Anomalous Event Detection via Reconstruction Error Scaling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training AI devices to handle anomalous events is challenging due to the rarity of such events, which limits the effectiveness of data-driven approaches and requires more sophisticated methods for detection and handling in applications like autonomous driving.
Innovation Solution
A system utilizing an encoding module to generate reconstruction data, identify reconstruction errors, and calculate anomaly scores, combined with a decoding module for action prediction and likelihood determination, to execute corrective actions based on scaled anomaly scores, thereby enhancing the detection and handling of anomalous events.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data-driven approaches are used to train AI devices, then the system can handle common events effectively, but it fails to detect rare anomalous events due to their scarcity in training data
Solution Approach 1:
Instead of trying to detect anomalous events directly from scarce positive examples, the patent inverts the approach by training the model to reconstruct normal events and then identifying anomalies as failures in reconstruction. The encoding module learns to compress and decode normal driving patterns, and any significant reconstruction error indicates an anomalous event, effectively detecting rare events through their deviation from normal patterns rather than requiring examples of them in training data
Solution Approach 2:
The patent introduces an intermediary reconstruction process between the input sensor data and the anomaly detection. The encoding module and decoding module act as intermediaries that learn normal patterns, and the reconstruction error serves as an intermediary metric that indirectly indicates the presence of anomalies. This intermediary mechanism allows the system to detect anomalous events without directly observing them during training
2Reliability
If the system uses reconstruction error to detect anomalies, then it can identify rare events, but it generates false positives that reduce detection reliability
Solution Approach 1:
The patent transforms the raw reconstruction error into a scaled anomaly score by changing the parameter representation. The scaling process adjusts the error values to account for the inherent uncertainty in reconstruction, converting absolute error measurements into normalized scores that better reflect true anomaly likelihood. This parameter transformation reduces false positives by distinguishing between acceptable reconstruction variations and genuine anomalies
Solution Approach 2:
The system incorporates feedback mechanisms where the decoding module's action predictions provide information back to refine the anomaly scoring. The likelihood values generated from action predictions feed into the scaling process, creating a feedback loop that continuously improves the distinction between true anomalies and false positives by learning from the relationship between reconstruction errors and predicted actions
3Measurement precision
If the system scales anomaly scores using likelihood values, then it reduces false positives, but it increases computational complexity
Solution Approach 1:
The patent merges the anomaly detection process with the action prediction process into a unified framework. The decoding module simultaneously performs action prediction and provides likelihood values for anomaly scoring, combining two functions into one computational pass. This merging reduces overall computational complexity compared to running separate detection and prediction systems while still achieving false positive reduction through the integrated likelihood scaling
Data Source
AI summary
According to one aspect, anomalous event detection based on deep learning may include a system for anomalous event detection for a device. The system includes a computing device having a processor, an encoding module, and a decoding module. The processor is configured to receive sensor data. The encoding module generates reconstruction data based on the sensor data, identifies at least one reconstruction error in the reconstruction data, and determines an anomaly score based on the at least one reconstruction error. The decoding module generates an action prediction based on the sensor data and determines a likelihood value based on the action prediction. The processor can then calculate a scaled anomaly score based on the anomaly score and the likelihood value and causes the processor to execute an action based on the scaled anomaly score.