VAE Anomaly Detection Training with Corrupted Sample Filtering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing systems face challenges in processing and consuming voluminous logs and other data sets due to their intricate nature and variability, making it difficult to identify irregular patterns that signify noteworthy occurrences such as system malfunctions or security infringements, necessitating automated anomaly detection systems.

Innovation Solution

An anomaly detection model is trained using a variational autoencoder to identify and remove corrupted samples from the training dataset based on estimated sample weights, ensuring the model is trained exclusively on normal samples.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional anomaly detection models are trained on datasets containing corrupted samples, then training efficiency is improved, but model accuracy deteriorates due to learning from imperfect data

Engineering Contradiction:
Improvetraining efficiencyVSAvoidanomaly detection accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system performs preliminary identification and removal of corrupted samples from the training dataset before training the anomaly detection model. By pre-filtering the training data to eliminate corrupted entries, the model trains exclusively on clean data, thereby maintaining high training efficiency while achieving superior anomaly detection accuracy without being exposed to harmful corrupted patterns during the learning process.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If corrupted samples are removed from the training dataset, then model accuracy is improved, but data processing complexity increases

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoiddata processing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system employs a self-service mechanism where the training data automatically identifies and flags corrupted samples through statistical analysis and anomaly detection algorithms during the preprocessing phase. This automated self-filtering process eliminates the need for manual data cleaning, reducing processing complexity while ensuring high accuracy by removing corrupted entries before model training begins.

Inventive Principle:
Principle #25Self-service

3Measurement precision

If manual cleaning of training data is performed to remove corrupted samples, then model accuracy is improved, but time consumption increases

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoiddata preparation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system replaces manual mechanical data cleaning processes with automated computational algorithms that can rapidly identify and remove corrupted samples through statistical analysis, pattern recognition, and machine learning-based detection. This substitution dramatically reduces the time required for data preparation while maintaining high anomaly detection accuracy, as the automated systems can process large datasets much faster than manual methods.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4664357A1Learning from imperfect data for anomaly detection
Publication Date: 2025.12.17 NOKIA SOLUTIONS & NETWORKS OY
  • EP4664357A1 patent drawingFigure 1
  • EP4664357A1 patent drawingFigure 2A
  • EP4664357A1 patent drawingFigure 2B

AI summary

Apparatus and method of training Machine Learning (ML) models. In an embodiment, the apparatus performs initial training of an anomaly detection model based on training samples of a training dataset over multiple epochs, where the anomaly detection model comprises a variational autoencoder (VAE). For each training sample during an epoch, the initial training comprises inputting an original data sequence of the training sample into the VAE encoder to output a multivariant distribution in latent space, sampling the multivariant distribution to generate multiple latent vectors, inputting the latent vectors into the VAE decoder to output reconstructed data sequences, and computing an estimated sample weight for the training sample. The apparatus identifies, after multiple epochs, corrupted samples from the training dataset based on the estimated sample weights, removes the corrupted samples to generate a filtered training dataset, and performs final training of the anomaly detection model based on the filtered training dataset.