Learning Apparatus for Machine Degradation Detection via Sound Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abnormality detection techniques in industrial machines primarily focus on detecting the occurrence of abnormalities rather than the degradation condition, failing to accurately identify subtle changes in machine performance that indicate impending failure.
Innovation Solution
A learning apparatus that utilizes a feature vector transformation function, optimized using a neural network like Variational Autoencoder (VAE), with tailored loss functions to differentiate between normal and degradation conditions, allowing for early detection of machine degradation by analyzing sound signals without direct contact with the machine.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional abnormality detection techniques are used, then abnormality occurrence can be detected, but degradation condition (subtle changes indicating impending failure) cannot be detected
Solution Approach 1:
The system performs preliminary learning to establish a baseline of normal machine sound patterns before actual operation. This preliminary model captures the characteristic features of normal operation, enabling subsequent detection of subtle degradation deviations that conventional methods would miss. The preliminary action creates a reference framework that enhances both detection precision and reliability.
Solution Approach 2:
The patent replaces conventional mechanical/threshold-based detection systems with a data-driven machine learning approach. Instead of using fixed thresholds or simple statistical methods, the system employs neural networks and pattern recognition algorithms to analyze sound signals, transforming the detection mechanism from mechanical rule-based to intelligent adaptive detection, thereby improving both precision and reliability.
2Productivity
If machine operation continues without early degradation detection, then productivity is maintained, but unexpected failures occur causing loss of time
Solution Approach 1:
The system implements continuous feedback by monitoring machine sound patterns in real-time and comparing them against the learned normal pattern. When subtle degradation deviations are detected, the system provides early warning feedback, enabling maintenance planning that prevents unexpected failures while maintaining continuous operation. This feedback loop bridges productivity and reliability by enabling proactive maintenance decisions.
Solution Approach 2:
By detecting degradation signs in advance through preliminary pattern recognition, the system enables scheduled maintenance actions before failure occurs. This preliminary detection allows planning and execution of maintenance during convenient downtime rather than forcing unexpected operational interruption, thus maintaining productivity while minimizing time loss.
3Measurement precision
If direct contact with machine is used for analysis, then accurate measurement is possible, but machine operation is interrupted
Solution Approach 1:
The system introduces sound waves as an intermediary medium to detect machine degradation without direct contact. The acoustic field serves as a non-intrusive mediator that carries information about machine condition, allowing precise measurement of degradation patterns while the machine continues uninterrupted operation. This intermediary approach eliminates the need for physical contact that would halt production.
Solution Approach 2:
The patent replaces direct mechanical contact-based measurement systems with acoustic field-based detection. Instead of using sensors that require physical attachment or disassembly operations, the system uses microphone-based acoustic analysis to capture degradation signatures, thereby maintaining both measurement precision and operational continuity simultaneously.
Data Source
AI summary
According to one embodiment, a learning apparatus includes a memory and a hardware processor connected to the memory which learns a transformation function to extract a feature value of an input signal. The hardware processor updates the transformation function based on a signal indicative of a first condition and a signal indicative of a second condition which is different from the first condition, using a first loss function on the signal indicative of the first condition and a second loss function on the signal indicative of the second condition. The second loss function is designed such that the second condition becomes distant from the first condition.