pith. sign in

arxiv: 1907.11778 · v1 · pith:EUOYMI53new · submitted 2019-07-26 · 💻 cs.LG · stat.ML

An Encoder-Decoder Based Approach for Anomaly Detection with Application in Additive Manufacturing

Baihong Jin , Yingshui Tan , Alexander Nettekoven , Yuxin Chen , Ufuk Topcu , Yisong Yue , Alberto Sangiovanni Vincentelli This is my paper

Pith reviewed 2026-05-24 15:33 UTC · model grok-4.3

classification 💻 cs.LG stat.ML
keywords anomaly detectionencoder-decoderadditive manufacturingunsupervised learningsequential datasensor datadeep learningmanufacturing process
0
0 comments X

The pith

Encoder-decoder model detects anomalies in sequential manufacturing sensor data without supervision.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops an unsupervised deep learning method based on an encoder-decoder architecture to detect anomalies in sequential sensor data from industrial manufacturing. The approach not only identifies whether an anomaly exists at a given time step but also predicts the next state in the process. When tested on image data from a real additive manufacturing testbed that includes both normal conditions and synthetic anomalies, the model locates the injected anomalies. It additionally surfaces previously unknown temperature non-uniformity across the testbed. This matters for manufacturing environments that produce large volumes of unlabeled sensor streams where collecting examples of every possible fault is difficult.

Core claim

The encoder-decoder model is able to identify the injected anomalies in a modern manufacturing process in an unsupervised fashion. In addition, it also gives hints about the temperature non-uniformity of the testbed during manufacturing, which is what we are not aware of before doing the experiment.

What carries the argument

Encoder-decoder architecture applied to sequential sensor data for unsupervised anomaly detection and next-step prediction.

Load-bearing premise

The synthetic anomalies created for the testbed are representative of the distribution of real manufacturing faults the model would encounter in deployment.

What would settle it

Apply the trained model to a manufacturing run that contains a genuine, previously unseen fault type and check whether the anomaly is flagged at a rate comparable to the synthetic cases.

Figures

Figures reproduced from arXiv: 1907.11778 by Alberto Sangiovanni Vincentelli, Alexander Nettekoven, Baihong Jin, Ufuk Topcu, Yingshui Tan, Yisong Yue, Yuxin Chen.

Figure 1
Figure 1. Figure 1: (a) LAMPS testbed [17]. (b) An example image [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An top-down illustration of the benchmark dataset. (Left) On the top we show examples of boresight images when the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Top-down views TABLE I: Experiment results on different layers Layers Off-nominal laser power (% of max value) Absolute power deviation (% of max value) Precision Recall A1 58 13 0.93 0.95 A2 56 11 0.90 0.99 A3 54 9 0.88 0.87 A4 50 5 0.81 0.65 A5 48 3 0.75 0.61 space representations can be extracted. In addition, in order to prevent the network from over-fitting, we add a “Dropout” layer to each group. To … view at source ↗
Figure 5
Figure 5. Figure 5: (a) The line-wise reconstruction and regression anomaly scores, averaged on each scan line, and (b) the de-trended and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The ROC curves of our learning-based model (darker [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

We present a novel unsupervised deep learning approach that utilizes the encoder-decoder architecture for detecting anomalies in sequential sensor data collected during industrial manufacturing. Our approach is designed not only to detect whether there exists an anomaly at a given time step, but also to predict what will happen next in the (sequential) process. We demonstrate our approach on a dataset collected from a real-world testbed. The dataset contains images collected under both normal conditions and synthetic anomalies. We show that the encoder-decoder model is able to identify the injected anomalies in a modern manufacturing process in an unsupervised fashion. In addition, it also gives hints about the temperature non-uniformity of the testbed during manufacturing, which is what we are not aware of before doing the experiment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes an unsupervised encoder-decoder architecture for anomaly detection and next-step prediction on sequential sensor/image data from additive manufacturing. It evaluates the approach on a real testbed dataset containing images from normal conditions and injected synthetic anomalies, claiming that the model successfully identifies the anomalies in an unsupervised manner and additionally reveals previously unknown temperature non-uniformity in the testbed.

Significance. If quantitative validation were provided, the work would offer a modest contribution by demonstrating the application of a standard autoencoder to sequential manufacturing sensor data and by surfacing an incidental observation about testbed temperature variation. The unsupervised framing and use of a physical testbed are strengths, but the absence of metrics, baselines, or characterization of the synthetic anomalies limits the result's broader significance for the field.

major comments (2)
  1. [Abstract] Abstract: the central claim that the encoder-decoder 'is able to identify the injected anomalies' is presented without any quantitative metrics (e.g., detection rates, AUC, or F1), baseline comparisons, or description of how reconstruction-error thresholds were selected or cross-validated. This absence directly undermines assessment of whether the unsupervised detection succeeded.
  2. [Abstract] Abstract / experimental evaluation: the dataset is described only as containing 'synthetic anomalies' with no characterization of the injection process, no distributional comparison (spatial, temporal, or thermal) to real manufacturing faults, and no hold-out test on naturally occurring defects. Because the paper positions the result as applicable to 'a modern manufacturing process,' this gap is load-bearing for the transferability of the reported detection performance.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments. We respond to each major point below and indicate planned revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the encoder-decoder 'is able to identify the injected anomalies' is presented without any quantitative metrics (e.g., detection rates, AUC, or F1), baseline comparisons, or description of how reconstruction-error thresholds were selected or cross-validated. This absence directly undermines assessment of whether the unsupervised detection succeeded.

    Authors: We agree that quantitative metrics are absent from the abstract and main evaluation. The original work relied on qualitative inspection of reconstruction error heatmaps to show anomaly localization. Since ground-truth labels for the injected anomalies are available, we will compute and report standard detection metrics (AUC, F1, precision-recall) together with the threshold-selection procedure and at least one baseline comparator in the revised manuscript. revision: yes

  2. Referee: [Abstract] Abstract / experimental evaluation: the dataset is described only as containing 'synthetic anomalies' with no characterization of the injection process, no distributional comparison (spatial, temporal, or thermal) to real manufacturing faults, and no hold-out test on naturally occurring defects. Because the paper positions the result as applicable to 'a modern manufacturing process,' this gap is load-bearing for the transferability of the reported detection performance.

    Authors: We will expand the experimental section to describe the synthetic anomaly injection procedure in detail, including the spatial, temporal, and thermal characteristics used. However, the collected dataset contains only normal runs and synthetically injected anomalies; no naturally occurring defects were recorded. Consequently, a hold-out evaluation on real manufacturing faults cannot be performed without new data collection. revision: partial

standing simulated objections not resolved
  • The study collected no naturally occurring manufacturing defects, so a hold-out test on real faults is not feasible with existing data.

Circularity Check

0 steps flagged

Standard unsupervised autoencoder with no circular reductions in derivation

full rationale

The paper applies a conventional encoder-decoder (autoencoder) architecture to sequential sensor/image data for anomaly detection. Training occurs unsupervised on normal-condition data; anomalies are flagged via reconstruction error or next-step prediction on held-out synthetic cases. No equations define a target quantity in terms of itself, no fitted parameters are relabeled as predictions, and no load-bearing self-citations or uniqueness theorems are invoked. The central performance claim rests on empirical evaluation against injected anomalies rather than any self-referential construction, rendering the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; all modeling choices (network depth, loss function, anomaly threshold) remain implicit and therefore un-audited.

pith-pipeline@v0.9.0 · 5678 in / 980 out tokens · 17215 ms · 2026-05-24T15:33:20.187547+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 8 internal anchors

  1. [1]

    Tofail, Elias P

    Syed A.M. Tofail, Elias P. Koumoulos, Amit Bandyopadhyay, Susmita Bose, Lisa O’Donoghue, and C.A. Charitidis. Additive manufacturing: Scientific and technological challenges, market uptake and opportunities. Materials Today, 21, 07 2017

    work page 2017

  2. [2]

    Anomaly detection: A survey

    Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomaly detection: A survey. ACM computing surveys (CSUR) , 41(3):15, 2009

    work page 2009

  3. [3]

    Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation

    Kyunghyun Cho, Bart Van Merri ¨enboer, Caglar Gulcehre, Dzmitry Bah- danau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 , 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  4. [4]

    Franc ¸ois Chollet et al. Keras. https://keras.io, 2015

    work page 2015

  5. [5]

    One class support vector machines for detecting anomalous windows registry accesses

    Katherine Heller, Krysta Svore, Angelos D Keromytis, and Salvatore Stolfo. One class support vector machines for detecting anomalous windows registry accesses. 2003

    work page 2003

  6. [6]

    Additive manufacturing: Current state, future potential, gaps and needs, and recommendations

    Yong Huang, Ming Leu, Jyoti Mazumder, and M Donmez. Additive manufacturing: Current state, future potential, gaps and needs, and recommendations. Journal of Manufacturing Science and Engineering , 137:014001, 02 2015

    work page 2015

  7. [7]

    Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. ArXiv, abs/1502.03167, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  8. [8]

    A One-Class Support Vector Machine Calibration Method for Time Series Change Point Detection

    Baihong Jin, Yuxin Chen, Dan Li, Kameshwar Poolla, and Alberto Sangiovanni-Vincentelli. A one-class support vector machine calibra- tion method for time series change point detection. arXiv preprint arXiv:1902.06361, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  9. [9]

    Detecting and Diagnosing Incipient Building Faults Using Uncertainty Information from Deep Neural Networks

    Baihong Jin, Dan Li, Seshadhri Srinivasan, See-Kiong Ng, Kameshwar Poolla, et al. Detecting and diagnosing incipient building faults using uncertainty information from deep neural networks. arXiv preprint arXiv:1902.06366, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  10. [10]

    ImageNet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. In Advances in neural information processing systems , pages 1097–1105, 2012

    work page 2012

  11. [11]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436, 2015

    work page 2015

  12. [12]

    Anomaly detection via online oversampling principal component analysis

    Yuh-Jye Lee, Yi-Ren Yeh, and Yu-Chiang Frank Wang. Anomaly detection via online oversampling principal component analysis. IEEE transactions on knowledge and data engineering , 25(7):1460–1470, 2012

    work page 2012

  13. [13]

    A data-driven strategy for detection and diagnosis of building chiller faults using linear discrimi- nant analysis

    Dan Li, Guoqiang Hu, and CostasJ Spanos. A data-driven strategy for detection and diagnosis of building chiller faults using linear discrimi- nant analysis. Energy and Buildings , 128:519–529, 2016

    work page 2016

  14. [14]

    Fault detection and diagnosis for building cooling system with a tree-structured learning method

    Dan Li, Yuxun Zhou, Guoqiang Hu, and Costas J Spanos. Fault detection and diagnosis for building cooling system with a tree-structured learning method. Energy and Buildings , 127:540–551, 2016

    work page 2016

  15. [15]

    LSTM-based Encoder-Decoder for Multi-sensor Anomaly Detection

    Pankaj Malhotra, Anusha Ramakrishnan, Gaurangi Anand, Lovekesh Vig, Puneet Agarwal, and Gautam Shroff. LSTM-based encoder-decoder for multi-sensor anomaly detection. arXiv preprint arXiv:1607.00148 , 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  16. [16]

    Anomaly Detection in Video Using Predictive Convolutional Long Short-Term Memory Networks

    Jefferson Ryan Medel and Andreas Savakis. Anomaly detection in video using predictive convolutional long short-term memory networks. arXiv preprint arXiv:1612.00390, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  17. [17]

    In-situ laser control method for polymer selective laser sintering (sls)

    Tim Phillips. In-situ laser control method for polymer selective laser sintering (sls). Master’s thesis, University of Texas at Austin, 2016

    work page 2016

  18. [18]

    A Comprehensive Survey of Data Mining-based Fraud Detection Research

    Clifton Phua, Vincent Lee, Kate Smith, and Ross Gayler. A compre- hensive survey of data mining-based fraud detection research. arXiv preprint arXiv:1009.6119, 2010

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  19. [19]

    Anomaly detection using autoen- coders with nonlinear dimensionality reduction

    Mayu Sakurada and Takehisa Yairi. Anomaly detection using autoen- coders with nonlinear dimensionality reduction. In Proceedings of the MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Analysis, page 4. ACM, 2014

    work page 2014

  20. [20]

    Estimating the support of a high-dimensional distribution

    Bernhard Sch ¨olkopf, John C Platt, John Shawe-Taylor, Alex J Smola, and Robert C Williamson. Estimating the support of a high-dimensional distribution. Neural computation , 13(7):1443–1471, 2001

    work page 2001

  21. [21]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  22. [22]

    Unsuper- vised learning of video representations using LSTMs

    Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsuper- vised learning of video representations using LSTMs. In International conference on machine learning , pages 843–852, 2015

    work page 2015

  23. [23]

    Extracting and composing robust features with denoising autoencoders

    Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning , pages 1096–1103. ACM, 2008

    work page 2008