Problems and Approaches to Storing and Analyzing Time Series under Conditions of Feature Redundancy

The paper examines modern approaches to efficient storage and analysis of time series arising in monitoring industrial equipment, in particular, CNC machines. The main problem is the growth of data amount, which requires both the optimization of storage systems and the reduction of the dimension of input features without loss of information content. To solve the problem, we conducted a comparative analysis of purpose-built database management systems (DBMS): ClickHouse, InfluxDB, TimescaleDB, and PostgreSQL. Based on the testing results for ingestion rate, scalability, and resource consumption, ClickHouse was chosen. In parallel, we studied feature selection methods based on machine learning: L1 regularization, random forest, and gradient boosting. By using a real dataset of 18 material (wax) milling experiments, it was shown that gradient boosting (LightGBM) provides the best accuracy and robustness in feature importance evaluation. As a result, 20 most significant features related to position, speed, current, and voltage along the X, Y, and spindle S axes were identified. Based on these features, the condition of the tool, the completion of machining, and visual inspection results can be predicted with high reliability. The proposed approach allows one to reduce the amount of stored data, improve query performance, and reduce hardware costs for storing and processing time series.