40 results about "Neural network topology" patented technology

The invention provides a combined wind power prediction method suitable for a distributed wind power plant, and the method comprises the steps: collecting the real-time meteorological data and historical power data of a wind power plant cluster, carrying out the cross-wind-plant data collaborative cleaning, and generating a time-space aligned standardized data set. Constructing an adaptive spatio-temporal feature extractor, outputting a spatio-temporal feature matrix, and inputting the spatio-temporal feature matrix into the spatio-temporal adaptive neural network, the graph attention prediction model and the physical constraint decision tree model to generate three prediction sequences. And the sequences are fused through a space-time collaborative attention mechanism to generate a dynamic weighted combination prediction result. And performing physical constraint correction on the result by using a space-time residual error correction network to generate a final prediction sequence. And updating the neural network topological structure based on the prediction error distribution, and outputting a prediction result with uncertainty evaluation to a power grid dispatching system. According to the method, the precision and reliability of wind power prediction of the distributed wind power plant can be improved, and the stability and economy of power grid dispatching are improved.

The invention relates to an encrypted traffic detection method based on multi-dimensional feature parallel fusion, and belongs to the technical field of network security. According to the method, when a multi-dimensional feature parallel fusion framework is constructed, the limitation of traditional statistical features on dynamic evolution characterization of encryption behaviors and the dependency of graph neural network topology modeling on computing resources are fully considered; through collaborative optimization of a Markov chain dynamic quantization protocol interaction state transition rule and a lightweight graph attention hierarchical compression mechanism, a detection model gives consideration to deep feature perception capability and efficient reasoning capability at the same time; based on the classification decision realized by the fusion mechanism, the recognition robustness and real-time defense efficiency of the encrypted malicious traffic are remarkably improved, and the active security protection level of the network is effectively enhanced.

Systems and methods are provided for analog hardware realization of neural networks. The method includes obtaining a neural network topology and weights of a trained neural network. The method also includes transforming the neural network topology to an equivalent analog network of analog components including operational amplifiers and resistors. Each operational amplifier represents an analog neuron of the equivalent analog network, and each resistor represents a connection between two analog neurons. The method also includes computing a weight matrix based on the weights of the trained neural network. The method also includes generating a resistance matrix for the weight matrix. The method also includes pruning the equivalent analog network to reduce the number of operational amplifiers or the resistors, based on the resistance matrix, to obtain an optimized analog network of analog components.

The invention belongs to the technical field of data processing, particularly relates to a server data processing method and system based on deep learning, and aims to solve the problems of low modal fusion efficiency, poor model adaptability and insufficient resource utilization rate in multi-modal data processing. The method comprises the following steps: firstly, preprocessing mixed to-be-processed data and dividing the mixed to-be-processed data into structured data, image data, text data and time series data subsets; a time-varying neural network topological structure is constructed, network edge weights are dynamically adjusted based on server loads and data features, and preliminary extraction and correlation capture of multi-modal features are achieved; generating a fusion feature vector by using a cross-modal feature extractor of the attention mechanism; a dynamic adaptive data processing module is used for matching the deep learning model for parallel processing; and finally, constructing a reinforcement learning reward function based on the processing efficiency, the resource occupancy rate and the accuracy rate, and carrying out joint iterative optimization on the model.

An integrated circuit includes an analog network of analog components fabricated by a method. The method includes obtaining a neural network topology and weights of a trained neural network. The method also includes transforming the neural network topology to an equivalent analog network of analog components including operational amplifiers and resistors. Each operational amplifier represents an analog neuron, and each resistor represents a connection between analog neurons. The method also includes computing a weight matrix for the equivalent analog network based on the weights of the trained neural network. The method also includes generating a resistance matrix for the weight matrix. The method also includes generating lithographic masks for fabricating a circuit implementing the equivalent analog network based on the resistance matrix. The method also includes fabricating the circuit based on the one or more lithographic masks using a lithographic process.

The invention relates to the technical field of medical data analysis, in particular to a spatio-temporal characteristic quantitative evaluation method for Parkinson's disease motion symptoms, which comprises the following steps: deploying an inertial measurement unit to collect three-dimensional acceleration angular velocity magnetic field data, constructing a human body connection structure to generate a connection neural network topological structure, the method comprises the following steps: calculating trajectory direction angle and angular velocity change adaptive weighting to judge stability, aggregating multiple rounds of convolution propagation of adjacent features to form a space-time fusion motion feature set, identifying tremor gait amplitude according to time sequence multi-head attention to extract a time sequence feature mode, calculating tremor gait coordination to generate a Parkinson's disease motion symptom quantitative evaluation result, and calculating a Parkinson's disease motion symptom quantitative evaluation result. According to the method, the limb coordination is captured by constructing sensor network motion data topological connection and fusing multi-dimensional part information, the remote association is captured by keeping the time sequence stable through adaptive weight attenuation and connection convolution depth aggregation features, and the multi-head attention fine recognition tremor frequency and gait change are combined. And the Parkinson's disease symptom identification and evaluation consistency is improved.

Systems and methods are provided for analog hardware realization of neural networks. The method incudes obtaining a neural network topology and weights of a trained neural network. The method also includes transforming the neural network topology into an equivalent analog network of analog components. The method also includes computing a weight matrix for the equivalent analog network based on the weights of the trained neural network. Each element of the weight matrix represents a respective connection between analog components of the equivalent analog network. The method also includes generating a schematic model for implementing the equivalent analog network based on the weight matrix, including selecting component parameter values for the analog components.

The invention discloses a data processing method and system, electronic equipment and a readable storage medium, the method is applied to the data processing system, the data processing system comprises an analog calculation module, a digital calculation module and an analog-to-digital converter, the analog-to-digital converter is connected with the analog calculation module and the digital calculation module, and the analog calculation module is connected with the digital calculation module. The method comprises the steps that a first matrix is calculated according to input data based on an analog calculation module, analog data are obtained, and the first matrix comprises a matrix corresponding to the weight of a neural network; converting the analog data into digital data based on the analog-to-digital converter; based on the digital calculation module, a second matrix is calculated according to the digital data, target data are obtained, and the second matrix comprises a matrix used for describing the neural network topological structure. The analog calculation module and the digital calculation module are combined for calculation, so that the calculation power consumption is greatly reduced under the condition that the calculation precision is hardly influenced.

Systems and methods are provided for generating libraries for hardware realization of neural networks. The method includes obtaining a plurality of neural network topologies. Each neural network topology corresponds to a respective neural network. The method also includes transforming each neural network topology to a respective equivalent analog network of analog components. The method also includes generating a plurality of lithographic masks for fabricating a plurality of circuits. Each circuit implements a respective equivalent analog network of analog components.

The invention provides a tensor-based memory offset calculation method for a memory overlapping degree. The tensor-based memory offset calculation method comprises the following steps: S1, sorting distribution priorities of intermediate tensors of a calculation graph according to the memory overlapping degree of the tensors; s2, searching a memory gap, and searching a corresponding intermediate tensor matched with a proper memory gap; and S3, overall memory allocation scheduling: performing complete scheduling of memory allocation on the neural network topological structure based on the offset calculation strategy of the memory overlapping degree of the tensor in the step S1 and the step S2. According to the method, the memory overlapping degree of the tensor is introduced into a memory offset calculation strategy, memory fragments can be solved to a certain extent, the size of a memory peak value is reduced, memory occupation of an inference engine is minimized, and memory gaps can be reduced to a certain extent.

This invention relates to the field of power grid metering technology, and more specifically, to an optimized neural network topology identification method and system based on a data acquisition terminal. This invention uses a BP neural network model for topology identification. It preprocesses data recorded by the target acquisition terminal itself and data read from all undetermined meters by the target acquisition terminal to obtain a preliminary result of the actual topology. This preliminary result is then used as input to a trained BP neural network model, improving the speed and accuracy of the model output. This solves the problem that existing methods for identifying electricity topology based on ordinary neural networks only involve simple data preprocessing, resulting in only a limited improvement in the speed and accuracy of network calculations.

The present application relates to the technical field of medical data analysis, in particular to a spatiotemporal feature quantitative evaluation method for Parkinson's disease motor symptoms, comprising: deploying an inertial measurement unit to collect three-dimensional acceleration angular velocity magnetic field data, constructing a human body connection structure to generate a connection neural network topology, calculating a trajectory direction angle and angular velocity change adaptive weighting to judge stability, aggregating adjacent features to form a multi-round convolution propagation to form a spatiotemporal fusion motion feature set, identifying tremor gait amplitude according to a time sequence multi-head attention to extract a time sequence feature mode, and calculating tremor gait coordination to generate a Parkinson's disease motor symptom quantitative evaluation result, wherein, in the present application, a sensor network motion data topology connection is constructed, multi-dimensional part information is fused to capture limb coordination, adaptive weight decay and connection convolution deep aggregation features are used to maintain time sequence stability and capture long-range associations, and multi-head attention is combined to distinguish tremor frequency and gait changes, thereby improving the consistency of Parkinson's symptom recognition and evaluation.

The present invention belongs to the fields of artificial intelligence and neurobiology, and particularly relates to a method for learning the constraints of a deep neural network topology. First, the rs-fMRI data of the subject is collected and preprocessed, the correlation coefficients between different brain regions are calculated to obtain the biological brain topology matrix, and a deep neural network model is constructed; then, the learning process of the neural network is constrained by the biological brain topology to train the model, and the backpropagation algorithm is used to update the neural network parameters. The loss function of the backpropagation algorithm simultaneously includes the negative log-likelihood loss and the topology matrix similarity loss. Therefore, after the model is trained, the topology matrix of the neural network will tend to be the biological topology matrix, thereby realizing the technology of directly integrating neurophysiological recordings into artificial neural networks. The present invention fills the technical gap in directly converting neurophysiological recordings into improvements in artificial neural networks, thereby improving the engineering performance of neural networks.

The invention relates to an optical neural network topology adaptive mode division multiplexing communication system and a training method, and the method provided by the invention is applied to a mode division multiplexing communication system, and is used for solving the problem that the transmission or calculation performance is reduced due to dynamic coupling crosstalk of a spatial mode caused by environmental disturbance. The method comprises the following steps: monitoring the output of an optical neural network in real time, and generating a state matrix representing mode crosstalk; extracting matrix features to construct an environment vector; through a pre-trained deep reinforcement learning network, a reconstruction action for controlling the adjustable photonic device is decided and generated according to the vector; the driving device dynamically adjusts the network physical topology to compensate crosstalk; and finally, optimizing the strategy network on line based on the reconstructed performance evaluation result to form a closed loop. According to the method, the optical neural network in the mode division multiplexing has online self-adaptive capability, the dynamic mode crosstalk can be continuously inhibited, and the stability and high performance of the system in actual deployment are guaranteed.

The invention discloses a method for predicting the optimal rotating speed of a motor by optimizing a neural network based on a whale optimization algorithm, and relates to the technical field of whale optimization algorithms. The method comprises the following steps: inputting data and an optimal rotating speed, and carrying out normalization processing on the data; designing a BP neural network topological structure, and initializing parameters of the BP neural network topological structure; initializing parameters of a whale optimization algorithm, improving the whale optimization algorithm by introducing a weight strategy and a nonlinear convergence factor, and optimizing the BP neural network by improving the whale optimization algorithm; outputting an optimal weight and a threshold value of the BP neural network, and carrying out a training test; and constructing a model for predicting the optimal rotating speed by optimizing the BP neural network. According to the method, the whale optimization algorithm is improved based on adaptive weight and nonlinear convergence, the optimization precision and the optimization speed are improved, errors are reduced, the BP neural network is optimized by using the improved whale optimization algorithm, and the accuracy of predicting the optimal rotating speed of the motor is improved.

The invention relates to the field of oil and gas transmission simulation, and discloses an oil and gas transmission system model correction method and system based on data-mechanism hybrid drive, and the method comprises the steps: constructing an oil and gas transmission system mechanism model, and achieving the transmission of each component, each component ratio, temperature, pressure and flow between the oil and gas transmission system mechanism models; constructing a data driving model of the oil and gas transmission system, determining a BP neural network topological structure by using the test set through the data test model, constructing a data prediction model according to the BP neural network topological structure, and obtaining an output characteristic value of the prediction set through the data prediction model; constructing a data mechanism mixed oil and gas transmission system model, and connecting the oil and gas transmission system mechanism model and the oil and gas transmission system data driving model to the data mechanism mixed oil and gas transmission system model to realize transmission of input characteristic values and output characteristic values between the models; and the flow, the gas phase flow and the temperature with the corrected characteristic values are reduced and output through a total output model.

The invention provides an aircraft structure crack damage prediction method based on a grey neural network under random uncertainty factors, and the method comprises the steps: firstly obtaining the structural crack historical data of a to-be-analyzed aircraft, and building a grey model for data prediction according to the structural crack historical data; and then utilizing the obtained grey model prediction sequence to output the structure crack prediction data of the next year through the trained neural network. According to the method, the grey theory and the neural network are combined, the damage of the aircraft structure is predicted through the grey theory under the small sample condition, then the neural network is used for conducting deep learning fitting on the predicted value of the grey theory, the neural network topological structure and the initial weight value are determined according to the grey theory, and therefore the prediction precision is improved.

Systems and methods herein disclose technology, architectures) and constructs of artificial (machine) brains having the feature of separated hardware and software wherein their full functionality is defined by software only. Implementations not only replicate all known functions of a human brain including consciousness but also establish universal, novel construction and characteristics of software-defined brains realizing any possible, known or yet unknown functionality of an intelligent computer. Additional aspects disclose classes of different Neural Network (NN) topologies, based on the universal representation of a software-defined nerons an d / or software-defined neural networks to implement any kind of known or yet unkno w n neural network via software implementations embodied on any universal hardware, which enable subjective and objective phenomena utilized for processing machine consciousness and machine subconsciousness. Various illustrative functional complexes are set forth via the disclosure of a. new class of neural networks, namely the Heterogeneous Hierarchical Model (HHM).

A transmission shaft life optimization method based on a digitalized generation structure comprises the steps that according to original geometry and material attributes, after the rigidity and fatigue life of an original shaft are obtained through finite element reference analysis, a preliminary CAD model is generated through neural network topology; and performing iterative finite element performance evaluation to obtain an optimization model of which the rigidity and the service life meet the requirements for manufacturing. According to the invention, through topology generation driven by the neural network and a multi-objective optimization strategy, rigidity and strength maintenance of the transmission shaft is realized, and structural parameters are automatically and intelligently regulated and controlled through a required life value input by a user, so that light weight is realized, and a test period is shortened.

The invention provides a solid propellant mechanical property prediction system based on artificial intelligence, and the system comprises a data input module which is used for achieving the multi-format import and visual preview of an experiment data file; the feature engineering module is used for executing data standardization, abnormal value elimination and key parameter extraction; the intelligent modeling module is used for configuring an adjustable neural network topological structure, including layer number setting, node configuration, dual-stage activation function selection and training parameter optimization; the prediction output module is used for dynamically displaying a mechanical property prediction value and a training error convergence curve; and the verification and evaluation module is used for providing test set performance index calculation and visual analysis functions. The prediction system supports localized deployment of edge equipment, can optimize a process window in the propellant production process in real time, and greatly improves the prediction precision compared with a traditional prediction method.

The invention relates to the technical field of neural network topology data processing, in particular to a public resource transaction big data analysis method based on heterogeneous data and intelligent portray.The method comprises the steps that the data of the public resource transaction big data are analyzed on the basis of the equity association strength between transaction subjects in the heterogeneous data, the equipment feature consistency and the spherical distance of the geographic position information of the transaction subjects; calculating a static space correlation degree between any two transaction subjects; according to the static space correlation degree and the Euclidean distance between the current feature vector and the historical feature mean vector, calculating a comprehensive risk index between transaction subjects in the current bid; the state transition of the transaction subject portrait label is judged through the comprehensive risk index, so that the accuracy of abnormal behavior recognition in public resource transaction can be effectively improved.

Systems and methods are provided for analog hardware realization of neural networks. The method incudes obtaining a neural network topology and weights of a trained neural network. The method also includes transforming the neural network topology to an equivalent analog network of analog components including a plurality of operational amplifiers and a plurality of resistors. Each operational amplifier represents an analog neuron of the equivalent analog network, and each resistor represents a connection between two analog neurons. The method also includes computing a weight matrix for the equivalent analog network based on the weights of the trained neural network. Each element of the weight matrix represents a respective connection. The method also includes generating a resistance matrix for the weight matrix. Each element of the resistance matrix corresponds to a respective weight of the weight matrix and represents a resistance value.

To provide tile location and / or cycle based weight set selection for base calling.SOLUTION: A system for base calling includes memory storing a topology of a neural network, a plurality of weights sets, and sensor data for a series of sensing cycles. Sequencing events span temporal progression of the base calling operation through subseries of sensing cycles, and spatial progression of the base calling operation through locations on a biosensor. A configurable processor is configured to load the topology on the configurable processor, select a weight set in dependence upon a subject subseries of sensing cycles and / or a subject location on the biosensor, load subject sensor data for the subject subseries of sensing cycles and the subject location on the processing elements, configure the topology using the selected weight set, and cause the neural network to process the subject sensor data to produce base call classification data for the subject subseries and the subject location.SELECTED DRAWING: Figure 1

A neural network topology is selected by generating a super-neural network embedding one or more of a plurality of candidate neural networks having different topologies, and generating at least one learnable transform comprising a plurality of trainable weights. At least one of the plurality of candidate neural networks is generated at least in part by applying the at least one learnable transform to the super-neural network. A candidate neural network is selected from among the plurality of candidate neural networks having different topologies for deployment based on performance metrics and topological constraints.

The invention relates to the technical field of actuator service life early warning, and particularly discloses an electric actuator part service life early warning system based on vibration-torque signal fusion, which comprises a data processing center connected with a sensor group, and the data processing center is configured to execute the following steps: responding to a real-time operation state of an electric actuator, and sending the real-time operation state of the electric actuator; and collecting an original vibration signal, an original torque signal and a multi-physics field auxiliary signal, and generating calibrated multi-physics field data based on multi-physics field data cross validation logic. According to the invention, a cross validation and self-calibration mechanism of multi-physical field data is utilized, the problem of false alarm caused by sensor drift in a severe environment is effectively solved, and the authenticity and reliability of input data are ensured; secondly, by sensing the fluctuation intensity of the working condition and adaptively adjusting the neural network topology structure, the degradation features are accurately captured under the dynamic load, and the interference of the non-stable working condition on fault feature extraction is eliminated.

The present invention provides a fuel cell modeling method based on a BP neural network, comprising the following steps: collecting the output voltages of a fuel cell under different operating conditions and dividing them into a training set and a validation set; performing data preprocessing on the training set and the validation set; respectively optimizing the neural network topologies of a single hidden layer and a double hidden layer by using a grid search method, and selecting the neural network with the smallest validation error; collecting the output voltages of the fuel cell under other operating conditions (operating conditions different from those of the training set and the validation set) and using them as a test set to test the prediction accuracy of the network. The present invention can achieve high-precision prediction of the output voltages of a fuel cell under different operating conditions and provide decision support for system control.

Efficient learning and use of topologies of neural networks in machine learning is disclosed. A mechanism for facilitating learning and application of neural network topologies in machine learning at autonomous machines is described. One method of embodiments as described herein includes monitoring and detecting structural learning of neural networks related to machine learning operations at computing devices having processors; and generating a recurrent generative model based on one or more topologies of one or more of the neural networks. The method can further include converting the generative model to a discriminative model.

The present invention relates to a method for adaptive gain control of a fiber Raman amplifier based on a two-stage neural network, comprising the following steps: obtaining training data; determining a two-stage neural network topology structure based on the training data; training the two-stage neural network using the training data; calculating the pump light power and wavelength at the amplifier's target gain value using the trained first-stage neural network; calculating a predicted amplifier gain value at the pump light power and wavelength output by the first-stage neural network using the trained second-stage neural network; subtracting the predicted gain value obtained by the second-stage neural network from the target gain value to obtain a gain error; and cyclically using a gradient descent method to update the neuron link weights based on the gain error, thereby improving the accuracy of the fiber Raman amplifier's output gain value. The present invention uses the first-stage neural network to calculate the pump light power and wavelength at the amplifier's target gain value, uses the second-stage neural network to calculate the amplifier's predicted gain value, and uses the gradient descent method to update the neuron link weights based on the gain error between the amplifier's predicted gain value and the target gain value, ultimately obtaining a highly accurate amplifier output gain value and the corresponding pump light power and wavelength values.

The invention discloses a metal pipeline corrosion and anticorrosive coating stripping test system, and relates to the technical field of metal pipeline detection, the system comprises: an acquisition unit configured to acquire multi-dimensional test data in real time; the first processing unit is used for constructing a grading test evaluation model; the second processing unit is used for mapping the multi-dimensional test data into feature vectors of nodes and edges; the third processing unit is used for carrying out distributed collaborative optimization and updating and optimizing parameters of each edge side test model; and the output unit is configured to obtain a grading test result according to the optimized model. According to the invention, by introducing a collaborative mechanism of multi-dimensional test perception, hierarchical evaluation modeling, graph neural network topological association and distributed collaborative optimization, high-precision, hierarchical and spatial coupling perception and test evaluation of metal pipeline corrosion and anticorrosive coating stripping are realized.

Some implementations provide devices, systems and / or methods for quantifying vulnerability of an artificial neural network (ANN) to poisoning attacks. Some implementations provide devices, systems and / or methods for reducing vulnerability of an artificial neural network (ANN) to poisoning attacks. Some implementations provide devices, systems and / or methods for detecting poisoning attacks in an ANN. An ANN is trained to generate inferences based on a function.