6 results about "Inductive analysis" patented technology

This invention discloses a method for reconstructing battery swapping station operation data based on adaptive truncation and feature correlation repair, belonging to the field of data reconstruction technology. First, a standardized two-dimensional spatiotemporal discrete model is constructed, and sequence mapping technology is used to align the discrete data stream to the date-time period operation space. Second, a zero-first truncation model based on the continuity of frequency distribution is established to adaptively remove outliers and noise. Then, Latin hypercube sampling and inverse cumulative distribution function transformation algorithms are introduced to fill in missing values ​​while ensuring that the reconstructed data conforms to the original probabilistic characteristics. Finally, the battery swapping frequency-electricity data is dimensionally reconstructed to facilitate the inductive analysis of battery swapping behavior characteristics. Numerical examples demonstrate that the method proposed in this invention effectively restores the statistical usability of the data, fully revealing the bi-peak tidal characteristics of battery swapping load and the distribution law of low-electricity anxiety, providing relatively reliable data support for load forecasting and regulation.

The invention belongs to the technical field of visual large model training. The invention provides a rapid autonomous evolution method for a large visual model of a ubiquitous embedded terminal. According to the embodiment of the invention, the most important layer is preferentially selected for updating on the premise that the cost requirement is met, and the optimal balance between the precision and the time delay can be realized. In order to evaluate layer time delay, in the reasoning process of the deep learning model, the calculation process relates to a plurality of layers in a large model, and input and output tensors with similar calculation sequences and logic structures are processed when each layer is executed. Through inductive analysis of forward propagation, reverse propagation and weight update delays of different layers, a model is established, different delay types are abstracted, and an accurate delay evaluation basis is provided for scheduler design. The intermediate activation result only depends on the input data and the output of the previous layer, and the same activation result is generated by the same input data, so that repeated calculation on a non-updated layer is avoided by adopting a memory-aware calculation multiplexing method, and redundant calculation is remarkably reduced.

This invention relates to the field of automated testing. To improve the efficiency and accuracy of test result analysis, it provides a method and system for automated test result summarization and analysis. By extracting error descriptions from semantic test results and combining them with domain mapping, the error data is transformed into structured information. Then, JSON examples are constructed based on error description classification. Using a large language model, standardized prompt words are used to summarize patterns and analyze root causes of various errors. Finally, structured results containing error indexes and root causes are output. This process significantly improves the automation and efficiency of error analysis, ensures the comprehensiveness and accuracy of error coverage, and can be flexibly adapted to new application scenarios by updating domain configuration files and prompt word templates.

The application provides a query method, device and electronic equipment. The method comprises the following steps: in response to obtaining a query text facing a first table, obtaining the first table, the query text comprising a selection text and a condition text, the first table adding a new column on the basis of a second table, the new column corresponding to a first expression, the first expression being used for inductive analysis of multiple lines of text in the second table; querying at least one line of target text in the first table that meets the condition text; determining a target column in the first table according to the selection text, the column name of the target column and the selection text belonging to the same semantics; and determining that the target value corresponding to the target column in the at least one line of target text is the result text corresponding to the query text, wherein the target value is determined according to the expression. Based on various types of query texts of the user, the application can obtain the corresponding result text, thereby improving the application range of table question and answer.

The application belongs to the technical field of large language model optimization, and discloses a large data driving-based large language model optimization method and system. The method comprises the following steps: inputting public text data into a test model to obtain test reply data, and calculating the test deviation between the test reply data and the reply data; when the test deviation meets a preset correction trigger condition, corresponding correction text data and correction reply data are identified; data supplement processing is performed based on the correction text data and the correction reply data to generate supplement data for model optimization; the data supplement processing comprises similarity comparison of the correction text data to divide the correction text data into data groups, generation of error registration data, and performance of a supplement operation based on a statistical analysis result. Through inductive analysis of the deviation and quality control of the supplement data, the application improves the reply accuracy, logical consistency and robustness of the target language model.

The invention belongs to the technical field of safety production management and accident analysis, and particularly discloses a limited space poisoning suffocation accident risk evolution path deduction method and system. According to the method, multiple historical cases are concluded and analyzed, and a multi-level system accident cause model is constructed; performing quantitative analysis on the historical accident case data by applying a related data mining algorithm so as to establish a quantitative association relationship among the cause factors; receiving one or more initial risk factors input by a user, and deducing and identifying a risk path according to the established quantitative association relationship; and comprehensively evaluating the identified risk path, determining a key risk factor combination, and generating a decision analysis report containing preventive management and control suggestions. The problems that traditional accident analysis is insufficient in depth and the process is difficult to quantify are solved, a scientific, quantitative and highly targeted basis is provided for safety management decision, and scientificity and effectiveness of limited space operation safety management are remarkably improved.