Adaptive Normal Sound Modeling for Stable Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection techniques for industrial equipment face a trade-off between cost and accuracy, with methods requiring large amounts of data for accurate model learning, and instability in learning without pair data in domain transformation.
Innovation Solution
The proposed solution involves a probability distribution learning apparatus that adaptively learns normal models for individual equipment using a common model, and employs adaptive batch normalization in Normalizing Flow to stabilize domain transformation without pair data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If a common normal model is learned using normal sound from multiple pieces of equipment, then the cost of data collection and learning is reduced, but the accuracy of anomaly detection deteriorates due to inability to capture slight differences for each equipment
Solution Approach 1:
The patent segments the learning process into two distinct phases: (1) learning a common normal model from multiple equipment to capture general characteristics, and (2) adaptively learning equipment-specific normal models by fine-tuning the common model using small amounts of equipment-specific data. This segmentation allows the system to benefit from both the cost efficiency of shared learning and the accuracy of individualized modeling.
Solution Approach 2:
The patent applies local quality by allowing each equipment to have its own adapted normal model parameters while sharing the overall model structure. The equipment-specific models capture local characteristics and slight differences for each individual equipment, while the common model provides global knowledge. This enables high-accuracy anomaly detection tailored to each equipment without requiring large amounts of data for every single device.
2Measurement precision
If normal models are learned individually for each piece of equipment using only normal sound from that equipment, then the accuracy of anomaly detection is improved, but the cost of data collection and learning increases with the number of equipment
Solution Approach 1:
The patent merges the learning processes by using a common normal model that is shared across multiple equipment. Instead of completely independent learning for each equipment, the system combines knowledge from multiple sources through the common model, then adapts it locally. This merging approach reduces redundant learning and data collection costs while maintaining individualized accuracy through adaptive fine-tuning.
Solution Approach 2:
The common normal model serves as a universal foundation that can be applied to multiple equipment of the same type. This universal model captures general characteristics that are common across all equipment, making it multi-functional. The model can then be efficiently adapted to specific equipment instances, reducing the overall cost of deploying anomaly detection systems across large numbers of devices.
3Ease of operation
If domain transformation is performed without pair data using conventional methods, then the requirement for data collection is reduced, but the learning stability deteriorates
Solution Approach 1:
The patent changes the parameters and objective functions used in domain transformation learning. Instead of conventional approaches that rely on pair data and adversarial training, the patent employs parameter-based domain transformation with carefully designed objective functions that ensure stable convergence. The transformation parameters are optimized to maintain distribution consistency while achieving domain adaptation, providing both ease of operation and learning stability.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
An anomaly detection technique which realizes high accuracy while reducing cost required for normal model learning is provided. An anomaly detection apparatus includes an anomaly degree estimating unit configured to estimate an anomaly degree indicating a degree of anomaly of anomaly detection target equipment from sound emitted from the anomaly detection target equipment (hereinafter, referred to as anomaly detection target sound) based on association between a first probability distribution indicating distribution of normal sound emitted from one or more pieces of equipment different from the anomaly detection target equipment and normal sound emitted from the anomaly detection target equipment (hereinafter, referred to as normal sound for adaptive learning).