Invertible Neural Network Training for Sensor Signal Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems struggle to reliably detect sensor signal outliers, which can lead to unpredictable and diminishing performance of machine learning systems, especially in automated devices.
Innovation Solution
A method for training an invertible neural network that maps sensor signals from a known 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
Engineering Contradiction Analysis
1Reliability
If statistical methods or standard machine learning systems are used to process sensor signals, then general processing capability is maintained, but reliability in detecting outliers and anomalies deteriorates
Solution Approach 1:
The system performs preliminary action by pre-defining a hypervolume in the latent space before actual anomaly detection occurs. This hypervolume is constructed from training data representations, establishing a reference region for normal data patterns. During inference, any data point falling outside this pre-established hypervolume is immediately identified as an anomaly, enabling reliable detection without complex real-time analysis.
Solution Approach 2:
The invention extracts the essential characteristic of normal data by isolating it within a specific hypervolume region in the latent space. By separating normal data representations from anomalous ones through this geometric boundary, the system achieves reliable anomaly detection. The hypervolume extraction captures the core pattern of normal operation, allowing straightforward classification of deviations.
2Measurement precision
If the neural network is trained to map all sensor signals to the hypervolume center, then normal signals are well-represented, but the ability to detect anomalies deteriorates
Solution Approach 1:
The loss function applies local quality by treating normal and anomalous data points differently in the latent space. For normal data points inside the hypervolume, the loss function encourages mapping toward the center (minimizing distance). For anomalous points outside the hypervolume, the loss function encourages mapping away from the center (maximizing distance). This localized differential treatment enables both precise representation of normal signals and reliable detection of anomalies.
Solution Approach 2:
The training approach introduces asymmetry by applying different optimization goals to different regions of the latent space. Data points within the hypervolume boundary are pulled toward the center, while points outside are pushed away. This asymmetric treatment creates a clear separation between normal and anomalous representations, enabling the network to maintain high precision for normal signals while simultaneously achieving reliable anomaly detection.
3Productivity
If standard training methods are used, then training speed is maintained, but the ability to distinguish normal from anomalous signals deteriorates
Solution Approach 1:
The invention changes the parameter space by working in the latent space rather than the original sensor signal space. By transforming data through the neural network to a latent representation and performing hypervolume-based training in this transformed space, the system achieves better anomaly discrimination. The parameter change from raw sensor values to latent representations enables more effective separation of normal and anomalous patterns while maintaining training efficiency through gradient-based optimization.
Data Source
AI summary
A computer-implemented method for training a neural network, wherein the neural network is an invertible neural network configured for accepting a sensor signal as input. The method includes: obtaining a first sensor signal from a dataset; sampling a value from a latent space of the neural network, wherein the value is sampled from a predefined hypervolume in the latent space; determining a second sensor signal from the sampled value by inversely mapping the sampled value through the neural network; determining a first latent representation by forward mapping the first sensor signal through the neural network; determining a second latent representation by forward mapping the second sensor signal through the neural network; determining a loss value from a loss function; and training the neural network based on a negative gradient of the loss value.