The results support neuromorphic acoustic anomaly detection as a practical candidate for low-power, persistent machine monitoring and support autoencoder-based acoustic anomaly detection on an Intel Loihi 2 neuromorphic processor.
Abstract
Persistent acoustic monitoring can detect machine faults without physical contact, but always-on inference is constrained by power, latency, and deployment complexity. We demonstrate autoencoder-based acoustic anomaly detection on an Intel Loihi 2 neuromorphic processor under clean and noisy conditions. Log-mel features are computed off chip; normalization, autoencoder inference, L1 reconstruction scoring, and thresholding run on chip. In a clean, microphone-position-invariant ToyADMOS ToyCar benchmark, the on-chip model achieves 0.9959 AUC and 0.9785 standardized pAUC at maximum false-positive rate 0.1. In the DCASE 2026 Task 2 ToyCar noisy benchmark, the model achieves source AUC 0.7990, target AUC 0.6466, and pAUC 0.6426, exceeding reported baseline metrics. Power profiling on a 16-chip Loihi 2 VPX system shows real-time throughput with 0.0406$\unicode{x2013}$0.0426 mJ dynamic energy per sample, two orders of magnitude lower than both a CPU and GPU. These results support neuromorphic acoustic anomaly detection as a practical candidate for low-power, persistent machine monitoring.
Deploying anomaly detection models on energy constrained industrial platforms is an open and pivotal challenge for manufacturing. This paper presents ALPS (Anomaly detection with Low Power Spiking neural networks), a lightweight neuro-morphic pipeline for acoustic fault classification in a reciprocating air compressor. Raw audio waveforms are decomposed by a 16-channel band pass filter bank and converted into spike trains, which are then processed by a feed-forward spiking neural network with one hidden layer composed of 128 leaky integrate-and-fire neurons. The full pipeline is deployed on the SynSense Xylo Audio 3 neuromorphic processor after 7-bit post-training quantization. On an eight-class benchmark dataset, the system achieves 0.93 macro F1 while consuming only 3.4 mW, roughly three orders of magnitude less than a 1D convolutional neural network running on a Raspberry Pi 4 at comparable accuracy. A robustness study with structured additive factory noise shows graceful performance degradation and no abrupt collapse. These results demonstrate that neuromorphic hardware is a viable, ultra-low-power alternative for always-on acoustic anomaly detection at the edge.
F. Aisa, Umberto Albertin, Mauro Martini et al.· 2026 IEEE/ASME International...· 0 citations
Radio-frequency (RF) classifiers for drone detection at critical energy facilities reach 99%+ accuracy in laboratory conditions, but the electromagnetic interference (EMI) of substations, refineries, and nuclear plants is absent from public benchmarks. We show that exposing the classifier to a parameterized synthetic industrial-EMI distribution at training time recovers most of the lost accuracy at a small cost in clean-condition performance. We assemble a five-profile synthetic EMI model (AWGN baseline, Middleton Class-A impulsive, narrowband SCADA/PLC, partial discharge, and co-channel WiFi) with parameters drawn from the power-systems measurement literature, and benchmark eleven classical and deep learning detectors on the public DroneRF dataset across seven SNR levels and three classification tasks. All experiments use recording-level grouped data partitions—every source recording is assigned entirely to one of the training, validation, or test sets before window selection—and each augmented model shares identical hyperparameters with its non-augmented baseline, so the comparisons isolate augmentation from both data leakage and model capacity. Under this protocol, noise-augmented training lifts binary-detection accuracy at SNR = 0 dB by up to 27.5 percentage points under SCADA interference (matched 1D-CNN-V2 backbone) and 23.9 percentage points under partial discharge for the capacity-matched random-forest pipeline (Holm–Bonferroni-corrected $p \le 0.011$ ), with average gains of 6–10 percentage points over the full −10 to + 20 dB sweep; the gains are stable across five independent grouped splits. Augmentation adds no inference-time cost. A leave-one-noise-out analysis shows the gains transfer fully to unseen impulsive interference and partially to broadband noise, but not to structurally distinct narrowband SCADA, and an EMI-only operating-characteristic study shows augmentation lowers both the per-window false-positive and miss probabilities, with zero event-level errors under temporal voting on held-out recordings. Synthetic-only EMI remains a deliberate limitation, with field validation the immediate next step; we release code and noise generators to support it.
Industrial acoustic anomaly detection in long-duration streaming monitoring suffers from nonstationary noise, cross-machine domain shifts, and transient acoustic disturbances, leading to fluctuating anomaly scores and unstable alarm behavior. Existing methods mainly focus on improving backbone-level score generation, but direct thresholding under a fixed operating point may still cause alarm flickering, fragmented alarm events, delayed responses, and missed anomalies. To address these challenges, this article proposes a deployment-oriented causal backend decision framework for streaming industrial acoustic anomaly detection under fixed false-positive-rate constraints. The proposed framework treats backbone networks as anomaly-score generators and introduces an inference-stage decision layer to stabilize score-to-alarm conversion. Anomaly probabilities are transformed into the logit domain and processed by causal filters, including moving average (MA), exponential MA (EMA), exponentially weighted MA (EWMA), median filtering, adaptive Kalman filter (AKF), and recursive least square (RLS). A normal-only initialization strategy and fixed-FPR threshold calibration are further employed to support practical deployment without requiring abnormal calibration data. Experiments on the MIMII dataset, including leave-one-ID-out evaluation, cross-machine validation, logit-domain ablation, and edge-device benchmarking, demonstrate that the proposed framework improves MissRate, ToggleRate, AlarmSeg/h, and AvgLatency while introducing negligible inference overhead, validating its effectiveness for reliable Industrial Internet of Things (IIoTs) edge monitoring.
C. Gong, Cheng-Yen Lin· IEEE Internet of Things Jour...· 0 citations
Industrial anomalous sound detection is important for machine condition monitoring, especially when anomalous recordings are unavailable and only a small number of normal recordings can be collected from the target machine. This paper proposes target-domain cosine prototype anchoring (TD-CPA), a lightweight and interpretable framework for limited-sample industrial anomalous sound detection. The method combines multi-resolution Log-Mel representations with global, temporal-delta, and segment-wise statistical features to characterize both stable operating patterns and short-duration acoustic variations. PCA is then used to obtain compact embeddings, while ten target-normal recordings are employed to construct a target acoustic prototype. Anomaly scores are calculated based on the cosine deviation between each test embedding and this prototype. Experiments covering seven industrial machine categories show that TD-CPA achieves mean AUC, pAUC, and HScore values of 0.6067, 0.1389, and 0.2095, respectively, representing the highest numerical mean performance among the evaluated statistical, distance-based, compact autoencoder, and frozen pretrained-embedding baselines. Ablation experiments confirm the complementary contributions of the statistical feature components and multiple time-frequency resolutions. The target-sample sensitivity experiment further demonstrates that broader target-normal coverage generally improves detection performance. Overall, TD-CPA provides a simple, normal-data-only, and interpretable solution for industrial anomalous sound detection under limited target-data conditions.
Autoencoder-based anomalous sound detection is attractive for machine condition monitoring because it can be trained using only normal recordings and yields an interpretable anomaly score from reconstruction error. Most prior work uses spectrogram autoencoders, but reconstructing detailed time--frequency patterns is sensitive to noise and transients, and models can reconstruct some anomalous inputs well, weakening normal--anomaly separation. We propose MemNMF, a constrained reconstruction method that operates on the Linear Predictive Coding spectrum, a compact estimate of the spectral envelope. MemNMF initializes a memory module from an NMF dictionary learned on normal LPC spectra and reconstructs each input as an attention-weighted combination of prototypical normal spectral patterns. Experiments on MIMII and DCASE 2020 Task 2 across multiple machine types and operating conditions show that LPC-spectrum inputs improve a standard autoencoder baseline and that MemNMF yields further gains, with especially strong robustness under noisy, non-stationary settings.
A novel self-supervised strategy for effective single-machine training based on classifying the distance between the monitored machine and each microphone sensor of a multi-channel recording system is introduced, providing a cost-efficient and privacy-preserving alternative while delivering competitive detection performance.
Erich Malan, Valentino Peluso, A. Calimera et al.· IEEE Access· 0 citations
