Centrala begrepp
A neuro-symbolic architecture that uses an online rule-learning algorithm to explain when a deep-learning-based anomaly detection model predicts failures in predictive maintenance applications.
Sammanfattning
This paper proposes a two-layer architecture for explainable predictive maintenance. The first layer uses an unsupervised deep-learning model, specifically an LSTM autoencoder, to detect anomalies and potential failures. The second layer learns interpretable regression rules that explain the outputs of the detection layer model.
The key highlights are:
The LSTM autoencoder is trained on normal operating data to learn the system's normal behavior. It signals an alarm when it receives data that deviates significantly from the normal, potentially indicating a failure.
In parallel, the rule learning system (AMRules) receives the input features and the reconstruction error from the autoencoder as the target variable. It learns a set of interpretable rules that map the input features to the reconstruction error.
The authors use an oversampling technique (ChebyOS) to focus the rule learning on the rare, high-value cases of the reconstruction error, which are the most relevant for predictive maintenance.
The proposed system can provide both global explanations (the set of learned rules) and local explanations (the specific rules triggered for a particular input) to help operators, technicians, and managers understand the causes of the detected anomalies and plan the appropriate maintenance actions.
The authors evaluate the approach on a real-world case study of predictive maintenance for the Metro do Porto system, demonstrating the benefits of the explanations provided by the neuro-symbolic architecture.
Statistik
The reconstruction error of the LSTM autoencoder is used as the target variable for the rule learning system.
Citat
"Fault detection is one of the most critical components of predictive maintenance. Nevertheless, predictive maintenance goes far behind predicting a failure, and it is essential to understand the consequences and the collateral damages of the failure."
"Explanations in predictive maintenance play a relevant role in identifying the causes of failure, e.g., the component in failure. This is the type of information required to define the repair plan."