Organizations: Department of Computer Science, University of Warwick, Coventry, UK · Department of Computer Science, University of Sheffield, Sheffield, UK · Advanced Manufacturing Research Centre, University of Sheffield, Sheffield, UK · Shanghai Research Institute for Intelligent Autonomous Systems, State Key Laboratory of Autonomous Intelligent Unmanned Systems, Department of Control Science and Engineering, Tongji University, Shanghai, China · School of Engineering, Queen Mary University of London, London, UK
We present LoRM (Language of Rotating Machinery), a self-supervised framework for multi-modal rotating-machinery signal understanding and real-time condition monitoring. LoRM is built on the idea that rotating-machinery signals can be viewed as a machine language: local signals can be tokenised into discrete symbolic units, and their future evolution can be predicted from observed multi-sensor context. Unlike conventional signal-processing methods that rely on hand-crafted transforms and features, LoRM reformulates multi-modal sensor data as a token-based sequence-prediction problem. For each data window, the observed context segment is retained in continuous form, while the future target segment of each sensing channel is quantised into a discrete token. Then, efficient knowledge transfer is achieved by partially fine-tuning a general-purpose pre-trained language model on industrial signals, avoiding the need to train a large model from scratch. Finally, condition monitoring is performed by tracking token-prediction errors as a health indicator, where increasing errors indicate degradation. In-situ tool condition monitoring (TCM) experiments demonstrate stable real-time tracking and strong cross-tool generalisation, showing that LoRM provides a practical bridge between language modelling and industrial signal analysis. The source code is publicly available at https://github.com/Q159753258/LormPHM.
Figures & tables
Figure 1: Training through token prediction: from ASR, LMs, and VLMs to LoRM. LoRM extends this paradigm to rotating machinery by predicting discrete future signal tokens from observed multi-sensor context.
Language modelling
LoRM for rotating machinery
Word / token
Discretised target token y(c)∈{1,…,K} for channel c
Vocabulary
Per-channel codebook of K clusters learned by k -means
Context
Continuous multi-sensor context segment Xctx
Token prediction
Predict p(y(c)∣Xctx) for each channel
Sequence (sentence)
Flattened context sequence from all channels (multi-sensor context)
Self-attention
Models interactions across time and channels within the unified sequence
Table I: Analogy between language-modelling concepts and LoRM.
Figure 2: Overall architecture and learning mechanism of LoRM. (a) A multi-channel signal window is divided into a context segment and a target segment, and the context segment is partitioned and flattened into a multi-sensor patch sequence. (b) The patch sequence is projected into the hidden space of a pre-trained language model with positional embeddings. (c) A partially fine-tuned pre-trained language model encodes temporal and cross-channel dependencies. (d) The contextualised patch features are mean-pooled into a global representation and transformed into channel-wise token probability distributions. (e) The target segment of each channel is independently tokenised by k -means clustering to generate ground-truth tokens. (f) A cross-entropy objective is applied between the predicted distributions and the generated tokens for self-supervised training.
Figure 3: Overview of the TCM case study. (a) Front view of the DMU 40 eVo CNC milling machine and the deployed sensing system, including a power sensor, tri-axis spindle accelerometer, a tri-axis plate accelerometer, a tri-axis force sensor and a microphone. (b) Examples of the collected multi-modal sensor signals. (c) Illustration of the dynamic milling strategy. (d) Evolution of the maximum tool wear measured from the four-flute cutters for Tool 1 to 3 using a microscope.
Metric
Notation
Formula
Accuracy
ACC
TP+TN+FP+FNTP+TN
Precision
P
TP+FPTP
Recall
R
TP+FNTP
F1-score
F1
2×P+RP×R
False positive rate
FPR
FP+TNFP
Table II: Classification Performance Metrics
Method
Backbone
Test Tool
Train on Tool1
Train on Tool2
Train on Tool3
Acc
P
R
F1
FPR
Error ↓ ( μ m)
Acc
P
R
F1
FPR
Error ↓ ( μ m)
Acc
P
R
F1
FPR
Error ↓ ( μ m)
Traditional
Statistics [ 6 ]
Tool1
0.98
1.00
0.91
0.95
0.00
1.21
0.96
1.00
0.79
0.88
0.00
5.62
0.79
0.00
0.00
0.00
0.01
185.95
Tool2
0.92
1.00
0.54
0.71
0.00
79.00
0.93
1.00
0.60
0.75
0.00
35.18
0.87
1.00
0.31
0.48
0.00
70.96
Tool3
0.88
1.00
0.53
0.70
0.00
66.23
0.98
1.00
0.94
0.97
0.00
11.26
0.99
1.00
0.95
0.97
0.00
1.15
Average
0.93
1.00
0.66
0.79
0.00
48.81
0.96
1.00
0.78
0.87
0.00
17.35
0.88
0.67
0.42
0.48
0.00
86.02
PLM -based
GPT2 (Pretrained) [ 22 ]
Tool1
0.90
0.68
1.00
0.81
0.13
23.99
0.97
1.00
0.86
0.92
0.00
1.96
0.84
1.00
0.27
0.42
0.00
7.51
Table III: Cross-Tool Prediction Performance of Different Methods
Figure 4: Field-test results of LoRM under cross-tool settings: (a) trained on T1 and tested on T1, T2, and T3; (b) trained on T2 and tested on T1, T2, and T3; and (c) trained on T3 and tested on T1, T2, and T3.
Deep learning is a new way for machinery fault diagnosis but requires extensive labeled data, a scarce resource in industrial settings. We propose Spectral Aliasing Pretext (SAP), a self-supervised learning method that pretrains models on unlabeled vibration data by exploiting spectral aliasing. We deliberately undersample signals to create folded spectrum, then train a Transformer to reconstruct the original unfolded spectrum. This pretext task forces the model to learn frequency-domain invariants characteristic of mechanical faults, without potentially destructive augmentations. Experiments on the CWRU dataset show that SAP learns stable and highly discriminative representations. In a linear probing setting, SAP quickly achieves very high classification performance with only a small fraction of labeled data and low variance. In contrast, full fine-tuning, including fully supervised training, does not lead to more stable or better results. Overall, these findings suggest that SAP combined with linear probing can be more effective and reliable than fully supervised training for fault diagnosis with limited labeled data.
Victor Gialis, Maxime Metz, David Esteve +1
LASPI, Univ. Jean Monnet, Roanne, 42300, France · Pellenc ST, Applied Research Group, Pertuis, 84120, France · LabCom Aioly, Artificial Intelligence and Optics Laboratory, Montpellier, 34196, France +1
Remaining Useful Life (RUL) prediction is essential for industrial predictive maintenance, yet many learning-based approaches rely on extensive feature engineering or large labeled datasets to train task-specific sequence models. In this work, we introduce a lightweight learning approach, in which we leverage a frozen pretrained time-series foundation model (TSFM) and combine it with a small regression head for RUL estimation from multivariate sensor streams. More specifically, we use Chronos-2 as a frozen backbone to extract context window features and train a lightweight regression neural network for RUL prediction. Experiments on real-world industrial sensor data from two device types show that Chronos-2 features consistently improve over recurrent, convolutional, Transformer-based, and gradient-boosting baselines under the same preprocessing and evaluation protocol. We further analyze the impact of context length and find that performance improves significantly with longer histories, indicating that TSFM representation offer a practical and data-efficient alternative for RUL estimation in industrial settings.
Amir El-Ghoussani, Michele De Vita, Ronald Naumann +1
Friedrich-Alexander-Universität Erlangen–Nürnberg (FAU), Germany · Nokia Solutions and Networks GmbH & Co. KG, Germany
Deep learning predictive maintenance models suffer from poor transferability across machines and operating conditions, especially when labelled data are scarce and signals span five orders of magnitude in sampling frequency (1 Hz to ~100 kHz). We propose FreqCondNorm, a Transformer-based architecture that introduces a FiLM-style frequency-conditioned normalization layer to unify heterogeneous time-series within a single model. The architecture is pretrained on five public predictive maintenance datasets (CWRU, MFPT, UOC18, PRONOSTIA, CMAPSS) using masked auto-encoding and contrastive learning with balanced domain sampling. On fault diagnosis, the model achieves 99.2% accuracy on CWRU (+6.4 pp over CNN) and 82.1% zero-shot accuracy on MFPT, demonstrating strong transfer across sampling frequencies. However, the approach does not improve remaining useful life prediction, suggesting a mismatch between pretraining and RUL objectives that warrants future investigation.
Zaynab Raounak, Camille LHermine, Zhiguo Zeng
Laboratoire Génie Industriel, CentraleSupélec, Université Paris-Saclay, France