PaperlayerView exact revision on arXiv

Source-linked AI summary

Unsupervised Maneuver-Aware Acoustic Fault Detection for Autonomous Drones

Ali M Ali, Ziyi Tang, Nurdaulet Nazarbay, Hashim A. Hashim

arXiv:2610.10517v1cs.SDcs.SYeess.SY

TL;DR

Acoustic fault detection is complicated by maneuver-dependent changes that can make healthy flights resemble faulty ones, while labeled faulty-flight data are difficult and potentially unsafe to collect. The paper combines denoising with maneuver-conditioned reconstruction trained on nominal flights. The full model achieves higher AUC and F1 than the strongest traditional baseline and runs at approximately 20 ms per segment on an embedded platform.

Abstract

from arXiv · show

This paper presents a maneuver-aware acoustic fault detection framework for autonomous drones that integrates Noise2Noise-inspired deep learning denoising with maneuver-conditioned reconstruction. A key practical constraint motivating this work is that labeled faulty-flight data are difficult and potentially unsafe to collect; the proposed framework therefore follows an unsupervised learning paradigm in which only nominal flight recordings are required for training. In flight environments, acoustic signals acquired from unmanned aerial vehicles are subject to variability arising both from environmental noise and from structured, maneuver-dependent aerodynamic effects. To address these challenges simultaneously, a two-stage learning architecture is developed. In the first stage, a Noise2Noise-inspired denoising model attenuates stochastic acoustic noise while preserving fault-relevant spectral-temporal structures, without requiring clean reference signals. In the second stage, a maneuver-Conditioned Convolutional AutoEncoder (maneuver-CCAE) is trained using maneuver-related labels including drone type and flight direction to model nominal acoustic behavior under varying operating conditions. Fault detection is subsequently performed using reconstruction error as an anomaly score. Experimental results demonstrate that the proposed maneuver-aware conditioning raises the area under the ROC curve (AUC) from $\AUCaeOnly$ (unconditioned baseline) to $\AUCfull$ (full model), validating the critical role of maneuver-dependent modeling. The complete framework is deployed on an NVIDIA Jetson Orin Nano Super embedded platform within a ROS2 pipeline, achieving an end-to-end fault detection latency of approximately $20\,\text{ms}$ per audio segment with a TensorRT half-precision (FP16) backend, confirming real-time viability for onboard UAV health monitoring.

I. INTRODUCTION

The introduction frames maneuver-induced changes as a key confound in acoustic fault detection and motivates modeling them explicitly. It situates this challenge against deterministic signal-processing and supervised-learning approaches with distinct limitations.

I. INTRODUCTION

  • Healthy drones banking or yawing can sound like faulty drones flying straight, motivating a framework that models maneuver-dependent variability while remaining robust to environmental noise.The considered faults include single-motor degradation and propeller damage.Source: §I · pp. 1–2

B. Supervised Learning for Acoustic Fault Diagnosis

  • Supervised methods can perform well in controlled settings but require labeled fault data that are costly, risky, and often impractical to collect for real-world drone operations.They formulate diagnosis as classification using engineered or learned features.Source: §I · p. 2, B. Supervised Learning for Acoustic… · p. 2

A. Deterministic Signal-Processing Based Acoustic FDD

  • Deterministic signal-processing approaches are computationally efficient but often rely on stationarity or repeatability assumptions violated during aggressive or sustained maneuvers.Such approaches use time-domain statistics, spectral features, and time-frequency representations.Source: A. Deterministic Signal-Processing… · p. 2

C. Unsupervised Anomaly Detection and Reconstruction-Based Methods

The paper motivates maneuver-aware unsupervised reconstruction as a response to structured variability in drone acoustics. It formulates the task using healthy-flight data and log-mel representations, then proposes a two-stage approach.

D. Maneuver-Dependent Variability: An Open Challenge

E. Contributions

  • The proposed two-stage architecture combines Noise2Noise-inspired denoising using maneuver-consistent surrogate pairs with conditional reconstruction through Conditional Batch Normalization.The complete pipeline, including training and threshold selection, uses only healthy-flight acoustic data; the threshold is the 95th percentile of nominal validation reconstruction error.Source: E. Contributions · p. 3

III. PROBLEM FORMULATION

  • The task is to detect motor degradation or propeller damage from onboard audio, learning nominal acoustic behavior across maneuvers from healthy-flight data and scoring deviations with reconstruction error.Collecting labeled faulty-flight data can be costly and unsafe.Source: §III · p. 3

A. Acoustic Feature Representation

  • Drone recordings are represented as log-mel spectrograms; the power spectrogram is mapped through overlapping triangular mel filters and logarithmic compression.The sampled audio is represented in 2-second windows at 16 kHz, or 32,000 samples per segment.Source: A. Acoustic Feature Representation · p. 3

B. Unsupervised Acoustic Anomaly Detection Objective

The objective uses nominal data to train a reconstruction model and reconstruction error to score anomalies. The section motivates a hybrid loss and maneuver-aware modeling because nominal acoustics vary systematically with operating conditions.

B. Unsupervised Acoustic Anomaly Detection Objective

  • The model is optimized over nominal data, and inference uses reconstruction loss as the anomaly score.The unsupervised setting precludes directly optimizing a class-discriminative loss; reconstruction error serves as a surrogate for deviation from the learned nominal manifold.Source: B. Unsupervised Acoustic Anomaly… · pp. 3–4
  • A homogeneous-variability assumption is inadequate because healthy drone acoustics vary structurally with maneuvering direction and operating conditions.The objective therefore needs to account for nominal maneuver-dependent variation.Source: B. Unsupervised Acoustic Anomaly… · pp. 3–4
  • The hybrid MSE+SSIM reconstruction loss penalizes pointwise spectral differences and structural dissimilarity, preserving harmonic patterns relevant to acoustic fault characterization.The loss is motivated by training without labeled faulty data.Source: B. Unsupervised Acoustic Anomaly… · p. 4

C. Maneuver-Conditioned Acoustic Anomaly Detection

The section models nominal acoustic behavior conditional on operating context, including drone identity and flight direction. This conditioning is intended to reduce anomaly-score sensitivity to benign maneuver variability.

C. Maneuver-Conditioned Acoustic Anomaly Detection

  • The model maps an acoustic observation and its maneuver condition to a reconstruction, with conditions including drone identity and flight direction.Examples of directions include forward, backward, left, right, clockwise, and counterclockwise.Source: C. Maneuver-Conditioned Acoustic… · p. 4
  • The maneuver-conditioned training dataset and parameter learning are defined for nominal operation.At inference, the anomaly score is evaluated for an acoustic observation together with its associated maneuver condition.Source: C. Maneuver-Conditioned Acoustic… · p. 4
  • Conditioning on maneuver state reduces anomaly-score sensitivity to benign operating-condition variability, so elevated errors are primarily associated with fault-related deviations.Source: C. Maneuver-Conditioned Acoustic… · p. 4
  • Fig. 3 compares healthy and anomalous acoustic signals in the time domain alongside their moving-average envelopes.Source: Figure 3 · p. 4

D. Training Data

The training data span multiple drone types, maneuvers, and controlled fault and noise conditions. The proposed architecture first denoises without clean reference signals, then reconstructs nominal acoustics using maneuver information.

D. Training Data

  • The study adopted the dronesound dataset [38], comprising recordings from three commercial quadrotors made with a RØDE Wireless Go2 microphone mounted atop the drone bodies in an anechoic chamber to eliminate wall reflections.The platforms were the Holy Stone HS720, MJX Bugs 12 EIS, and ZLRC SG906 Pro2.Source: D. Training Data · p. 4
  • The dataset covers six motion directions, controlled propulsion defects, and background noise mixed at signal-to-noise ratios from −5 to +5 dB.Defects included partial propeller cuts and motor-cap deformation; Fig. 4 describes spectral spreading and additional sidebands in the anomalous condition.Source: D. Training Data · p. 5, Figure 4 · p. 5

IV. PROPOSED ARCHITECTURE

  • The two-stage architecture suppresses stochastic acoustic corruption without clean references, then uses a maneuver-conditioned convolutional autoencoder to model nominal behavior and detect faults through reconstruction error.Source: §IV · p. 5

A. Noise2Noise-Inspired Denoising Network

  • The denoiser uses maneuver-consistent surrogate targets because independent noise realizations of the same in-flight signal are unavailable, assuming spectral-temporal structure is approximately stable within a maneuver condition.The authors characterize this as a weak-reference Noise2Noise-inspired strategy, not perfect recovery of a single clean realization.Source: A. Noise2Noise-Inspired Denoising… · pp. 5–6
  • Maneuver information is represented by one-hot type and direction vectors, projected into embeddings that condition the denoising network through CBN.The concatenated maneuver embedding is 16-dimensional when the embedding dimension is 8.Source: A. Noise2Noise-Inspired Denoising… · pp. 5–6
  • The denoising loss combines SSIM, which encourages preservation of local structure and harmonics, with MSE, which captures pointwise spectral-energy redistribution.Source: A. Noise2Noise-Inspired Denoising… · p. 6

B. Maneuver-Conditioned Convolutional Autoencoder

The maneuver-CCAE is a shallow, parameter-efficient conditional convolutional autoencoder that reconstructs denoised spectrograms to model nominal drone acoustics across operating conditions.

B. Maneuver-Conditioned Convolutional Autoencoder

1) Conditional Input Embedding:

Maneuver context is supplied both alongside the acoustic features and through conditional normalization in the encoder and decoder.

1) Conditional Input Embedding:

2) Conditional Normalization:

3) Reconstruction and Anomaly Score:

The maneuver-conditioned autoencoder reconstructs denoised inputs under their associated maneuver conditions, and reconstruction error provides the anomaly score. The implementation summary describes the two processing stages and lists architecture and computational-complexity tables.

3) Reconstruction and Anomaly Score:

  • At inference, the denoised test segment is reconstructed under its associated maneuver condition, and the Stage 2 input-to-reconstruction error is the anomaly score.The score compares the test input presented to Stage 2 with its own reconstruction; no training quantity enters the test-time score.Source: 3) Reconstruction and Anomaly Score · p. 6

C. Implementation Summary

  • The implementation summary denoises the input spectrogram, then conditions the autoencoder on maneuver information to reconstruct it and compute an anomaly score.The denoising steps include projecting maneuver labels to an embedding and applying feature-wise conditioning; the autoencoder encodes and decodes the denoised input.Source: C. Implementation Summary · p. 6
  • The implementation summary includes tables for the denoising-network and maneuver-conditioned autoencoder architectures, plus computational complexity for the two-stage pipeline.The complexity caption specifies parameter counts, FLOPs for one 128 × 63 log-mel segment, and mean TensorRT FP16 latency over 1,000 inference calls on the Jetson Orin Nano Super.Source: Table II · p. 7, Table I · p. 7, Table III · p. 7

V. RESULTS AND DISCUSSION

The section describes the dataset partition and evaluation protocol, including training on normal segments and selecting the anomaly threshold from nominal validation errors. Reported detection metrics use the evaluation split; the anomalous-only test split is not used for those metrics.

A. Experimental Setup

  • The dataset covers three quadrotor types, six maneuver directions, and normal and faulty conditions, with recordings segmented into non-overlapping two-second windows.Fault conditions include partial propeller-blade cuts and motor-cap mechanical deformation.Source: A. Experimental Setup · p. 7
  • The predefined partition uses normal-only training data and a separate evaluation split containing normal and anomalous segments; the anomalous-only test split is not used for reported metrics.All reported detection metrics are computed on the disjoint evaluation split, while the test split admits neither AUC nor false-positive rate.Source: A. Experimental Setup · p. 7
  • The anomaly threshold is the 95th percentile of nominal validation reconstruction errors, selected without inspecting faulty segments or labels.The authors describe this as a conservative operating choice; sensitivity analysis reports stable detection performance from the 90th through 99th percentile.Source: A. Experimental Setup · p. 8

B. Detection Performance

The full denoising-plus-conditioning model achieves the strongest reported overall AUC and F1-score, while detection results show separability across fault categories, drone types, and maneuvers. The comparisons also indicate that denoising benefits multiple detectors and that operating-condition variability is important to model.

B. Detection Performance

  • FNR falls from 0.707 (CCAE without denoising) to 0.093 (full model), while FPR rises from 0.046 to 0.296 and recall increases from 0.293 to 0.907.The paper describes this operating point as more suitable for UAV monitoring, where missed detections are generally more critical than false alarms.Source: B. Detection Performance · p. 9, Table V · p. 9
  • Across six maneuver directions, AUC ranges from 0.831 to 0.914 and F1-scores from 0.782 to 0.856; clockwise is most challenging due to aerodynamic noise masking fault signatures.Type C has the highest recall (0.972) and lowest FNR (0.028).Source: B. Detection Performance · p. 9
  • Both injected fault categories are detected across all three platforms, indicating response to more than one fault signature.The categories are propeller blade cuts and motor-cap deformations; Table VI reports the row-normalized confusion matrix.Source: B. Detection Performance · p. 9, Table VI · p. 9

C. Comparison with Unsupervised Baselines

  • Denoising improves AUC across several detector paradigms, while denoised GMM-18 reaches AUC = 0.865 and surpasses every single global model in the comparison.The evaluation compares methods using the same train/evaluation partition, log-mel features, and 95th-percentile threshold.Source: C. Comparison with Unsupervised… · pp. 9–10, Table VII · p. 10
  • AUC = 0.884 and F1-score = 0.824 make the full CCAE with denoising the top evaluated method, surpassing GMM-18 (AUC = 0.865, F1 = 0.745).Table VII compares the unsupervised methods under a shared evaluation protocol.Source: Table VII · p. 10, C. Comparison with Unsupervised… · p. 10

B. Detection Performance

  • Figure 6 shows separated nominal and faulty reconstruction-error distributions, compact nominal clusters by maneuver condition, and stable training-loss convergence.The full model’s ROC curve has AUC = 0.884.Source: B. Detection Performance · pp. 8–9, Figure 6 · p. 11

D. Embedded Real-Time Implementation

The framework runs as a two-node ROS2 pipeline on a Jetson Orin Nano Super, with sequential TensorRT FP16 denoising and conditioned detection. Its measured end-to-end latency is approximately 20 ms per segment; INT8 quantization remains future work.

D. Embedded Real-Time Implementation

  • Deployment uses an NVIDIA Jetson Orin Nano Super with TensorRT FP16 inference engines executed on the GPU Tensor Cores.The prototype uses ROS2 Humble and JetPack 6.1.Source: D. Embedded Real-Time Implementation · p. 11
  • INT8 post-training quantization is not implemented in the prototype and remains future work that could further reduce latency and memory footprint.Source: D. Embedded Real-Time Implementation · p. 11
  • The ROS2 pipeline segments audio into 2-second windows, computes log-mel features, then runs denoising and CCAE sequentially using audio and maneuver inputs.The fault detector publishes an anomaly score and binary fault flag.Source: D. Embedded Real-Time Implementation · pp. 11–12
  • Approximately 20 ms end-to-end latency is well below the 2-second audio-window duration, supporting real-time operation without buffer accumulation.Latency was averaged over 1,000 TensorRT FP16 inference calls with engines pre-loaded into GPU memory.Source: D. Embedded Real-Time Implementation · p. 12

VI. CONCLUSION

  • The conclusion describes a two-stage unsupervised approach combining Noise2Noise-inspired denoising with maneuver-conditioned reconstruction learning.The conclusion states that both stages condition on drone type and flight direction.Source: §VI · p. 12
Loading 2610.10517v1…