Deep Latent Feature Extractors With Self-Supervised Sensor Pretraining

The patent applies preliminary action by pre-training the neural network on large amounts of unlabeled industrial sensor data using self-supervised learning objectives (reconstruction, prediction, classification). This preliminary training phase equips the network with domain-specific knowledge and feature extraction capabilities before the actual task-specific training, thereby reducing the amount of labeled data needed for the final application.

The patent implements self-service through self-supervised learning where the neural network learns from unlabeled data by creating its own training objectives. The network performs reconstruction of input data, prediction of future states, or classification without external labels, thereby serving itself to acquire knowledge from the abundant unlabeled industrial sensor data available in the domain.

The neural network performs self-service by automatically learning from unlabeled industrial sensor data through self-supervised learning objectives. It creates its own training signals through reconstruction, prediction, or classification tasks without requiring external human labeling, thereby eliminating the need for expensive and time-consuming expert annotation while still achieving high-quality feature representations.

The patent introduces an intermediary self-supervised learning framework that mediates between the unlabeled industrial sensor data and the final task-specific application. This intermediary training phase with reconstruction and prediction objectives acts as a bridge, transforming raw unlabeled data into meaningful feature representations that can be subsequently fine-tuned for specific tasks with minimal labeled data.

The patent applies dynamics by using a deep neural network with learnable parameters that can adapt to different tasks. Unlike static hand-crafted features, the neural network's features are dynamic and can be retrained or fine-tuned for different industrial automation tasks. The network architecture allows flexible adaptation to new tasks while maintaining the benefit of reduced labeled data requirements through its self-supervised pre-training capability.

The patent implements universality through a multi-functional neural network framework that can handle multiple industrial automation tasks. The same pre-trained network serves as a foundation for various applications including anomaly detection, predictive maintenance, and process optimization. The network's universal feature extraction capability, developed through self-supervised learning on diverse industrial sensor data, enables it to be adapted to different tasks without requiring task-specific feature engineering.