Vibration Signal Reconstruction by Frequency Band for Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing equipment anomaly diagnosis technologies using vibration signals struggle to accurately detect and classify anomalies in equipment using unsupervised learning, particularly in the time domain, due to the complexity of vibration data and the need for effective threshold determination.
Innovation Solution
A training apparatus and method that utilizes a signal reconstruction model, such as autoencoders or LSTM autoencoders, to filter vibration signals by frequency bands, calculate reconstruction errors, and set threshold values based on the three-sigma rule, enabling precise anomaly detection and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If unsupervised learning type AI model is used to diagnose equipment abnormality using vibration data, then the ability to detect anomalies without labeled data is improved, but the accuracy of anomaly detection deteriorates due to the complexity of vibration data and difficulty in determining effective thresholds
Solution Approach 1:
The vibration signal is segmented into multiple frequency bands using filters. Each frequency band is processed independently through separate autoencoders, allowing the system to capture anomaly patterns at different frequency levels. This segmentation improves detection accuracy by focusing on specific frequency characteristics rather than treating the entire spectrum uniformly.
Solution Approach 2:
The patent transforms the vibration signal from time domain to frequency domain through filtering operations. By changing the parameter representation (from raw time-series to frequency-band components), the system can better identify anomaly patterns that are frequency-specific, thereby improving detection accuracy while maintaining automated operation.
2Measurement precision
If vibration signal is filtered by frequency bands and processed through multiple autoencoders, then the accuracy of anomaly detection is improved, but the device complexity increases due to multiple filters and models
Solution Approach 1:
The system divides the vibration signal processing into multiple independent frequency band channels. Each channel contains a filter and an autoencoder that operate independently. This modular segmentation allows for improved accuracy through frequency-specific analysis while keeping each individual processing unit relatively simple and manageable.
Solution Approach 2:
The patent employs multiple autoencoders that share the same architectural structure and training methodology across different frequency bands. This universal approach allows the system to handle multiple frequency bands with a standardized processing framework, reducing the complexity that would otherwise arise from designing custom models for each band.
3Measurement precision
If reconstruction error values are calculated for each frequency band with frequency-specific thresholds, then the accuracy of anomaly classification is improved, but the difficulty of detecting and measuring increases due to the need to manage multiple error values and thresholds
Solution Approach 1:
The threshold determination is segmented into independent processes for each frequency band. Each autoencoder's reconstruction error is evaluated separately, and thresholds are established independently for each band based on its specific statistical characteristics. This segmentation improves classification accuracy by accounting for frequency-specific variations while managing complexity through modular, independent threshold management.
Solution Approach 2:
The system automatically determines thresholds for each frequency band based on the statistical properties of the reconstruction errors from normal operation data. The three-sigma rule is applied automatically to each band's error distribution, eliminating the need for manual threshold setting and reducing the difficulty of managing multiple thresholds while maintaining high classification accuracy.
Data Source
AI summary
Provided is a training apparatus, a training method, a diagnosis apparatus for diagnosing equipment anomaly. A training apparatus includes a processor, and a memory configured to store instructions executable by the processor. The instructions cause the training apparatus to obtain a vibration signal of a time domain obtained by measuring a vibration that occurs when equipment is operating normally, obtain a vibration signal for each frequency band filtered according to defined frequency bands by inputting the vibration signal to at least one filter, train a signal reconstruction model using the vibration signal for each frequency band, obtain a reconstruction signal corresponding to each frequency band, determine a reconstruction error value showing a difference between the vibration signal for each frequency band and the reconstruction signal, and determine a threshold value for anomaly detection based on the reconstruction error value determined for each of the defined frequency bands.