8 results about "Continuous distributions" patented technology

The invention discloses a causal graph inference method and a causal graph inference system which are oriented to strong nonlinear continuous variable distribution and based on RealNVP and micro NOTEARS constraints. According to the method, a structure parameter matrix is constructed to represent candidate causal connection, parent variable condition input is constructed for each variable based on structure parameters, and accurate likelihood modeling is performed on the condition density of each variable by adopting a condition RealNVP normalization flow model. By constructing an objective function containing a conditional log-likelihood item, a sparse regular item and a NOTEARS style differentiable acyclic constraint item, a sparse causal structure meeting acyclic constraint is obtained by utilizing gradient optimization, and a causal graph is output. Further, based on the learned causal graph and a conditional RealNVP model, anti-fact inference is executed under a given intervention condition, and an anti-fact result is output. The method is suitable for strong nonlinear relation and complex continuous distribution scenes, and has the advantages of being stable in structure inference, high in interpretability and capable of supporting anti-fact analysis.

The invention discloses a dam engineering leakage channel entrance-channel-escape full view identification method and system. The method comprises the following steps: acquiring multi-source monitoring data of a dam engineering area; constructing a background seepage flow field reflecting a macroscopic hydraulic gradient and an anisotropic diffusion tensor field reflecting local permeability characteristics based on multi-source monitoring data; performing bidirectional probability field evolution calculation by using an unsteady convection-diffusion equation set and respectively taking the inlet geometrical characteristics and the escape temperature anomaly as sources to obtain an inlet reachability field and an escape reachability field which are continuously distributed in the dam body space; and determining the three-dimensional full-view probability distribution of the leakage channel based on the space interaction overlapping characteristics of the two fields. According to the method, physical self-consistent reconstruction from discrete monitoring data to a continuous channel entity is realized by constructing a different-field coupling feedback and tensor guiding mechanism, and the topological fuzzy problem of leakage path identification in a complex medium is effectively solved.

The invention relates to a wafer defect analysis method and system and a readable storage medium. The method comprises the following steps: dividing a wafer into a plurality of regions with different importance; evaluating and endowing the areas with different weight values so as to establish a comprehensive evaluation index system for specific areas with the weight values greater than a set value; obtaining original measurement data of the wafer, and performing weighting and dimension processing on the measurement data of the corresponding area according to the weight value to obtain optimized measurement data; processing the optimized measurement data by adopting a Kriging interpolation algorithm to generate a profile distribution diagram capable of reflecting the continuous distribution condition of the wafer surface parameters; and performing defect identification and analysis on the specific region based on the profile distribution diagram. Different weight values are given to the regions with different importance, so that the Kriging interpolation algorithm can more sensitively capture the process variation of the specific region with a higher weight value, and the effect and efficiency of defect analysis of the specific region are improved.

The present disclosure relates to deriving one or more of a precision, a recall, or an F-score metric. In accordance with the disclosed techniques a set of data is acquired or accessed. The data is continuously distributed. Based on an expected signal and an observed signal, one or more of a precision, a recall, or an F-score are determined for one or more automated routines configured to operate on the set of data. One or more of the precision, the recall, or the F-score provide a comparative measure of the accuracy of the one or more automated routines.

This invention discloses a continuous passive domain adaptive fault diagnosis method, comprising: pre-training a fault diagnosis model using a labeled source domain dataset; constructing an adaptive network based on an average teacher architecture, including a student model and a teacher model; during the adaptation process, using a confidence-weighted knowledge consistency regularization strategy, selecting high-confidence samples and combining hybrid augmentation to provide reliable pseudo-supervision for the student model; using a mutual information maximization strategy to enhance the plasticity adaptation of the student model to the current target domain distribution; and simultaneously employing a stochastic recovery strategy and a progressive hyperparameter annealing strategy to constrain the teacher model parameter updates to prevent catastrophic forgetting and dynamically optimize the balance between stability and plasticity. This invention can effectively solve the problem of continuous distribution shift caused by dynamic operating conditions of rotating machinery equipment under strict data privacy constraints, without accessing the source domain and historical data, thus achieving robust fault diagnosis.

The application discloses a non-iterative design method of a non-uniform continuous distribution composite pupil function, solves the problem that the design method of an existing pupil function is relatively complex, including expansion of the pupil function, matrix calculation and a calculation process of solving a high-resolution gain expression, the non-uniform continuous distribution composite pupil function is composed of a Sonine polynomial, a Zernike polynomial and a Fourier-Bessel series, numerical mapping relationships between the Sonine polynomial, the Zernike polynomial, the Fourier-Bessel series and an imaging complex amplitude distribution at a radial focal plane are respectively constructed, coefficients of the Sonine polynomial, the Zernike polynomial and the Fourier-Bessel series are solved, a function formula of the composite pupil function is finally determined, and the mapping relationship is relatively simple and operable.

The present application relates to a kind of gravity background field construction method based on iterative optimization kriging, comprising the following steps: S1. obtain local gravity measurement data as known point set;S2. based on the known point set, preliminary interpolation is carried out to the grid point of grid background map using kriging method, and the kriging estimated value of each grid point is obtained;S3. by introducing local gravity anomaly covariance and kriging variation function, the random uncertainty of kriging estimated value in interpolation process is evaluated, and each grid point is divided into reference point and to be modified point based on random uncertainty;S4. reference point is combined with known point set to construct continuous distribution constraint surface, and the continuous distribution constraint surface is locally continuously modified based on the obtained reference point, and the continuous distribution constraint surface is iteratively extended based on the locally modified constraint surface until the continuous distribution constraint surface is continuously modified, to realize the construction of gravity background field.

The invention relates to a large model mixing precision quantitative reasoning method and device based on a rotation matrix. The method comprises the following steps: acquiring a target activation tensor; a rotation matrix is adopted to carry out rotation transformation on the target activation tensor, lexical element dimension mixing precision quantification is carried out, and activation values with the same precision in the quantized activation tensor are continuously distributed in a memory; the activation tensor of the first precision is stored in an outlier buffer area allocated in a memory; calling a tensor core, and performing matrix multiplication operation on the quantized activation tensor and the quantized weight tensor to obtain a quantization operation result; and carrying out inverse quantization on a quantization operation result, and taking an inverse quantization result as a matrix multiplication operation result of the target activation tensor and the target weight tensor. According to the method and the device, the activation values with the same precision are continuously distributed in the memory, and the activation tensor with the first precision is stored in the outlier buffer area, so that the tensor core can be efficiently calculated, and the capability and relatively high accuracy of the model can be kept while model reasoning is accelerated.