PHM Sensor Fusion Using Deep Neural Networks for Fault Detection

The system performs self-service by automatically learning features from raw sensor data through deep neural networks, eliminating the need for external domain knowledge. The network autonomously identifies relevant features and patterns, making the system self-sufficient and adaptable to any application domain without requiring pre-existing expert knowledge.

The patent replaces the manual mechanical process of hand-crafting features by domain experts with an automated computational system. Deep neural networks automatically extract features from sensor data, substituting the manual expert process with an automated learning-based approach that scales across different applications without requiring domain-specific expertise.

The patent merges multiple heterogeneous sensor modalities into a unified deep neural network model. Different sensor types (vibration, temperature, pressure, etc.) are integrated as input features to the network, which automatically learns how to combine them. This combining approach achieves comprehensive monitoring while managing complexity through the network's automated feature integration capabilities.

The deep neural network serves as a universal processing framework that handles multiple sensor modalities simultaneously. The same network architecture can process different types of sensor data (vibration, temperature, pressure, acoustic emissions) and apply the same learning algorithms, providing a multi-functional solution that reduces overall system complexity despite handling diverse data sources.

The patent applies preliminary action by using unsupervised pre-training followed by supervised fine-tuning. The deep neural network is first pre-trained using autoencoders to learn general feature representations from unlabeled data, then fine-tuned with labeled fault data. This two-stage approach simplifies the overall training process by breaking it into manageable steps and reduces the difficulty of training from scratch.

The patent introduces an intermediary layer in the form of autoencoders that serve as a bridge between raw sensor data and the final classification task. The autoencoder intermediate layer learns compressed representations of the input data, which then serve as features for the subsequent supervised classification. This intermediary structure simplifies training by creating a hierarchical learning process where general features are learned first, then task-specific features are refined.