Efficient pattern search in large-scale time series

In experimental physics, diagnostic systems produce signals that reflect reproducible physical behavior. Under similar operating conditions, recurring physical processes manifest as structurally similar signal shapes, establishing a direct correspondence between physical behavior and the temporal structure of measured signals. These recurring structures, referred to as structural patterns, constitute the natural entry point for large-scale data analysis, as they encode the essential characteristics of the underlying physical phenomena. In long-pulse nuclear fusion experiments, thousands of diagnostics continuously generate high-resolution time series data during the production of discharges. Identifying all occurrences of a given structural pattern (corresponding to a specific physical event) in large datasets is a critical but challenging task. Traditional workflows based on manual visual inspection or discharge-oriented queries do not scale to modern data volumes and limit the statistical exploitation of reproducible behavior. This work presents a scalable framework for shape-based similarity search in large-scale time-series distributed collections. The proposed method combines Haar wavelet-based dimensionality reduction, sliding-window vectorization, amplitude-invariant normalization, cosine similarity, and FAISS (Facebook AI Similarity Search) exact indexing to enable efficient retrieval of structurally similar patterns. The framework was validated on a dataset of synthetic signals containing approximately 1.8 million samples each. Experimental results show that the proposed approach preserves retrieval accuracy while reducing query time to 34 ms on a dataset of approximately 2.62 million indexed subsequences (10 signals). Scalability experiments over the complete dataset (690 signals, approximately 181 million indexed subsequences) demonstrate near-linear growth in query time, confirming the suitability of the proposed approach for large-scale fusion diagnostic data analysis.