Are Time-Series Foundation Models Ready for E-Nose Data? An Empirical Assessment of Their Embeddings

# Are Time-Series Foundation Models Ready for E-Nose Data? An Empirical Assessment of Their Embeddings
Source: [https://arxiv.org/html/2606.27672](https://arxiv.org/html/2606.27672)
###### Abstract

Inspired by advances in natural language processing and computer vision, “time\-series foundation models” \(TSFMs\) have recently been introduced with the promise of strong generalization across diverse time\-series tasks, including forecasting, classification, and anomaly detection, as well as across domains such as healthcare, climate science, and manufacturing\. However, their utility for*gas\-sensing*data remains largely unexplored\. To address this gap, this paper systematically evaluates recent TSFMs on electronic nose \(E\-Nose\) data\. In particular, we investigate whether embeddings produced by representative TSFMs, includingChronos\-2and MOMENT, provide effective representations for gas identification and concentration prediction\. Specifically, we show that fine\-tuning is necessary to achieve satisfactory performance on E\-Nose data, and fusing TSFM embeddings with representations learned by specialized predictive models can further improve the performance, suggesting both the potential and limitations of current TSFMs for gas\-sensing applications\.

## IIntroduction

Advanced time\-series foundation models \(TSFMs\) have recently been introduced to develop general\-purpose models capable of supporting predictive tasks across diverse time\-series domains\[Goswamiet al\.,[2024](https://arxiv.org/html/2606.27672#bib.bib3), Ansariet al\.,[2025](https://arxiv.org/html/2606.27672#bib.bib2), Garzaet al\.,[2023](https://arxiv.org/html/2606.27672#bib.bib5)\]\. By pretraining on large\-scale collections of time\-series data, these models are expected to learn reusable temporal representations that can be applied to downstream tasks in zero\-shot settings or adapted through fine\-tuning with limited task\-specific data\. Recent studies have reported promising results in domains such as healthcare\[Zhanget al\.,[2025b](https://arxiv.org/html/2606.27672#bib.bib6), Liet al\.,[2026](https://arxiv.org/html/2606.27672#bib.bib7)\], climate systems\[Rangarajet al\.,[2025](https://arxiv.org/html/2606.27672#bib.bib9)\], and manufacturing\[Zhanget al\.,[2025a](https://arxiv.org/html/2606.27672#bib.bib8)\], suggesting that TSFMs, including Chronos\-2\[Ansariet al\.,[2025](https://arxiv.org/html/2606.27672#bib.bib2)\]and MOMENT\[Goswamiet al\.,[2024](https://arxiv.org/html/2606.27672#bib.bib3)\], can capture transferable temporal patterns from pretraining data and generalize to previously unseen time\-series data\.

However, the generalizability of TSFMs to Electronic Nose \(E\-Nose\) data remains largely unexplored, while E\-Nose systems have been widely adopted to detect gases and odors and estimate concentration levels in practical domains such as environment and food safety monitoring\[Rossi and Brunelli,[2014](https://arxiv.org/html/2606.27672#bib.bib11), Sberveglieriet al\.,[2014](https://arxiv.org/html/2606.27672#bib.bib10)\]\. In E\-Nose, arrays of cross\-reactive metal\-oxide \(MOX\) sensors are configured to exhibit different sensitivity patterns toward target gases, including volatile organic compounds \(VOCs\), thereby producing cross\-sensor response patterns for gas analysis\. As a result, the generated data are inherently multivariate and temporal, making them a natural candidate for TSFMs\.

![Refer to caption](https://arxiv.org/html/2606.27672v1/figures/concept_pipeline.png)Figure 1:Illustration of the designed pipeline used to examine the effectiveness of embeddingsrrextracted by each time\-series encoderϕ\\phi\(e\.g\.,Chronos\-2\)\. Classification and regression heads are trained onrrto predict the gas typeygy\_\{g\}and concentrationycy\_\{c\}, respectively\.To address this gap, we investigate whether embeddings produced by pretrained TSFMs encode informative representations of E\-Nose time\-series signals \(cf\.[Fig\. 1](https://arxiv.org/html/2606.27672#S1.F1)\)\. We select Chronos\-2\[Ansariet al\.,[2025](https://arxiv.org/html/2606.27672#bib.bib2)\]and MOMENT\[Goswamiet al\.,[2024](https://arxiv.org/html/2606.27672#bib.bib3)\]as representative TSFMs and use them as encoders, followed by task\-specific prediction heads\. The evaluation is conducted in a practical multi\-task scenario involving gas\-type identification as a classification task and concentration\-level prediction as a regression task\.

To the best of our knowledge, this is the first study to investigate TSFMs for E\-Nose sensor time\-series data\. Since publicly available descriptions of these foundation models do*not*indicate pretraining on chemical gas sensor data, this study examines their out\-of\-domain generalizability and transferability to gas\-sensing tasks\. We design experiments to address the following research questions:

1. 1\.Do pretrained TSFMs offer useful embeddings for gas\-type and concentration prediction without fine\-tuning?
2. 2\.Can TSFM\-based embeddings outperform domain\-specific models for gas sensing?
3. 3\.Does combining TSFM\-based embeddings with domain\-specific ones further improve performance?

Weighted Avg\.B1B2B3B4Acc\.F1Acc\.F1AccF1Acc\.F1Acc\.F1MLP\.704\.704\.654\.654\.896\\mathbf\{\.896\}±\.028\\pm\.028\.894\\mathbf\{\.894\}±\.028\\pm\.028\.923\\mathbf\{\.923\}±\.034\\pm\.034\.918\\mathbf\{\.918\}±\.037\\pm\.037\.517\.517±\.028\\pm\.028\.393\.393±\.037\\pm\.037\.258\.258±\.006\\pm\.006\.167\.167±\.015\\pm\.015DBFE\.667\.667\.632\.632\.507\.507±\.033\\pm\.033\.472\.472±\.056\\pm\.056\.777\.777±\.055\\pm\.055\.747\.747±\.081\\pm\.081\.735\.735±\.124\\pm\.124\.714\.714±\.148\\pm\.148\.629\.629±\.142\\pm\.142\.556\.556±\.143\\pm\.143Chronos\-2Frozen\.496\.496\.404\.404\.479\.479±\.015\\pm\.015\.321\.321±\.009\\pm\.009\.579\.579±\.093\\pm\.093\.552\.552±\.115\\pm\.115\.271\.271±\.018\\pm\.018\.139\.139±\.031\\pm\.031\.817\.817±\.075\\pm\.075\.802\.802±\.100\\pm\.100Fine\-Tuned\.845\.845\.839\.839\.881\.881±\.036\\pm\.036\.879\.879±\.038\\pm\.038\.813\.813±\.015\\pm\.015\.806\.806±\.014\\pm\.014\.890\\mathbf\{\.890\}±\.031\\pm\.031\.891\\mathbf\{\.891\}±\.029\\pm\.029\.750\.750±\.142\\pm\.142\.723\.723±\.170\\pm\.170MOMENTFrozen\.474\.474\.387\.387\.602\.602±\.016\\pm\.016\.515\.515±\.016\\pm\.016\.583\.583±\.046\\pm\.046\.515\.515±\.042\\pm\.042\.296\.296±\.018\\pm\.018\.191\.191±\.027\\pm\.027\.354\.354±\.006\\pm\.006\.271\.271±\.009\\pm\.009Fine\-Tuned\.453\.453\.361\.361\.619\.619±\.023\\pm\.023\.558\.558±\.030\\pm\.030\.379\.379±\.033\\pm\.033\.289\.289±\.042\\pm\.042\.356\.356±\.066\\pm\.066\.245\.245±\.037\\pm\.037\.463\.463±\.031\\pm\.031\.341\.341±\.012\\pm\.012MLP \+ Chronos\-2\.876\\mathbf\{\.876\}\.876\\mathbf\{\.876\}\.888\.888±\.005\\pm\.005\.887\.887±\.004\\pm\.004\.854\.854±\.029\\pm\.029\.845\.845±\.032\\pm\.032\.888\.888±\.084\\pm\.084\.884\.884±\.090\\pm\.090\.875\\mathbf\{\.875\}±\.037\\pm\.037\.871\\mathbf\{\.871\}±\.004\\pm\.004TABLE I:Gas\-type identification performance for each held\-out sensing unit \(B1–B4\)\. Overall results are reported as sample\-weighted averages across sensing units, and smaller values below each unit\-wise result denote standard deviations across three random splits\.
## IIStudy Design

### II\-AProblem Formulation

We formulate a practical multi\-task learning problem in which a model simultaneously performs gas\-type identification and concentration prediction\. Formally, each sample is represented as a multivariate time\-series matrix withτ\\tautime steps:𝐗=\[𝐱1,𝐱2,…,𝐱τ\]⊤∈ℝτ×m\\mathbf\{X\}=\[\\mathbf\{x\}\_\{1\},\\mathbf\{x\}\_\{2\},\\ldots,\\mathbf\{x\}\_\{\\tau\}\]^\{\\top\}\\in\\mathbb\{R\}^\{\\tau\\times m\}, where𝐱t=\[xt\(1\),xt\(2\),…,xt\(m\)\]∈ℝm\\mathbf\{x\}\_\{t\}=\[x^\{\(1\)\}\_\{t\},x^\{\(2\)\}\_\{t\},\\ldots,x^\{\(m\)\}\_\{t\}\]\\in\\mathbb\{R\}^\{m\}denotes the multichannel response vector captured by themmMOX sensors at time steptt\. As in[Fig\. 1](https://arxiv.org/html/2606.27672#S1.F1), an encoderϕ\\phimaps each sample𝐗\\mathbf\{X\}to a compact vector embeddingr∈ℝdr\\in\\mathbb\{R\}^\{d\}, which is then processed by a classifierℱ\\mathcal\{F\}for identification of the gas typeyg∈\{0,,…,n−1\}y\_\{g\}\\in\\\{0,,\.\.\.,n\-1\\\}and a regressor𝒢\\mathcal\{G\}for prediction of the concentrationyc∈ℝy\_\{c\}\\in\\mathbb\{R\}, wherennis the number of available gas types\. The regressor additionally uses the output ofℱ\\mathcal\{F\}to condition its concentration prediction on the predicted gas type\.

The performance of both classification and regression tasks is evaluated to determine whether each encoder captures informative patterns in the learned representation, which serves as the main input to the task\-specific prediction heads\.

### II\-BTSFM\-Based Encoding

In this work, we assess the capability of TSFMs as the encoderϕ\\phi\. Specifically, we employChronos\-2\[Ansariet al\.,[2025](https://arxiv.org/html/2606.27672#bib.bib2)\]and MOMENT\[Goswamiet al\.,[2024](https://arxiv.org/html/2606.27672#bib.bib3)\]to investigate whether their embedding spaces, learned from large and diverse time\-series datasets, can produce informative representationsrrfor gas\-sensing tasks\.

Chronos\-2 was pretrained for forecasting on synthetic time\-series datasets, using both temporal attention, which captures dependencies across time within each channel, and group attention, which enables information sharing across different channels\. In contrast, MOMENT was pretrained on Time\-Series Pile\[Goswamiet al\.,[2024](https://arxiv.org/html/2606.27672#bib.bib3)\], a large collection of public time\-series datasets, for reconstruction of randomly masked patch sequences from the remaining context\.

Both Chronos\-2 and MOMENT first divide the input sequence into non\-overlapping temporal patches, each of which is finally mapped to a representationrrthrough a series of network, including Transformer\-based encoders, and pooling\[Ansariet al\.,[2025](https://arxiv.org/html/2606.27672#bib.bib2), Goswamiet al\.,[2024](https://arxiv.org/html/2606.27672#bib.bib3)\]\. To obtain a compact representation for downstream tasks \(i\.e\., gas\-type and concentration prediction\), we apply global average pooling across both the sensor \(variable\) and temporal patch dimensions, resulting in a fixed\-dimensional representationr∈ℝdr\\in\\mathbb\{R\}^\{d\}for each input sample\.

Weighted Avg\.B1B2B3B4RMSEMAPERMSEMAPERMSEMAPERMSEMAPERMSEMAPEMLP71\.1971\.1976\.6076\.6044\.8644\.86±4\.30\\pm 4\.3031\.85\\mathbf\{31\.85\}±\.53\\pm\.5384\.6084\.60±\.92\\pm\.9263\.98\\mathbf\{63\.98\}±\.54\\pm\.5480\.2680\.26±\.8\.67\\pm\.8\.67137\.68137\.68±19\.09\\pm 19\.0978\.8878\.88±4\.48\\pm 4\.4869\.13\\mathbf\{69\.13\}±9\.09\\pm 9\.09Chronos\-2 Fine\-Tuned61\.9561\.9570\.8170\.8161\.0861\.08±10\.37\\pm 10\.3761\.6561\.65±6\.43\\pm 6\.4374\.9074\.90±2\.79\\pm 2\.7976\.6576\.65±6\.06\\pm 6\.0648\.6748\.67±2\.35\\pm 2\.3559\.2159\.21±12\.63\\pm 12\.6364\.35\\mathbf\{64\.35\}±8\.92\\pm 8\.92100\.65100\.65±6\.20\\pm 6\.20MLP \+ Chronos\-248\.52\\mathbf\{48\.52\}57\.21\\mathbf\{57\.21\}43\.57\\mathbf\{43\.57\}±2\.01\\pm 2\.0132\.2032\.20±4\.00\\pm 4\.0054\.93\\mathbf\{54\.93\}±3\.48\\pm 3\.4873\.1173\.11±10\.57\\pm 10\.5735\.83\\mathbf\{35\.83\}±6\.38\\pm 6\.3848\.06\\mathbf\{48\.06\}±12\.67\\pm 12\.6770\.9570\.95±5\.03\\pm 5\.0393\.7293\.72±5\.72\\pm 5\.72TABLE II:Concentration prediction performance for each held\-out sensing unit \(B1–B4\)\. The reporting format follows[Table I](https://arxiv.org/html/2606.27672#S1.T1)\.

## IIIExperiments

### III\-AGas\-Sensing Dataset

We use the Twin Gas Sensor Arrays dataset\[Fonollosaet al\.,[2016](https://arxiv.org/html/2606.27672#bib.bib4)\], which has been widely adopted for evaluating machine learning models under cross\-device generalization settings\[Yanget al\.,[2026](https://arxiv.org/html/2606.27672#bib.bib1)\]\. The dataset contains640640recordings collected from five individual E\-Nose units \(B1–B5\), each equipped with an array of eight MOX sensors \(m=8m=8\), including TGS2611, TGS2612, TGS2610, and TGS2602 sensors\. Each unit was exposed to four gases \(n=4n=4\) at ten concentration levels:12\.512\.5–125\.0125\.0ppm for ethanol and ethylene, and25\.025\.0–250\.0250\.0ppm for methane and carbon monoxide\. For each session, responses from the eight MOX sensors were recorded for600600s at100100Hz\.

Following the protocol in\[Yanget al\.,[2026](https://arxiv.org/html/2606.27672#bib.bib1)\], we use only the first300300s of each recording, as the second half corresponds to the purge stage of the sensing chamber\. The selected window is then downsampled to300300time points for model input \(τ=300\\tau=300\)\. Although the five sensing units were designed to be identical, hardware variations and differences in data collection days introduced device\-dependent drifts in the collected signals\. While this study does not focus on drift adaptation, such variations offer a useful testbed for evaluating the generalizability of learned embeddings across different sensor conditions\.

### III\-BImplementation Details & Baselines

We evaluate different models as the encodersϕ\\phi, while using the same prediction heads for the classifierℱ\\mathcal\{F\}and regressor𝒢\\mathcal\{G\}\. The classifier comprises a fully connected layer with512512nodes, followed by ReLU activation, and a four\-node output layer for gas\-type classification\. The regressor also consists of a fully connected layer followed by a one\-dimensional output layer with sigmoid activation\. Since the concentration ranges differ across gas types, we apply min–max normalization to the concentration values for each gas type so that the normalized targets range from0to11\. The predicted concentration values are then re\-scaled back to their original ppm ranges for evaluation\.

To evaluate the generalizability of the extracted representations across sensing units, we construct four cross\-device test settings by holding out one sensing unit, B1–B4, at a time as the test set\. The data from the remaining units are used for training and validation with a9:19\{:\}1split\. Following\[Yanget al\.,[2026](https://arxiv.org/html/2606.27672#bib.bib1)\], all methane recordings from B5 are discarded due to reported data corruption, and B5 is therefore not used as a held\-out test unit\. Moreover, all reported results are averaged over three random training–validation splits for each held\-out test unit\.

For comparative evaluation, Chronos\-2\[Ansariet al\.,[2025](https://arxiv.org/html/2606.27672#bib.bib2)\]and MOMENT\[Goswamiet al\.,[2024](https://arxiv.org/html/2606.27672#bib.bib3)\]are compared with the following baseline encoders:

- •MLP: A multilayer perceptron consisting of two connected layers of1,0241,024and512512hidden nodes with ReLU activation\. Each sample is flattened into an input vector\.
- •DBFE\[Yanget al\.,[2026](https://arxiv.org/html/2606.27672#bib.bib1)\]: State\-of\-the\-art convolution\-based model designed to capture temporal dynamics and cross\-sensor correlations in “gas” sensor data\. Since no official implementation is available, we implemented it based on the manuscript\. Meta learning\-specific losses introduced in\[Yanget al\.,[2026](https://arxiv.org/html/2606.27672#bib.bib1)\]are excluded to focus only on representational capability of the learned embeddings\.
- •MLP\+Chronos\-2: Representations from MLP and Chronos\-2 under fine\-tuning are concatenated\.

MLP and DBFE are selected to represent domain\-specific encoders trained directly on E\-Nose data\. For TSFM\-based encoders, we use Chronos\-2 Small\[Ansariet al\.,[2025](https://arxiv.org/html/2606.27672#bib.bib2)\]and MOMENT Small\[Goswamiet al\.,[2024](https://arxiv.org/html/2606.27672#bib.bib3)\]to obtain512512\-dimensional embeddings \(d=512d=512\)\. Each TSFM is tested under frozen and fine\-tuned settings\. All models are trained for up to11K epochs with a learning rate of10−410^\{\-4\}by minimizing both the cross\-entropy loss and mean squared error \(MSE\)\. Early stopping is applied if the validation MSE does not improve for5050consecutive epochs\. The checkpoint with the lowest validation MSE is used for final evaluation\. Our code and data splits will be released upon acceptance\.

### III\-CResults

#### III\-C1Gas\-type Classification

[Table I](https://arxiv.org/html/2606.27672#S1.T1)reports the accuracy and F1 scores for gas\-type classification\. Overall, performance varies substantially across held\-out sensing units, indicating distributional differences among the units\. For example, MLP achieves the best performance on B1 and B2, with accuracy scores above\.896\.896, but its accuracy drops to\.517\.517and\.258\.258on B3 and B4, respectively\. Similarly, DBFE, a state\-of\-the\-art model for E\-Nose data, does not consistently generalize to unseen sensing units when used only as an encoder without the meta\-learning strategy for which it was originally designed\.

More importantly, frozenChronos\-2and MOMENT also fail to offer consistently strong performance across the test sets\. Yet,Chronos\-2shows substantial improvement after*fine\-tuning*, achieving sample\-weighted F1 scores that are107%107\\%and28%28\\%higher than frozenChronos\-2and MLP, respectively\. This implies thatChronos\-2representations become useful for gas\-sensing tasks only after task\-specific learning\.

Furthermore,MLP\+Chronos\-2outperforms fine\-tunedChronos\-2on all test units except B3, suggesting that TSFM\-based representations can provide*complementary*information when fused with embeddings learned by task\-specific models\.

#### III\-C2Concentration Prediction

[Table II](https://arxiv.org/html/2606.27672#S2.T2)reports the root mean squared error \(RMSE\) and mean absolute percentage error \(MAPE\) for concentration prediction\. Since the regressor conditions its prediction on the gas\-type classification output, models that did not achieve promising classification performance in[Table I](https://arxiv.org/html/2606.27672#S1.T1)are excluded from this evaluation\.

A pattern similar to the classification results is observed\.MLP\+Chronos\-2achieves the best overall performance, whereas MLP performs the worst among the evaluated models\. For example, fine\-tunedChronos\-2reduces the sample\-weighted RMSE by13%13\\%compared with MLP, andMLP\+Chronos\-2further improves the result by19%19\\%\.

These results suggest that fine\-tunedChronos\-2provides useful embeddings for concentration prediction\. Its strong classification performance also contributes to regression accuracy, since the predicted gas class determines both the input condition to the regressor and the inverse scaling from the sigmoid\-normalized output to the original ppm range\.

## IVConclusion & Future Work

This paper examined TSFMs as feature encoders for gas\-sensing time\-series data from E\-Nose systems\. The results show that pretrained TSFM representations are limited effectiveness, while fine\-tuning substantially improves performance\. In addition, combining TSFM\-based representations with domain\-specific encoders further improves gas\-type identification and concentration prediction\. Future work will explore larger TSFMs and more diverse gas\-sensing datasets

## Acknowledgment

The authors thank Subeen Choi for conducting preliminary E\-Nose research and assisting with manuscript editing\. This work was partially supported by the NSF \(2502025\) and by the Georgia Peanut Commission \(KSU\-1\-26/26\)\.

## References

- A\. F\. Ansari, O\. Shchur, J\. Küken, A\. Auer, B\. Han, P\. Mercado, S\. S\. Rangapuram, H\. Shen, L\. Stella, X\. Zhang,et al\.\(2025\)Chronos\-2: from univariate to universal forecasting\.arXiv preprint arXiv:2510\.15821\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p1.1),[§I](https://arxiv.org/html/2606.27672#S1.p3.1),[§II\-B](https://arxiv.org/html/2606.27672#S2.SS2.p1.2),[§II\-B](https://arxiv.org/html/2606.27672#S2.SS2.p3.2),[§III\-B](https://arxiv.org/html/2606.27672#S3.SS2.p3.1),[§III\-B](https://arxiv.org/html/2606.27672#S3.SS2.p4.5)\.
- J\. Fonollosa, L\. Fernandez, A\. Gutiérrez\-Gálvez, R\. Huerta, and S\. Marco \(2016\)Calibration transfer and drift counteraction in chemical sensor arrays using direct standardization\.Sensors and Actuators B: Chemical236,pp\. 1044–1053\.Cited by:[§III\-A](https://arxiv.org/html/2606.27672#S3.SS1.p1.9)\.
- A\. Garza, C\. Challu, and M\. Mergenthaler\-Canseco \(2023\)TimeGPT\-1\.arXiv preprint arXiv:2310\.03589\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p1.1)\.
- M\. Goswami, K\. Szafer, A\. Choudhry, Y\. Cai, S\. Li, and A\. Dubrawski \(2024\)Moment: a family of open time\-series foundation models\.arXiv preprint arXiv:2402\.03885\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p1.1),[§I](https://arxiv.org/html/2606.27672#S1.p3.1),[§II\-B](https://arxiv.org/html/2606.27672#S2.SS2.p1.2),[§II\-B](https://arxiv.org/html/2606.27672#S2.SS2.p2.1),[§II\-B](https://arxiv.org/html/2606.27672#S2.SS2.p3.2),[§III\-B](https://arxiv.org/html/2606.27672#S3.SS2.p3.1),[§III\-B](https://arxiv.org/html/2606.27672#S3.SS2.p4.5)\.
- H\. Li, B\. Deng, C\. Xu, Z\. Feng, V\. Schlegel, Y\. Huang, Y\. Sun, J\. Sun, K\. Yang, Y\. Yu,et al\.\(2026\)Mira: medical time series foundation model for real\-world health data\.Advances in Neural Information Processing Systems38,pp\. 98657–98685\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p1.1)\.
- R\. Rangaraj, J\. Shi, A\. Shirali, R\. Paudel, Y\. Wu, and G\. Narasimhan \(2025\)How effective are large time series models in hydrology? a study on water level forecasting in everglades\.arXiv preprint arXiv:2505\.01415\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p1.1)\.
- M\. Rossi and D\. Brunelli \(2014\)Ultra low power mox sensor reading for natural gas wireless monitoring\.IEEE Sensors Journal14\(10\),pp\. 3433–3441\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p2.1)\.
- G\. Sberveglieri, G\. Zambotti, M\. Falasconi, E\. Gobbi, and V\. Sberveglieri \(2014\)MOX\-nw electronic nose for detection of food microbial contamination\.InSENSORS, 2014 IEEE,pp\. 1376–1379\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p2.1)\.
- Q\. Yang, J\. Liu, Y\. Shi, Y\. Wang, and H\. Men \(2026\)MDFE\-net: a meta\-learning driven dual\-branch feature extraction network for e\-nose sensor drift adaptation\.ACS sensors11\(5\),pp\. 4057–4067\.Cited by:[2nd item](https://arxiv.org/html/2606.27672#S3.I1.i2.p1.1),[§III\-A](https://arxiv.org/html/2606.27672#S3.SS1.p1.9),[§III\-A](https://arxiv.org/html/2606.27672#S3.SS1.p2.3),[§III\-B](https://arxiv.org/html/2606.27672#S3.SS2.p2.1)\.
- J\. Zhang, F\. Farbiz, M\. Jafary\-Zadeh, and S\. L\. Sing \(2025a\)From detection to forecasting: utilizing time\-series foundation models to anticipate defects in metal additive manufacturing\.Journal of Manufacturing Processes150,pp\. 1040–1052\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p1.1)\.
- X\. Zhang, S\. Sridharan, N\. H\. B\. Zahrin, N\. Venkataraman, and G\. S\. Hiong \(2025b\)Benchmarking a time\-series foundation model \(timegpt\) for real\-world forecasting applications\.Machine Learning with Applications,pp\. 100801\.Cited by:[§I](https://arxiv.org/html/2606.27672#S1.p1.1)\.