Unsupervised Anomaly Detection via Contrastive Latent Density Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Industrial dynamic systems face challenges in anomaly detection due to noisy sensor signals and high dimensionality in time-series data, making it difficult to apply supervised machine learning or accurately estimate probability density, especially when anomalies are rare.
Innovation Solution
A method using deep learning to transform multivariate time-series data into signature vectors in an embedding space, training a neural network to estimate probability distributions, and computing anomaly scores through contrastive density learning, allowing for unsupervised anomaly detection without labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised machine learning or classification approaches are applied, then anomaly detection accuracy can be improved, but the approach becomes inappropriate when the number of failures is very small
Solution Approach 1:
The patent inverts the traditional supervised learning approach by using unsupervised learning. Instead of training on labeled anomaly data (which is scarce), the system trains on normal operational data only, learning what constitutes normal behavior. Anomalies are then detected as deviations from this learned normal distribution, making the approach suitable for rare anomaly scenarios where labeled data is unavailable.
Solution Approach 2:
The system performs self-service by automatically learning the probability distribution of normal data without requiring external labeled anomaly data. The neural network autonomously identifies patterns in normal operational data and uses this knowledge to detect anomalies, eliminating the need for manual annotation of rare failure cases.
2Measurement precision
If probability density estimation is performed on noisy time-series data, then anomaly detection can be achieved, but the high dimensionality makes precise estimation challenging
Solution Approach 1:
The patent applies dimensionality reduction techniques to transform the high-dimensional noisy time-series data into a lower-dimensional embedding space. This transformation preserves the essential characteristics of the data while reducing complexity, making probability density estimation feasible and accurate in the reduced dimensional space.
Solution Approach 2:
The system extracts the essential features and patterns from the noisy high-dimensional time-series data by learning a probability distribution model. By focusing on the underlying distribution rather than the raw high-dimensional data, the system achieves accurate anomaly detection while filtering out noise and redundant information.
Data Source
AI summary
Anomaly detection in industrial dynamic process can include receiving a set of multivariate time series data representative of sensor data obtained over time. The set of multivariate time series data can be transformed into a set of signature vectors in an embedding space. A neural network can be trained to estimate a probability distribution of the set of signature vectors in the embedding space. Streaming data can be received. The streaming data can be appended with a previously stored time series data. The appended streaming data can be transformed into an embedding. The embedding can be input into the trained neural network, the trained neural network outputting a first probability distribution score. A second probability distribution score associated with the embedding can be determined based on a given proposed probability distribution. Anomaly score can be determined based on the first probability distribution score and the second probability distribution score.