Machine Sound Anomaly Detection With Timbre Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current anomaly detection methods using unsupervised machine learning fail to distinguish between different types of anomalies, treating all outliers similarly and lacking the ability to differentiate between relevant and irrelevant anomalous sounds in acoustic data from machines.
Innovation Solution
The use of a computer program product that employs a target autoencoder and an environmental autoencoder to reconstruct waveforms, with a neural network classifier determining whether an input waveform represents anomalous behavior of a target device by distinguishing between target device waveforms and unrelated sound events, enabling automatic response to anomalies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If unsupervised machine learning is used for anomaly detection, then the system can detect abnormal waveforms, but it cannot distinguish between different types of anomalies
Solution Approach 1:
The patent divides the anomaly detection task into two separate autoencoders: one trained on target device waveforms and another on environmental sound waveforms. Each autoencoder specializes in reconstructing its specific type of waveform, enabling the system to distinguish between device-generated anomalies and environmental noise by comparing reconstruction errors from both models.
2Ease of manufacture
If a single anomaly detection model is used, then the system is simple to implement, but it treats all anomalies the same without distinguishing relevant from irrelevant sounds
Solution Approach 1:
The patent introduces an intermediary classification layer that takes the reconstruction errors from both autoencoders as inputs and determines whether an anomaly is relevant (device-related) or irrelevant (environmental). This intermediary component preserves anomaly relevance information while building upon the simple autoencoder framework.
3Measurement precision
If acoustic data is used to monitor machine condition, then insights about inner workings are obtained, but irrelevant environmental sounds create false anomalies
Solution Approach 1:
The patent extracts and isolates environmental sound patterns by training a separate autoencoder exclusively on environmental waveform data. This dedicated model reconstructs environmental sounds, allowing the system to identify and filter out environmental noise from the target device monitoring, thereby reducing false anomalies while preserving genuine machine condition insights.
Data Source
AI summary
Methods, systems, and computer program products for detecting anomalous behavior include reconstructing a waveform of a target device from an input waveform using a target autoencoder. A waveform of unrelated sound events is reconstructed from the input waveform using an environmental autoencoder. The input waveform is classified to determine that the input waveform is produced by anomalous behavior of the target device using a classifier, based on the reconstructed waveform of the target device and the reconstructed waveform of the unrelated sound events. An automatic response to the anomalous behavior is generated.