Deep Belief Network Learning for Equipment Failure Signatures
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large manufacturers face challenges in integrating scattered data silos across different departments, making it difficult to predict equipment failures effectively due to differences in data measurement scales and proprietary coding schemes, which hinders asset performance improvement.
Innovation Solution
A method using failure signature information and a learning agent to detect equipment failures by identifying patterns in sensor data, with a global equipment taxonomy for classification and transfer learning, enabling the prediction of equipment failures across similar assets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data from multiple departments are integrated to improve equipment failure prediction, then prediction accuracy is improved, but data integration complexity increases due to different measurement scales and proprietary coding schemes
Solution Approach 1:
The patent transforms data from different measurement scales and proprietary coding schemes into a unified parameter space. By applying parameter transformation techniques, the system converts control system data (seconds), maintenance data (calendar time), and financial data (fiscal periods) into comparable parameters, enabling integration while managing complexity.
Solution Approach 2:
The patent introduces an intermediary layer that mediates between disparate data sources. This intermediary component handles the conversion and normalization of data from different departments, acting as a buffer that simplifies the integration process while maintaining prediction accuracy.
2Reliability
If population-based learning with deep belief networks is used to learn from multiple equipment units, then prediction reliability is improved, but computational complexity increases
Solution Approach 1:
The patent segments the learning process into distinct stages using deep belief networks. The network is divided into multiple layers that process information hierarchically, allowing the system to learn from population data in manageable segments rather than attempting to process all data simultaneously, thus reducing computational complexity while maintaining reliability.
Solution Approach 2:
The patent applies preliminary unsupervised pre-training to deep belief networks before fine-tuning with supervised learning. This preliminary action allows the network to learn useful feature representations from unlabeled population data first, reducing the computational burden during subsequent supervised training and improving overall prediction reliability.
Data Source
Figure 1A
Figure 1B
Figure 2A
AI summary
A plant asset failure prediction system and associated method. The method includes receiving user input identifying a first target set of equipment including a first plurality of units of equipment. A set of time series waveforms from sensors associated with the first plurality of units of equipment are received, the time series waveforms including sensor data values. A processor is configured to process the time series waveforms to generate a plurality of derived inputs wherein the derived inputs and the sensor data values collectively comprise sensor data. The method further includes determining whether a first machine learning agent may be configured to discriminate between first normal baseline data for the first target set of equipment and first failure signature information for the first target set of equipment. The first normal baseline data of the first target set of equipment may be derived from a first portion of the sensor data associated with operation of the first plurality of units of equipment in a first normal mode and the first failure signature information may be derived from a second portion of the sensor data associated with operation of the first plurality of units of equipment in a first failure mode. Monitored sensor signals produced by the one or more monitoring sensors are received. The first machine learning agent is then and activated, based upon the determining, to monitor data included within the monitored sensor signals.