Machine Learning System for Sensor Signal Evaluation

The machine learning system is divided into two distinct sub-systems: a first sub-system for extracting latent representations from measurement signals, and a second sub-system for determining operating state variables from these representations. This segmentation allows the first sub-system to be trained unsupervised on abundant unlabeled data, while the second sub-system requires only minimal labeled data for supervised training, thereby resolving the contradiction between accuracy and data quantity requirements.

Latent representations serve as an intermediary between the raw measurement signals and the operating state variables. The first sub-system transforms measurement signals into compressed latent representations through unsupervised learning, which then serve as input to the second sub-system. This intermediary layer enables the system to leverage large amounts of unlabeled signal data while requiring only minimal labeled data for the final variable determination task.

By segmenting the learning process into two stages with different data requirements, the system can perform unsupervised pre-training on large amounts of unlabeled data quickly, followed by efficient supervised fine-tuning of the second sub-system using minimal labeled data. This segmentation dramatically reduces the total training time and computational resources compared to training a single system on large labeled datasets.

The first sub-system performs preliminary unsupervised learning to extract latent representations from measurement signals before the second sub-system performs supervised learning to determine operating state variables. This preliminary action leverages the abundance of unlabeled data to build a robust foundation quickly, reducing the need for extensive supervised training and thereby reducing overall training time and resource consumption.

The patent replaces conventional physical sensors with a virtual sensor system that uses machine learning models to determine operating state variables from existing measurement signals. Instead of adding complex physical hardware, the system substitutes a computational approach where the machine learning model processes existing sensor data to infer additional variables, thereby reducing system complexity and cost while maintaining measurement capabilities.

Latent representations act as an intermediary that bridges existing measurement signals and the target variables of interest. Rather than directly measuring difficult-to-obtain variables with complex physical sensors, the system uses the intermediary latent representations to transform readily available measurement data into useful information about operating state variables, simplifying the overall system architecture.