73 results about "Recurrent neural nets" patented technology

The invention relates to the technical field of computer vision and multi-sensor data fusion, in particular to a multi-sensor image fusion obstacle real-time detection and tracking system, which comprises the following steps of: firstly, fusing data of a camera, a millimeter wave radar and a laser radar, and extracting and fusing multi-modal features; performing multi-target tracking based on a recurrent neural network: generating a target state through space-time modeling, associating a target with a historical track by using an attention mechanism, and maintaining track consistency; and the system performs semantic classification and interaction analysis on the obstacle, predicts the movement track of the obstacle, realizes deep semantic understanding, and finally outputs the identity label and the complete historical track of the obstacle.

The invention belongs to the technical field of aircraft control, particularly relates to a range extender control system for an unmanned aerial vehicle and a controller, and aims to solve the problems of limited endurance, poor flight stability and the like caused by discontinuous energy supply and power response lag. The system comprises a range extender power source module, an electric energy conversion and distribution module, a flight state sensing module, a load power prediction module, a multi-target optimization scheduling module and a closed-loop feedback execution module. Through flight state real-time perception and load power prediction based on a recurrent neural network, an optimal power generation instruction is generated in combination with multi-target optimization scheduling, and precise rotating speed tracking is realized through adaptive PID control. A fuel consumption, power supply smooth switching and battery health combined cost function is introduced, dynamic weighting is carried out according to task types, and the cruising ability and the system robustness are improved; the system has the functions of fault emergency response, power battery hot plug and model online updating, and safety and maintainability are enhanced.

The invention discloses a time-varying and time-lag multi-CSTR system adaptive control method and system based on a neural network, and the method comprises the steps: collecting key operation parameters of a multi-CSTR system in real time, constructing a time-varying and time-lag dynamic model through a recurrent neural network, and updating a time-lag parameter estimation value. And designing a hierarchical neural network structure comprising an upper global coupling model and a lower local compensation model, designing an adaptive algorithm based on a stability theory to adjust the weight of the neural network, and generating a control instruction to drive an execution mechanism. The system correspondingly comprises a data acquisition layer, a time-varying time-delay estimation module and the like. According to the scheme, the problems of time varying, time lag and coupling of the multi-CSTR system are solved, and stable operation of the system is guaranteed.

The invention provides a three-coordinate measuring machine adaptive dynamic error compensation method, and relates to the field of three-coordinate measuring machines, and the method comprises the steps: obtaining multi-source state data of a three-coordinate measuring machine, and calculating a real-time dynamic error, the state data comprising a motion state, a dynamic response and environment disturbance data; based on the real-time dynamic error, using a recurrent neural network to construct a virtual measuring machine model, and performing offline training to obtain an initial error prediction model; and acquiring real-time error feedback data, and performing fine adjustment on the initial error prediction model by using an incremental learning algorithm and a sliding time window mechanism to obtain a prediction model capable of dynamically evolving and aging adaptive parameters. The method is used for overcoming the defect that in the prior art, a linear model or a fixed compensation parameter is difficult to accurately describe and compensate all dynamic errors sometimes.

The invention relates to the technical field of coal-fired unit energy flow analysis and prediction, in particular to a coal-fired unit boiler energy flow analysis and prediction method under rapid variable load, which specifically comprises the following steps of: firstly, combining a thermodynamic physical framework with a recurrent neural network, and representing a function relationship between a time-varying parameter and a state variable in a mechanism model through a neural unit; rapid fluctuation of energy in the boiler is accurately captured; secondly, a multi-mode switching strategy based on an attention mechanism is introduced, flexible switching is carried out between a strong coupling state and a weak coupling state, and the prediction precision under high-speed load change is guaranteed; and finally, introducing an error compensation model, and dynamically correcting a hybrid model prediction residual error. According to the method, physical constraints of mechanism modeling, multi-mode switching of an attention mechanism and an error compensation model are integrated and fused, so that the problems that most of hybrid modeling studies in the prior art mainly focus on steady-state or finite transient prediction, and studies on energy flow dynamic characteristics under rapid load change are still relatively deficient are solved.

Computer-implemented methods and systems provide a persistent fixed-size recurrent memory that supports long or unbounded sequence processing with substantially constant compute and bounded memory. A recurrent model maintains a matrix state updated per chunk from key, value, gate, and optional control signals; a gated error between value and a key-weighted state yields a rank-1 proposal. Chunk proposals are reconciled coordinate-wise by an order-invariant convex combiner; solitary proposals pass unchanged. Queries can be answered from proposals without materializing a full updated state. Multiple asynchronous agents emit sparse updates with overwrite strengths that are merged using sparse convex or max selection and optional optimizer preprocessing. A feedback projection writes higher layer activations into lower layer state to consolidate durable skills and reduce forgetting. The approach enables durable retention, parallel inference-time adaptation, scalable linear-attention approximation, and persistent fixed memory use over arbitrarily long sequences.

The invention discloses a composite wing unmanned aerial vehicle intelligent obstacle avoidance method and system based on deep learning, and the method comprises the steps: responding to an obstacle avoidance task signal, collecting multi-modal environment perception data, constructing a lightweight multi-modal fusion neural network model which comprises an encoder and a decoder, and carrying out the recognition of the multi-modal environment perception data through the encoder and the decoder; fusing the image sequence, the three-dimensional point cloud data and the circumferential sensing data through an encoder to obtain a fused feature tensor, decoding the fused feature tensor through a decoder, outputting a risk result based on an obstacle, inputting the risk result based on the obstacle and flight state parameters into a recurrent neural network controller together, and outputting the flight state parameters. And outputting the unmanned aerial vehicle flight control instruction vector, thereby obtaining an unmanned aerial vehicle control signal, controlling the unmanned aerial vehicle to avoid the obstacle, and obtaining an obstacle avoidance task completion result. The invention provides a composite wing unmanned aerial vehicle intelligent obstacle avoidance technology capable of realizing deep fusion of multi-mode perception, end-to-end intelligent decision and platform customization control.

The invention relates to the technical field of industrial automatic measurement and control communication, and particularly discloses multi-protocol fusion intelligent measurement and control host equipment, which comprises a data acquisition module, a health assessment module, a thermal simulation module and a cooperative control module. The equipment adopts a dual-core hardware architecture with a bottom plate and a top plate, integrates seven communication interfaces, realizes multi-protocol parallel processing through a hardware accelerator and a FreeRTOS system, the protocol switching time does not exceed 10ms, and the health assessment module adopts an online incremental learning recurrent neural network to calculate the health state index of the super capacitor module in real time, so that the super capacitor module can be used as the health state index of the super capacitor module. The thermal simulation module predicts the thermal runaway risk level based on an embedded finite volume method, and the cooperative control module dynamically adjusts the rotating speed of a fan and the switching frequency of a power supply by adopting a Nash equilibrium optimization algorithm, so that early performance attenuation recognition, thermal runaway risk prediction and intelligent cooperative control of the super-capacitor module in an extreme environment are realized. And the reliability and the safety of the system are obviously improved.

The invention relates to the technical field of intelligent question answering, and discloses an enterprise regulation intelligent question answering system and method based on semantic understanding, and the system comprises a data collection layer which is used for obtaining a multi-level complex regulation query statement input by a user; the semantic understanding layer is used for performing word segmentation, dependency syntax analysis and semantic role labeling on a query statement; and the knowledge reasoning layer is used for carrying out multi-step evidence fusion and answer verification through a logical reasoning engine. The method comprises the following steps: S1, acquiring a multi-level complex regulation query statement input by a user; s2, performing word segmentation, dependency syntax analysis and semantic role labeling on the query statement; and S3, analyzing the nested structure statement by using a hierarchical recurrent neural network. By introducing a multi-level semantic understanding mechanism and combining dependency syntactic analysis, semantic role labeling and a level recurrent neural network, deep structure analysis of the composite query statement is achieved, and the real intention which is not shown by a user can be accurately recognized.

A method, computer program product, and computer system of artificial intelligence (AI) intervention to detect and mitigate an abnormality associated with an object moving during performance of a process. A trained recurrent neural network (RNN) determines, from sensor data, a probability (Pr1) that the abnormality existed at a first time, where Pr1 exceeds a threshold T1 and in response, an alternative generative adversarial network (AGAN) determines a probability (Pa1) that the abnormality existed at the first time. A score S1, which is computed as a function of Pr1 and Pa1, exceeds T1 and in response the RNN and the AGAN determines, from the senso data, a probability (Pr2) and a probability (Pa2), respectively, that the abnormality existed at a second time. A score S2, which is computed as a function of at least Pr2, Pa2, and (S1-T1), exceeds a threshold (T2) and in response, the abnormality is mitigated.

A system includes a machine-learning model implemented on a data processing apparatus, which features a parallel processor with a memory hierarchy. The machine-learning model is a recurrent neural network (RNN) with a multi-head architecture, comprising multiple sub- vectors that process parallel data streams. The RNN's weight matrix is structured as a block-diagonal matrix, allowing for parallel processing of the sub-vectors. A fused computational kernel executes an entire time-series processing loop for the multi-head RNN, maintaining the weight matrix blocks in on-chip memory and performing matrix multiplications and element-wise operations for each sub-vector in a single kernel execution.

Two-pass automatic speech recognition (ASR) models can be used to perform streaming on-device ASR to generate a text representation of an utterance captured in audio data. Various implementations include a first-pass portion of the ASR model used to generate streaming candidate recognition(s) of an utterance captured in audio data. For example, the first-pass portion can include a recurrent neural network transformer (RNN-T) decoder. Various implementations include a second-pass portion of the ASR model used to revise the streaming candidate recognition(s) of the utterance and generate a text representation of the utterance. For example, the second-pass portion can include a listen attend spell (LAS) decoder. Various implementations include a shared encoder shared between the RNN-T decoder and the LAS decoder.

The application discloses a GIS equipment state research and judgment method and device based on IMF energy entropy, and the method comprises the following steps: acquiring GIS switch state data and corresponding state label values, and constructing a training set; constructing a time-frequency memory recurrent neural network based on IMF energy entropy; designing a cross-entropy loss function by considering a classification result and a real state label value; training the neural network, stopping the training when the training round reaches a maximum training round or the value of the cross-entropy loss function reaches a minimum, so that a trained neural network is obtained; inputting real-time collected GIS switch state data into the trained neural network to obtain a GIS equipment state research and judgment result; and the application has the advantages that the accuracy of GIS equipment state research and judgment is improved.

This invention relates to a data-driven intelligent logistics supply chain digital management system. The system includes a demand forecasting and analysis unit, an inventory optimization unit, a supplier management unit, a logistics route optimization unit, a system integration and feedback unit, and a real-time logistics marketing unit. By analyzing historical orders and market data through recurrent neural networks and long short-term memory networks, accurate forecasting of short-term and medium-to-long-term demand is achieved. The inventory optimization unit combines a Bayesian dynamic linear model and a convolutional neural network to achieve dynamic inventory management. The supplier management unit optimizes supplier selection using a multilayer perceptron and fuzzy clustering algorithm. The logistics route optimization unit achieves optimal scheduling of logistics routes based on genetic algorithms and an adaptive large-scale neighborhood search algorithm (ALNS). By integrating multiple functional units, this invention effectively improves the operational efficiency and response speed of the logistics supply chain.

The invention discloses a pipeline evaluation and design optimization method, system, equipment and medium integrating flow simulation and machine learning, and belongs to the technical field of oil and gas pipeline safety monitoring and operation optimizing.The method comprises the steps that pipeline operation historical data are collected, a pipeline operation historical data model is established through a recurrent neural network, and feature extraction is conducted; obtaining a first type of features, a second type of features and a third type of features, and outputting a multi-dimensional feature vector; constructing a hybrid model, inputting the multi-dimensional feature vector into the hybrid model, training the hybrid model, constructing a hybrid neural network, performing multi-class risk identification, and outputting a risk prediction result; based on a risk prediction result, establishing a double-layer optimization structure, and performing parameter optimization and self-evolution on the hybrid neural network, and by deeply fusing flow simulation and machine learning and combining multi-dimensional feature extraction with space-time diagram convolution and an attention mechanism, the problem that a traditional data driving method is high in misjudgment rate under extreme working conditions is solved.

The invention provides a machine learning underwater sound propagation prediction method based on parabolic equation hard constraint, and the method comprises the following steps: 1), building a PE model of underwater sound propagation, and expressing the sound propagation as an initial value problem suitable for stepping solving; 2) constructing a recurrent neural network architecture which comprises a PE constraint module and a network correction module; 3) acquiring an underwater sound field data set in an initial distance range as training data; 4) only training the correction network module by using the training data, and learning a correction residual error between the physical model and a real sound field; and 5) recursively predicting sound fields of one or more extrapolation areas outside the initial distance range in a stepping manner by using the trained model starting from a known sound field. Through the unique architecture of'hard constraint physics + data-driven correction ', physical consistency is fundamentally ensured, error accumulation is remarkably inhibited, and high precision and stability superior to those of an existing soft constraint model are shown on a long-distance extrapolation task.

The application discloses a kind of flight ground support data interpolation method and system, method includes: initial flight ground support data is preprocessed, and the adjacent matrix of each node of flight ground support network is obtained;Dimension expansion is carried out to ground flight support data using fully connected neural network to obtain characteristic tensor matrix, first feature extraction and dimension reduction are carried out to characteristic tensor matrix using interpolation recursive neural network, and first characteristic vector is obtained;Second feature extraction is carried out to first characteristic vector using recursive neural network, and the characteristic vector in hidden layer is obtained;The feature vector in hidden layer is down-sampled by transposed convolution neural network, and the original dimension is restored;Depth neural network parameters are trained using loss function, and the interpolation of data is realized using trained depth neural network, and complete flight ground support data is output.The quality of flight support data is improved, the interpolation accuracy of data is more accurate, and the integrity of flight ground support data is guaranteed.

Disclosed are system and methods of dynamically balancing power and performance of a processing unit. The method includes receiving workloads to be processed by the processing unit; classifying, using a recurrent neural network, the workloads according to expected resources to be expended by different functional blocks; identifying a critical path workload based on the classification of workloads; determining a temperature-independent operating frequency for the critical path workload; determining a temperature-dependent operating frequency for the critical path workload based on a junction temperature of the processing unit; adjusting the temperature-dependent operating frequency to a target operating frequency having a value within a predetermined difference from the temperature-independent operating frequency by actively cooling the processing unit; setting a critical path supply voltage based on the target operating frequency; and setting a non-critical path voltage for remaining ones of the workloads at a value less than the critical path supply voltage.

A method, computer program product, and computer system of artificial intelligence (AI) intervention to detect and mitigate an abnormality associated with an object moving during performance of a process. A trained recurrent neural network (RNN) determines, from sensor data, a probability (Pr1) that the abnormality existed at a first time, where Pr1 exceeds a threshold T1 and in response, an alternative generative adversarial network (AGAN) determines a probability (Pa1) that the abnormality existed at the first time. A score S1, which is computed as a function of Pr1 and Pa1, exceeds T1 and in response the RNN and the AGAN determines, from the senso data, a probability (Pr2) and a probability (Pa2), respectively, that the abnormality existed at a second time. A score S2, which is computed as a function of at least Pr2, Pa2, and (S1-T1), exceeds a threshold (T2) and in response, the abnormality is mitigated.

The application provides a grid synchronization control method, device, equipment and medium of a photovoltaic inverter, comprising: collecting three-phase voltage signals at a grid point of common connection in real time, and preprocessing the three-phase voltage signals to obtain q-axis voltage residuals; serializing sampling the q-axis voltage residuals to construct a short-time phase deviation time sequence; inputting the short-time phase deviation time sequence into a pre-trained recurrent neural network model to obtain a predicted phase angle increment; based on the predicted phase angle increment, combining a photovoltaic grid-connected working condition, and calculating a synchronization phase angle; and based on the synchronization phase angle, controlling the grid synchronization phase of the photovoltaic inverter. The phase deviation is predicted through the recurrent neural network, the phase error problem caused by the lag of the traditional PI control is overcome, and the accuracy and stability of the grid synchronization of the photovoltaic inverter are effectively improved.

A calculating unit calculates a semantics set relating to an entirety of state of a recursive neural network satisfying a specification. A determining unit determines whether or not the recursive neural network that is an object of checking satisfies the specification, on the basis of the semantics set and an initial state of the recursive neural network that is the object of checking.

The invention relates to the technical field of coal cutter fault diagnosis, in particular to a coal cutter bearing early degradation detection method and system adopting a spiking neural network. The method comprises the following steps: collecting a bearing acoustic signal in real time and carrying out blind source separation and band-pass filtering preprocessing; converting the preprocessed signal into a logarithmic Mel time-frequency spectrogram, and coding the logarithmic Mel time-frequency spectrogram into a pulse event sequence through peak detection; inputting the pulse sequence into a pulse neural network model comprising a pulse convolutional layer and a pulse recurrent neural network layer, and extracting state features; and judging whether the bearing is degraded early based on the classification or deviation calculation. According to the method, by simulating biological auditory sense and a nerve processing mechanism, utilizing the characteristics of high sensitivity and low power consumption of the pulse neural network to time sequence signals and combining targeted noise reduction and coding, sensitive and accurate recognition of the weak characteristics of early degradation of the bearing in a strong noise environment is achieved, and the method is particularly suitable for intelligent edge monitoring of underground equipment.

A computer-implemented method for customizing a recurrent neural network transcriber (RNN-T) is provided. The computer-implemented method includes synthesizing first domain audio data from first domain text data and feeding the synthesized first domain audio data into a trained encoder of a recurrent neural network transcriber (RNN-T) with an initial condition, wherein the encoder is updated using the synthesized first domain audio data and the first domain text data. The computer-implemented method also includes synthesizing second domain audio data from second domain text data and feeding the synthesized second domain audio data into the updated encoder of the recurrent neural network transcriber (RNN-T), wherein a prediction network is updated using the synthesized second domain audio data and the second domain text data. The computer-implemented method further includes restoring the updated encoder to the initial condition.

The present application provides a control device, more appropriately reduces noise of a sensor signal, and suppresses performance reduction of control based on the sensor signal. The control device as one example of the present disclosure has: a noise reduction processing section that acquires a sensor signal containing noise, and reduces the noise contained in the sensor signal based on a recurrent neural network, wherein the sensor signal is a signal based on an output from a sensor that detects time series data, and the recurrent neural network is trained in a manner of learning a correspondence between a first signal and a second signal, wherein the first signal is a signal corresponding to the sensor signal containing the noise, and the second signal represents the first signal from which the noise is removed; and a control processing section that controls an actuator based on an output from the noise reduction processing section.

The invention discloses an intelligent detection method and system for abnormal SQL statement judgment, and the method comprises the steps: designing a CNN-LSTM dual-channel architecture and a cross-modal attention fusion layer through the introduction of a combined design of an abstract syntax tree, byte pair coding and a recurrent neural network, enabling the two networks to capture local and global features respectively, and carrying out the recognition of abnormal SQL statements. The feature weight is dynamically adjusted through an attention mechanism, deep interactive fusion is realized, a heterogeneous base learner cluster is constructed, a meta learner is constructed by extracting multi-dimensional scene features and introducing a multi-layer perceptron, the base learner weight adaptive to the scene is dynamically generated, and meanwhile, a judgment threshold value is adjusted in combination with actual requirements, so that scene self-adaptive accurate decision is realized; and finally, abnormal SQL detection is achieved. According to the method, the feature fusion defect is solved through a cross-modal dynamic attention mechanism, scene adaptive optimization is realized through a meta-learning-driven dynamic decision framework, and the diversity of the model is enhanced through cooperation of multi-scale embedding and differentiated training strategies.

The invention provides a photovoltaic power prediction method and system based on a recursive Fourier kernel KAN network, and belongs to the technical field of photovoltaic power generation power prediction. Through deep fusion of a recurrent neural network and a KAN network architecture, Fourier series is used as a learnable activation function, and a substantial leap of prediction precision is realized. Specifically, the RNN structure integrated in the model can effectively capture the short-term and long-term time dependence relationship in the photovoltaic power sequence, and overcomes the defects of time sequence data processing of a traditional static model. By utilizing the inherent periodicity of the Fourier series, the model can directly and efficiently learn and fit strong laws such as a daily period and an annual period in the photovoltaic power, and compared with a traditional spline function or a fixed activation function, the extraction of periodic characteristics is more accurate and efficient. The introduced Fourier series provides powerful field priori knowledge for the model, so that the model does not need to learn the basic mode from zero, and the training convergence process is greatly accelerated.

This invention provides a method, apparatus, and system for predicting the microreaction activity of catalysts in catalytic cracking units, belonging to both the chemical industry and the field of artificial intelligence. The method includes: preprocessing variable data to obtain a processed sample set; constructing an initial prediction model for the microreaction activity of catalytic cracking catalysts based on the sample set; optimizing the initial prediction model using a fast descent algorithm to obtain a target prediction model for the microreaction activity of catalytic cracking catalysts; and using the target prediction model to predict the preprocessed test data to obtain predicted values ​​for catalyst microreaction activity. Addressing the problem of limited and low-frequency data related to catalyst microreaction activity in existing systems, this invention utilizes a local anomaly factor algorithm based on feature attributes to identify anomalous data and compensate for missing data, and establishes a catalyst microreaction activity prediction model based on a recursive RBF neural network, thus achieving the prediction of catalyst microreaction activity.

This invention relates to a method for determining system state variables (x) in a fuel cell system (1) at successive analysis moments. (i+1) A computer-implemented method for the fuel cell system (1) wherein the following steps are performed at each current analysis time (i+1): - Determine (S2) one or more operating state variables (y) of the fuel cell system (1) at the current analysis time (i+1). (i+1) ); - Using a trained recurrent neural network, based on the internal state vector (h) of the recurrent neural network at the previous analysis time (i) (i) And based on the running state variable (y) determined at the current analysis time (i+1) (i+1) To determine (S3) the current system state variable (x) (i+1) The internal state vector describes the internal state of the fuel cell system (1); and - based on the currently determined system state variable (x) (i+1) (S3) to run the fuel cell system (1).

This application relates to building units for machine learning models for denoising images and systems and methods using the same. In some instances, a machine learning model can be trained to denoise an image. In some instances, the machine learning model can identify noise in an image of a sequence based at least in part on at least one other image of the sequence. In some instances, the machine learning model can include a recurrent neural network. In some instances, the machine learning model can have a modular architecture including one or more building units. In some instances, the machine learning model can have a multi-branch architecture. In some instances, the noise can be identified and removed from the image through an iterative process.

A computer-implemented method for training a recurrent neural network using forward propagation rather than back propagation through time. The method is particularly suited to training sparse recurrent neural networks, and may be implemented on specialized hardware.