PaperlayerView exact revision on arXiv

Source-linked AI summary

Continuous Online Fault Detection for Mobile Robots via Adaptive Edge Models

Jordan Levy, Nicolas Verstaevel, Vincent Talon, Benoit Gaudou

arXiv:2609.29194v1cs.ROcs.LG

TL;DR

Mobile-robot fault detection must adapt to changing conditions, but deep models are costly to adapt continuously on edge hardware. The paper distills an offline Teacher into an adaptive Student, whose VUS-PR rises to 0.75 under domain shifts, compared with approximately 0.26 for the Teacher and 0.35 for a static Student.

  • Problem

    Mobile robots need fault detection that adapts to wear and environmental shifts, while deep models make continuous edge learning computationally costly.

    Source: §I · p. 1

  • Method

    An offline Teacher labels unlabeled robotic time-series data, and a lightweight Student approximates it for onboard adaptation with operator feedback.

    Source: §I · p. 2

  • Results

    VUS-PR reached 0.75 for uncertainty-driven adaptive MiniRocket, versus approximately 0.26 for the offline Teacher and 0.35 for the static Student.

    Source: 2) Online Learning · p. 7

  • Takeaways & Limitations

    Offline-to-online distillation deploys deep-model representational power on constrained robotic edge CPUs without deep backpropagation.

    Source: §I · p. 2

    Long-term resilience to catastrophic forgetting is constrained by the lack of continuous, multi-hour robotic benchmarks with sequential fault transitions.

    Source: §VI · p. 8

Abstract

from arXiv · show

Mobile robots require robust, real-time fault detection capable of continuous adaptation on constrained edge hardware. While deep time-series models excel at unsupervised anomaly detection, their computational cost prohibits high-frequency onboard execution. This paper bridges this gap via a Teacher-Student distillation framework. An offline foundation model (TSPulse) generates pseudo-labels from unlabeled time series augmented with fault injections. A lightweight MiniRocket Student, adapted with a Recursive Least Squares estimator, approximates this complex decision boundary to execute real-time inference onboard. Evaluations on the TSB-AD benchmark and a physical mobile robot demonstrate the Student achieves a 4.30 ms CPU inference latency. During real-world domain shifts, online adaptation enables the Student to recover from unseen mechanical degradation, improving VUS-PR scores from 0.26 to 0.75 without catastrophic forgetting. Crucially, an uncertainty-guided active learning strategy minimizes operator cognitive load, requesting sparse interventions only when encountering novel fault distributions. These results validate the deployment of state-of-the-art anomaly detection on resource-constrained robotics through offline-to-online distillation.

I. Introduction

Mobile-robot fault detection must adapt to changing conditions, but deep models are costly to adapt continuously on edge hardware. The paper proposes offline-to-online distillation, with sparse uncertainty-guided operator feedback supporting a lightweight adaptive Student.

I. Introduction

  • Deep models' computational cost makes continuous edge learning difficult, while wear and environmental shifts require robots to adapt beyond static models.Edge backpropagation is also memory-intensive and prone to catastrophic forgetting.Source: §I · pp. 1–2
  • The Teacher-Student framework distills an offline deep model into a lightweight Student that adapts onboard without deep backpropagation.The Teacher generates pseudo-labels from historical data, while sparse uncertainty-triggered operator feedback drives Student adaptation.Source: §I · p. 2
  • The paper adapts MiniRocket by replacing its static head with a Recursive Least Squares estimator for low-latency inference, uncertainty quantification, and online learning.Source: §I · p. 2
  • An uncertainty-guided active-learning loop requests sparse operator feedback and regularizes imbalanced streams to prevent forgetting during online adaptation.Source: §I · p. 2

II. Background

Data-driven fault detection avoids exhaustive physical modeling, but mobile robotics demands both low-latency execution and adaptation to wear. Deep time-series detectors offer strong detection capabilities, yet their computational requirements constrain edge deployment.

A. Data-Driven Fault Detection

  • Data-driven fault detection identifies operational deviations from learned representations of nominal behavior, but mobile robots require both low latency and adaptation to mechanical wear.Source: A. Data-Driven Fault Detection · p. 2
  • Prior onboard methods include UAV feature selection and drill-telemetry subspace analysis, while other approaches use world models [9] or Koopman operators.The described real-time architectures remain static after offline training and lack continuous learning.Source: A. Data-Driven Fault Detection · p. 2
  • The proposed framework extends anomaly-detector distillation to high-frequency multivariate time series, synthetic fault-boundary generation, and continuous online learning.Source: A. Data-Driven Fault Detection · p. 2

B. Unsupervised Time Series Anomaly Detection

C. Extrinsic Time Series Regression

MiniRocket efficiently maps temporal features to regression scores, but existing lightweight approaches lack native online adaptation. The framework separates fixed temporal feature extraction from an adaptive head and trains a Student against an unsupervised Teacher using predominantly nominal robotic data.

C. Extrinsic Time Series Regression

  • Extrinsic Time Series Regression maps an input time series to a continuous score, and MiniRocket extracts fixed-kernel temporal features for ridge regression.Source: C. Extrinsic Time Series Regression · p. 3
  • Existing lightweight streaming regressors struggle to extract temporal features, while adaptive deep models require costly edge backpropagation; the proposed model pairs fixed feature extraction with an adaptive head.Source: C. Extrinsic Time Series Regression · p. 3

III. Overall Framework

  • The framework uses an offline unsupervised Teacher to score augmented sequences, trains a lightweight supervised Student, then adapts it onboard through event-triggered operator feedback.Source: §III · p. 3

A. Problem Setup

  • The problem setup treats training data as overlapping windows from a multivariate sensor stream, predominantly nominal even if some anomalies occur.The detector maps each window to a continuous anomaly score, with high scores indicating potential faults.Source: A. Problem Setup · p. 3

B. Teacher-Student Pipeline

The Teacher assigns pseudo-labels to unlabeled windows, and the Student learns a weighted regression approximation. Synthetic injections add anomalous examples so the Student can learn boundaries beyond nominal operation.

B. Teacher-Student Pipeline

  • The offline Teacher scores unlabeled windows, and its volatile scores are smoothed and standardized before Student regression.Exponential sample weights prioritize severe anomalies while keeping nominal windows near a baseline weight of 1.Source: B. Teacher-Student Pipeline · p. 3
  • The Student minimizes weighted regression error to approximate the Teacher, while training only on normal data would collapse predictions toward normality and risk false negatives.Source: B. Teacher-Student Pipeline · p. 3, C. Data Augmentation · p. 3

C. Data Augmentation

  • Randomly positioned spikes, cutoffs, drifts, and noise are injected into sequences, then Teacher-scored and weighted for inclusion in the Student's training corpus.Source: C. Data Augmentation · pp. 3–4

D. Online Execution and Continuous Adaptation

The Student processes streaming time series onboard and continuously learns from feedback while meeting real-time constraints. The implementation uses an offline Teacher and a selected foundation model to generate pseudo-labels.

D. Online Execution and Continuous Adaptation

  • The deployed Student sequentially scores streaming time series, with inference and online learning required to meet hard latency constraints.Feedback may be discrete nominal/anomalous targets or continuous values for partial degradation or uncertainty.Source: D. Online Execution and Continuous… · p. 4

IV. System Implementation

  • The framework is model-agnostic, while this implementation selects specific Teacher and Student models for validation.The Teacher acts as an offline Oracle, free of robotic-edge latency and hardware constraints.Source: §IV · p. 4, A. The Teacher Model · p. 4

A. The Teacher Model

  • TSPulse is selected as Teacher for its zero-shot generalization, using a fixed context window of L = 512 time steps.The passage also characterizes TSPulse as a time-series foundation model.Source: A. The Teacher Model · p. 4

B. The Embedded Student

The Student uses MiniRocket features with an RLS head for lightweight, continuous adaptation. Predictive uncertainty determines when to update the model and request operator feedback.

B. The Embedded Student

  • MiniRocket applies fixed convolutional kernels to each time-series window and forms features from the proportion of positive values.The implementation replaces its static ridge head with an RLS estimator, updating decision weights and inverse covariance sequentially from pseudo-labels.Source: B. The Embedded Student · p. 4
  • RLS confines training overhead to the linear weights while keeping MiniRocket’s feature extraction frozen.This avoids backpropagation for continuous online adaptation.Source: B. The Embedded Student · p. 4
  • Predictive uncertainty combines feature-space novelty with dynamic noise variance; only when their product reaches τ does the system update the covariance matrix and request operator feedback.Otherwise, the model state is frozen; this gates the O(k^2) matrix update.Source: B. The Embedded Student · p. 4

V. Experimentation

The evaluation first tests the framework’s foundational properties, then assesses online adaptation and resilience on a mobile-robotics dataset. The setup uses TSPulse and MiniRocket, and evaluates range-based anomaly detection with VUS-PR.

V. Experimentation

  • The evaluation has two phases: foundational tests of data injection and computational speed, followed by mobile-robotics tests of online adaptation and resilience to catastrophic forgetting.The second phase deploys the complete architecture on a mobile robotics dataset.Source: §V · p. 4

A. Experimental Setup

  • The setup uses pre-trained TSPulse as the offline oracle and MiniRocket with an RLS estimator using λ = 0.99.The MiniRocket implementation uses aeon with Numba and NumPy for CPU execution.Source: A. Experimental Setup · p. 4
  • VUS-PR evaluates range-based time-series detection by penalizing late detections and prolonged false alarms while tolerating minor temporal shifts.The source code is reported as available on GitHub and random seeds are fixed.Source: A. Experimental Setup · p. 4

B. Foundational properties of the framework

1) Evaluation of the Data Augmentation:

The experiments assess augmentation strategies and edge-compute costs. Injection augmentation outperforms the alternatives, while optimized MiniRocket configurations achieve low inference and update latency on the Jetson Orin Nano.

1) Evaluation of the Data Augmentation:

  • The augmentation ablation compares zero augmentation, injection, VAE, and hybrid regimes while varying naug; it filters for datasets where TSPulse exceeds 0.30 VUS-PR.This filtering is intended to measure Student approximation of a competent Teacher.Source: 1) Evaluation of the Data Augmentation · p. 5
  • Injection augmentation consistently outperforms the baseline Student and VAE and hybrid alternatives; without augmentation, MiniRocket averages 0.27 VUS-PR below TSPulse.VAE and hybrid augmentation fail to meaningfully improve Student performance.Source: 1) Evaluation of the Data Augmentation · p. 5
  • With injection, performance increases up to naug = 6, then plateaus around a −0.2 VUS-PR deficit relative to the Teacher as MiniRocket’s feature space saturates.Beyond naug = 6, additional augmented samples are described as redundant within established linear decision boundaries.Source: 1) Evaluation of the Data Augmentation · p. 5

2) Computational Efficiency and Edge Deployability:

  • On NVIDIA Jetson Orin Nano, CPU peak OS memory for TSPulse and MiniRocket converges around ∼550 MB; Teacher GPU acceleration initializes CUDA context and inflates total process footprint to 804.87 MB.Source: 2) Computational Efficiency and Edge… · pp. 5–6
  • At L = 100, the Student reaches 4.30 ms latency, while injection gives a +0.14 average VUS-PR improvement over the L = 512 baseline.At L = 512, Student latency is 25.00 ms; the Teacher’s ARM CPU latency is 343.45 ms.Source: 2) Computational Efficiency and Edge… · p. 6
  • Reducing MiniRocket from k = 10,000 to k = 2,000 kernels lowers RLS update latency from 330.27 ms to under 15 ms.The passage reports a negligible benchmark performance drop at k = 2,000.Source: 2) Computational Efficiency and Edge… · p. 6

C. Real-World Robotic Application

The robot was tested in uncontrolled campus navigation, with physical faults introduced to evaluate adaptation. Training used uncurated sensor logs, and the physical fault datasets were replayed sequentially.

1) Dataset Generation:

  • Seven missions lasting 2 to 11 minutes provided 63 continuous channels sampled at 50 Hz; despite visual supervision and possible hidden anomalies, the unsupervised Teacher established a nominal baseline from uncurated logs.Source: 1) Dataset Generation · p. 6
  • Three physical faults—excessive power consumption, overweight, and GNSS jamming—were recorded across two missions per fault with annotated temporal ground truth.The faults were induced using electrical resistors, added load, and aluminium foil over the GNSS sensor, respectively.Source: 1) Dataset Generation · p. 6

C. Real-World Robotic Application

  • The adaptive MiniRocket was distilled from TSPulse scores using six injected-fault datasets and 2,000 kernels, then evaluated by sequential replay of physical fault datasets.The configuration followed the optimum identified in the preceding section.Source: 2) Online Learning · p. 6

2) Online Learning:

Sequential physical-fault evaluation compared adaptive MiniRocket with static and simpler online-learning baselines. The adaptive model improved detection across domain shifts, while threshold selection traded detection performance against operator interventions.

2) Online Learning:

  • The evaluation compared adaptive MiniRocket with per-time-step RLS and tested uncertainty-triggered feedback at τ = 0.01 against random feedback with a 1% sampling probability.VUS-PR was computed every 500 time steps to track novel-fault acquisition and resistance to catastrophic forgetting.Source: 2) Online Learning · p. 7
  • VUS-PR rose to 0.75 for uncertainty-driven adaptive MiniRocket, compared with approximately 0.26 for the offline Teacher and 0.35 for the static Student.Naive RLS reached 0.41 after catastrophic forgetting, while random feedback reached 0.52 and required approximately 12,000 human interventions.Source: 2) Online Learning · p. 7
  • For τ ∈[0.18, 0.27], detection remained stable with 1,224–1,833 operator interventions, while lowering τ from 0.18 to 0.16 reduced VUS-PR from 0.58 to 0.54 despite increasing interventions from 1,833 to 1,902.The reported operational lifecycle requires more operator feedback at domain-shift onset, then progressively less as uncertainty decays.Source: 2) Online Learning · p. 7

VI. Conclusion

The paper distills a static unsupervised anomaly detector into a lightweight model that can adapt online through uncertainty-driven feedback. It identifies limits in long-term evaluation, interpretability, and synthetic fault generation, and proposes corresponding future work.

VI. Conclusion

  • The framework uses offline teacher-generated pseudo-labels, fault injection, and an RLS readout to distill a static unsupervised detector into a lightweight model for online adaptation.The conclusion describes uncertainty-driven feedback as the mechanism for continuous online adaptation.Source: §VI · p. 7
  • Long-term resilience to catastrophic forgetting is constrained by the lack of continuous, multi-hour robotic benchmarks with sequential fault transitions.The authors also note opaque high-dimensional temporal features that leave operators without diagnostic guidance and simple statistical perturbations that may miss complex physical dynamics.Source: §VI · p. 8
  • Future work proposes an open lifelong-learning benchmark, explainability for temporal features driving uncertainty, and simulation-driven, physics-informed fault injection.These directions target prolonged drift testing, lower operator cognitive load, and more physically consistent anomaly signatures.Source: §VI · p. 8
Loading 2609.29194v1…