Abstract
This paper describes and compares different forecasting techniques used to build a real-world Industry 4.0 application using concepts of Digital Twins. For this experiment, real data collected from a temperature sensor during the initial stages of a manufacturing process is used. This raw data from the sensors is preprocessed using state-of-the-art time series techniques for gap removal, normalization, and interpolation. The processed data are then used as input for the selected forecasting techniques for training, forecasting, and tests. Finally, the rates of the different techniques are compared using accuracy measures to determine the most accurate technique to be used in the application to support its forecasting use cases. This paper also explores different areas that can be used as topics for future work.
Introduction
Digital Twins (DTs) serve as virtual replicas of physical assets, enabling real-time monitoring, simulation, and optimization. In the context of Industry 4.0, a critical functionality of a DT is its ability to predict future states. Accurate forecasting allows for proactive maintenance and energy optimization, preventing costly downtime.
However, industrial sensor data is frequently "noisy," characterized by outliers and gaps due to sensor malfunctions or network disturbances. Furthermore, selecting the appropriate forecasting algorithm—balancing simplicity, statistical rigor, and computational complexity—remains a significant challenge for engineers. This study provides a comparative benchmark of various methods using real-world manufacturing data to determine the most effective approach for DT integration.
Methodology
The research framework follows a robust three-stage pipeline designed for industrial time-series data:
1. Preprocessing
Raw sensor data underwent rigorous cleaning to ensure model stability. Mean Absolute Deviation (MAD) was utilized for outlier detection, while linear interpolation addressed data gaps. Finally, Min-Max normalization was applied to scale the data for model compatibility.
2. Validation Strategy
To simulate real-world deployment, the study employed Walk-Forward Validation with a sliding window. Unlike standard cross-validation, this approach preserves temporal order, ensuring the model is tested on "future" data relative to its training set, mirroring actual operational conditions.
3. Forecasting Models
Three categories of models were evaluated:
- Baselines: Naive (predicting the next value equals the last) and Naive Drift (incorporating a linear trend).
- Statistical Models: Auto ARIMA, which automates the identification of optimal parameters for Autoregressive Integrated Moving Average modeling.
- Deep Learning: Long Short-Term Memory (LSTM) networks, a variant of Recurrent Neural Networks (RNNs) designed to capture long-term temporal dependencies. Hyperparameters were optimized using the Optuna framework.
Experimental Setup
The dataset comprised temperature readings from an industrial sensor during a manufacturing process:
- Timeframe: June 1, 2022, to June 29, 2022.
- Volume: 88,741 total readings (averaging ~3,060 per day).
- Data Split: 80% training (70,992 readings) and 20% testing (17,749 readings).
- Stationarity: The series was confirmed stationary via the Augmented Dickey-Fuller (ADF) test, which simplified the modeling process as statistical attributes remained constant over time.
Discussion of Findings
A pivotal discovery was the "data hunger" of deep learning models. While LSTMs are powerful, they underperformed compared to statistical methods when training sets were small. The study identified a threshold of approximately 500 sensor readings before the LSTM consistently outperformed traditional methods.
Additionally, the stationarity of the data allowed for a significant computational shortcut: models did not require re-fitting for every window movement. This finding suggests that for stable industrial processes, computational load can be significantly reduced without sacrificing real-time accuracy.
Comparative Results
Performance was measured using Root Mean Square Error (RMSE) across forecasting horizons ranging from 1 to 10 steps ahead:
- LSTM: The overall winner, outperforming Auto ARIMA by an average of 4.83% across all horizons.
- Auto ARIMA: A robust statistical performer, beating Naive baselines by an average of 21.30%.
- Naive Methods: Showed nearly identical performance, likely due to the high sampling frequency of the sensor data.
For a 1-step ahead forecast, the RMSE results were:
- LSTM: 0.016199
- Auto ARIMA: 0.018433
- Naive: 0.020107
Conclusion and Future Work
The study concludes that while LSTMs offer superior accuracy for industrial temperature forecasting within Digital Twins, their complexity and data requirements must be weighed against simpler statistical models like Auto ARIMA. Future research directions include:
- Multivariate Analysis: Incorporating multiple sensor types (e.g., pressure, vibration) and geolocations to enhance predictive power.
- Automated Training Optimization: Developing algorithms to automatically determine the optimal training window length based on data volatility.
- Multi-step Strategies: Comparing recursive versus direct multi-step forecasting strategies for longer-range predictive maintenance.
Summary updated