Anomaly Detection Model Using Shared Weights and Segmented Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods in industrial settings face challenges in efficiently updating deep learning models when data patterns change, particularly in handling complex, non-linear patterns and distinguishing anomalies from normal data over time.
Innovation Solution
A data processing method utilizing artificial intelligence that generates an anomaly detection model comprising multiple sub-models trained on different data subsets grouped by time intervals and domains, with shared weights and dimensionality reduction/restore networks, to determine anomalies based on reconstruction errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep learning model is trained on all past data to maintain high detection accuracy, then anomaly detection precision is improved, but training time and computational resources increase significantly
Solution Approach 1:
The patent segments the training data into multiple time-based subsets (first time interval and second time interval) and trains separate anomaly detection sub-models for each subset. This allows the system to maintain multiple specialized models without requiring to retrain all models on complete historical data, significantly reducing training time while preserving detection precision for different time periods.
Solution Approach 2:
The system dynamically updates anomaly detection models by training new sub-models on recent data subsets while preserving knowledge from previous models through weight sharing. This dynamic approach enables continuous adaptation to changing data patterns without the need to retrain all models from scratch, balancing precision maintenance with reduced training time.
2Adaptability or versatility
If multiple anomaly detection sub-models are trained on different time intervals to capture changing patterns, then adaptability to data pattern changes is improved, but model complexity increases
Solution Approach 1:
The patent implements weight sharing across multiple anomaly detection sub-models, where the first network function weights are shared between the first and second anomaly detection sub-models. This universal weight structure allows different sub-models to function independently for their specific time intervals while maintaining a unified underlying representation, reducing overall model complexity despite having multiple sub-models.
Solution Approach 2:
The system nests multiple anomaly detection sub-models within a unified anomaly detection model framework. Each sub-model is trained on a specific time interval subset but shares common network function weights with other sub-models. This nested structure enables the system to handle multiple data patterns and time intervals while maintaining a compact, manageable model architecture.
3Measurement precision
If the model is updated frequently to adapt to new data patterns, then detection accuracy is improved, but the risk of forgetting previously learned knowledge increases
Solution Approach 1:
The patent trains anomaly detection sub-models on different time interval subsets in a predetermined sequence (first time interval subset then second time interval subset). This preliminary structured training approach ensures that knowledge from earlier time periods is captured before introducing newer patterns, allowing the system to adapt to new data while preserving previously learned knowledge through the sequential training strategy.
Solution Approach 2:
The system uses weight sharing as a feedback mechanism where the first network function weights trained on the first time interval subset serve as initial weights for the second anomaly detection sub-model. This feedback loop allows the model to retain previously learned patterns while adapting to new data, preventing catastrophic forgetting and maintaining detection accuracy across different time periods.
Data Source
AI summary
According to an exemplary embodiment of the present disclosure, disclosed is a computer program stored in a computer readable storage medium. When the computer program is executed in one or more processors, the computer program performs the following method for anomaly detection of data using a network function, and the method includes: generating an anomaly detection model including a plurality of anomaly detection sub models including a trained network function using a plurality of training data sub sets included in the training data set; calculating input data using at least one of the plurality of generated anomaly detection sub models; and determining whether there is an anomaly in the input data based on output data for input data of at least one of the plurality of generated anomaly detection sub models and the input data.