Recurrent Neural Networks with Long Term Temporal Dependencies in Machine Tool Wear Diagnosis and Prognosis

Binil Starly; Jianlei Zhang

arxiv: 1907.11848 · v1 · pith:SNFVY4GGnew · submitted 2019-07-27 · 📡 eess.SP

Recurrent Neural Networks with Long Term Temporal Dependencies in Machine Tool Wear Diagnosis and Prognosis

Jianlei Zhang , Binil Starly This is my paper

Pith reviewed 2026-05-24 15:07 UTC · model grok-4.3

classification 📡 eess.SP

keywords LSTMrecurrent neural networkstool wearprognosismachine condition monitoringvibration signalsremaining useful lifemilling

0 comments

The pith

LSTM recurrent networks model tool wear from vibration signals better than simple RNNs for prediction and remaining life estimation.

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

The paper applies long short-term memory networks to predict how cutting tools wear down during machining. It uses vibration data collected near the workpiece to model how the tool's condition changes over time and how that condition appears in the sensor readings. This avoids needing detailed physics equations for each type of wear. The LSTM approach captures longer sequences of data than ordinary recurrent networks, leading to more accurate forecasts of tool state one or two steps ahead and estimates of how much life remains in the tool. If the method works, factories could monitor tools without building custom mathematical models for every machine setup.

Core claim

An LSTM-based recurrent neural network can learn both the system transition function and the system observation function from vibration signals to predict cutting tool wear states and remaining useful life, outperforming a simple RNN in one-step and two-step ahead predictions on milling machine experiments.

What carries the argument

LSTM architecture that maintains long-term dependencies in sequential vibration data to model tool degradation without analytic wear models.

If this is right

One-step and two-step look ahead predictions of tool wear state become feasible from indirect vibration measurements.
Remaining useful life of cutting tool inserts can be estimated using a generative RNN approach.
The method eliminates the need for domain-specific analytic models required by HMMs, Kalman filters, and particle filters.
Performance improves over simple RNNs when long-term temporal dependencies are present in the data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Factories could integrate this into existing CNC machines with minimal sensor additions.
Similar LSTM modeling might apply to other degradation processes like bearing wear or battery health if vibration-like signals are available.
Online monitoring could reduce unplanned downtime by scheduling tool changes based on predicted RUL.
The approach might generalize to multi-sensor fusion if more signals are added.

Load-bearing premise

Vibration signals collected near the workpiece fixtures carry enough information about gradual tool wear to let the network learn accurate transition and observation functions.

What would settle it

If retraining the LSTM on new milling runs produces higher prediction error than a simple RNN or than a physics-based model on the same data, the claim that LSTM captures the dependencies better would not hold.

read the original abstract

Data-driven approaches to automated machine condition monitoring are gaining popularity due to advancements made in sensing technologies and computing algorithms. This paper proposes the use of a deep learning model, based on Long Short-Term Memory (LSTM) architecture for a recurrent neural network (RNN) which captures long term dependencies for modeling sequential data. In the context of estimating cutting tool wear amounts, this LSTM based RNN approach utilizes a system transition and system observation function based on a minimally intrusive vibration sensor signal located near the workpiece fixtures. By applying an LSTM based RNN, the method helps to avoid building an analytic model for specific tool wear machine degradation, overcoming the assumptions made by Hidden Markov Models, Kalman filter, and Particle filter based approaches. The proposed approach is tested using experiments performed on a milling machine. We have demonstrated one-step and two-step look ahead cutting tool state prediction using online indirect measurements obtained from vibration signals. Additionally, the study also estimates remaining useful life (RUL) of a machine cutting tool insert through generative RNN. The experimental results show that our approach, applying the LSTM to model system observation and transition function is able to outperform the functions modeled with a simple RNN.

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.

Desk Editor's Note private letter to a colleague

This applies LSTM to tool wear prediction from vibration signals and claims it beats a simple RNN on short forecasts and RUL, but supplies no numbers or validation details to judge the result.

read the letter

The main takeaway is that the authors use an LSTM RNN to learn both system transition and observation functions directly from minimally intrusive vibration signals, then show one-step and two-step predictions plus RUL estimates on milling machine data. They position this as better than a plain RNN and as a way to skip analytic degradation models required by HMMs or Kalman filters. That framing is practical for cases where physics-based models are difficult to build. The real-machine experiments give a concrete testbed for the idea. What the work does reasonably is demonstrate the online indirect measurement setup and the generative RNN for RUL. Avoiding domain-specific assumptions is a clear operational benefit in this narrow industrial setting. The soft spots are straightforward. No quantitative metrics, error values, dataset sizes, or validation procedures appear in the description, so the outperformance claim cannot be checked. LSTMs were already standard for sequential data by 2019, making the contribution an application rather than a new method. The risk that reported gains are mostly in-sample fitting is real when details are missing. This is for engineers working on predictive maintenance in machining who want an applied RNN example. A reader already focused on industrial prognostics might pick up the sensor placement or the transition-observation split, but the paper will not shift broader thinking on recurrent networks. I would not send it for peer review in its current form; the central empirical claim is too thinly supported to justify referee time.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes an LSTM-based RNN to model system transition and observation functions from minimally intrusive vibration signals for milling tool wear diagnosis and prognosis. It claims this approach enables accurate one-step and two-step ahead state predictions plus RUL estimation, outperforms a baseline simple RNN, and avoids the analytic modeling assumptions of HMMs, Kalman filters, and particle filters. Experiments are performed on a milling machine.

Significance. If the empirical outperformance holds under proper validation, the work would demonstrate a practical data-driven alternative for tool wear monitoring that leverages long-term temporal dependencies without requiring domain-specific physics models. The use of generative RNN for RUL is a positive aspect, but the lack of reported quantitative metrics, dataset sizes, cross-validation procedures, or error bars in the provided description limits assessment of whether the central claim is load-bearing or reproducible.

major comments (2)

[Abstract] Abstract: the claim that the LSTM approach 'is able to outperform the functions modeled with a simple RNN' for one/two-step prediction and RUL estimation is presented without any numerical results, RMSE/MAE values, statistical significance tests, or dataset details; this prevents evaluation of whether the reported advantage is meaningful or merely due to hyperparameter tuning.
[Experimental results] The experimental section (inferred from abstract) does not appear to report independent test-set performance, cross-validation strategy, or comparison against non-RNN baselines; without these the outperformance claim cannot be assessed for circularity or generalization beyond the training distribution.

minor comments (2)

Notation for the system transition and observation functions should be defined explicitly with equations rather than described only in prose.
Clarify the exact architecture (number of layers, hidden units, training procedure) and whether the same hyperparameter search was performed for both LSTM and simple RNN baselines.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed comments. We address each major comment point-by-point below and indicate revisions to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the LSTM approach 'is able to outperform the functions modeled with a simple RNN' for one/two-step prediction and RUL estimation is presented without any numerical results, RMSE/MAE values, statistical significance tests, or dataset details; this prevents evaluation of whether the reported advantage is meaningful or merely due to hyperparameter tuning.

Authors: We agree that the abstract would be strengthened by including quantitative results. In the revised manuscript we will add specific RMSE and MAE values for the one- and two-step predictions and RUL estimates, along with a brief mention of dataset size, to allow readers to assess the magnitude of the reported outperformance over the simple RNN. revision: yes
Referee: [Experimental results] The experimental section (inferred from abstract) does not appear to report independent test-set performance, cross-validation strategy, or comparison against non-RNN baselines; without these the outperformance claim cannot be assessed for circularity or generalization beyond the training distribution.

Authors: The experimental section describes the milling-machine data collection and the direct comparison against the simple RNN baseline. To improve clarity we will explicitly state that an independent test set was held out, describe the cross-validation procedure used, and note that the simple-RNN comparison serves as the primary data-driven baseline; we acknowledge that additional non-RNN machine-learning baselines are not included and will add a short discussion of this scope limitation. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical ML comparison is self-contained

full rationale

The paper presents a data-driven LSTM model trained on vibration signals to predict tool wear and RUL, with an empirical claim of outperformance versus simple RNN on milling-machine experiments. No load-bearing mathematical derivation, self-definitional equations, fitted-input predictions, or self-citation chains appear in the abstract or described claims. Performance metrics are measured on experimental data and remain falsifiable on held-out sequences, satisfying the criteria for an independent result.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that vibration signals suffice for learning long-term dependencies and on the neural network's ability to model transition and observation functions; the model itself contains a large number of fitted parameters.

free parameters (1)

LSTM network weights, biases, and hyperparameters
All network parameters are fitted to the experimental vibration data during training.

axioms (1)

domain assumption Vibration signals near workpiece fixtures contain the necessary information to capture long-term tool wear dynamics
Invoked when the paper states that indirect measurements from a minimally intrusive sensor can replace analytic degradation models.

pith-pipeline@v0.9.0 · 5732 in / 1298 out tokens · 24587 ms · 2026-05-24T15:07:55.746987+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages · 4 internal anchors

[1]

and Schmidhuber, J., 2001

Hochreiter, S., Bengio, Y., Frasconi, P. and Schmidhuber, J., 2001. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies

work page 2001
[2]

and Schmidhuber, J., 1997

Hochreiter, S. and Schmidhuber, J., 1997. Long short -term memory. Neural computation, 9(8), pp.1735-1780

work page 1997
[3]

and Krishnakumar, K., 2008

Wang, X., Wang, W., Huang, Y., Nguyen, N. and Krishnakumar, K., 2008. Design of neural network- based estimator for tool wear modeling in hard turning. Journal of intelligent manufacturing, 19(4), pp.383-396

work page 2008
[4]

and Karpat, Y., 2005

Özel, T. and Karpat, Y., 2005. Predictive modeling of surface roughness and tool wear in hard turning using regression and neural networks. International Journal of Machine Tools and Manufacture, 45(4), pp.467-479

work page 2005
[5]

and Kumara, S.R.T., 1995

Kamarthi, S.V. and Kumara, S.R.T., 1995. A new neural network architecture for continuous estimation of flank wear in turning. In Proc. First World Congr. Intelligent Manufacturing Processes and Systems (Vol. 2, pp. 1145-1156). 13

work page 1995
[6]

and Langari, R., 1997, June

Luetzig, G., Sanchez -Castillo, M. and Langari, R., 1997, June. On tool wear estimation through neural networks. In Neural Networks, 1997., International Conference on (Vol. 4, pp. 2359 -2363). IEEE

work page 1997
[7]

Generating Sequences With Recurrent Neural Networks

Graves, A., 2013. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850

work page internal anchor Pith review Pith/arXiv arXiv 2013
[8]

and Courville, A., 2016

Goodfellow, I., Bengio, Y. and Courville, A., 2016. Deep learning. MIT Press

work page 2016
[9]

Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

Chung, J., Gulcehre, C., Cho, K. and Bengio, Y., 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555

work page internal anchor Pith review Pith/arXiv arXiv 2014
[10]

Adam: A Method for Stochastic Optimization

Kingma, D. and Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

work page internal anchor Pith review Pith/arXiv arXiv 2014
[11]

and Lee, Y.S., 2017

Zhang, J., Starly, B., Cai, Y., Cohen, P.H. and Lee, Y.S., 2017. Particle learning in online tool wear diagnosis and prognosis. Journal of Manufacturing Processes

work page 2017
[12]

Multi-Sensor Prognostics using an Unsupervised Health Index based on LSTM Encoder-Decoder

Malhotra, P., TV, V., Ramakrishnan, A., Anand, G., Vig, L., Agarwal, P. and Shroff, G., 2016. Multi- sensor prognostics using an unsupervised health index based on ls tm encoder-decoder. arXiv preprint arXiv:1608.06154

work page internal anchor Pith review Pith/arXiv arXiv 2016
[13]

and Salakhutdinov, R., 2014

Srivastava, N., Hinton, G.E., Krizhevsky, A., Sutskever, I. and Salakhutdinov, R., 2014. Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning research, 15(1), pp.1929-1958. *Text Figure(s) Table(s) Click here to view linked References Figure(s) Click here to download high resolution image Figure(s) Click here to d...

work page 2014

[1] [1]

and Schmidhuber, J., 2001

Hochreiter, S., Bengio, Y., Frasconi, P. and Schmidhuber, J., 2001. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies

work page 2001

[2] [2]

and Schmidhuber, J., 1997

Hochreiter, S. and Schmidhuber, J., 1997. Long short -term memory. Neural computation, 9(8), pp.1735-1780

work page 1997

[3] [3]

and Krishnakumar, K., 2008

Wang, X., Wang, W., Huang, Y., Nguyen, N. and Krishnakumar, K., 2008. Design of neural network- based estimator for tool wear modeling in hard turning. Journal of intelligent manufacturing, 19(4), pp.383-396

work page 2008

[4] [4]

and Karpat, Y., 2005

Özel, T. and Karpat, Y., 2005. Predictive modeling of surface roughness and tool wear in hard turning using regression and neural networks. International Journal of Machine Tools and Manufacture, 45(4), pp.467-479

work page 2005

[5] [5]

and Kumara, S.R.T., 1995

Kamarthi, S.V. and Kumara, S.R.T., 1995. A new neural network architecture for continuous estimation of flank wear in turning. In Proc. First World Congr. Intelligent Manufacturing Processes and Systems (Vol. 2, pp. 1145-1156). 13

work page 1995

[6] [6]

and Langari, R., 1997, June

Luetzig, G., Sanchez -Castillo, M. and Langari, R., 1997, June. On tool wear estimation through neural networks. In Neural Networks, 1997., International Conference on (Vol. 4, pp. 2359 -2363). IEEE

work page 1997

[7] [7]

Generating Sequences With Recurrent Neural Networks

Graves, A., 2013. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850

work page internal anchor Pith review Pith/arXiv arXiv 2013

[8] [8]

and Courville, A., 2016

Goodfellow, I., Bengio, Y. and Courville, A., 2016. Deep learning. MIT Press

work page 2016

[9] [9]

Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

Chung, J., Gulcehre, C., Cho, K. and Bengio, Y., 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555

work page internal anchor Pith review Pith/arXiv arXiv 2014

[10] [10]

Adam: A Method for Stochastic Optimization

Kingma, D. and Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

work page internal anchor Pith review Pith/arXiv arXiv 2014

[11] [11]

and Lee, Y.S., 2017

Zhang, J., Starly, B., Cai, Y., Cohen, P.H. and Lee, Y.S., 2017. Particle learning in online tool wear diagnosis and prognosis. Journal of Manufacturing Processes

work page 2017

[12] [12]

Multi-Sensor Prognostics using an Unsupervised Health Index based on LSTM Encoder-Decoder

Malhotra, P., TV, V., Ramakrishnan, A., Anand, G., Vig, L., Agarwal, P. and Shroff, G., 2016. Multi- sensor prognostics using an unsupervised health index based on ls tm encoder-decoder. arXiv preprint arXiv:1608.06154

work page internal anchor Pith review Pith/arXiv arXiv 2016

[13] [13]

and Salakhutdinov, R., 2014

Srivastava, N., Hinton, G.E., Krizhevsky, A., Sutskever, I. and Salakhutdinov, R., 2014. Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning research, 15(1), pp.1929-1958. *Text Figure(s) Table(s) Click here to view linked References Figure(s) Click here to download high resolution image Figure(s) Click here to d...

work page 2014