Noise reduction auto-encoder-based anomaly detection model training method

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing abnormal-detection models face challenges in distinguishing abnormal data due to identity mapping and poor generalization ability, particularly when limited abnormal samples are available for training.

Innovation Solution

The method employs an improved denoising autoencoder to generate a rectangular frame with a preset resolution ratio, occlude original images, fill random noise, and perform constraint learning to enhance global feature learning and improve model performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional autoencoder is used for anomaly detection, then the model can learn normal sample patterns, but identity mapping occurs and generalization ability deteriorates

Engineering Contradiction:

Improveanomaly detection accuracyVSAvoidmodel generalization ability

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent segments the input image into multiple patches and applies different occlusion strategies to each patch (random occlusion, rectangular occlusion, or no occlusion). This segmentation approach prevents the model from learning identity mapping by ensuring that different regions are treated differently, thereby improving generalization ability while maintaining anomaly detection accuracy.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies occlusion masks to images before feeding them into the model during training. This preliminary action of occluding certain regions forces the model to learn robust features that are invariant to occlusions, preventing identity mapping and improving generalization to unseen anomaly patterns.

Inventive Principle:

Principle #10Preliminary action

2Adaptability or versatility

If occlusion is applied to prevent identity mapping, then generalization improves, but training complexity increases

Engineering Contradiction:

Improvemodel generalization abilityVSAvoidtraining process complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent dynamically selects occlusion strategies and adjusts occlusion parameters during training. Different occlusion modes (random, rectangular, or none) are applied with varying probabilities, and occlusion ratios are adjusted based on training progress. This dynamic approach maintains model simplicity while effectively preventing identity mapping and improving generalization.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS12387466B2Noise reduction auto-encoder-based anomaly detection model training method

Publication Date: 2025.08.12 INSPUR SUZHOU INTELLIGENT TECH CO LTD

AI summary

Disclosed is a method for training an abnormal-detection model based on an improved denoising autoencoder, including: acquiring an original image; generating a rectangular frame according to a preset range of a resolution ratio, by the improved denoising autoencoder, and occluding the original image by the rectangular frame, wherein the resolution ratio is a ratio of a resolution of an occlusion area formed by the rectangular frame to a resolution of the original image; filling random noise in the rectangular frame to obtain a noised image, by the improved denoising autoencoder; and performing constraint learning on the original image and the noised image, by the abnormal-detection model, to implement training of the abnormal-detection model. Because a learning task is more complex, it is helpful to alleviate identity mapping, and detection performance of the model is improved. The present application further provides an apparatus, a device, and a readable storage medium thereof.