Optimized Time Series Feature Selection for Manufacturing AI: Reducing Complexity and Improving Classification Accuracy

Although AI enhances productivity and quality in manufacturing, real-time validation remains difficult due to the complexity and high-dimensionality of time series data under strict inference time requirements. To address this, we propose two time series feature selection strategies. First, we extract meaningful variables by analyzing the time series patterns of non-defective and defective products. Through STL decomposition, we eliminate high-frequency noise and seasonal effects, and transform conventional feature-wise PCA into a production-wise approach by independently analyzing each sensor across products. We further quantify the pattern differences between non-defective and defective products for each sensor using DTW similarity, and perform clustering to systematically select significant variables. Second, we reconstruct lightweight statistical features through ANOVA and correlation analysis to further reduce feature dimensionality. In experiments on real-world manufacturing datasets, our techniques improve predictive performance and achieve reduced inference latency, successfully satisfying strict real-time production constraints.