Performance Evaluation of Machine Learning—Based NL2SQL Models on Time-Series Industrial Data

Abstract

Industrial automation systems generate large volumes of time-series data that are typically stored in relational databases. Accessing this data using Structured Query Language (SQL) requires technical expertise, limiting usability for non-expert users. This paper presents a comparative performance evaluation of machine learning–based Natural Language to SQL (NL2SQL) models for querying time-series industrial datasets using natural language. Two large language models, SQLCoder7B-2 and Qwen2-1.5B, are evaluated on a synthetically generated relational database containing industrial parameters such as temperature, pressure, timestamp, and data validity status. The models are tested using diverse natural language queries involving temporal filtering, aggregation, and conditional operations. The performance of two models is assessed using metrics including accuracy, precision, recall, F1-score, and query latency. Experimental results show that Qwen2-1.5B outperforms SQLCoder-7B-2 in terms of accuracy of 88%, recall of 0.88, and significantly lower latency. The findings demonstrate that large language models can effectively enable natural language interaction with industrial databases, improving accessibility and efficiency in data retrieval tasks.