Correlation Matrix Memory for Novel Fault Detection in Asset Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Asset monitoring systems face challenges in detecting potential faults, especially those that have not been foreseen during asset development, as they lack training data and require complex pattern recognition techniques to identify deviations from normal behavior.
Innovation Solution
The AURA Alert system utilizes a Correlation Matrix Memory (CMM) to store binary patterns of asset states over time, allowing for quick comparison of current states to historical data, even if individual variables are within their normal range, using binning processes and kernel operations to convert measurements into binary vectors for efficient pattern matching.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If complex pattern recognition techniques are used to detect novel faults, then detection capability is improved, but device complexity increases
Solution Approach 1:
The system segments the complex pattern recognition task into distinct functional modules: a binning process that converts continuous measurements into discrete binary vectors, a correlation matrix memory structure that stores historical patterns, and a recall mechanism that compares current states against stored patterns. This segmentation transforms an intractable complex problem into manageable discrete operations.
Solution Approach 2:
The system changes the parameter representation from continuous measurement values to discrete binary vectors through the binning process. Each measurement is converted into a binary vector where bits represent different ranges or states, transforming the problem from continuous pattern matching to discrete correlation operations, which are computationally more efficient.
2Measurement precision
If millions of historical states are stored for comparison, then detection accuracy is improved, but memory requirements and processing time increase
Solution Approach 1:
The system creates simplified binary copies of historical asset states and stores them in correlation matrix memory. Instead of storing and processing full continuous measurement datasets, the system stores compact binary vector representations that capture the essential pattern information, enabling rapid comparison operations.
Solution Approach 2:
The system replaces complex mechanical or computational pattern matching mechanisms with a correlation matrix memory structure that uses simple binary operations. The recall process substitutes intensive continuous optimization algorithms with efficient binary vector comparisons, dramatically reducing processing time while maintaining detection accuracy.
3Productivity
If binning processes and kernel operations are applied to convert measurements, then pattern matching efficiency is improved, but measurement precision may be lost
Solution Approach 1:
The binning process intentionally changes measurement parameters from continuous to discrete binary values, accepting a controlled loss of precision in exchange for dramatic gains in processing efficiency. The kernel operation further transforms these binary vectors into a form optimized for correlation memory storage and retrieval, enabling scalable pattern matching.
Solution Approach 2:
The system applies partial binning where not all measurement dimensions are converted to binary form, or where the binary representation uses multiple bits per measurement to preserve more information. This partial action approach balances the trade-off between processing efficiency and precision retention, converting only the most critical measurements or using excessive binary representation to maintain accuracy.
Data Source
AI summary
To record events, a vector A is taken of n measurements An. Each measurement An is converted into a binary vector Bn using a binning process. Measurements in vector Bn, are transformed using a kernel operation, to make a vector Cn. All vectors Cn are concatenated into a single vector D. A binary vector M is created with P bits set. A date and vector A are recorded against vector M in a database Z. Vector D is associated into a correlation matrix memory R against vector M. To recall events, the above steps are repeated up to concatenating all vectors Cn into single vector D. Vector D is then used to access a correlation matrix memory R to recall an association vector S, and vector S is applied directly or indirectly to database Z to recover the or each date associated with vector S. This may facilitate the monitoring of assets such as generators, gas turbines, motors, and larger assets such as trains, boats and planes, to ensure that they continue to operate reliably.