Anomaly Detection Using Reconstruction Error Cycles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods using principal component analysis often misidentify accidental noise as anomalies due to local increases in reconstruction error, leading to false positives and missed detections when noise is present, as they rely solely on reconstruction error magnitude without considering cyclic properties of signals.
Innovation Solution
An anomaly detection apparatus that calculates reconstruction errors and cycle information from time-series sensor signals, using smoothing and weighting techniques to differentiate between noise and anomaly signals, and employs machine learning models to determine the presence of anomalies based on both reconstruction error and cyclic patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If principal component analysis is used for anomaly detection relying solely on reconstruction error magnitude, then the detection method is simple and easy to implement, but accidental noise is erroneously detected as anomaly signals and anomaly signals may not be detected when reconstruction error locally increases due to noise
Solution Approach 1:
The patent introduces a new dimension of analysis by calculating cycle information from the reconstruction error signal. Instead of relying solely on the magnitude of reconstruction error (one-dimensional approach), the system now analyzes both the magnitude and cyclic characteristics (periodicity, frequency) of the error signal. This dimensional expansion allows the system to distinguish between noise-induced errors and true anomaly signals, resolving the contradiction between simplicity and reliability.
Solution Approach 2:
The patent changes the parameters used for anomaly detection from solely reconstruction error magnitude to include cycle information derived from the reconstruction error signal. By extracting cyclic characteristics (such as period, frequency, or spectral properties) from the reconstruction error, the system transforms the detection basis from a single parameter to multiple parameters, enabling better discrimination between noise and anomalies while maintaining operational simplicity.
2Productivity
If reconstruction error magnitude is used as the sole criterion for anomaly detection, then the detection process is fast and computationally simple, but false positives increase due to local increases in reconstruction error caused by accidental noise
Solution Approach 1:
The patent adds another dimension to the detection process by analyzing the cyclic properties of the reconstruction error signal. This allows the system to maintain fast detection speeds while improving precision, as the cycle information analysis can be performed efficiently on the existing reconstruction error data without requiring additional complex computations.
Solution Approach 2:
The patent uses cycle information as an intermediary feature that bridges the reconstruction error signal and the anomaly detection decision. By extracting cyclic characteristics from the reconstruction error, this intermediary allows the system to filter out false positives caused by noise while maintaining the speed of the original detection process.
3Use of energy by stationary object
If only the magnitude of reconstruction error is analyzed, then the detection method requires minimal computational resources, but the system cannot distinguish between noise-induced errors and true anomalies
Solution Approach 1:
The patent extends the analysis to include the cyclic dimension of the reconstruction error signal. This additional dimensional analysis can be performed with minimal extra computational resources by utilizing spectral analysis or period detection methods on the existing error signal, thereby improving reliability without significantly increasing energy consumption.
Solution Approach 2:
The patent changes the detection parameters from solely magnitude-based to include cyclic characteristics such as period, frequency, or spectral content. These parameter changes enable the system to distinguish between noise and anomalies with minimal additional computational overhead, as the cycle information can be extracted from the reconstruction error using efficient signal processing techniques.
Data Source
AI summary
According to one embodiment, an anomaly detection apparatus includes a processing circuit configured to calculate a reconstruction error of an input signal being a time-series signal, calculate cycle information indicating cyclic property of the reconstruction error, and determine presence/absence of an anomaly signal in the input signal on the basis of the cycle information or the cycle information and the reconstruction error.