Forecast Model Retraining by Residual Gradient for Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing time-series anomaly detection methods are inefficient and inaccurate, particularly in non-stationary signals, often requiring frequent retraining of simplified models and failing to identify anomalies that start before exceeding confidence intervals.
Innovation Solution
Implementing a method that enforces gradient constraints on forecasting models, allowing for more complex models by intermittently retraining only when the gradient of the residual signal exceeds a predetermined range, thereby improving accuracy and reducing computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Duration of action of moving object
If standard forecasting methods are used to produce confidence levels, then forecast duration can be extended, but prediction accuracy deteriorates and the methods are limited to simpler time-series data trends
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the gradient threshold parameter that controls when retraining occurs. By changing this parameter, the system can adapt between prioritizing forecast duration (higher threshold, less retraining) versus prediction accuracy (lower threshold, more frequent retraining), resolving the contradiction for different operational requirements.
Solution Approach 2:
The system implements dynamics by making the model structure adaptive through conditional retraining based on gradient monitoring. The forecasting model transitions from a static to a dynamic state where complexity increases only when necessary (when gradient exceeds threshold), allowing extended forecast duration while maintaining accuracy through on-demand adaptation.
2Productivity
If simplified forecasting models are used, then computational resources and training time are reduced, but the models fail to accurately capture complex non-linear time-series trends
Solution Approach 1:
The patent applies partial action by using simplified models for most operations (normal forecasting) and only applying complex model retraining partially when necessary (when gradient threshold is exceeded). This avoids the continuous computational burden of complex models while maintaining accuracy when needed, resolving the contradiction between training efficiency and prediction accuracy for non-linear trends.
Solution Approach 2:
The system changes the model complexity parameter dynamically based on the gradient threshold condition. When the gradient exceeds the threshold, the system transitions to a more complex retrained model; otherwise, it operates with the simpler existing model, thus achieving both efficiency and accuracy for non-linear trends.
3Measurement precision
If frequent retraining of forecasting models is performed, then prediction accuracy is maintained, but computational resources and processing time are excessively consumed
Solution Approach 1:
The patent implements feedback by continuously monitoring the gradient of the residual signal and using this feedback to control when retraining occurs. The gradient threshold acts as a feedback mechanism that triggers retraining only when prediction accuracy is genuinely compromised, avoiding unnecessary retraining and thus reducing computational resource consumption while maintaining accuracy.
Solution Approach 2:
The system applies periodic action through interval-based gradient monitoring and conditional retraining. Instead of continuous or frequent retraining, the system checks at regular intervals whether the gradient threshold is exceeded, performing retraining periodically only when necessary, thus balancing accuracy maintenance with computational efficiency.
4Reliability
If confidence envelope methods are used for anomaly detection, then anomalies can be identified, but anomalies that start before exceeding the confidence interval are missed and false positives occur in safety-critical applications
Solution Approach 1:
The patent applies preliminary action by performing gradient-based anomaly detection before the confidence envelope is exceeded. The system proactively identifies anomalies by monitoring gradient changes in the residual signal, allowing detection of anomalies at their onset rather than waiting for them to exceed confidence thresholds, thus improving both reliability and precision in safety-critical applications.
Data Source
AI summary
Predicted observations within a first window are forecast by a predictor and a residual signal is calculated based on a difference between predicted observations and true observations. A gradient of the residual signal is determined indicative of a degree of divergence. When the gradient is outside a predetermined range, the predictor is retrained. Based on further predicted observations forecast by the predictor, a confidence envelope for a second window is generated and, using the confidence envelope, it is determined whether an outlier portion exists within the second window.