Invertible Neural Network Training for Sensor Signal Anomaly Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing systems struggle to reliably detect sensor signal outliers, which can lead to unpredictable and diminishing performance, especially in automated systems relying on machine learning for perception and self-assessment.

Innovation Solution

A method for training an invertible neural network that maps sensor signals from a training dataset to a hypervolume center while pushing other signals outside, using a loss function with negative gradients to enhance anomaly detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If statistical methods or traditional machine learning systems are used to process sensor signals, then the system can extract useful information and perform classification, but the system cannot reliably detect outlier sensor signals, leading to unpredictable and diminishing performance

Engineering Contradiction:
Improveanomaly detection capabilityVSAvoidperformance with unseen data
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network exclusively on normal sensor signals before deployment. This one-class training approach prepares the system in advance to recognize only normal patterns, enabling it to automatically detect anomalies when encountering unseen or abnormal data during operation, thus resolving the contradiction between reliability for known patterns and adaptability to unknown patterns

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If the neural network is trained on diverse data to improve adaptability, then it can handle various sensor signals, but it loses the ability to reliably detect outliers as anomaly detection requires specialized one-class training

Engineering Contradiction:
Improvehandling various sensor signalsVSAvoidoutlier detection accuracy
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent applies the taking out principle by extracting and isolating only the normal class data for training purposes. By removing abnormal or outlier data from the training set and training exclusively on normal sensor signals, the neural network learns to recognize the boundaries of normal behavior, thereby achieving reliable anomaly detection while maintaining the ability to handle various normal sensor signal types

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4654045A1Device and method for training a one-class invertible neural network
Publication Date: 2025.11.26 ROBERT BOSCH GMBH
  • EP4654045A1 patent drawingFigure 1
  • EP4654045A1 patent drawingFigure 2
  • EP4654045A1 patent drawingFigure 3

AI summary

Computer-implemented method for training a neural network (60), wherein the neural network (60) is an invertible neural network configured for accepting a sensor signal as input, wherein the method comprises the steps of: • Obtaining a first sensor signal (xi) from a dataset (T); • Sampling a value (z̃i) from a latent space (s) of the neural network (60), wherein the value (z̃i) is sampled from a predefined hypervolume (h) in the latent space (s); • Determining a second sensor signal (x̃i) from the sampled value (z̃i) by inversely mapping the sampled value (z̃i) through the neural network (60); • Determining a first latent representation (z1) by forward mapping the first sensor (xi) signal through the neural network (60); • Determining a second latent representation (z2) by forward mapping the second sensor signal (x̃i) through the neural network (60); • Determining a loss value (l) from a loss function ( ), wherein the loss function ( ) comprises a first term that characterizes a difference of the first latent representation (z1) to a center (c) of the hypervolume (h) and wherein the loss function ( ) further comprises a second term that characterizes a negative difference of the second latent representation (z2) to the center (c) of the hypervolume (h); • Training the neural network (60) based on a negative gradient of the loss value (l).