, extending the data via the methods, such as the geometric transformation, the contrast enhancement, an analog noise adding method, etc.; acquiring the insulator samples, aiming at the insulators of different types, classifying to acquire; determining a to-be-trained model structure; inputting the samples in the to-be-trained model, and continuously adjusting the weights and the bias parameters by the forward propagation and backward propagation methods, and finally determining an optimal model parameter, based on the trained model, taking a to-be-detected image as an input signal, and by the network multi-layer convolution, pooling and full-connection operations, obtaining a final detection identification result. According to the present invention, by a deep learning method, the insulator characteristics are learned continuously, a learning network model is determined, the different insulators are identified under different background environments, and support is provided for the electric power maintenance decisions.

A method of predicating ultra-short-term wind power based on self-learning composite data source includes following steps. Model parameters of an autoregression moving average model are obtained by inputting data. A predication result is obtained by inputting data required by wind power predication into the autoregression moving average model. A post-evaluation is performed to the predication result by analyzing error between the predication result and measured values, and performing model order determination and model parameters estimation again while the error is greater than an allowable maximum error.

The invention discloses a federated learning method and device based on a block chain. The method comprises the steps: determining the block chain; enabling the coordinator node to create a federatedlearning task according to the model original data sent by each participant node; receiving training data obtained by local training of the participant nodes; sending the to-be-updated parameters to other participant nodes according to the training data, so as to enable the other participant nodes to update own model parameters according to the to-be-updated parameters; and after model training iscompleted, issuing reward resources according to training data provided by each participant node in the training process, and writing rewards into the block chain. Compared with a traditional mode, the mutual trust problem of all parties is effectively solved; all parties participating in federated learning negotiate together to generate a coordinator node, so that the transparency of the processis improved; federated learning whole-process data is recorded in a block chain, so that the traceability of data operation is ensured; all parties are encouraged to actively participate through rewarding resources, and the enthusiasm of participants is improved.

A computational methodology that improves the accuracy of model parameters in a compact model uses methods and algorithms to self-consistently match independently developed base and stress models by re-fitting the stress model to the data set that generates the base model. The re-fitting algorithm removes any discrepancy between the base model and the stress model as the stress model is applied to the data set obtained from a dimension-scaling macro. Stress offsets for dimension-scaling macro devices are calculated to fit the measured values of the model parameters for the same devices. The process of fitting the model parameters to the data set from the dimension-scaling macro calculates constant, linear, and quadratic coefficients for the model parameters, which are employed to increase the accuracy of the model parameters and of the compact model used in circuit simulations and optimization.

The invention discloses a parametric modeling method of a rigid-flexible coupled model. The method includes the steps of 1, establishing a rigid component model, 2, establishing an APDL macro file of a flexible body model, 3, establishing a rigid-flexible coupled system parametric graphical user interface, 4, inputting flexible body parameters in the interface and reading, altering and updating the APDL macro file, 5, executing the updated APDL macro file to build the flexible body model, 6, inputting rigid body model parameters, reconstructing a rigid body model, importing a flexible body modal neutral file and generating a command file containing a rigid-flexible coupled dynamics simulation model, 7, importing the command file, loading a virtual prototype model, conducting dynamics simulation analysis and outputting a simulation result and 8, outputting the optimum simulation data and storing a data file according to the simulation result. According to the rigid-flexible coupled model parametric modeling method, automatic creation of the parametric modeling of a flexible body and the rigid-flexible coupled model can be achieved, manual intervention is not needed in the whole process, the problem that it is difficult to parameterize the flexible body in the engineering practice is solved, the modeling efficiency is improved, and the method has very good engineering application value.

The invention discloses a method and a device for neural network machine learning model training, and the method comprises the steps: applying the method and device to a distributed computing frame which comprises a plurality of computing nodes, segmenting training data into training data slices in advance, and enabling the number of the segmented slices to be the same as the number of the computing nodes participating in the calculation; a computing node obtaining training data slices and trains local model parameters; the computing node transmits the trained local model parameters to a parameter server; and the computing node updating local model parameters according to the global model parameters returned by the parameter server and continues to train the local model parameters. According to the method, the calculation acceleration ratio at multiple nodes can almost reach a linear ideal value, and the period of model training is greatly shortened.

A class is to be characterized of a probabilistic classifier or clustering system that includes probabilistic model parameters. For each of a plurality of candidate words or word combinations, divergence of the class from other classes is computed based on one or more probabilistic model parameters profiling the candidate word or word combination. One or more words or word combinations are selected for characterizing the class as those candidate words or word combinations for which the class has substantial computed divergence from the other classes.

The invention discloses an inclination-angle-constraint-based kinematic calibration method for a Stewart parallel robot. The method comprises the following steps of: firstly, theoretically establishing a novel kinematic constraint, which is an inclination angle constant constraint, namely keeping two inclination angles of a motion platform of the Stewart parallel robot relative to a horizontal plane constant group by group; secondly, using a servo regulating way to physically realize the established kinematic constraint in a high-precision way; thirdly, establishing a calibration model on the basis of a principle of least square according to the kinematic constraint; and finally, identifying a model parameter by solving a nonlinear least square optimizing problem and compensating for the model parameter in a robot control software. By making full use of a characteristic that a repeating precision of a measuring instrument is superior to a position measuring precision of the measuring instrument, the method has the advantages of good calibration effect, simple measurement, high automatism in the calibrating process and the like.

The invention discloses a decentralized federated learning method under privacy protection. The decentralized federated learning method comprises a system initialization step, a request model and local parallel training step, a model parameter encryption and model sending step, a model receiving and recovering step and a system updating step. Decentralization is achieved by using a strategy of randomly selecting participants as parameter aggregators, and the defects that existing federated learning is easily attacked by DoS, a parameter server has a single point of failure and the like are overcome; a PVSS verifiable secret distribution protocol is combined to protect participant model parameters from model inversion attacks and data member reasoning attacks. Meanwhile, it is guaranteed that parameter aggregation is carried out by different participants in each training task, when an untrusted aggregator occurs or the aggregator is attacked, the aggregator can recover to be normal by itself, and the robustness of federated learning is improved; while the functions are achieved, the federated learning performance is guaranteed, the safety training environment of federated learning is effectively improved, and wide application prospects are achieved.

The invention provides a remote sensing image semantic segmentation method based on regional attention multi-scale feature fusion. The remote sensing image semantic segmentation method comprises the following steps of: S1, constructing a network model for a remote sensing image semantic segmentation network; S2, constructing a training data set, and preprocessing the collected data set for training; and S3, inputting the data set for training into a network model for training, and predicting a result after acquiring training parameters. According to the remote sensing image semantic segmentation method, the idea of an image cascade network is introduced, and the model parameter quantity is greatly reduced; meanwhile, coding features and decoding features are optimized by using an attentionmechanism, a regional attention module and a multi-scale group fusion module are constructed, feature maps of different scales are extracted and fused, training is guided by using multi-scale semantic tags and boundary tags, and the model performance is effectively improved under the condition that the parameter quantity of the model is only 8.4 M.

The present invention relates to the field of traffic safety, in particular to a judgment and prediction method of driving behavior intention based on the implicit Markov model (HMM), aiming to overcome the defect that the existing driving behavior intention recognition and prediction technology does not take into account the dynamics and continuity of driving behavior, as well as complex behaviors such as lane changing, car following and braking and the like. The judgment and prediction method of driving behavior intention obtains time series segmentation data from cluster analysis of dynamic driving data, the linear direction HMM, lateral HMM and speed classification model are trained respectively, and the obtained identification results are regarded as the observation sequence of behavior recognition layer; Then, off-line training is performed to deal with normal or emergency braking, normal or emergency lane change, normal or dangerous driving behavior, and multi dimensional discrete HMM model; according to the model parameters and the observation sequence, the next time step driving behavior can be predicted. The judgment and prediction method of driving behavior intention takes the complexity and continuity of driving behavior into account and can dynamically judge and predict the driving behavior intention and warn the dangerous behavior, and accordingly can be applied to driving behavior evaluation and driving assistance system.

A method of optimising a model Mx suitable for use in data analysis and determining a prognostic outcome specific to a particular subject (input vector x), the subject comprising a number of variable features in relation to a scenario of interest for which there is a global dataset D of samples also having the same features relating to the scenario, and for which the outcome is known is disclosed. In one implementation, the method includes: (a) determining what number and a subset Vx of variable features will be used in assessing the outcome for the input vector x; (b) determining what number Kx of samples from within the global data set D will form a neighbourhood about x; (c) selecting suitable Kx samples from the global data set which have the variable features that most closely accord to the variable features of the particular subject x to form the neighbourhood Dx; (d) ranking the Vx variable features within the neighbourhood Dx in order of importance to the outcome of vector x and obtaining a weight vector Wx for all variable features Vx; (e) creating a prognostic model Mx, having a set of model parameters Px and the other parameters from (a)-(d); (f) testing the accuracy of the model Mx at e) for each sample from Dx; (g) storing both the accuracy from (f), and the model parameters developed in (a) to (e); (h) repeating (a) and/or (b) whilst applying an optimisation procedure to optimise Vx and/or Kx, to determine their optimal values, before repeating (c)-(h) until maximum accuracy at (f) is achieved.

The invention discloses a federated modeling method, device and equipment based on a block chain, and a storage medium. The method comprises the following steps: a training initiator issues configuration information to all training clients based on client information corresponding to the training clients when the number of the training clients is monitored to reach a preset number; the training initiator uploads a to-be-trained model to a main chain; and the training initiator determines a target model based on aggregation gradient and the to-be-trained model. Modeling of federated learning isrealized through the block chain; on the premise of protecting the privacy of federated learning data, the accuracy of federated learning is not affected, the training effect and model precision of federated learning are improved, model parameters such as gradients in transmission do not need to be modified, and the balance between privacy protection of the model parameters such as gradients andmodel convergence or model precision is achieved; and information leakage can be completely prevented, so the safety of data samples in federated learning is improved.

The invention discloses a model method of a traffic prediction. The method comprises the following steps of: (1) establishing a road traffic basic facility geographical information database, storing and updating traffic planning schemes, and establishing intersection delay determination; (2) establishing an urban planning construction geographical information database, storing regulatory detailed plans of each area, and utilizing an original unit method to construct traffic generation cases; (3) calling data of the road traffic basic facility geographical information database and the urban planning construction geographical information database, and utilizing a gravity model method to predict parallel distributed matrixes; (4) constructing a multi-element logit model, and determining the probability distribution for a single person selecting a travelling mode in specific travelling information; and (5) utilizing a capacity-limited multi-path distribution method to carry out traffic distribution. The modeling purpose, method to realization are overall planed, modeling parameter selection, basic data sources and calculation program design are coordinated, and the convenient application and operation of a traffic prediction model system are ensured.

A relay feedback based AC servo system automatic setting method belonging to the motion control technology field comprises a speed ring setting process and a position ring setting process. The method comprises the following steps: setting the parameters of saturated relay with delay; bringing the relay identification link into the closed speed ring (or position ring) to cause the system to establish a limit ring quickly; evaluating the system delay and the limit oscillation period parameter; identifying the AC servo system model; and setting the PID control parameter automatically by the PID parameter setting principle such as pole assignment method and amplitude phase margin method according to the identified model parameters; By combining the saturated relay feedback and he PID parameter setting principle such as ITAE method and amplitude phase margin method, the invention can accomplish the parameter automatic setting to the controller of AC servo motor.

The invention provides a facial recognition model training and facial recognition method, system and device and a medium. The training method comprises the steps of constructing a face recognition model, wherein the model comprises a feature extraction layer, a full connection layer and a loss function layer, the full connection layer comprises a first feature output layer, a second feature outputlayer and at least one attribute output layer, and the loss function layer comprises a feature extraction loss function layer and at least one attribute loss function layer; and training the model byadopting a training set. The neural network model is constructed to be used for face feature extraction and face attribute analysis at the same time. After the image is input into the neural networkmodel, the feature for recognition and the attribute of the image can be output at the same time, the calculated amount in image recognition is reduced, and the image recognition efficiency is improved. In the model training process, feature extraction and attribute analysis are trained at the same time, a plurality of loss functions influence each other, and better model parameters can be obtained through training.

The invention discloses a privacy protection method and system based on federated learning, and a storage medium, and the method comprises the steps: carrying out the encryption of a global model through employing a parameter encryption algorithm, and obtaining a ciphertext model; training on the ciphertext model by using local data, decrypting the obtained ciphertext gradient information and noise item to obtain a parameter gradient, updating the global model by using the parameter gradient, and repeating the steps until the model converges or reaches a specified iteration number to obtain amodel parameter; encrypting the model parameters to obtain encrypted model parameters, and updating the global model by adopting the encrypted model parameters to obtain a global encrypted model; andperforming local training on the encrypted global model to realize privacy protection. According to the method, semi-trusted federated learning participants can be effectively prevented from acquiringreal parameters of the global model and an output result of the intermediate model, and meanwhile, the participants can acquire a real prediction result by utilizing the finally trained encryption model.

The invention provides a network intrusion detection method based on an improved convolutional neural network. The method comprises the steps of acquiring a data set, performing numeralization, performing normalization processing, improving convolutional neural network model training, performing iterative optimization of model parameters, acquiring five classifiers, outputting a five-dimensional confusion matrix to serve as a classification result, and evaluating the classification result; the network intrusion detection method based on the improved convolutional neural network provided by theinvention is based on an improved convolutional neural network model, model training is performed by using a preprocessed original sample data set in combination with a cross-layer design manner, themodel achieves a good convergence effect through continuous feature extraction and iterative optimization, and then a classification test is performed by using the trained classifier, so that the feasibility and effectiveness of an intrusion detection effect can be improved by the method.