HARMONY: Hierarchical Anchor Retrieval on Manifold for Oblivious-source Machinery Acoustic Anomaly Detection

Yuhang Zhang; Yang Li; Qiang Wu; Mingyuan Li

cs.SDOct 22, 2026

HARMONY: Hierarchical Anchor Retrieval on Manifold for Oblivious-source Machinery Acoustic Anomaly Detection

Authors: Yuhang Zhang, Yang Li, Qiang Wu, Mingyuan Li

Organizations: School of Software and Microelectronics, Peking University, Beijing, China · Independent Researcher · Collaborative Innovation Center for Western Ecological Safety, Lanzhou University, Lanzhou, China · ELLIS Institute of Finland ·

Abstract

In standard acoustic anomaly detection (ASD), models are usually trained separately for each machine or domain using explicit metadata. However, in realistic deployments, machine identifiers are incomplete, unreliable, or simply unavailable. Thus, metadata-free machinery ASD asks a single model to monitor mixed machine populations without machine IDs. In this paper, we argue that the main failure mode is not a weak detector family, but representation entanglement. The heterogeneous normal modes overlap in frozen pre-trained feature spaces, making anomaly scoring and approximate retrieval unreliable. We study this problem directly and propose HARMONY (Hierarchical Anchor Retrieval on Manifold for Oblivious-source acoustic aNomalY detection), an anchor-guided geometric repartitioning framework that reorganizes mixed-source features into compact Voronoi regions and reuses this structure for hierarchical retrieval. On DCASE 2020, HARMONY improves mean AUC from 75.40% to 90.00% (14.60 percentage points) over the strongest unified baseline and from 83.90% to 90.00% over a strong foundation-feature baseline. On a single CPU, its hierarchical retrieval achieves a 7.8x speedup over exhaustive search with only a 0.20 percentage-point AUC drop. These results suggest that incorporating explicit geometric partitioning can improve both detection accuracy and retrieval efficiency for ASD. Code is available at https://github.com/Zzyh0208/HARMONY.

Figures & tables

Figure 1: Overview of HARMONY. Spectral anchors define the coarse topology, anchor-guided projection compacts mixed-source features, and the same geometry is reused for hierarchical retrieval.

Method	Unified Setting		Per-Region Setting		Cost Transition (Uni $\to$ Reg)
Method	AUC	pAUC	AUC ( $\Delta$ )	pAUC	Train Time (h)	Latency (ms)
PatchCore	62.10	56.20	91.80 +29.70	78.90	0 $\to$ 0	41.8 $\to$ 30.6
Audio-PC	83.90	71.30	93.10 +9.20	80.40	0 $\to$ 0	67.5 $\to$ 52.1
EfficientAD	66.30	58.70	88.90 +22.60	76.10	2.4 $\to$ 18.5	6.3 $\to$ 6.3
Audio-MAE	86.10	72.80	92.40 +6.30	79.60	14.2 $\to$ 108.0	82.4 $\to$ 82.4
MCDD	78.30	66.90	89.30 +11.00	76.80	7.1 $\to$ 55.6	54.7 $\to$ 44.9

Table 1: Unified vs. per-region training (%). Baseline AUC gains (

\Delta

) under per-region training expose mixed-source geometry as the core bottleneck, yet partitioning severely inflates costs. HARMONY natively achieves the optimal accuracy-efficiency balance.

Methods	Machine Types						Average
Methods	ToyCar	ToyConv	Fan	Pump	Slider	Valve	Average
PatchCore	62.80	60.60	62.20	57.20	67.80	61.50	62.10
Classic Adaptation
SimpleNet	63.20 +0.40	62.10 +1.50	56.70 -5.50	58.00 +0.80	66.30 -1.50	62.20 +0.70	61.40 (-1.1%)
EfficientAD	65.80 +3.00	65.60 +5.00	65.70 +3.50	68.50 +11.30	66.50 -1.30	65.60 +4.10	66.30 (+6.8%)
Foundation Embeddings

Table 2: Metadata-free machinery anomaly detection accuracy (AUC, %) on DCASE 2020. Absolute sub-metric changes ( + / - ) and relative average improvements (%) are computed against the PatchCore baseline. All methods share the same pre-trained features.

Method	MIMII	DCASE 2022	Target
Audio-PC	$79.73\pm 0.44$	$77.84\pm 0.46$	$73.72\pm 0.59$
UniFlow	$71.24\pm 0.79$	$72.69\pm 0.81$	$68.43\pm 0.98$
Dinomaly	$72.06\pm 0.71$	$73.21\pm 0.73$	$68.97\pm 0.91$
HARMONY	$84.57\pm 0.58$	$82.36\pm 0.52$	$77.99\pm 0.68$

Table 3: Cross-dataset AUC (%, mean

\pm

standard deviation over five seeds). MIMII uses the

-6

dB setting. DCASE 2022 is evaluated with machine metadata withheld from the model.

Variant	Machine Types						Average	$\Delta$
Variant	ToyCar	ToyConv	Fan	Pump	Slider	Valve	Average	$\Delta$
Full Model	98.00	89.70	88.70	89.70	87.60	90.00	90.00	-
Ablated Components
w/o Soft Assignment	96.50	87.00	85.00	87.50	85.20	84.40	87.60	-2.40
w/o Spectral Init	89.50	76.50	78.20	82.50	79.00	83.10	81.40	-8.60
w/o Projection	64.10	56.50	58.80	69.20	62.30	63.50	62.40	-27.60

Table 4: Component ablation study on DCASE 2020 (%). The full flat-search model serves as the baseline. Performance drops (

\Delta

) are highlighted in red . The final deployed HARMONY (Hierarchical) maintains near-lossless accuracy.

Panel of Figure 2 — Figure 2: Robustness and stability. Left: HARMONY remains stable as SNR drops to $-6$ dB, unlike degrading baselines. Right: Silhouette Score tracks AUC, enabling unsupervised model selection.

Feature Space	QE ( $\downarrow$ )	Margin ( $\uparrow$ )	Ratio ( $\uparrow$ )	Silh. Score ( $\uparrow$ )	AUC ( $\uparrow$ , %)
Raw BYOL-A	0.842	0.418	0.496	0.331	62.40
HARMONY	0.181	0.852	4.707	0.824	90.00

Table 5: Empirical validation of geometric metrics (measured via aligned cosine distance).

Groups	Selected $K$	Selected AUC	Oracle gap
2 types	4 / 2–6	92.84	0.27
3 types	6 / 4–8	91.97	0.32
4 types	8 / 6–12	91.31	0.27
5 types	8 / 8–12	90.64	0.37
6 types	8 / 8–12	90.07	0.19

Table 6: Normal-only anchor selection across mixtures of

C

machine types. Counts show mode / range over five seeds; AUC entries are five-seed means (%). The test oracle is used only for retrospective comparison.

Retrieval Method	Accuracy (%)		Efficiency (CPU)
Retrieval Method	AUC ( $\uparrow$ )	pAUC ( $\uparrow$ )	Latency ( $\downarrow$ )	Speedup ( $\uparrow$ )
HARMONY (Flat)	90.00	74.90	42.3 ms	1.0 $\times$
HARMONY (IVF)	89.80 -0.20	74.70 -0.20	5.4 ms (-87.2%)	7.8 $\times$

Table 7: Efficiency versus accuracy trade-off on CPU. Hierarchical retrieval (IVF) offers substantial latency reduction ( green ) with negligible accuracy drop ( red ) compared to exhaustive search (Flat).

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Definition
$\phi:\mathcal{X}\rightarrow\mathcal{Z}$	Pre-trained frozen encoder mapping audio to feature space.
$L$ / $K$	Latent normal-mode count / learned anchor count (not necessarily equal).
$D$	Ambient dimension of the full input feature space $\mathcal{Z}$ .
$\phi_{\#}\mathcal{P}$	Pushforward measure of probability distribution $\mathcal{P}$ under mapping $\phi$ .
$g_{\theta}:\mathcal{Z}\rightarrow\mathcal{H}$	Trainable projection network mapping features to hypersphere $\mathbb{S}^{d-1}$ .
$d^{\prime}$ / $\eta$	Intrinsic dimensionality / Cumulative energy threshold for spectral filtering.

Appendix

Table 8: Summary of key mathematical notations.

Figure 3: Anchor occupancy statistics across varying machine complexities. (Left) The Entropy Ratio remains near 1.0, indicating uniform utilization. (Right) The maximum occupancy percentage tightly tracks the ideal uniform distribution boundary, showing nonempty anchors in these evaluated configurations.

Figure 4: Smooth optimization trajectory of the alignment loss

\mathcal{L}_{align}

over 20 epochs, showing the observed loss trajectory.

Group	Count	Mean	Median	95th Pctl	Min
Normal	5,399	0.013	0.011	0.022	0.003
Anomalous	5,469	0.187	0.164	0.361	0.021

Appendix

Table 9: Distribution of Anomaly Scores

s(q)

for normal vs. anomalous test queries.

Figure 5: t-SNE [ 19 ] visualization of manifold disentanglement. Projections transition from entangled raw features (Left) to separated Voronoi regions (Right), illustrating the learned feature organization.

Dataset / method	NMI	Purity	Hom./Comp.
DCASE / Raw KMeans	$0.42\pm 0.03$	$0.63\pm 0.03$	0.49 / 0.38
DCASE / MCDD	$0.66\pm 0.03$	$0.81\pm 0.02$	0.74 / 0.57
DCASE / HARMONY	$0.78\pm 0.02$	$0.89\pm 0.01$	0.85 / 0.65
MIMII / MCDD	$0.61\pm 0.03$	$0.77\pm 0.03$	0.69 / 0.51
MIMII / HARMONY	$0.72\pm 0.03$	$0.85\pm 0.02$	0.80 / 0.59

Appendix

Table 10: Post-hoc alignment with machine types. Hom./Comp. lists homogeneity/completeness; NMI and purity are mean

\pm

standard deviation over five seeds.

Backbone	Model Type	Mean AUC (%)
Audio-MAE	Masked SSL	92.60
CLAP	Contrastive	87.40
PANNS	Supervised Semantic	84.40
BYOL-A (Default)	Audio SSL	90.00

Appendix

Table 11: Performance verification across different pre-trained backbones.

Initialization Method	Assignment Method	Mean AUC (%)
Spectral + GMM	Boltzmann	89.40
Spectral + vMF Mixture	Boltzmann	90.70
Spectral + K-Means++ (Default)	Boltzmann	90.00

Appendix

Table 12: Robustness test replacing the topological initialization method.

Assignment Method	Mean AUC (%)	$\Delta$ AUC
Boltzmann (Ours)	90.00	-
Hard Nearest	87.60	-2.40

Appendix

Table 13: Ablation of the assignment mechanism (Soft vs. Hard). Metrics are in percentages (%).

Retrieval Strategy	Mean AUC (%)	Latency (CPU)	Speedup
Flat Search (Exact)	90.00	42.3 ms	1.0 $\times$
Hierarchical Index (IVF)	89.80	5.4 ms	7.8 $\times$

Appendix

Table 14: Efficiency vs. accuracy trade-off between retrieval strategies (measured on CPU).

Methods	Machine Types						Average
Methods	ToyCar	ToyConv	Fan	Pump	Slider	Valve	Average
PatchCore	62.80 $\pm$ 0.1	60.60 $\pm$ 0.2	62.20 $\pm$ 0.1	57.20 $\pm$ 0.2	67.80 $\pm$ 0.2	61.50 $\pm$ 0.1	62.10 $\pm$ 0.2
Classic Adaptation
SimpleNet	63.20 $\pm$ 0.7	62.10 $\pm$ 0.9	56.70 $\pm$ 0.6	58.00 $\pm$ 0.8	66.30 $\pm$ 0.7	62.20 $\pm$ 0.6	61.40 $\pm$ 0.7
EfficientAD	65.80 $\pm$ 0.4	65.60 $\pm$ 0.5	65.70 $\pm$ 0.3	68.50 $\pm$ 0.4	66.50 $\pm$ 0.3	65.60 $\pm$ 0.4	66.30 $\pm$ 0.4
Foundation Embeddings

Appendix

Table 15: Multi-seed stability for universal anomaly detection (AUC, %) on DCASE 2020. Metrics are reported as Mean

\pm

Std over 5 random seeds.

Methods	Machine Types						Average
Methods	ToyCar	ToyConv	Fan	Pump	Slider	Valve	Average
PatchCore	57.60 $\pm$ 0.1	56.00 $\pm$ 0.1	57.10 $\pm$ 0.2	52.30 $\pm$ 0.1	59.40 $\pm$ 0.1	54.90 $\pm$ 0.2	56.20 $\pm$ 0.1
Classic Adaptation
SimpleNet	55.00 $\pm$ 0.6	53.20 $\pm$ 0.5	49.60 $\pm$ 0.4	52.90 $\pm$ 0.7	56.80 $\pm$ 0.6	56.00 $\pm$ 0.5	53.90 $\pm$ 0.5
EfficientAD	58.20 $\pm$ 0.3	57.40 $\pm$ 0.4	57.60 $\pm$ 0.2	59.30 $\pm$ 0.3	60.80 $\pm$ 0.2	58.30 $\pm$ 0.3	58.70 $\pm$ 0.3
Foundation Embeddings

Appendix

Table 16: Multi-seed stability for universal anomaly detection (pAUC, %) on DCASE 2020. Metrics are reported as Mean

\pm

Std over 5 random seeds.

Method Variant	Machine Types						Avg. AUC
Method Variant	ToyCar	ToyConv	Fan	Pump	Slider	Valve	Avg. AUC
Raw Feat + Exact kNN	64.10	56.50	58.80	69.20	62.30	63.50	62.40
PCA + KMeans + Ret.	66.50	58.20	61.00	71.50	64.80	66.20	64.70
Raw Feat + IVF/HNSW	63.50	56.10	58.40	68.50	61.90	63.10	61.90
Proj + Random Init	89.50	76.50	78.20	82.50	79.00	83.10	81.40
Euclidean Proj ( $\mathbb{R}^{128}$ )	86.20	72.10	74.50	80.10	76.80	79.50	78.20

Appendix

Table 17: Direct control experiments validating the necessity of individual architectural choices. Metrics are in percentages (%).

Method	Routed AUC	Recall	Latency
PCA+KMeans	$64.71\pm 0.29$	$87.94\pm 0.83$	$6.81\pm 0.28$
MCDD+local kNN	$81.47\pm 0.87$	$94.18\pm 0.62$	$7.26\pm 0.33$
DeepCluster+local kNN	$84.62\pm 0.72$	$95.73\pm 0.54$	$6.32\pm 0.27$
vMF soft prototypes	$85.71\pm 0.61$	$96.21\pm 0.49$	$6.19\pm 0.26$
HARMONY hard	$87.34\pm 0.49$	$96.77\pm 0.43$	$5.57\pm 0.23$
HARMONY rep.+HNSW	$89.71\pm 0.32$	$97.26\pm 0.34$	$5.16\pm 0.22$

Appendix

Table 18: Matched controls on DCASE 2020. AUC and recall are percentages; latency is milliseconds.

Figure 6: Performance degradation under severe manifold perturbations. (Left) Non-linear distortion eroding the geometric margin. (Right) PCA-aligned anisotropic noise violating orthogonal assumptions. HARMONY exhibits graceful decay compared to the baseline.

Method	Original	$+6$ dB	0 dB	$-6$ dB	Dynamic	Unseen
Audio-PC	$83.90\pm 0.20$	$80.82\pm 0.34$	$75.47\pm 0.51$	$66.11\pm 0.76$	$71.84\pm 0.64$	$69.52\pm 0.71$
UniFlow	$75.40\pm 0.60$	$72.17\pm 0.74$	$66.32\pm 0.91$	$58.41\pm 1.18$	$63.18\pm 1.02$	$61.02\pm 1.11$
Dinomaly	$75.30\pm 0.50$	$72.64\pm 0.63$	$67.28\pm 0.79$	$60.17\pm 1.04$	$64.22\pm 0.89$	$62.34\pm 0.97$
HARMONY	$90.00\pm 0.30$	$88.73\pm 0.42$	$85.64\pm 0.62$	$79.42\pm 0.89$	$83.12\pm 0.75$	$81.37\pm 0.82$

Appendix

Table 19: AUC under recorded environmental noise (%, mean

\pm

standard deviation).

Figure 7: Hyperparameter sensitivity analysis. The framework demonstrates robust performance plateaus across broad ranges of anchor count

K

, intrinsic dimensionality

d^{\prime}

, and softmax temperature

\tau

. The default configurations are situated within optimal stable regions.

Figure 8: Global correlation between the unsupervised Silhouette Score and test Average AUC across all hyperparameter variations and extreme stress tests. The strong linear alignment (Pearson

\rho=0.973

) validates the use of the Silhouette Score as a highly reliable surrogate metric for label-free model selection.

$C$	Selected $K$	Oracle $K$	Selected / oracle AUC	$\Delta_{oracle}$	$\rho_{s}$
2	4 / 2–6	4 / 4–8	92.84 / 93.11	0.27	0.91
3	6 / 4–8	6 / 6–8	91.97 / 92.29	0.32	0.93
4	8 / 6–12	8 / 8–12	91.31 / 91.58	0.27	0.94
5	8 / 8–12	12 / 8–12	90.64 / 91.01	0.37	0.96
6	8 / 8–12	8 / 8–16	90.07 / 90.26	0.19	0.97

Appendix

Table 20: Normal-only selection for mixtures of

C

machine types (AUC, %).

Setting	$K$ (mode / range)	AUC (%)	$\Delta_{oracle}$
DCASE count transferred	8 / 8–8	$84.21\pm 0.63$	$0.63\pm 0.28$
Normal-only selection	12 / 8–16	$84.57\pm 0.58$	$0.27\pm 0.19$
Test oracle	12 / 12–16	$84.84\pm 0.51$	$0.00\pm 0.00$

Appendix

Table 21: MIMII selection and retrospective oracle comparison.

$IR$	$K/\tau$	Macro AUC	Tail AUC	Head–Tail	Entropy	$\Delta_{oracle}$
1	8 / 0.10	$90.00\pm 0.30$	$89.78\pm 0.42$	$0.34\pm 0.31$	$0.98\pm 0.01$	$0.18\pm 0.14$
3	8 / 0.10	$89.31\pm 0.44$	$88.16\pm 0.68$	$1.88\pm 0.55$	$0.96\pm 0.02$	$0.28\pm 0.20$
5	8 / 0.10	$88.54\pm 0.57$	$86.63\pm 0.87$	$3.23\pm 0.73$	$0.93\pm 0.02$	$0.37\pm 0.25$
10	16 / 0.05	$86.96\pm 0.82$	$83.41\pm 1.24$	$5.76\pm 1.08$	$0.89\pm 0.03$	$0.58\pm 0.34$
20	16 / 0.05	$84.83\pm 1.06$	$79.36\pm 1.72$	$8.97\pm 1.46$	$0.83\pm 0.04$	$0.84\pm 0.47$

Appendix

Table 22: Joint selection under long-tailed training mixtures. AUC and gaps use percentage-point units; entropy is dimensionless.

$IR$	$\tau=0.05$	$\tau=0.10$	Difference
1	$89.34\pm 0.35$	$90.00\pm 0.30$	$-0.66$
10	$86.44\pm 0.79$	$85.91\pm 0.91$	$+0.53$
20	$84.11\pm 1.04$	$82.86\pm 1.27$	$+1.25$

Appendix

Table 23: Fixed-

K

temperature comparison (AUC, %;

K=8

Method / $IR$	$K$	Rare-type anchors	Absorption	Rare-type AUC (%)
HARMONY / 1	8	$2.67\pm 0.31$	$0.05\pm 0.02$	$89.78\pm 0.42$
HARMONY / 5	8	$2.14\pm 0.26$	$0.14\pm 0.04$	$86.63\pm 0.87$
HARMONY / 10	16	$1.87\pm 0.24$	$0.18\pm 0.04$	$83.41\pm 1.24$
HARMONY / 20	16	$1.19\pm 0.22$	$0.29\pm 0.05$	$79.36\pm 1.72$
MCDD / 20	16	$0.87\pm 0.21$	$0.47\pm 0.07$	$66.21\pm 1.48$
PCA-KMeans / 20	8	$0.23\pm 0.18$	$0.64\pm 0.08$	$53.92\pm 1.57$

Appendix

Table 24: Rare-type behavior (five-seed mean

\pm

standard deviation).

Category	Parameter	Symbol	Value
Data	Sample Rate	$sr$	16,000 Hz
	FFT / Hop	-	1024 / 160
	Mel Bins	$N_{mels}$	64
Model	Input Dim	$D$	3072
	Latent Dim	$d$	128
Train	Epochs	$E$	20 (DCASE)

Appendix

Table 25: Default hyperparameter settings.

Machine Type	Machine IDs	Train (Normal)	Test (Normal)	Test (Anomaly)
ToyCar	01, 02, 03, 04	4,000	1,400	1,059
ToyConveyor	01, 02, 03	3,000	2,399	1,110
Fan	00, 02, 04, 06	3,675	400	1,475
Pump	00, 02, 04, 06	3,349	400	456
Slider	00, 02, 04, 06	2,804	400	890
Valve	00, 02, 04, 06	3,291	400	479

Appendix

Table 26: Detailed statistics of the DCASE 2020 Universal Dataset.

Machine Type	Machine IDs	Train (Normal)	Test (Normal)	Test (Anomaly)
Fan	00, 02, 04, 06	3,261	814	1,475
Pump	00, 02, 04, 06	3,000	749	456
Slider	00, 02, 04, 06	2,566	638	890
Valve	00, 02, 04, 06	2,954	737	479
Total	16 IDs	11,781	2,938	3,300

Appendix

Table 27: Detailed statistics of the MIMII -6dB Dataset.

Type	ID	AUC	pAUC	Time	Type	ID	AUC	pAUC	Time
ToyCar	01	97.78	88.89	5.39 ms	Pump	00	89.08	74.53	5.34 ms
	02	97.97	90.51	5.12 ms		02	85.50	65.39	5.11 ms
	03	97.77	88.80	5.46 ms		04	89.40	80.68	5.64 ms
	04	98.42	92.42	5.33 ms		06	90.77	79.51	5.57 ms
ToyConveyor	01	92.15	75.47	5.06 ms	Slider	00	93.52	80.53	5.65 ms
	02	88.94	71.41	5.16 ms		02	89.51	72.26	5.62 ms

Appendix

Table 28: Per-ID performance on DCASE 2020. Metrics are reported in percentages (%).

Type	ID	AUC	pAUC	Time	Type	ID	AUC	pAUC	Time
Fan	00	84.54	65.68	5.63 ms	Slider	00	90.72	78.32	5.24 ms
	02	83.96	64.73	5.92 ms		02	86.24	67.76	5.39 ms
	04	81.16	63.10	5.93 ms		04	84.23	69.59	5.38 ms
	06	70.21	61.68	5.21 ms		06	82.13	66.08	5.79 ms
Pump	00	90.23	72.59	5.48 ms	Valve	00	80.72	62.17	5.18 ms
	02	89.20	73.27	5.61 ms		02	83.95	67.34	5.10 ms

Appendix

Table 29: Per-ID performance on MIMII -6dB. Metrics are reported in percentages (%).

Method	Macro AUC	Macro pAUC	Worst-ID AUC
PatchCore	$58.74\pm 0.41$	$53.92\pm 0.36$	$49.18\pm 0.83$
SimpleNet	$57.93\pm 0.87$	$52.84\pm 0.72$	$47.62\pm 1.34$
EfficientAD	$62.18\pm 0.62$	$56.73\pm 0.51$	$51.37\pm 1.06$
Gen-Rep	$77.86\pm 0.48$	$65.82\pm 0.41$	$63.94\pm 0.92$
Audio-PC	$79.73\pm 0.44$	$67.11\pm 0.38$	$65.72\pm 0.87$
CADA	$68.41\pm 0.92$	$59.28\pm 0.76$	$54.13\pm 1.42$

Appendix

Table 30: MIMII

-6

dB baseline comparison (%).

Method	Overall	Target	Worst section
Audio-PC	$77.84\pm 0.46$	$73.72\pm 0.59$	$64.27\pm 0.99$
UniFlow	$72.69\pm 0.81$	$68.43\pm 0.98$	$58.92\pm 1.48$
Dinomaly	$73.21\pm 0.73$	$68.97\pm 0.91$	$59.41\pm 1.39$
HARMONY	$82.36\pm 0.52$	$77.99\pm 0.68$	$68.42\pm 1.12$

Appendix

Table 31: DCASE 2022 comparison (AUC, %).

Method	0%	0.5%	1%	2%	5%
Raw+1NN	$62.18\pm 0.23$	$59.84\pm 0.38$	$57.63\pm 0.51$	$54.21\pm 0.74$	$50.47\pm 1.06$
PatchCore	$61.91\pm 0.25$	$60.42\pm 0.36$	$58.73\pm 0.48$	$55.97\pm 0.69$	$52.86\pm 0.94$
Audio-PC	$83.74\pm 0.27$	$82.16\pm 0.39$	$80.34\pm 0.53$	$76.88\pm 0.77$	$69.41\pm 1.12$
MCDD	$78.09\pm 1.03$	$77.18\pm 1.08$	$75.93\pm 1.17$	$73.47\pm 1.31$	$68.26\pm 1.57$
HARMONY	$89.65\pm 0.35$	$89.04\pm 0.43$	$88.29\pm 0.57$	$86.82\pm 0.76$	$83.31\pm 1.13$

Appendix

Table 32: Training contamination on DCASE 2020. Columns are contamination rates; entries are AUC (%, mean

\pm

standard deviation).

Anomaly Detection

Anomaly detection finds rare or unexpected patterns in data, from defective products and system faults to fraud and security threats.

See all 652 papers

How Much Does Machine Identity Matter in Anomalous Sound Detection at Test Time?
Feb 18, 2026Kevin Wilkinghoff, Keisuke Imoto, Zheng-Hua TanAnomaly Detection Industrial Anomaly Detection
Aalborg University, Denmark · Pioneer Centre for Artificial Intelligence, Denmark · Kyoto University, Japan
UD-ASD: A Unified Diffusion Model for Anomalous Sound Detection
Jul 14, 2026Pengxiang Gao, Yu Qiu, Yanzhi SongAudio Diffusion Models Fault Detection
University of Science and Technology of China, China
Sub-Band Spectral Matching with Localized Score Aggregation for Robust Anomalous Sound Detection
Mar 14, 2026Phurich Saengthong, Takahiro ShinozakiAnomaly Detection Unsupervised Anomaly Detection
Department of Information and Communications Engineering, Institute of Science Tokyo, Japan

Feb 18, 2026J/K move · Enter open · S save

How Much Does Machine Identity Matter in Anomalous Sound Detection at Test Time?

Kevin Wilkinghoff, Keisuke Imoto, Zheng-Hua Tan

Aalborg University, Denmark · Pioneer Centre for Artificial Intelligence, Denmark · Kyoto University, Japan

Anomaly Detection Industrial Anomaly Detection

Open paper AskPDF arXiv

Anomalous sound detection (ASD) benchmarks typically assume that the identity of the monitored machine is known at test time and that recordings are evaluated in a machine-wise manner. However, in realistic monitoring scenarios with multiple known machines operating concurrently, test recordings may not be reliably attributable to a specific machine, and requiring machine identity imposes deployment constraints such as dedicated sensors per machine. To reveal performance degradations and method-specific differences in robustness that are hidden under standard machine-wise evaluation, we consider a minimal modification of the ASD evaluation protocol in which test recordings from multiple machines are merged and evaluated jointly without access to machine identity at inference time. Training data and evaluation metrics remain unchanged, and machine identity labels are used only for post hoc evaluation. Experiments with representative ASD methods show that relaxing this assumption reveals performance degradations and method-specific differences in robustness that are hidden under standard machine-wise evaluation, and that these degradations are strongly related to implicit machine identification accuracy.

Abstract

Anomaly Detection

Explore similar work

How Much Does Machine Identity Matter in Anomalous Sound Detection at Test Time?

UD-ASD: A Unified Diffusion Model for Anomalous Sound Detection

Sub-Band Spectral Matching with Localized Score Aggregation for Robust Anomalous Sound Detection