Sound Anomaly Detection Using Augmented Waveform Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods for sound data face challenges in distinguishing anomalous sounds from normal sounds, particularly due to the limitations of geometric transformation when applied to audio samples, and the inefficiency of binary classification with imbalanced datasets.
Innovation Solution
A neural network is trained using data augmentation techniques on sound waveforms, generating labeled normal and abnormal samples to identify the type of data augmentation performed, with abnormal samples contributing differently to the training loss function, enhancing the model's ability to detect anomalies through confidence-based scoring.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If geometric transformation principles are applied to audio samples, then the detection capability may be improved, but the performance deteriorates because geometric transformation is designed for images not audio
Solution Approach 1:
Instead of applying image-based geometric transformation to audio (which fails), the patent inverts the approach by using data augmentation techniques specifically designed for audio signals. The system applies transformations such as pitch shifting, time stretching, and noise addition to audio waveforms, which are appropriate for the audio domain while still achieving the goal of enhancing anomaly detection capability.
Solution Approach 2:
The patent transforms audio samples by modifying key audio parameters including pitch, timing, and noise characteristics through data augmentation. This changes the parameter space in which the neural network operates, allowing it to learn more robust features for anomaly detection without applying inappropriate geometric transformations.
2Device complexity
If binary classification is used with imbalanced datasets, then the classification process is simple, but the detection accuracy worsens due to the imbalance between normal and abnormal samples
Solution Approach 1:
The patent applies data augmentation to abnormal samples before training the neural network, creating synthetic abnormal samples that balance the dataset. This preliminary action of generating additional abnormal samples through pitch shifting, time stretching, and noise addition ensures that the neural network receives balanced training data, improving detection accuracy without significantly increasing classification complexity.
3Quantity of substance
If data augmentation is applied to both normal and abnormal samples, then the training dataset size increases, but the labeling complexity increases because abnormal samples must be labeled differently
Solution Approach 1:
The patent uses a feedback mechanism where the neural network is trained to identify the type of data augmentation applied to each sample. During training, abnormal samples are augmented and labeled with both the original abnormal class and the applied augmentation type. This feedback loop allows the system to learn robust features while automatically managing the labeling complexity through the network's classification capability.
Data Source
AI summary
Methods and computer program products for training a neural network perform multiple forms of data augmentation on sample waveforms of a training dataset that includes both normal and abnormal samples to generate normal data augmentation samples and abnormal data augmentation samples. The normal data augmentation samples are labeled according to a type of data augmentation that was performed on each respective normal data augmentation sample. The abnormal data augmentation samples are labeled according to a type of data augmentation other than that which was performed on each respective abnormal data augmentation sample. A neural network model is trained to identify a form of data augmentation that has been performed on a waveform using the normal data augmentation samples and the abnormal data augmentation samples.