Source-linked AI summary
Anomalous sound detection based on interpolation deep neural network
Kaori Suefusa, Tomoya Nishida, Harsh Purohit, Ryo Tanabe, Takashi Endo, Yohei Kawaguchi
TL;DR
Non-stationary machinery sounds can make conventional autoencoder reconstruction errors large and variable even without anomalies, limiting reliable detection. The paper proposes interpolation deep neural network detection and reports that it outperformed conventional approaches, especially for non-stationary sounds.
Problem
Non-stationary machinery sounds can make conventional autoencoder reconstruction errors large and variable even without anomalies, limiting reliable anomalous-sound detection.
Method
The interpolation deep neural network removes a spectrogram’s center frame from its input and predicts that frame, using the interpolation error for anomaly detection.
Results
The proposed approach outperformed conventional approaches, especially on non-stationary machinery sounds.
Takeaways & Limitations
Avoiding edge-frame prediction is relevant to anomalous-sound detection for non-stationary machinery sounds.
Takeaways & Limitations
The approach relies on the assumption that avoiding edge-frame prediction improves detection performance.
Abstract
from arXiv · showhide
As the labor force decreases, the demand for labor-saving automatic anomalous sound detection technology that conducts maintenance of industrial equipment has grown. Conventional approaches detect anomalies based on the reconstruction errors of an autoencoder. However, when the target machine sound is non-stationary, a reconstruction error tends to be large independent of an anomaly, and its variations increased because of the difficulty of predicting the edge frames. To solve the issue, we propose an approach to anomalous detection in which the model utilizes multiple frames of a spectrogram whose center frame is removed as an input, and it predicts an interpolation of the removed frame as an output. Rather than predicting the edge frames, the proposed approach makes the reconstruction error consistent with the anomaly. Experimental results showed that the proposed approach achieved 27% improvement based on the standard AUC score, especially against non-stationary machinery sounds.
1. INTRODUCTION
The paper addresses the difficulty of maintaining industrial machinery with fewer skilled workers by proposing IDNN, which detects anomalies from interpolation errors rather than conventional reconstruction errors. IDNN removes a spectrogram’s center frame, predicts it from surrounding frames, and outperformed the conventional approach, especially for non-stationary machinery sounds.
- Motivation: Decreasing labor availability has made it difficult to maintain machinery using skilled technicians, motivating automatic diagnosis from operating sounds.Factory machinery failures cause significant costs, while conventional diagnosis relies on technicians listening to machines.
- Limitations of conventional methods: Conventional unsupervised detection uses autoencoders that reconstruct multiple spectrogram frames and identify anomalies from reconstruction errors.These approaches can achieve high performance but remain problematic for non-stationary machine sounds.
- Limitations of conventional methods: For non-stationary machinery sounds, reconstruction errors can be large regardless of anomalies because edge frames are difficult to predict.The resulting reconstruction-error variation can also be large.
- Proposed approach: IDNN removes the center frame from multiple spectrogram frames, predicts the missing frame, and detects anomalies from the interpolation error.The interpolation error is the difference between the predicted frame and the true frame.
- Experimental result: Experiments with real-life industrial machine sounds showed that IDNN outperformed the conventional approach, especially for non-stationary machinery sounds.The comparison evaluated the proposed approach against the conventional one.
2. CONVENTIONAL APPROACHES
Conventional unsupervised anomalous sound detection commonly uses deep neural networks with autoencoders or variational autoencoders. Despite high performance, these approaches face reconstruction errors unrelated to anomalies in non-stationary sounds and inefficiencies from reconstructing multi-frame inputs.
- Conventional models: Deep neural-network approaches for unsupervised anomalous sound detection include autoencoders (AEs) and variational autoencoders (VAEs).VAEs are used in a manner similar to AEs.
- Conventional models: AEs are trained on normal data to minimize reconstruction errors, defined as differences between original inputs and reconstructed outputs.The reconstruction error is used to detect anomalies after learning normal data.
- Limitations: Conventional approaches have three reported issues: non-stationary sounds inflate reconstruction errors, whole-input reconstruction increases parameters, and input-inclusive prediction can yield trivial solutions.The first issue arises from difficulty predicting edge frames.
3. PROPOSED APPROACH
The proposed interpolation DNN (IDNN) predicts a removed center spectrogram frame from surrounding consecutive frames, avoiding edge-frame prediction. The study also evaluates prediction-based and variational autoencoder variants to test the underlying hypothesis and compare network designs.
- Interpolation DNN: IDNN removes the center frame from consecutive input frames and predicts it as an interpolation output.The method is designed to avoid predicting edge frames, which motivated its name, interpolation DNN.
- Prediction DNN: PDNN was tested as an alternative that predicts the next frame from consecutive multiple input frames, directly examining whether avoiding edge-frame prediction improves detection.Its architecture and loss function were evaluated to verify the key assumption behind IDNN.
- Interpolation DNN: IDNN predicts only the center frame, reducing parameters for easier optimization and preventing trivial autoencoder solutions by embedding normal spectrotemporal structure.The removed target frame cannot be copied directly from the input, encouraging the model to learn normal-sound structure.
- Model variants: Four approaches were evaluated: IDNN and VIDNN using AE or VAE networks, and PDNN and VPDNN as their prediction-based counterparts.IDNN and VIDNN share the same input feature vector but use AE and VAE networks, respectively.
4. EXPERIMENT
The experiment evaluated IDNN and related models on real-life machinery sounds using AUC, with results showing advantages for non-stationary valve and slider sounds. Restoration examples further indicate that IDNN produced anomaly-consistent errors and was more robust when edge-frame prediction was difficult.
- The dataset contained 24,490 normal and 5,620 anomalous sound segments across machine types, each comprising seven individual machines.
- Models used log-Mel spectrogram inputs and were trained separately for each machine type, with performance evaluated by receiver-operating-characteristic AUC over three iterations per machine.The spectrogram used frame size 1024, hop size 512, and 64 Mel filter banks.
- IDNN achieved significantly higher AUC than AE and PDNN for valve sounds, while IDNN and PDNN both exceeded AE for slider sounds; the methods performed similarly for fan and pump sounds.
- IDNN outperformed the conventional approach on non-stationary valve and slider sounds, motivating subsequent analysis using valve sounds.
- For anomalous valve sounds, IDNN produced a properly large anomaly score, whereas AE produced a smaller error and PDNN produced a large error regardless of the sound.IDNN’s reconstructed spectrogram resembled the normal valve sound, indicating accurate training for non-stationarity.
- IDNN was more robust when short-duration sound changes made predicting edge frames difficult.
5. CONCLUSION
The paper proposes interpolation error from AE/VAE as an anomaly score that avoids predicting edge frames and outperforms conventional approaches, particularly for non-stationary sounds. The study used four input frames and one output, while the effects of these parameters on detection rate remain for future assessment.
- 5. CONCLUSION: The proposed approach uses AE/VAE interpolation error as an anomaly score, avoiding the difficulty of predicting edge frames.The study set the number of input frames to four and the output to one.
- 5. CONCLUSION: The approach outperformed conventional approaches particularly for non-stationary sounds.
- 5. CONCLUSION: Further studies are needed to assess how the number of input frames and output affects the detection rate.