Time-series foundation models put to the test on factory data

Time-series foundation models forecast sensor readings without a model trained separately for each machine. They could therefore simplify the management of plant data. Four studies published in 2026 have put them to the test. In energy consumption forecasting, foundation models achieved accuracy comparable to that of task-specific models trained on plant data. In remaining useful life estimation they outperformed task-specific models. On the signals of a milling machine and in anomaly detection, however, task-specific models were more accurate. Among the developments identified by the same studies are the joint processing of signals from different sensors and forecasts accompanied by a measure of their reliability, which could help to detect anomalies and missing measurements.

Forecasting sensor measurements without a model for each sensor

In an industrial plant, thousands of sensors record values continuously and at high frequency (Varverakis et al., 2026). Each sensor produces a time series, that is, a sequence of measurements taken at regular intervals, such as the temperature of a motor measured every second. Forecasts of these series are used to estimate energy consumption, to plan maintenance and to detect abnormal machine behaviour.

Normally, a task-specific model is trained and updated for each sensor. This is a program that estimates the sensor's future values from its past measurements. According to Varverakis et al. (2026), with thousands of sensors this approach requires excessive computing resources and becomes difficult to manage.

Time-series foundation models draw on large language models (Varverakis et al., 2026), such as those underlying conversational assistants, which are trained on large collections of text and then used to answer new questions. In the same way, a foundation model is trained once on large sets of time series from different sectors and is then applied to new forecasting problems. One example is Chronos-2, developed by Amazon Web Services with several European and US universities (Ansari et al., 2025).

In zero-shot forecasting, the model forecasts the future values of a series it has never seen during training, using only recent measurements and without any modification (Ansari et al., 2025). For a sensor, this means obtaining an estimate of future values without training a task-specific model. Alternatively, a foundation model can be adapted to a plant by fine-tuning it on the plant's data (Calisir, Leszek, 2026).

According to Varverakis et al. (2026), during training a foundation model learns general patterns, such as seasonality, trend and variability, and then identifies them in new series. For those who manage plant data, zero-shot forecasting would therefore simplify the processing of measurements, reduce the effort required to deploy and maintain models, and allow new sensors to be integrated quickly.

Benchmarks and factory data

Assessing these benefits requires knowing how accurate foundation models are on plant measurements. Their performance is usually measured on benchmarks, that is, shared collections of time series on which different systems are compared. One of these is GIFT-Eval, which includes more than 144,000 series from seven sectors (Aksu et al., 2024).

The authors of GIFT-Eval report that some models, including Chronos, the predecessor of Chronos-2, had already been trained on part of the series in the benchmark. Their results may therefore appear better than they would be on new series. Manufacturing, moreover, is not among the sectors covered by GIFT-Eval (Aksu et al., 2024).

Understanding how these models behave on plants and machines requires tests on industrial data. In 2026, four studies carried out such tests, complementing factory data with some related cases, such as building energy consumption and signals from simulated engines. Two studies, one from the Technical University of Berlin (Calisir, Leszek, 2026) and one from the University of Nicosia (Varverakis et al., 2026), have passed peer review. Peer review is the evaluation by other researchers that precedes publication in a journal or in conference proceedings. The other two, from the University of Erlangen-Nuremberg (El-Ghoussani et al., 2026) and the National Taiwan University of Science and Technology (Wen, Chen, 2026), are preprints, that is, papers released on the arXiv repository before this review.

Energy consumption and remaining useful life

The first factor in the choice between a foundation model and a task-specific model is what needs to be forecast. The Berlin study by Calisir and Leszek (2026) compares three open-source foundation models with task-specific statistical or machine learning models, that is, algorithms that derive forecasting rules from examples. The cases examined are a steelworks and a computer numerical control (CNC) milling machine.

In the steelworks, the models had to forecast energy consumption for the following day. The forecast error is the average deviation between forecast and measured values. Without fine-tuning, Chronos-2 and GTT, the latter also trained on industrial time series, achieved a lower error than the task-specific models that use only past consumption. After fine-tuning on the plant data, the error of the two foundation models decreased further (Calisir, Leszek, 2026).

Photo: DDEKK, Wikimedia Commons (CC0).

The comparison changes when the task-specific model also receives information known in advance, such as the hour of the day and the day of the week. In this case CatBoost, a task-specific machine learning model, achieved a lower error than Chronos-2 without fine-tuning (Calisir, Leszek, 2026).

Outside the factory, the Taiwan study by Wen and Chen (2026) reports a result similar to that of Chronos-2 and GTT in the steelworks, on the hourly electricity consumption of 12 building meters. The TimesFM foundation model, without fine-tuning, forecast consumption for the following hour with the lowest error, ahead of a task-specific machine learning model trained on the meter data.

The Erlangen-Nuremberg study addresses remaining useful life, that is, the number of days left before the next failure. It is based on real measurements from two devices supplied by Nokia Solutions and Networks, which employs one of the authors (El-Ghoussani et al., 2026). Each device has dozens of monitored variables.

For this estimate, Chronos-2 is used without fine-tuning. The model condenses the measurements of the last few hours of each device into a numerical summary of their behaviour. A small neural network, a type of machine learning model, is trained on the failure dates recorded in the maintenance logs and then estimates the remaining days from this summary. On the first device the average error is 44 days, half the 88 days of the best of the task-specific models compared. On the second device, where the most accurate task-specific model has an average error of 95 days, the error is 64 days. According to El-Ghoussani et al. (2026), the approach is a viable alternative to task-specific models trained from scratch.

Machining and anomalies

On the signals of the milling machine in the Berlin study, the results are reversed. The models had to forecast the current drawn by the spindle while machining wax blocks. Without further information, only GTT, after fine-tuning on the machine data, achieved a lower error than a forecast that repeats the mean value. When CatBoost also receives the machining parameters set on the CNC controller, such as programmed speeds, feed rate and clamping pressure, its error is more than a third lower than that of the fine-tuned GTT. According to the authors, the result of GTT, which was also trained on industrial series, indicates that training data related to the sector can be decisive (Calisir, Leszek, 2026).

The Taiwan study also examined anomaly detection, that is, the identification of behaviour that departs from the normal operation of a machine. The MOMENT foundation model, used without fine-tuning, was compared with small task-specific models trained only on measurements of normal operation. The tests were carried out on sound recordings of industrial pumps, intact or faulty, and on aircraft engines simulated by the US space agency (NASA). For the engines, the last 30 operating cycles before failure are treated as anomalous. In both cases the task-specific models were more accurate. The authors note that larger versions of MOMENT or fine-tuning could change the outcome, and overall they consider the advantages of foundation models in their study to be limited (Wen, Chen, 2026).

Continuity of measurements

Information available before the forecast is made also matters. Real plant measurements are often incomplete or corrupted. Sensors fail or lose power, the network drops some readings and electromagnetic interference produces anomalous spikes (Varverakis et al., 2026). In standard procedures, before forecasting, missing values can be reconstructed and corrupted ones filtered, operations that the authors of the Nicosia study consider costly. If a foundation model remained accurate even on raw measurements, this stage could be simplified.

To test this, Varverakis et al. (2026) fed three foundation models, without fine-tuning, with measurements in which values had been deliberately removed or altered, without reconstructing them. The measurements were collected on an industrial water-circulation test bench with eight sensors recording one value per second. Each forecast covers the following minute and is based on the last 10 minutes of measurements. Part of the recordings contains real process faults.

Compared with forecasts based on complete measurements, the random loss of up to 20% of the measurements and isolated spikes, that is, single anomalous values, increase the average error by at most 3%. A gap of 400 consecutive measurements corresponds to almost seven of the 10 minutes on which the forecast is based. In this case the average error across the eight sensors increases by between 7.8% and 11.1% in recordings without faults. For TimesFM the deterioration is greater in recordings with real faults and on one sensor with low-variability values. For prolonged gaps, the authors therefore consider it justified to reconstruct the missing measurements before forecasting (Varverakis et al., 2026).

The same group also evaluated prediction intervals, that is, the ranges of values within which the actual measurement should fall, in this case eight times out of ten. Two of the three models provide reliable intervals, whereas those of Chronos-2 turn out to be too narrow on stable, fault-free signals (Varverakis et al., 2026).

Computing resources

The choice between the two types of model also depends on the computing resources required. In the Taiwan study, MOMENT takes a median of 13.4 milliseconds to process a block of consecutive measurements, compared with 0.3 milliseconds for a small task-specific model. The task-specific model's file is 14 kilobytes in size, MOMENT's 138 megabytes. The authors caution that the ratio between the two times should not be regarded as a stable value (Wen, Chen, 2026).

According to Calisir and Leszek (2026), for control systems that require high-frequency forecasts, models with lower computing requirements may offer a better trade-off between accuracy and processing time.

Findings and next steps

Used without fine-tuning, foundation models achieved accuracy comparable to that of task-specific models in energy consumption forecasting. Chronos-2, combined with a small network trained on failure dates, achieved the lowest error in estimating the remaining useful life of the two devices (El-Ghoussani et al., 2026). Task-specific models were more accurate when they had access to information known in advance, such as the time of day in the steelworks and the machining parameters on the milling machine, and in anomaly detection (Calisir, Leszek, 2026; Wen, Chen, 2026). In the Taiwan study they also required less processing time and memory, and its authors consider the advantages of foundation models in their tests to be limited.

Sparse data losses had little effect on the accuracy of foundation models, whereas prolonged gaps reduced it more markedly. Two of the three models evaluated provided reliable prediction intervals (Varverakis et al., 2026).

A first research direction identified by the studies is the joint processing of signals from different sensors. El-Ghoussani et al. (2026), who already used dozens of monitored variables, make it an objective of future work. Varverakis et al. (2026) analysed each sensor separately and point to this as a limitation. Chronos-2, one of the models in their study, is already designed to handle several series and information known in advance together (Ansari et al., 2025).

The second concerns prediction intervals. El-Ghoussani et al. (2026) aim to estimate forecast uncertainty more reliably. Varverakis et al. (2026) intend to verify whether intervals can be used to detect a gap in measurements or an anomaly.

The third concerns the preparation of measurements. Varverakis et al. (2026) will study a less costly reconstruction of missing values, triggered only when the system detects a prolonged gap.

Sources

Ansari, A. F. et al. (2025): Chronos-2: From Univariate to Universal Forecasting, arXiv preprint 2510.15821, 17 October 2025

Aksu, T. et al. (2024): GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation, arXiv preprint 2410.10393, version of 11 November 2024

Calisir, C., Leszek, S. (2026): A comparison of open time-series foundation models for industrial manufacturing applications, ESANN 2026 proceedings, Bruges, April 2026

Wen, G.-H., Chen, K.-Y. (2026): Do Time-Series Foundation Models Pay Off for Industrial Monitoring? A Cost-Aware Empirical Study, arXiv preprint 2608.22968, 24 August 2026

El-Ghoussani, A., De Vita, M., Naumann, R., Belagiannis, V. (2026): Time-Series Foundation Model Embeddings for Remaining Useful Life Estimation, arXiv preprint 2606.11990, version of 11 August 2026

Varverakis, E., Michael, P., Trihinas, D. (2026): Evaluating the Robustness of Zero-Shot Time Series Models Against Data Degradation in Industrial Sensor Streams, Future Internet, 18(10), 550, 13 October 2026