Abstract
In the context of Industry 4.0, proactive maintenance is essential to minimize downtime and ensure product quality. This paper presents a modular "forecaster–analyzer" architecture designed to predict industrial alarms before they occur. By decoupling time-series forecasting (using LSTM networks) from alarm classification (using ResNet), the system provides operators with a critical action-margin. The approach is validated on real-world data from an industrial extruder, demonstrating high accuracy even for complex alarms with unknown triggering rules.
Introduction
Modern manufacturing environments are increasingly sensor-rich, yet traditional alarm systems remain largely reactive, triggering only after a threshold is breached. This delay often results in "bad quality" production or costly machine failure. To address this, predictive systems must navigate the complexities of multivariate sensor data and the non-stationary nature of industrial processes, where "concept drift" can degrade model performance over time.
This study introduces a flexible, two-stage framework capable of predicting various alarm types across multiple time horizons (5, 10, and 15 minutes), allowing for proactive intervention.
The Forecaster–Analyzer Methodology
The core of the proposed system is the separation of sensor value prediction from alarm detection.
The Forecaster
A Long Short-Term Memory (LSTM) Recurrent Neural Network serves as the forecaster. LSTMs are uniquely suited for capturing long-term dependencies in multivariate time-series data. In this implementation, the model processes data from 11 distinct sensors simultaneously to project future sensor states.
The Analyzer
The predicted sensor data is fed into a Residual Neural Network (ResNet) classifier. Unlike simple rule-based systems, ResNets excel at identifying the intricate patterns that precede an alarm. This is particularly valuable for "unknown rule" alarms, where the exact physical cause of failure is not explicitly defined by human experts.
Robustness and Adaptation
To handle real-world uncertainties, the system incorporates OpenMax for Open Set Recognition, enabling it to identify novel sensor patterns it hasn't encountered during training. Furthermore, the system monitors its own prediction error; a significant spike in this "anomaly score" indicates concept drift, signaling that the model requires retraining to adapt to new environmental conditions.
Experimental Setup and Dataset
The system was tested using data from a Capital Equipment Manufacturer (CEM) specifically involving an industrial extruder used in plastic bottle production.
The dataset includes readings from 11 sensors monitoring parameters such as melting temperature, zone temperatures, die temperatures, pressures, and motor consumption. Data was collected at a 1Hz frequency over a four-month period. Preprocessing involved Piecewise Aggregate Approximation (PAA) for dimensionality reduction and Discrete Fourier Transform (DFT) for noise filtering.
Results and Discussion
The LSTM forecaster demonstrated superior performance over statistical baselines like ARIMA, maintaining a low Root Mean Squared Error (RMSE) of 0.01737 at a 15-minute horizon.
The ResNet analyzers achieved exceptional alarm detection accuracy:
- Melting Temperature Alarm: 0.99937 AUC-ROC.
- Incorrect Temperature Alarm: 0.99270 AUC-ROC.
- Die Zone 2 Alarm (Complex/Unknown Rule): 0.97381 AUC-ROC.
The high performance on the "Die Zone 2" alarm is particularly noteworthy, as the underlying failure pattern was unknown to experts, yet successfully learned by the deep learning model.
Conclusion
The forecaster–analyzer approach offers a modular and robust solution for proactive industrial monitoring. By providing a reliable "action-margin," the system empowers operators to prevent failures before they occur. Future work will focus on automating the learning of novel alarm categories and optimizing computational efficiency for real-time stream analytics.
Summary updated