Anomaly Detection Learning Device Using Cycle Difference Waveforms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly indication detection techniques in power systems, such as those described in Patent Literature 1, may not accurately detect malfunctions caused by unique waveform characteristics resulting from events like fallen trees or bird nests in mountainous areas, as they rely on specific frequency spectrum characteristics that are not always present.
Innovation Solution
A learning device that generates learned data for anomaly indication detection by performing preprocessing to subtract average values from normal data over preceding cycles, and using waveform similarity analysis to determine normal waveforms and a normality threshold, enabling more accurate detection of anomalies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If frequency spectrum analysis with specific characteristics is used for anomaly detection, then detection accuracy for known fault patterns is improved, but detection accuracy for anomalies with different waveform characteristics deteriorates
Solution Approach 1:
The patent transforms the detection approach by changing from frequency domain parameters to time domain parameters. Specifically, it uses cycle-to-cycle difference values of instantaneous values directly from the time domain, avoiding frequency spectrum transformation. This parameter change allows the system to detect various waveform characteristics without being constrained by specific frequency patterns, thereby improving both detection accuracy and adaptability across different fault types.
2Measurement precision
If supervised learning with neural network model is used, then detection accuracy for training-pattern anomalies is improved, but detection capability for unseen anomaly patterns deteriorates
Solution Approach 1:
The patent extracts the essential feature of anomaly detection by taking out the need for complex supervised learning models. Instead of using neural networks that require extensive training data and struggle with unseen patterns, it extracts only the critical comparison between current cycle and previous cycle difference values. This simplified extraction approach maintains high detection accuracy while significantly improving generalization to new anomaly patterns that were not part of the training set.
3Measurement precision
If complex preprocessing and multiple analysis steps are applied, then detection precision for specific patterns is improved, but system complexity and processing time increase
Solution Approach 1:
The patent applies the extraction principle by removing unnecessary preprocessing steps and complex analysis procedures. It retains only the essential operation of calculating cycle-to-cycle difference values of instantaneous values and comparing them against thresholds. This drastic simplification reduces system complexity and processing time while maintaining or even improving detection precision by focusing on the most critical anomaly indicators without distraction from redundant processing steps.
Data Source
AI summary
A learning device generates learned data to be used for anomaly indication detection. The learning device includes a preprocessing unit that subtracts, from a value at each of points in one cycle of normal data, an average of values at corresponding points in normal data from one to N cycles ago in computing difference values with respect to the preceding N cycles, the one cycle corresponding to a specified time length, N being an integer greater than or equal to 2, the normal data being data from a normal state; and a first waveform analysis unit that generates, through waveform similarity analysis using the difference values with respect to the preceding N cycles, normal waveforms and a normality determination threshold to be used in determining whether or not there is normality as the learned data.