Adversarial Learning Network for Multi-Type Data Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection algorithms face challenges in accurately handling multi-type data, as they either require all normal samples for approximation-based methods, leading to increased operational pressure with scaled data, or fail to capture the variety of normal modes in real-life scenarios, limiting the accuracy of model-based methods.
Innovation Solution
An adversarial learning network is trained to fit the distribution of normal samples and learn potential modes, constructing an anomaly evaluation function based on reconstruction errors, which is then used to build an anomaly detection model for improved accuracy in multi-type data scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If approximation-based anomaly detection algorithm is used, then anomaly detection can be performed using degree of approximation between test sample and normal sample, but operational pressure increases significantly when data scale expands
Solution Approach 1:
The patent applies preliminary action by pre-training a generative model on normal samples before actual anomaly detection. The model learns the distribution characteristics of normal data in advance, creating a reference framework that enables efficient anomaly detection without requiring all normal samples during the detection phase. This pre-processing step resolves the contradiction by shifting computational burden from the detection phase to the training phase.
Solution Approach 2:
The patent uses copying by generating synthetic normal samples through the trained generative model. Instead of comparing test samples against all original normal samples, the model creates copies (synthetic samples) that represent the normal distribution, enabling efficient approximation-based anomaly detection while maintaining accuracy. This resolves the operational pressure issue by replacing large-scale sample storage and processing with compact model-based generation.
2Productivity
If model-based anomaly detection algorithm is used, then operational efficiency is improved by learning from normal data, but accuracy deteriorates when normal samples contain multiple potential modes that are ignored
Solution Approach 1:
The patent applies segmentation by decomposing the complex multi-mode normal distribution into multiple potential modes. The generative model is trained to identify and learn distinct modes (e.g., different body shapes, sexes in X-ray images) separately rather than treating normal data as a single unified distribution. This segmentation enables the model to capture the diversity of normal variations, improving anomaly detection accuracy while maintaining operational efficiency through the learned structured representation.
Solution Approach 2:
The patent uses parameter changes by modifying the generative model's parameters during training to adapt to different potential modes of normal samples. The model dynamically adjusts its parameters to fit the characteristics of various normal modes, enabling it to represent multi-mode distributions accurately. This allows the model to maintain high operational efficiency while capturing the complexity of real-world normal data variations.
3Device complexity
If unified mode learning is applied to normal samples, then model complexity is reduced, but the ability to capture effective modes different from abnormal samples deteriorates
Solution Approach 1:
The patent applies dynamics by creating a dynamic and flexible model structure that can adapt to different normal modes. Rather than using a rigid unified mode, the generative model employs dynamic parameters that can adjust to capture various potential modes of normal data. This dynamic approach maintains reasonable model complexity while significantly improving the ability to distinguish normal variations from anomalies, thereby enhancing detection reliability.
Solution Approach 2:
The patent uses dimensionality change by introducing latent space dimensions that represent different potential modes of normal samples. The generative model maps normal samples into a latent space where different modes are represented as distinct regions or clusters. This additional dimensional representation allows the model to capture multi-mode characteristics without proportionally increasing model complexity, improving reliability by better separating normal variations from abnormal patterns.
Data Source
AI summary
The present disclosure provides an anomaly detection method and apparatus for multi-type data. According to the anomaly detection method for multi-type data, an adversarial learning network is trained, so that a generator in the adversarial learning network fits a distribution of a normal training sample and learns a potential mode of the normal training sample, to obtain an updated adversarial learning network, an anomaly evaluation function in the updated adversarial learning network is constructed according to a reconstruction error generated during training, and the updated adversarial learning network is constructed into an anomaly detection model, to perform anomaly detection on inputted detection data by the anomaly detection model, to obtain an anomaly detection result. A mode classifier is introduced to effectively resolve difficult anomaly detection when a distribution of detected data is similar to that of normal data, further improving the accuracy of anomaly detection.