3results about How to "Few training samples" patented technology

A battery discharge end point prediction method based on an interpretable physical driving neural network model, current data and voltage data during battery discharge are collected; an interpretable physical driving neural network based on a second-order RC equivalent circuit model is established; the collected current data are taken as features, and the voltage data are taken as labels, and the interpretable physical driving neural network is trained; after the training of the interpretable physical driving neural network is completed, given current data, the trained interpretable physical driving neural network outputs a voltage prediction curve, and the discharge end is predicted based on the voltage prediction curve.

The application mainly relates to a strip plate shape classification method, characterized in that the method comprises the following steps: acquiring actual plate shape deviation values of each detection area of a reference strip and acquiring process parameter data of each detection area of the reference strip; based on the actual plate shape deviation values, determining plate shape defect types corresponding to each detection area of the strip, taking the plate shape defect types as training label data, taking the process parameter data as training feature data, and obtaining multiple groups of training sample data; constructing an initial plate shape classification model, training the initial plate shape classification model based on the multiple groups of training sample data, and obtaining a plate shape classification model; and based on process parameter data of a target area in a to-be-detected strip, determining a plate shape defect type of the target area in the to-be-detected strip through the plate shape classification model. The application can solve the problems of experience-based errors and omissions in manual strip plate shape classification to a certain extent.

The invention provides a text classification model training method, system and product based on mixed sampling, and relates to the technical field of deep learning, and the method comprises the steps: predicting the probability that data in an unlabeled text data set belongs to each text category through a text classification model, and obtaining a prediction result; determining a labeling progress; in the early stage of the labeling progress, based on a prediction result, sampling a subset by a class balance sampling strategy, and labeling to obtain a labeled subset; in the middle stage of the labeling progress, based on a prediction result, a subset is sampled through an uncertainty and diversity mixed sampling strategy, and a labeling subset is obtained through labeling; at the later stage of the labeling progress, based on a prediction result, sampling a subset by an edge sampling strategy or an uncertainty and diversity mixed sampling strategy, and labeling to obtain a labeled subset; and when each annotation subset is obtained, constructing a corresponding augmented training set, and training the text classification model for one time by using the constructed augmented training set. The objective of the invention is to improve sample labeling and model training efficiency.