312 results about "Feedforward neural network" patented technology

The invention discloses a large language model accelerator architecture based on a three-dimensional NAND flash memory, and belongs to the technical field of calculation, reckoning or counting. The architecture comprises a three-dimensional NAND flash memory used for executing feedforward neural network calculation; the auxiliary calculation unit is used for executing attention mechanism calculation; the DRAM chip is used for storing attention mechanism related weights and KV cache; and the interconnection resource is used for realizing data interaction among the components. Wherein the three-dimensional NAND flash memory comprises a logic chip and an NAND array chip, the logic chip is used for controlling and executing calculation, and the NAND array chip is used for storing weights and participating in calculation. The invention further provides a scheduling method based on KV cache awareness. The scheduling method comprises decomposition and dynamic allocation of calculation tasks. Through collaborative design of the hardware architecture and the scheduling method, the memory wall bottleneck in large language model calculation can be relieved, the energy consumption is reduced, and the overall calculation performance is improved.

The invention discloses an energy storage rapid compensation method and system based on track overhead line system voltage fluctuation feature recognition, and relates to the technical field of track traffic traction power supply. The method comprises the following steps: synchronously acquiring voltage and current signals of the overhead line system, and extracting a composite feature vector containing disturbance root attributes and transient change rate; on the basis of the vector, a transient energy vacancy sequence covering the ultra-short-term future is output in real time through a prediction model fusing dynamic phasor analysis and a feedforward neural network; according to the sequence, a cooperative control strategy fusing overshoot and active damping is generated in combination with a proximity compensation principle and a voltage recovery state so as to drive an along-line energy storage unit; the strategy is executed in advance, compensation energy is injected, and finally double closed-loop correction is conducted based on the voltage residual error. According to the method, the problems of response lag, unclear disturbance identification, poor collaboration and lack of adaptive ability in the prior art are solved, and advanced, accurate, collaborative and self-optimized rapid suppression of the voltage fluctuation of the overhead line system is realized.

The invention provides an optical element surface type simulation method based on a physical information neural network, and relates to the technical field of element design and simulation, and the method comprises the steps: constructing a feedforward neural network; generating a deformation control equation model library; constructing a three-dimensional simulation database; forming a machine learning parameter library; training the PINN deformation simulation model in combination with a deformation control equation model library, a three-dimensional simulation database and a machine learning parameter library, and screening out a target PINN deformation simulation model meeting a preset condition; inputting the design parameters, the support structure parameters and the given external load parameters of different combinations into the target PINN model to obtain a key performance prediction result of the optical element; according to the method, the key performance of the element under various parameter combinations can be rapidly evaluated, the design iteration period is shortened, the trial and error cost is reduced, direct data support is provided for supporting structure optimization design, and the bottleneck problem that a traditional method restricts the research and development efficiency of a high-precision optical system is effectively solved.

The invention discloses a flow field prediction-oriented full-process automatic joint reduced-order modeling method, which comprises the following steps of: specifying a target physical field parameter space, and randomly generating a sample space according to a Latin hypercube sampling method; constructing a full-process automatic simulation tool chain, driving target physical field numerical calculation and generating a training data set; carrying out singular value decomposition-based intrinsic orthogonal decomposition on the output physical field data, and only retaining first r main feature components to construct a reduced-order data set; constructing a multi-input multi-output full-connection feedforward neural network, and modeling and training a nonlinear mapping relation between input parameters and reduced-order features; new working condition parameters are input, reduced-order features are predicted through the trained neural network, distribution of a target physical field is reconstructed according to a singular value decomposition reduction matrix, and more flexible and reliable technical support is provided for reducing the training cost of a reduced-order model and improving simulation efficiency.

The invention relates to a bearing fault diagnosis method based on variational mode decomposition and time sequence partitioning cross attention fusion, which comprises the following steps: acquiring an original vibration acceleration signal of a rolling bearing, and constructing a standardized original data set; segmenting the standardized original data into a plurality of data blocks, and generating a time domain embedding feature; based on the time domain embedded features, extracting high-order global time domain features by using a multi-head self-attention mechanism, residual connection and a feedforward neural network; decomposing the standardized original data into a plurality of intrinsic mode functions, and extracting frequency domain distribution features through a convolutional neural network; taking the frequency domain distribution characteristics as query vectors, retrieving and matching related fault context information in the global time sequence characteristics, realizing weighted fusion of time-frequency modes, inputting fused fault representation vectors into a classifier, and calculating a result of a bearing health state; and constructing a loss function containing label smoothing and a dynamic learning rate scheduling strategy, and carrying out iterative optimization on model parameters until the model converges.

The invention discloses an electric power system large-scale language model tuning method, system and device based on LoRA technology optimization and a medium, and belongs to the technical field of electric power system large-scale language models, and the method comprises the steps: obtaining historical language data formed in the operation process of an electric power system, and converting the historical language data into a training sample; constructing a task data set based on the training sample; constructing a pre-training language model; receiving a to-be-processed input text, extracting semantic features of the input text, and generating an output sequence corresponding to the input text. According to the method, the low-rank weight path is inserted in the feed-forward neural network in a parallel form, and an independent path is bound for each type of tasks, so that differential fine tuning of the electric power semantic tasks is realized on the premise of not modifying the structure and parameters of an original model, the overall tuning cost of the model is reduced, and the stability of the original model is kept.

The invention provides a power short-term load prediction system based on multi-source information adaptive fusion and a working method thereof, and the system comprises a multi-view data construction module which is used for organizing heterogeneous input data into a plurality of independent semantic views; the spatial dependence coding module is used for learning spatial dependence among nodes through a graph attention network and generating spatial embedding representation; the time dynamic modeling module is used for processing the time sequence view through a recurrent neural network and generating time feature embedded representation; the auxiliary information modeling module is used for processing other auxiliary views through independent feedforward neural network processing and generating auxiliary feature embedded representation; the cross-view attention fusion module is used for carrying out adaptive weighted fusion on all the embedded representations so as to inhibit representation degradation caused by low-quality views; and the load prediction module is used for generating a load prediction result based on the fusion representation and carrying out end-to-end training by taking a mean square error as a target.

The invention belongs to the technical field of image recognition, and provides a low-illumination target detection method, which comprises the following steps of: improving a model structure by taking YOLOv11n as a reference model, and designing a noise frequency band sensing spectrum feedforward neural network module, an aliasing sensing frequency reconstruction up-sampling module and an illumination sensing multi-attention fusion detection head; wherein the noise frequency band perception spectrum feedforward neural network module introduces a frequency domain adaptive modulation and illumination perception double-gating mechanism, the aliasing perception frequency reconstruction up-sampling module introduces a frequency perception and reconstruction mechanism, and the illumination perception multi-attention fusion detection head introduces an illumination perception gating mechanism; using an ExDark low illumination data set to train an improved network model; and performing low-illumination image detection by using the trained model. According to the method, the high reasoning speed can be kept in a low-light environment, and the detection capability and the positioning accuracy of weak visible targets, fuzzy edge targets and small-scale targets are improved.

The invention discloses a star-sky dim and small target detection method fusing coordinate attention and a convolutional feedforward neural network, and belongs to the technical field of dim target detection. The method comprises the steps of obtaining a starry sky image data set, outputting an enhanced feature map by applying a coordinate attention mechanism, constructing a network model fusing the coordinate attention mechanism and a ConvFFN module through the ConvFFN module of a convolutional feedforward neural network, training a verification model based on simulation data, and finally outputting the centroid position and bounding box information of a target. According to the method, features are extracted in the transverse and longitudinal directions through a coordinate attention mechanism, the feature expression ability of a weak and small target in a low signal-to-noise ratio environment is effectively enhanced, the detection precision is remarkably improved, the omission ratio is reduced, the calculation efficiency is improved by combining the lightweight structure of ConvFFN and deep convolution optimization, and the method is suitable for large-scale popularization and application. A reliable and efficient detection solution is provided for astronomical observation and space debris monitoring.

The invention discloses a large language model machine forgetting algorithm based on representation spatial offset. The algorithm comprises the following steps: constructing a forgetting set and a retention set, executing causal tracking on a large model by using a knowledge exploration data set, and identifying a feedforward neural network having a significant contribution to correct prediction of the model to determine a forgetting layer; estimating an input feature space by using the reserved set approximation, and constructing a null-space projection matrix based on the feature space; inputting a forgetting set and a retention set, obtaining representation output of the training model and the original model in the last forgetting layer, and exporting a parameter updating gradient through a loss function; and carrying out null-space projection transformation on the updated gradient and then updating the last linear transformation matrix of the feedforward neural network in all forgetting layers. According to the method, the knowledge storage characteristics of the feedforward neural network in the large model are utilized, knowledge removal is guided in the representation space, interference on non-target knowledge is effectively suppressed while accurate forgetting is achieved, and the effectiveness and controllability of the forgetting process of a large model machine are remarkably improved.

The invention discloses a multi-task electromagnetic model based on a hybrid expert network, and belongs to a wireless communication technology. The model comprises a preprocessing module, a feature extraction module, a task output module and a pre-training-fine tuning learning strategy. The preprocessing module carries out standardization processing on the multi-source electromagnetic signals; the feature extraction module is based on a Transform structure, introduces a hybrid expert network to replace part of a traditional feedforward neural network, and dynamically selects an expert sub-network through a task specific routing mechanism; the task output module configures a special structure according to different task targets; in the pre-training stage, a mask auto-encoder is used for pre-training large-scale label-free data, and a downstream task is subjected to full-amount fine adjustment through small-scale label data. According to the method, multi-task collaborative learning and differential expression are realized, and the recognition performance, robustness and processing efficiency of the model in a complex electromagnetic environment are improved.

The invention discloses a current prediction control method and device for a permanent magnet motor, and the method achieves the real-time prediction of a voltage error caused by a dead zone and other non-ideal factors through the powerful nonlinear fitting capability of a feedforward neural network in combination with the learning of historical current time sequence characteristics. And the prediction error is further fed forward and compensated to a future current prediction model to improve the accuracy of current prediction, so that the cost function evaluation is more accurate, the optimal voltage vector is selected, and the steady-state current error and the low-speed torque ripple are effectively inhibited. In this way, precise control over the current of the motor is achieved, and the control precision and operation stability of the driving system are comprehensively improved.

The invention relates to the technical field of sodium-ion batteries, and discloses a sodium-ion battery state-of-charge estimation method and device and a readable storage medium. The method comprises the following steps: acquiring the state of the sodium-ion battery and determining an initial charge state; collecting actual operation data, preprocessing the actual operation data to obtain battery data, and determining battery change data according to the battery data; determining the actual available capacity and the corrected internal resistance of the sodium ion battery according to the temperature data; determining a charge state estimation value according to the initial charge state, the actual available capacity and the current data; according to the battery data, the battery change data, the actual available capacity and the corrected internal resistance, utilizing a pre-trained feedforward neural network to predict a charge state correction value; and determining a final charge state according to the charge state estimation value and the charge state correction value. The precision, the real-time performance and the adaptability of sodium ion battery state-of-charge estimation are effectively improved, and the requirements of various application scenes are met.

The invention relates to an unmanned aerial vehicle attitude control method based on motor state disturbance compensation, belongs to the technical field of unmanned aerial vehicle control, and solves the problems of hysteresis and limited control effect of an existing unmanned aerial vehicle attitude control method. The position controller calculates to obtain expected tension, a roll angle and a pitch angle; the attitude feedback controller calculates a feedback torque according to the expected attitude, the actual Euler angle and the body angular velocity; the load identification module calculates an estimated disturbance torque based on the target rotating speed, the actual rotating speed and the current of each motor; the feedforward neural network module adopts a trained RBF neural network to calculate torque residual error estimation, and torque synthesis is carried out to obtain a torque feedforward compensation amount; adding the feedback torque and the feed-forward compensation amount to obtain a total control torque; and the mixing controller calculates the target rotating speed of each motor according to the expected tension and the total control torque, and finally the control distributor drives the motors to execute. The attitude stability and the control response speed of the unmanned aerial vehicle in the wind disturbance environment are remarkably improved.

The invention discloses a three-dimensional visual positioning method based on uncertainty relation perception Transform, and the method comprises the steps: inputting three-dimensional point cloud data, extracting point cloud features through a Point Net + + backbone network, and generating seed points; generating voting points by adopting an uncertainty perception voting strategy, and obtaining candidate areas in combination with a probability offset refinement strategy; the prompt statement is input into a prompt word learning module for semantic enhancement, text features are extracted, and corresponding text categories are generated; constructing a relationship perception Transform module, modeling a multi-dimensional spatial relationship between the candidate regions, realizing cross-modal fusion of the candidate regions of the visual end and text features of the semantic end, and outputting fusion features; and predicting the confidence of the candidate region and the position of the bounding box based on the feedforward neural network, and outputting a target positioning region which is most matched with the semantics. According to the method provided by the invention, the accuracy and robustness of three-dimensional visual positioning in a multi-element complex scene can be effectively improved.

The invention provides an emotion recognition method, device and equipment based on electroencephalogram signals and a medium. The emotion recognition method based on the electroencephalogram signals comprises the steps that original electroencephalogram signals under all electrode positions are obtained; electroencephalogram characteristic data are extracted according to the original electroencephalogram signals; extracting frequency band and spatial features from the electroencephalogram feature data to obtain a space-frequency feature map; a bidirectional Mama network is adopted to extract time sequence characteristics of the space-frequency characteristic pattern in each time period, and a time sequence characteristic sequence is obtained; fusing each time sequence feature in the time sequence feature sequence on a time period dimension to obtain a fused time sequence feature; and recognizing the fusion time sequence features by using a feedforward neural network to obtain an emotion recognition result. According to the emotion recognition method and device based on the electroencephalogram signals, the equipment and the medium, deep feature fusion can be achieved, hidden features related to emotion changes in the electroencephalogram can be fully mined, and the accuracy of emotion recognition is improved.

Some implementations relate to methods, systems, and computer-readable media for adapting simulated character interactions to different morphologies and interaction scenarios. The system accesses a graph representing a control policy for a simulated character's movements in a virtual environment. This graph undergoes encoding and processing through a graph neural network to generate latent embeddings for the graph. A fixed-length latent vector is determined from the latent embeddings. This vector is input to a feedforward neural network, generating control signals for the character's actions. Through a reinforcement learning loop, the character's motions are continuously refined by iteratively adjusting the graph based on evaluating the actions of the simulated character via a reward function, adapting the control policy to different character morphologies and / or interaction scenarios.

The invention discloses an electric power intra-day price prediction method based on dynamic holiday weight and multi-source fusion, and the method comprises the steps: obtaining multi-source historical data, and carrying out the time synchronization processing; holiday and festival time information is acquired, and a dynamic weight is generated based on the influence of holidays and festivals and upstream and downstream dates on the power load and the electricity price; constructing a multi-dimensional predictive factor matrix based on the multi-source historical data after time synchronization processing; and according to the multi-dimensional predictive factor matrix, on the basis of a collaborative optimization multi-model combination comprising a feedforward neural network model and a bagged regression tree model, intra-day joint prediction is executed, the intra-day joint prediction refers to a process of predicting the power load and the electricity price hourly, and an hourly prediction result of the power load and the electricity price is output. A dynamic holiday weight mechanism is introduced, a multi-source fused high-dimensional predictive factor matrix is constructed, and a multi-model combined predictive strategy of collaborative optimization of a feedforward neural network and a bagged regression tree is adopted, so that the precision and stability of intra-day electricity price and load prediction of the electricity market are improved.

The invention discloses a mechanical product multi-field simulation and prediction system based on digital twinning, and relates to the technical field of thermal error digital twinning. A main shaft load current, a rotating speed, a bearing temperature, a workpiece surface temperature and a cooling liquid parameter are obtained in real time; rapidly predicting a first thermal error according to the load current and the rotating speed by using a preset multi-layer feed-forward neural network, dynamically calculating a thermal inertia time coefficient in combination with the cooling liquid efficiency and the thermal state of a workpiece, and filtering and extracting a slowly-changing thermal state from the temperature of a main shaft bearing according to the thermal inertia time coefficient, thereby obtaining second thermal error prediction; a fusion weight is generated according to the load current change rate and the thermal inertia time coefficient, weighted fusion is carried out on the two predicted values, and a compensation instruction is output; and by monitoring the stability of the fusion weight, the thermal inertia time is reversely and dynamically adjusted to suppress oscillation. According to the method, high-robustness and self-adaptive accurate compensation of complex thermal errors is realized in field processing with changeable environments and working conditions, and the long-term stability of processing precision is remarkably improved.

The invention discloses a parameter optimization method and system based on an agent model and multi-algorithm gradient collaboration, and relates to the technical field of nuclear engineering and nuclear security, and the method comprises the steps: firstly determining a nuclear facility parameter combination, carrying out the high-fidelity calculation through a Monte Carlo program to obtain a Keff value, and constructing a training set; training a micro-proxy model capable of predicting a Keff value and gradient information through a feedforward neural network in combination with a back propagation algorithm, taking the micro-proxy model as a target function, taking the minimum absolute error of a Keff predicted value and a target value as an optimization target, and generating candidate parameter solutions through global exploration of a hybrid optimization algorithm; and an optimal parameter combination candidate solution is obtained through gradient-assisted local optimization, and finally, through Monte Carlo program verification, if the error reaches the standard, the optimal parameter combination candidate solution is used as a final parameter combination solution. The problem that a parameter optimization method in the prior art cannot give consideration to calculation efficiency and precision is solved.

The invention provides a cola production process-oriented Chinese herbal medicine active component deep learning optimization extraction control method, and belongs to the technical field of adaptive control systems. The method comprises the following steps: establishing a staged sequential extraction model, and respectively designing distillation and water decoction processes for volatile and non-volatile components; a hybrid intelligent controller based on LSTM and a feedforward neural network is constructed, and real-time optimization control of process parameters is realized; designing a multi-objective comprehensive reward function, fusing extraction efficiency, product quality, energy consumption and taste indexes, and performing model training through reinforcement learning; a layered online detection system is configured, and dynamic monitoring and data fusion of key indexes are achieved; executing staged sequential optimization control, and adjusting process parameters based on real-time data to realize efficient extraction of active components; and establishing an edge and cloud collaboration mechanism, and continuously optimizing a control strategy through federal learning. According to the method, the extraction efficiency, the control precision and the product quality are remarkably improved.

The invention relates to the field of reinforcement learning, in particular to a reinforcement learning-based large language model antagonism fine-tuning enhancement system. The state prediction module is used for acquiring system state data and robot action data; based on the robot action data, generating a prediction state vector through a neural network prediction model, and comparing the prediction state vector with the system state data to obtain a residual vector; the action decoding module inputs the residual vector into a feedforward neural network, codes the residual vector to generate a natural language state report, and performs autoregression decoding through a large language model to obtain a loss value; the strategy optimization module converts a penalty function value into a reinforcement learning reward signal through a nonlinear function, and the PPO algorithm calculates the total loss according to the reinforcement learning reward signal; and calculating, optimizing and updating the gradient of the low-rank adaptive parameter in the large language model based on the total loss. According to the invention, through the feedforward neural network and the PPO algorithm, the tiny dynamic deviation is captured, and the predictability and safety of the system are improved.

The invention provides a residual error enhanced lithium ion battery health state large model evaluation method and system, and relates to the technical field of lithium ion battery state evaluation. The method comprises the following steps: acquiring original monitoring data of a full life cycle charging and discharging process of the lithium ion battery, and performing denoising and standardization preprocessing to obtain a high-quality time sequence data set; constructing a hierarchical time sequence token coding mechanism, and converting the multi-dimensional battery cycle characteristic sequence into a semantic token sequence which can be processed by a large model; constructing a residual enhanced feature mapping network; constructing a time sequence reasoning trunk based on a large language model architecture, and capturing a long-range degradation dependency relationship by using rotation position coding and a multi-head self-attention mechanism; and outputting a health state evaluation value of the lithium ion battery through the feedforward neural network and the regression head. According to the method, the defects of a traditional deep learning method in the aspect of capturing long-range dependence and generalization ability are overcome, end-to-end accurate SOH estimation is achieved, and the lithium ion battery health state evaluation value can be directly and accurately output.

The invention discloses a centrifuge full life cycle intelligent analysis and operation and maintenance method. The method comprises the following steps: constructing a health scoring model which is based on an SAE-SOM neural network and comprises wavelet denoising and future parameter prediction functions; constructing an equipment life prediction model; a data set obtained from detection equipment is subjected to problem sample screening and normalization operation, and then a feedforward neural network is trained. And the hyper-parameter of the feedforward neural network is determined through a genetic algorithm. And a fault tracing model is constructed. The system is composed of a feature extraction network, a middle layer, a relation measurement network and a fault classification network. And establishing an equipment operation environment cross validation module. According to the invention, through mutual cooperation of the equipment health assessment module, the whole machine life estimation module, the fault alarm module, the early warning and traceability module, the operation environment cross validation module and the warning module, fusion analysis is carried out on key operation parameters of the centrifugal machine and external environment data; the method can be widely applied to water plants and other complex industrial scenes with high requirements for the operation reliability of key equipment.

The invention discloses an epileptic seizure detection system based on a multi-dimensional hypergraph fusion network, which comprises a multi-modal data collector, a signal preprocessor, a feature extractor, a multi-dimensional hypergraph builder, a hypergraph feature extractor, a feature fusion device, a classifier and a time sequence prediction corrector, the multi-dimensional hypergraph constructor constructs three hypergraph structures, namely an intra-modal hypergraph, an inter-modal hypergraph and a time sequence hypergraph; the feature fusion device integrates the features extracted by the three hypergraphs to generate a comprehensive feature vector; the classifier classifies the fusion features through a two-layer feedforward neural network, and outputs probability prediction of epileptic seizure; the time sequence prediction corrector applies a majority voting smoothing filter to carry out post-processing on a classification result, isolated misclassification is eliminated, and the time sequence consistency of prediction is enhanced. The invention provides an innovative multi-dimensional hypergraph fusion network, and the robustness and accuracy of epileptic seizure detection are remarkably improved by simultaneously modeling high-order relationships in modals, between modals and time sequence dimensions.

The invention relates to the technical field of industrial automation control, in particular to an automation control device based on a neural network algorithm model, which comprises a signal acquisition module, a neural network control module, a driving output module and an online learning module, one end of the signal acquisition module is detachably connected with an industrial field sensor, the other end of the signal acquisition module is electrically connected with the neural network control module and the online learning module respectively, and the signal acquisition module is used for acquiring a process variable PV of a controlled object and a set value SP set by a user in real time; according to the method, a nonlinear time-varying system can be effectively adapted without establishing an accurate mathematical model of a controlled object through the nonlinear mapping characteristic of the multi-layer feedforward neural network. Experiments prove that in a thermostat temperature control scene, compared with a traditional PID controller, the control precision of the device is improved by 35%, the steady-state error is smaller than or equal to + / -0.2 DEG C, the overshoot is smaller than or equal to 3%, and the steady-state error of + / -0.5 DEG C and the overshoot of 8% of the PID controller are far better.

The invention relates to a commodity sales volume prediction method and device based on multi-modal feature fusion and a storage medium, and is applied to the technical field of store commodity sales volume prediction, and the method comprises the steps: obtaining static and dynamic multi-modal features affecting the store commodity sales volume, and carrying out the unique coding representation of a store and a commodity; the feature processing layer converts unique coding representation into dynamic embedded features and then performs weighted fusion on the dynamic embedded features and static features through feature gating, the feature gating has the advantage of dynamic weight distribution, and adaptive adjustment cannot be performed according to input sample features due to the fact that static weight distribution is adopted in a traditional method; then, the dynamic features continue to be fused through a cross attention mechanism, and soft alignment and weighted fusion of cross-modal features are achieved through the attention mechanism; then, the incidence relation between fusion features is learned through a multi-layer feedforward neural network of a calculation layer; and inputting the final fusion feature into a task layer to obtain a sales prediction result.