50 results about "Negative log likelihood" patented technology

The present application relates to the technical field of data processing, and particularly relates to a power monitoring method and system for remote electric energy meter, the method comprising: constructing a time series factor graph model with voltage and current phasor as hidden variables and original electric parameter data as observation nodes; performing synchronous compression wavelet transform on current phasors in the electric energy state sequence to generate a time-frequency energy distribution graph and obtain a monitoring feature vector sequence; inputting the monitoring feature vector sequence into a deep auto-encoder pre-trained on normal power consumption working condition data to calculate a reconstruction error, simultaneously calculating a Lyapunov index of the sequence within a preset time window, and obtaining a negative log-likelihood probability according to a pre-established Gaussian mixture model describing the distribution of the index under normal working condition; and weighting and fusing the reconstruction error and the negative log-likelihood probability to generate a comprehensive abnormality index for judging power consumption events. The present application can realize high-precision and low-false-alarm-rate detection of power consumption events.

A chemical fertilizer recommendation interpretation generation method based on a large language model comprises the steps of obtaining historical interaction data and comment texts of a user for chemical fertilizer commodities, and extracting explicit preferences and implicit preferences of the user; encoding the commodity text information and the user behavior sequence by using a pre-training language model, and constructing commodity embedding and user interest representation; inputting the representation into an expert hybrid model for multi-modal feature alignment, and generating collaborative semantic embedding; the embedded information is injected into each layer of attention mechanism of a large language model; adopting a multi-stage generation strategy to sequentially generate abstract description, demand alignment, detail supplementation and feasible action suggestions, and outputting a structured recommendation explanation text; and the training language model is optimized by combining a character-level negative logarithm likelihood loss function and a low-rank self-adaptive fine tuning method in a training process. The method improves the individuation, readability and scientificity of recommendation explanation, has good deployment efficiency and practical application value, and is suitable for an intelligent agricultural recommendation system.

The invention discloses a quantitative rainfall estimation method based on a Hurdle-IMDL framework. The method comprises the following steps: acquiring historical rainfall measurement data and meteorological satellite observation data; constructing a biased quantitative rainfall estimation probability model based on a Hurdle model, correcting a biased rainfall estimation probability sub-model according to an IMDL method, substituting into empirical distribution to construct an empirical quantitative rainfall estimation probability model, and deriving a negative logarithm likelihood function of the empirical quantitative rainfall estimation probability model; constructing an AI model, and optimizing the AI model by taking a negative log-likelihood function as a loss function; and inputting meteorological satellite observation data of a to-be-inverted region into the trained AI model to obtain an estimated value of a parameter of the empirical quantitative precipitation estimation probability model, and estimating the precipitation according to conditional expectation. According to the invention, the Hurdle model is used to solve the zero expansion problem, the IMDL learning method is used to cope with the long tail problem, and the inversion accuracy of extreme rainfall is improved.

The invention discloses a time sequence neural network tire force estimation method and system under an unstructured road surface, and belongs to the technical field of vehicle dynamics control. The method comprises the steps of collecting vehicle state information under a non-structural road surface by constructing a vehicle information collection platform; carrying out cleaning, interpolation, normalization, time sequence completion and mask preprocessing on the vehicle state information; an LSTM model fused with the attention mechanism is constructed, and an output layer of the model outputs an estimated mean value and an estimated variance of tire force at the same time to quantify estimation uncertainty; performing network training by adopting a negative log-likelihood loss function; and automatically optimizing the network hyper-parameters by using a Bayesian optimization method. The method effectively solves the estimation problem caused by strong time sequence dependence, obvious nonlinearity and obvious hysteresis effect of the tire force under the unstructured road surface, does not need to additionally install a tire internal force sensor, is low in cost, high in precision and strong in generalization ability, and remarkably improves the performance of a vehicle chassis control system under the complex road condition.

The application discloses a rotating machinery fault diagnosis method and system based on a Gaussian boundary constraint network, relates to the technical field of fault diagnosis, and obtains vibration signal data of rotating machinery under a single source domain working condition to form a source domain training set with labels; based on the source domain training set, a Gaussian boundary constraint network is trained to obtain a trained Gaussian boundary constraint network; the Gaussian boundary constraint network comprises a feature extractor, a feature decoupling module, a Gaussian boundary constraint module and a known category prediction head, the Gaussian boundary constraint module comprises a Gaussian classifier, and the Gaussian classifier trains the input first features through a negative log-likelihood loss and a confidence-aware boundary loss; vibration signal data of the rotating machinery is obtained, input into the trained Gaussian constraint network for fault diagnosis, and a fault diagnosis result is output. The application can process various label inconsistency scenarios, and realizes accurate identification of shared categories and robust rejection of unknown categories.

The invention provides a large language model staged pre-training method and system. The method comprises the following steps: training a Transform model by using a basic data set, optimizing a negative logarithm likelihood target, and adopting an AdamW optimizer and a cosine attenuation learning rate; continuing training by using the universal knowledge data set based on the first-stage parameters; and weighting the professional data in the training field by adopting an oversampling strategy. By structuring a training target and a data type, the model can efficiently learn language basis, general knowledge and professional skills in stages. Experiments show that according to the method, the training efficiency of the model in the basic stage is improved by 40%, the overall training time is shortened by 30%, and meanwhile the accuracy rate of tasks in the professional field is 15%-20% higher than that of traditional end-to-end training. Finally, model parameters are evaluated through professional ability, and both universal language understanding and domain specialty are achieved.

A system includes a controller configured to receive, at a trained transformer model, one or more run-time logs indicating information associated with interaction between a unique device and a server, output a user-score associated with the unique device and the one or more run-time logs in response to determining a negative log-likelihood of the one or more run-time logs in a next-log prediction probability distribution modeled by the trained transformer model, output a server-score utilizing at least normal-cluster centers associated with the trained transformer model and the one or more run-time logs, and in response to a sum of the user-score and server-score exceeding a threshold, outputting an indication of a cyber-threat associated with the one or more run-time logs.

This invention discloses a spacecraft pose estimation and uncertainty modeling method based on intrinsic space, belonging to the field of spacecraft pose estimation technology. The method involves directly constructing a hierarchical Bayesian probability model on the rotating manifold SO(3) and translation space, using Fisher and Gaussian distributions respectively, and introducing conjugate priors to achieve the fundamental decomposition and quantification of accidental and cognitive uncertainties. An end-to-end multi-task neural network is used to jointly learn the pose probability model parameters, key points, and segmentation information, and an iterative optimization module is employed to improve estimation accuracy. During the training phase, marginal negative log-likelihood and evidence regularization are jointly optimized; during the inference phase, the probability distribution of pose prediction is obtained through analytical marginalization, and the two types of uncertainty are distinguished. This invention improves pose estimation accuracy while outputting well-calibrated uncertainties, providing a reliable basis for the autonomous and safe operation of spacecraft in orbit.

The invention discloses a single cell chemical perturbation transcription response prediction method based on a generative neural network and application thereof, the generative neural network of a prediction model constructs a learnable conditional mapping function, and multi-modal embedding is constructed during prediction model training. A training result is optimized by adopting a zero-expansion Gaussian negative logarithm likelihood loss function; the multi-modal embedding is the key input of a conditional mapping function and comprises gene expression semantic features, multi-source biological priori knowledge and drug molecular structure semantic features, so that the learning ability of the conditional function on the true disturbance law of the known drug structure on the gene expression of the known cell type can be remarkably improved; the trained prediction model can break through the limitation that traditional drug reaction research is highly dependent on experimental conditions, low in flux and high in cost, and prediction single cell perturbation transcription response spectrums without drugs are obtained according to structural semantics of new drugs or expression semantics of new cells in a zero sample scene.

This invention discloses a method and system for assessing tumor content in liquid biopsies. The method performs quality control on high-throughput sequencing data from body fluid samples, filtering germline variations, clonal hematopoietic-related variations, and sequencing errors to obtain copy number variation (CNV) information and somatic mutation characteristic information. A weighted mean squared error loss function is constructed based on the CNV information, and the tumor content is solved using a constrained gradient descent method. When the CNV signal does not meet preset judgment conditions, the system switches to the mutation characteristic module, constructing a negative log-likelihood function based on a binomial distribution based on the principal clonal cluster, and solving for the tumor content through iterative optimization. In extreme cases, a backoff mechanism is triggered. This invention combines the stability of CNV at high tumor content with the sensitivity of mutational VAF at low content, achieving adaptive assessment from low to high tumor content.

This application discloses a method and system for detecting circuit board defects based on multi-band impedance matching, relating to the field of circuit board quality inspection technology. The method includes: determining the corresponding test bias voltage of a reconfigurable impedance matching network at multiple test frequencies based on a frequency bias voltage relationship table; applying a test bias voltage matching the test frequency to each adjustable element for each test frequency, and injecting an RF signal into the circuit board under test to acquire the corresponding reflection coefficient through a test probe; reconstructing a three-dimensional scattering image inside the circuit board based on a multi-frequency sparse dictionary and compressed sensing algorithm; calculating the negative log-likelihood distribution corresponding to the three-dimensional scattering image, and combining the spatial information of each circuit physical layer to obtain the hierarchical anomaly score of the corresponding circuit physical layer, thereby screening at least one abnormal circuit physical layer. Thus, by using signal feature reconstruction instead of high-precision physical imaging, hierarchical spatial identification of circuit board defects is achieved.

The invention discloses a rotating machine fault diagnosis method and system based on a Gaussian boundary constraint network, and relates to the technical field of fault diagnosis, and the method comprises the steps: obtaining vibration signal data of a rotating machine under a single source domain working condition, and forming a marked source domain training set; training the Gaussian boundary constraint network based on the source domain training set to obtain a trained Gaussian boundary constraint network; the Gaussian boundary constraint network comprises a feature extractor, a feature decoupling module, a Gaussian boundary constraint module and a known category prediction head, the Gaussian boundary constraint module comprises a Gaussian classifier, and the Gaussian classifier trains an input first feature through negative logarithm likelihood loss and confidence perception boundary loss; and acquiring vibration signal data of the rotating machine, inputting the vibration signal data into the trained Gaussian constraint network for fault diagnosis, and outputting a fault diagnosis result. The method can deal with various label inconsistent scenes, and realizes accurate identification of shared categories and robust rejection of unknown categories.

A computer-implemented multispectral imaging method for use in analysis of a sample comprising a plurality of types of fluorescent label, each of the plurality of types of fluorescent label having a respective emission spectrum is disclosed. The method comprises: receiving multi-channel image data, each channel in the multi-channel image data comprising image data derived from an unfiltered image of the sample and having a respective spectral content; for each channel: i) forming a vector of measured quantum particle counts from the image data for the channel, the vector having an entry for each pixel in the image; ii) iteratively generating, for each pixel in the image, a vector of possible values having entries for the contribution made by each of the plurality of types of fluorescent label to the unfiltered image, and for each iteration, calculating a vector of expected quantum particle counts, having an entry for each pixel in the image, by multiplying the vector of possible values for each pixel by a mixing matrix defining the relationship between the unfiltered image and the multi-channel image data; and iii) selecting the vectors of possible values for which a negative log-likelihood function describing the probability of a vector of measured quantum particle counts being generated given a corresponding vector of expected quantum particle counts is a minimum; and for each of the plurality of types of fluorescent label in the sample, constructing a corresponding data structure comprising image data in which, for each pixel, the data structure includes the entry for the contribution made by the type of fluorescent label from the vector of possible values for the pixel, each data structure thereby being useable to reconstruct an image of the sample with a spectral content corresponding to the respective emission spectrum of the type of fluorescent label for which the data structure was constructed.

The application discloses an anti-reward hacker recommendation system optimization method and an anti-reward hacker recommendation system, and the method comprises the following steps: a dynamic interval mechanism is constructed, a dynamic interval item is calculated based on a user historical sequence and a positive and negative sample pair, and the dynamic interval item is used for quantifying the preference difference intensity between the positive and negative samples; the dynamic interval item is subjected to cross-sample statistical normalization to obtain a normalized interval item; the dynamic interval is decoupled into an optimization constraint item through a stop gradient operation, model training is carried out in combination with a preference optimization loss; a negative log-likelihood loss is introduced, and the generation probability of the positive sample is directly maximized; the preference optimization loss and the negative log-likelihood loss are jointly optimized to obtain a total loss function, which is used for inhibiting the reward hacker behavior in the recommendation system. The application dynamically adjusts the interval threshold based on the user behavior intensity, ensures that the model differentiates modeling of the fine-grained preference and the coarse-grained demand of the final positive and negative samples in the recommendation sequence, and alleviates the recommendation deviation caused by data sparseness and diversity.

The invention belongs to the technical field of image defect detection, and discloses a self-supervised multi-category industrial product image defect detection method and device based on a twin flow model. The method comprises the following steps: firstly, generating a simulated artificial defect sample by randomly sampling a normal sample and combining Poisson editing; then, a pre-trained feature extractor is used for extracting deep and shallow features, and the deep and shallow features are respectively input into a twin flow model for distribution estimation; fusing multi-scale feature output, and generating a feature distribution diagram of the normal sample; and finally, high-precision defect detection is realized through confrontation optimization of negative logarithm likelihood loss and self-supervised segmentation loss. Through the lightweight twin flow model architecture, the consumption of calculation and storage resources is remarkably reduced, and meanwhile, the accuracy and efficiency of multi-class defect detection are improved in combination with a self-supervision method. The method is suitable for real-time defect detection in an industrial assembly line.

The invention discloses a trajectory prediction system and method in a heterogeneous traffic environment based on a double-layer LSTM and a 4D graph, and belongs to the field of vehicle automatic driving, and the method comprises the following steps: S1, collecting the driving information of heterogeneous traffic participants, and carrying out the preprocessing of the information, and obtaining a data set which defines the instance information, category information and time-space interaction relation of the traffic participants; s2, constructing a 4D graph based on the data set in the step S1, wherein the 4D graph is used for capturing space-time interaction and category generality information of traffic participants; s3, constructing a double-layer LSTM model based on the 4D graph constructed in the step S2, and performing training by using the data set obtained in the step S1; s4, track prediction is carried out by using the trained double-layer LSTM model containing the model parameters, and the model parameters are updated; the model outputs bivariate Gaussian distribution parameters (a mean value, a standard deviation and a correlation coefficient) of future positions of traffic participants through cooperation of an instance layer and a category layer, and prediction of future trajectories in heterogeneous traffic scenes is realized based on optimization of a negative log-likelihood loss function.

The application discloses an estimation method for an attacker's estimation of patrol strategy switching time in an anti-environment, and belongs to the field of unmanned system patrol strategy analysis. The method comprises the following steps: the attacker continuously observes a patrol area, and records a state observation sequence of a patroller; time points in the sequence satisfying a sliding window condition are taken as candidate change point time points, observation subsequences before and after the window are extracted, and a local state transition matrix is estimated; a difference value of the matrix before and after each candidate change point time point is calculated, a difference signal sequence is constructed, the sequence presents a peak value at a state transition matrix switching position; an interval cost function based on a negative log-likelihood is defined; a global optimization objective function with a penalty term is introduced, an optimal change point time point set is solved by using a pruning accurate linear time dynamic programming framework; and intervals are divided according to the optimal change point set, and state transition matrices of each section are independently estimated. The application can automatically determine the position of the switching time point only by relying on the observation sequence, and provides support for the attacker to identify a patrol blind area.

The application discloses a clock error drift modeling method and system based on statistical likelihood optimization. The method first acquires clock error data in a clock synchronization system and performs standardization processing; then, clock speed and clock drift are calculated, and the clock drift is modeled as a random process conforming to a Gaussian process; next, a neural network model is constructed, the negative log-likelihood of the clock drift is taken as a loss function for model training, and the loss function is optimized to minimize the clock error drift error; finally, the system generates correction instructions by using the trained model, and high-precision clock synchronization is maintained when GNSS signals are missing. The application effectively improves the robustness and precision of clock synchronization, and is particularly suitable for environments where GNSS signals are unstable or unavailable.

The invention provides a probabilistic neural network mineral product prediction method based on spatial awareness, and relates to the technical field of mineral product prediction, and the method comprises the steps: extracting mineral control elements from multi-source geological data as feature data, extracting mineral production area data as label data, and randomly dividing the feature data and the label data into a training set and a test set; constructing a mineral product prediction model based on the neural network, introducing a space kernel function to perform output category weighting on a neural network output layer, outputting a neural network regression coefficient and bias, and introducing Dirichlet distribution to model output; training the model by using the training set, constructing a loss function through negative logarithm likelihood of Dirichlet distribution and KL divergence regularization, and testing the trained model by using the test set; and performing mineral product prediction by using the tested model to obtain a prediction probability, and calculating the uncertainty of each sample. According to the invention, non-stationary prediction and uncertainty measurement and simulation of a predicted target space are realized.

The invention discloses a spacecraft pose estimation and uncertainty modeling method based on an intrinsic space, and belongs to the technical field of spacecraft pose estimation. The method comprises the following steps: directly constructing a hierarchical Bayesian probability model on a rotating manifold SO (3) and a translation space, respectively using Feisnow distribution and Gaussian distribution, and introducing conjugate prior to realize principle decomposition and quantification of accidental uncertainty and cognitive uncertainty. Through an end-to-end multi-task neural network, pose probability model parameters, key points and segmentation information are jointly learned, and an iterative optimization module is adopted to improve the estimation precision. In the training stage, joint optimization is carried out through edge negative logarithm likelihood and evidence regularization; in the reasoning stage, probability distribution of pose prediction is obtained by analyzing marginalization, and two types of uncertainty are distinguished. According to the method, the pose estimation precision is improved, meanwhile, the uncertainty of good calibration can be output, and a reliable basis is provided for on-orbit autonomous safe operation of a spacecraft.

The application discloses a method for detecting machine-generated Chinese text, comprising: splitting the Chinese text to be detected according to a set step length to obtain a list of N text paragraphs; traversing the list of N text paragraphs, sampling each text paragraph at a set sampling rate, masking M times to obtain M masked text paragraphs, and inputting the M masked text paragraphs into a T5 model in sequence for decoding to obtain a list of unmasked text paragraphs; calculating the confidence score of each text paragraph according to the negative log-likelihood function score of each text paragraph and the negative log-likelihood function score of each element; comparing the confidence score with a set threshold value to determine whether the text paragraph is artificially written or machine-generated; and comparing the average confidence score with a set threshold value to determine whether the Chinese text to be detected is machine-generated or artificially written. The method is simple to implement and can quickly and accurately detect whether the Chinese text is machine-generated.

This application discloses a salinity prediction method that couples a Bayesian optimization deep learning network with Gaussian likelihood. The method includes: acquiring a salinity observation dataset; preprocessing the data in the salinity observation dataset; inputting the preprocessed salinity observation dataset into a salinity prediction model to obtain an initial salinity prediction result; training the salinity prediction model based on a Bayesian optimization parameter optimization method to obtain a trained salinity prediction model; obtaining an optimized salinity prediction result based on the trained salinity prediction model; performing likelihood estimation on the optimized salinity prediction result based on the Gaussian likelihood method and constructing a negative log-likelihood loss function to obtain the likelihood estimation result and the corresponding probability prediction interval; and using the optimized salinity prediction result and the corresponding probability prediction interval as the final prediction result. This application can improve the efficiency, robustness, and real-time performance of salinity prediction.

The invention relates to the field of ocean geotechnical engineering, in particular to a pile shoe foundation penetration risk early warning method and system based on the Bayesian theory and the optimization strategy and a medium, and the early warning method comprises the steps that prior distribution of soil layer uncertainty parameters is established, and real-time observation data in the penetration process is collected; screening an active parameter set by calculating the sensitivity of the model predicted value to the parameters; constructing a negative log-likelihood loss function, determining an active parameter optimal value by combining a Bayesian theory and an optimization search strategy, and updating posterior distribution of the active parameter optimal value; and a bearing capacity prediction value and a prediction interval are generated by adopting Monte Carlo sampling and Gaussian regression methods. Compared with the prior art, monitoring data can be dynamically integrated, the uncertainty of soil layer parameters is effectively reduced, continuous and accurate prediction of the bearing capacity in the whole penetration process is achieved, the pile pitching risk is controlled, a decision basis is provided for offshore pile standing, and the operation safety of a self-elevating offshore wind power installation platform is remarkably improved.

The invention discloses a runoff probability forecasting method based on an improved deep integration strategy, belongs to the field of runoff probability forecasting, and solves the problem that an existing deep integration strategy can only integrate a single network and arithmetic average distribution weight. The method comprises the following steps: 1, selecting LSTM, GRU and SWM as an integrated sub-network according to an inclusive strategy, and constructing a prediction framework of an input step length 4, an output step length 1 and a hidden layer number 3; 2, a sigma unit is added to an output layer of the sub-network, and softplus is used for activating function constraint, so that uncertainty can be predicted in a quantized mode; 3, taking a negative logarithm likelihood function based on normal distribution as a training scoring rule; 4, after the sub-networks are trained in parallel, a training set is split through three-fold cross validation, an integrated weight is dynamically optimized through a genetic algorithm, and an MNDE model is constructed; the model is superior to a single deep learning model and a traditional machine learning model in certainty, interval, probability prediction and reliability, and is suitable for runoff probability prediction.

The invention discloses a power transmission line fault positioning system and positioning method based on traveling wave recording, and the system is characterized in that a traveling wave recording starting device is installed in a transformer substation, collects the abnormal data information of all lines and buses in the transformer substation, and uploads the abnormal data information to a safety platform of the Internet of Things; and the Internet of Things security platform performs security processing on the data information and then transmits the data information to the power transmission line fault traveling wave positioning system. The power transmission line fault traveling wave positioning system preprocesses abnormal information, adopts a same-station multi-line traveling wave data superposition algorithm of a fixed time window to form traveling wave characteristic data based on a time sequence, and adopts a double-tower neural network and a negative log-likelihood loss function optimization model to realize hidden danger early warning and accurate fault positioning of a power transmission line. And the operation state and the hidden danger early warning and fault positioning result are sent to the power transmission operation and maintenance platform, so that information support is provided for efficient operation and maintenance of the power transmission line.

A system includes a controller configured to receive, at a trained transformer model, one or more run-time logs indicating information associated with interaction between a unique device and a server, output a user-score associated with the unique device and the one or more run-time logs in response to determining a negative log-likelihood of the one or more run-time logs in a next-log prediction probability distribution modeled by the trained transformer model, output a server-score utilizing at least normal-cluster centers associated with the trained transformer model and the one or more run-time logs, and in response to a sum of the user-score and server-score exceeding a threshold, outputting an indication of a cyber-threat associated with the one or more run-time logs.

The invention discloses a transformer state signal analysis method, and relates to the technical field of transformers, and the method comprises the steps: converting dissolved gas analysis data into physical characteristic data based on a preset prior knowledge base; judging a current sample mode of the target transformer according to the physical characteristic data; if the current sample mode is a known mode, inputting the physical characteristic data into a pre-trained multivariate classifier to obtain a basic fault physical mode label of the target transformer; if the current sample mode is an unknown mode, inputting the physical feature data and the equipment parameters into a pre-trained semantic mapping model to obtain distribution representation data in a semantic space, the distribution representation data being a probability distribution parameter set; determining the negative logarithm likelihood of the distribution representation data and the plurality of composite fault prototypes; and determining a composite fault physical mode label of the target transformer according to the composite fault prototype corresponding to the minimum one of the plurality of negative logarithm likelihoods. According to the invention, the accuracy of transformer oil state monitoring is improved.

The invention relates to the technical field of signal processing and data fusion, and discloses an electric detection-radar heterogeneous signal positioning method based on an encoder-decoder network, and the method comprises the steps: obtaining a radar and an electric detection data sequence which are aligned in a space-time manner; performing deep interactive fusion by using a hybrid feature encoder, and outputting a fused feature sequence; decoding the fusion feature sequence into a probabilistic result containing a position mean value and a covariance matrix by using a probability positioning decoder; carrying out supervised training on the coding and decoding network model by adopting a negative log-likelihood loss function; and inputting real-time data into the trained model, and directly outputting a positioning result through forward propagation. Through deep adaptive fusion and probabilistic decoding, the problems of shallow fusion level and weak interference resistance are solved, and the precision, robustness and reliability of heterogeneous signal positioning are improved.

The present application relates to the technical field of hallucination suppression, and discloses a hallucination suppression training method and system of a large model fused with real-time fact library verification, comprising the following steps: constructing a real-time fact library graph, calculating the discrete Ricci curvature and in-degree of nodes, combining input query to retrieve candidate documents and generating a neutral search distribution; calculating the weight using geometric debiasing parameters, reweighting and normalizing the distribution to obtain a target evidence distribution; encoding the candidate documents and generating an evidence usage distribution through a gating network, and calculating the negative log-likelihood loss; calculating the optimal transmission loss based on the shortest path of the nodes as the cost, and constructing the tail index loss based on the tail index of the in-degree; finally, weighting and summing the three types of losses, and jointly optimizing the model hyperparameters using the gradient descent method. The method realizes geometric debiasing of fact distribution and dynamic calibration of search results, effectively suppresses the structured hallucination bias of large models, and improves the authenticity and reliability of generated content.

The application discloses a kind of non-structural road surface under timing neural network tire force estimation method and system, belong to vehicle dynamics control technical field.Pass through the construction vehicle information acquisition platform, collect vehicle state information under non-structural road surface;Vehicle state information is washed, interpolation, normalization and timing is filled up and mask preprocessing;LSTM model of fusion attention mechanism is constructed, the output layer of this model simultaneously outputs tire force estimation mean and estimation variance, to quantify estimation uncertainty;Network training is carried out using negative log-likelihood loss function;Network hyperparameter is automatically optimized using Bayesian optimization method.The application effectively solves the estimation problem caused by strong timing dependence, significant nonlinearity and obvious hysteresis effect of tire force under non-structural road surface, without additional installation tire internal force sensor, low in cost, high in precision, strong in generalization, significantly improve the performance of vehicle chassis control system under complex road conditions.