Neural Network Reinjection for Unlabeled Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Anomaly detection in data streams using deep artificial neural networks is challenging due to the difficulty in obtaining labeled training data, especially for anomalous data, which is often scarce, leading to sub-optimal performance and inefficiencies in distinguishing regular from anomalous data.
Innovation Solution
A method involving a trained artificial neural network that implements an auto-associative function to replicate input data samples, with reinjection operations to compute distance parameters, comparing these to thresholds to classify data as anomalous or regular, and optionally performing adversarial data detection using classification functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If supervised learning approach is used to train the system to recognize regular and anomalous data, then the system can detect anomalies, but it requires extensive labeled training data which is difficult and costly to obtain
Solution Approach 1:
The system uses self-supervised learning where the model trains itself by predicting missing portions of input data sequences without requiring external labeled anomaly data. The model generates its own training signals by learning to reconstruct normal data patterns, enabling autonomous anomaly detection capability.
Solution Approach 2:
The patent changes the learning paradigm from supervised learning with labeled data to self-supervised learning with unlabeled data. By modifying the training objective to predict missing data points rather than classify labeled examples, the system achieves anomaly detection without requiring scarce labeled anomaly training data.
2Productivity
If supervised learning approach is used with class imbalance (more regular data than anomalous data), then the system can be trained on available data, but performance becomes sub-optimal due to the imbalance
Solution Approach 1:
The model generates its own balanced training signals by learning to reconstruct normal data sequences. By using self-supervised learning objectives on unlabeled normal data, the system creates equivalent learning pressure for detecting anomalies without being constrained by the actual distribution of labeled anomaly examples.
Solution Approach 2:
Instead of trying to detect anomalies directly through classification, the patent inverts the approach by training the model to reconstruct normal data sequences. Anomalies are detected when the reconstruction fails, effectively detecting what the model cannot predict rather than directly classifying anomaly labels.
3Measurement precision
If more labeled training data is collected to improve anomaly detection, then detection accuracy improves, but the cost and time required to obtain and process labeled data increases
Solution Approach 1:
The system eliminates the need for manual data labeling by using self-supervised learning. The model automatically generates training signals from unlabeled data sequences through reconstruction tasks, completely removing the time-consuming manual annotation process while maintaining detection accuracy.
Data Source
AI summary
The present disclosure relates to a method of anomaly detection using a trained artificial neural network (502) configured to implement at least an auto-associative function for replicating an input data sample at one or more outputs (A), the method comprising: a) injecting an input data sample into the trained artificial neural network (502) in order to generate a first replicated sample at the one or more outputs (A); b) performing at least one reinjection operation; c) computing a first parameter based on a distance between a value of an nth replicated sample present at the one or more outputs and a value of one of the previously injected or reinjected values; and d) comparing the first parameter with a first threshold (δ), and processing the input data sample as an anomalous data sample if the first threshold is exceeded.