Improving Anomalous Sound Detection through Pseudo-anomalous Set Selection and Pseudo-label Utilization under Unlabeled Conditions

TL;DR

The paper tackles anomalous sound detection when both suitable similar-machine data and detailed operation labels are scarce. It introduces three synergistic components: (1) pseudo-anomalous set selection from external data to augment normal-like samples, (2) triplet-loss–driven pseudo-labeling for unlabeled data, and (3) iterative learning to refine external-data selection and pseudo-labels across cycles. In unlabeled settings, the approach yields average AUC gains exceeding $6.6$ points and narrows the gap to fully supervised methods, while in labeled settings external data can provide additional boosts. The work demonstrates a practical ASD workflow that scales to diverse industrial contexts with limited annotation, providing robust deployment guidance and insights into data-volume trade-offs and loss-function choices.

Abstract

This paper addresses performance degradation in anomalous sound detection (ASD) when neither sufficiently similar machine data nor operational state labels are available. We present an integrated pipeline that combines three complementary components derived from prior work and extends them to the unlabeled ASD setting. First, we adapt an anomaly score based selector to curate external audio data resembling the normal sounds of the target machine. Second, we utilize triplet learning to assign pseudo-labels to unlabeled data, enabling finer classification of operational sounds and detection of subtle anomalies. Third, we employ iterative training to refine both the pseudo-anomalous set selection and pseudo-label assignment, progressively improving detection accuracy. Experiments on the DCASE2022-2024 Task 2 datasets demonstrate that, in unlabeled settings, our approach achieves an average AUC increase of over 6.6 points compared to conventional methods. In labeled settings, incorporating external data from the pseudo-anomalous set further boosts performance. These results highlight the practicality and robustness of our methods in scenarios with scarce machine data and labels, facilitating ASD deployment across diverse industrial settings with minimal annotation effort.

Figures (3)

• Figure 1: Overview of the proposed method, illustrating the integration of three key components within an iterative learning framework: (1) Selection of the pseudo-anomalous set from external data using a feature extractor, (2) Assignment of pseudo-labels to unlabeled original data via the same feature extractor, and (3) Iterative learning, where the model is retrained over multiple cycles using updated training data $\mathcal{X}_{\text{train}}$ derived from both the pseudo-anomalous set and original data to progressively improve performance.
• Figure 2: Overview of the process for selecting the pseudo-anomalous set from external data, consisting of three steps: (1) Training a baseline model on the original dataset with the subspace loss $\mathcal{L}_{\text{ss}}$, (2) Processing external data through the baseline model to compute anomaly scores, which are then sorted and filtered using machine-specific thresholds $a^\text{thr}_{\text{machineA}}, a^\text{thr}_{\text{machineB}}, \ldots$ to select the pseudo-anomalous set, and (3) Retraining the baseline model on the combined dataset to refine predictions with the subspace loss $\mathcal{L}_{\text{ss}}$.
• Figure 3: Overview of the proposed method for assigning pseudo-labels to unlabeled data, consisting of three steps: (1) Training a baseline model on the original dataset by passing it through the function $f$ to produce category outputs $z^{\text{cat}}$, optimized using the loss function $\mathcal{L}_{\text{mlt}}$, (2) Using the trained baseline model to predict and assign pseudo-labels to unlabeled data, generating $x, l^{(\text{pseudo})}$, and (3) Retraining the model on the pseudo-labeled data $x, l^{(\text{pseudo})}$ to iteratively improve predictions via the loss function $\mathcal{L}_{\text{mlt}}$.