Fraunhofer Institute for Manufacturing Engineering and Automation IPA
AI-Based Anomaly Detection and Predictive Maintenance
Detect deviations – prevent failures
Unplanned downtime, gradual energy losses, and unexplained quality issues are costly disruptions in any production environment. They are often caused by subtle deviations from normal process behavior – anomalies that remain undetected by conventional monitoring methods. To address this, we use the energy consumption of your equipment as a highly sensitive indicator of process stability and equipment health.
Our core expertise lies in energy-based anomaly detection. We use the energy and process data already available in your operations to learn the unique “energy signature” of your processes. Our AI systems detect even the smallest deviations from this normal state in real time. This allows us to identify operator errors, mechanical wear, or leaks long before they lead to costly defects or production downtime.
Our approach is highly cost-efficient, as it primarily relies on existing data streams and minimizes the need for additional sensor infrastructure.
Our methodological approach: From theory to industrial application
Anomaly detection scenarios: Selecting the right strategy for your data
- Unsupervised scenario (the industry standard): In most cases, little or no labeled fault data is available. Our systems learn normal behavior directly from your data and identify any significant deviation as an anomaly. This approach is well suited for real-world applications, as it avoids time-consuming manual annotation.
- Supervised scenario: When labeled data for normal and faulty states is available, we formulate the problem as a classification task. This approach achieves high detection rates for known fault types.
- Semi-supervised scenario: We combine a small amount of labeled data with a large volume of unlabeled data. Even a limited number of labeled examples can significantly improve detection performance and model robustness.
Methods: Our toolbox for high-precision anomaly detection
We cover the three main classes of anomaly detection methods and combine them into hybrid, high-performance solutions:
- Reconstruction-based methods: These methods are particularly suitable for complex, high-dimensional time series. Models such as autoencoders and transformers (e.g., LSTM-AE, TranAD) learn to reconstruct normal energy patterns. Observations with high reconstruction error are flagged as anomalies. They are especially effective at capturing subtle, non-linear patterns.
- Distance-based methods: Measurements are represented in a multidimensional feature space. Algorithms such as k-means or DBSCAN identify clusters of normal behavior. Data points with large distances to the nearest cluster are classified as anomalies. These methods are often fast and efficient, particularly when the data exhibits clear structure.
- Classification-based methods: In supervised settings, we use machine learning models (e.g., XGBoost) to distinguish between normal operation and anomalies. Performance is evaluated using standard metrics such as precision, recall, and F1 score.
Our approach: From data to operational deployment in four phases
Phase 1: Data acquisition and preprocessing
Robust anomaly detection starts with high-quality data. We aggregate energy and process data from your measurement systems (e.g., power meters, PLCs), synchronize time series, and clean the data. Missing values are handled using advanced statistical (e.g., ARIMA) or machine learning methods (e.g., LSTM).
Phase 2: Modeling and training
The choice of algorithm depends on the specific scenario and data characteristics. In unsupervised settings, we model normal behavior. Reconstruction-based methods such as autoencoders are particularly effective, as they can learn complex, non-linear patterns in energy data. We define a precise anomaly score and rigorously validate detection performance.
Phase 3: Detection and alerting
In operation, the trained model evaluates new measurements in real time. Deviations from learned normal behavior are quantified (e.g., via reconstruction error). These insights are translated into alerts or recommended actions and passed on to downstream systems.
Phase 4: Integration and feedback
Results are seamlessly integrated into your system landscape (e.g., energy management, maintenance, MES). Dashboards present findings and help identify root causes. Expert feedback can be incorporated as new labels, enabling continuous model improvement through regular retraining.
Use Cases: Added Value for Industry
Our expertise in anomaly detection using energy flows delivers direct, measurable benefits:
- Early-Warning System for Machinery (Predictive Maintenance): Abnormal energy signatures indicate bearing failures, motor malfunctions, or wear and tear. This enables predictive maintenance, reduces unplanned downtime, and extends the service life of equipment.
- Leak Detection in Compressed Air and Heating/Cooling Networks: Unexpected sustained consumption or atypical loads at night indicate precise leak locations. Rapid localization minimizes energy losses and prevents consequential damage.
- Quality Control in Production: Energy consumption often correlates directly with product quality. In a project with Lorch Schweißtechnik, our system detected operator errors during welding based solely on current and voltage data, thereby helping to prevent defective batches.
- Equipment Diagnostics: Monitoring pumps, inverters, or cooling systems identifies abnormal operating cycles, blockages, or overloads and enables rapid intervention.
FAQ – frequently asked questions on energy monitoring and NILM
Why are simple threshold values not sufficient?
Static thresholds for anomaly detection are prone to false alarms, as they cannot capture complex, dynamic patterns. AI-based methods identify subtle changes in context (e.g., seasonal variations or different product variants) and adapt to changing operating conditions.
How is the quality of anomaly detection evaluated?
In research, performance is often assessed using standard metrics such as precision or F1 score, as fully labeled fault data is typically available. In industrial practice, however, this is rarely the case. In unsupervised settings, we therefore apply a human-in-the-loop approach: results are validated through expert sampling and compared with historical maintenance data. Ultimately, what matters is the operational value – such as measurable reductions in costs or downtime.
How do you handle limited labeled fault data?
This is the typical situation in industry. Unlike research, which often relies on fully labeled datasets (supervised learning), such data is rarely available in production environments. Our methods are therefore designed for unsupervised and semi-supervised scenarios. Models learn normal behavior directly from operational data and detect deviations automatically – without the need for extensive manual annotation.
What data is relevant beyond energy consumption data?
While our focus is on cost-efficient energy data, integrating contextual information significantly improves detection performance. Relevant data includes process variables (e.g., pressure, temperature), production plans, material batches, and machine states from PLC systems. These “exogenous variables” help the model better understand the causes of changes in energy consumption.
Is this approach suitable for small and medium-sized enterprises?
Yes. Our systems are modular, primarily use existing measurement data, and do not require costly additional hardware. The benefits – reduced downtime, lower energy costs, and more predictable maintenance windows – typically result in a rapid return on investment.