Abnormal Point Simulation in Time Series Data

The patent applies preliminary action by simulating abnormal points in advance within the evaluation data before model evaluation. The system identifies normal data blocks, simulates various types of abnormal points (sudden changes, gradual changes, outliers) at appropriate locations, and creates augmented evaluation datasets. This pre-processing ensures that when models are evaluated, they are tested against realistic anomaly scenarios, significantly improving evaluation accuracy without requiring changes to the model architecture itself.

The patent uses copying by creating simulated copies of abnormal points based on patterns learned from historical data. Instead of requiring real anomaly data for evaluation, the system generates synthetic abnormal point instances that replicate the statistical characteristics and temporal patterns of actual anomalies. These copied abnormal patterns are inserted into normal evaluation data, providing a scalable solution that improves measurement precision without proportionally increasing data collection complexity.

The patent applies segmentation by dividing the time series evaluation data into discrete data blocks and analyzing the traits of each block individually. The system identifies characteristics such as mean, variance, and temporal patterns for each segment. This segmentation allows the simulation process to tailor abnormal points to the specific characteristics of each data block, ensuring appropriateness. The modular approach also enables parallel processing, reducing the overall time penalty of the analysis.

The patent implements local quality by making the abnormal point simulation adaptive to local data block characteristics rather than applying a uniform simulation approach. The system adjusts the type, magnitude, and frequency of simulated abnormal points based on the specific traits of each data block (e.g., high-variance blocks receive different treatment than low-variance blocks). This localized adaptation significantly improves the appropriateness and realism of simulated anomalies, directly enhancing evaluation precision.

The patent applies merging by combining adjacent data blocks that share similar traits into unified simulation units. After clustering data blocks based on their characteristics (variance, trend, seasonality), the system merges blocks within the same cluster to create homogeneous regions for simulation. This merging reduces the number of individual simulation decisions required while maintaining adaptability to different data patterns. The complexity is managed through automated clustering algorithms that group blocks based on statistical similarity, making the process scalable.

The patent implements parameter changes by dynamically adjusting simulation parameters (abnormal point magnitude, frequency, type distribution) based on the clustered data block characteristics. Different clusters receive different simulation parameter sets tailored to their statistical properties. For example, clusters with high variance receive simulations with larger magnitude abnormalities, while seasonal clusters receive simulations that respect temporal patterns. This parameter adaptation enhances simulation versatility without requiring manual configuration for each data block.