Abstract
In complex industrial settings, it is common practice to monitor the operation of machines in order to detect undesired states, adjust maintenance schedules, optimize system performance or collect usage statistics of individual machines. In this work, we focus on estimating the power output of a Combined Heat and Power (CHP) machine of a medium-sized company facility by analyzing the total facility power consumption. We formulate the problem as a time-series classification problem where the class label represents the CHP power output. As the facility is fully instrumented and sensor measurements from the CHP are available, we generate the training labels in an automated fashion from the CHP sensor readings. However, sensor failures result in mislabeled training data samples which are hard to detect and remove from the dataset. Therefore, we propose a novel multi-task deep learning approach that jointly trains a classifier and an autoencoder with a shared embedding representation. The proposed approach targets to gradually correct the mislabelled data samples during training in a self-supervised fashion, without any prior assumption on the amount of label noise. We benchmark our approach on several time-series classification datasets and find it to be comparable and sometimes better than state-of-the-art methods. On the real-world use-case of predicting the CHP power output, we thoroughly evaluate the architectural design choices and show that the final architecture considerably increases the robustness of the learning process and consistently beats other recent state-of-the-art algorithms in the presence of unstructured as well as structured label noise.
Introduction
Monitoring individual machines in industrial environments is vital for operational efficiency and predictive maintenance. However, the high cost of per-device instrumentation often necessitates Non-Intrusive Load Monitoring (NILM)—disaggregating total facility power consumption to estimate the state of specific units, such as Combined Heat and Power (CHP) machines.
A significant barrier to training effective NILM models is label noise. While training labels can be generated automatically from temporary sensors, these sensors frequently fail, producing erroneous data. Standard deep neural networks tend to memorize this noise, degrading their generalization capabilities. Unlike existing solutions that require a predefined noise ratio, this research introduces a framework that dynamically identifies and corrects label noise during the training process.
The SREA Framework
The researchers propose SREA (Self-Re-Labeling with Embedding Analysis), a multi-task deep learning architecture designed for robustness against noisy labels. The core of SREA is a shared low-dimensional embedding space that feeds into three distinct modules:
- Autoencoder: Acts as a self-supervised anchor, focusing on signal reconstruction to learn robust features independent of potentially incorrect labels.
- Classifier: Predicts the power output level based on the learned embeddings.
- Constrained Clustering: Ensures that different power states are geometrically separated in the embedding space, facilitating clearer class boundaries.
The training process follows a strategic three-phase schedule. It begins with a Warm-up phase for unsupervised feature learning. During the Re-labeling phase, the model compares provided labels with its internal cluster assignments and predictions, gradually "correcting" labels where disagreements occur. Finally, a Fine-tuning phase uses these cleaned labels to reach peak performance.
Experimental Setup and Datasets
The SREA framework was validated using two primary sources:
- UCR Time-Series Archive: Ten diverse datasets used to benchmark against state-of-the-art algorithms.
- Real-World CHP Dataset: 78 days of measurements from an industrial facility, utilizing total load ($P_{tot}$), water temperature ($T_{water}$), and ambient temperature ($T_{amb}$) to predict CHP power output.
To simulate real-world sensor failures, the researchers introduced various noise types: Symmetric (random), Asymmetric (class-shifting), and Flip noise (simulating a sensor stuck at zero).
Key Results and Insights
SREA demonstrated superior robustness compared to standard Cross-Entropy training and other noise-robust methods:
- Benchmark Performance: SREA outperformed baselines in the majority of UCR experiments, particularly under asymmetric noise conditions.
- CHP Robustness: In the real-world CHP use case, SREA maintained a high F1-score of ~0.95 even when 40% of training labels were corrupted. Standard models saw significant performance drops under similar conditions.
- Dynamic Correction: A key advantage of SREA is its ability to discover noise levels dynamically, removing the need for users to estimate the "noise ratio" beforehand.
Conclusion
SREA provides a robust solution for industrial time-series classification where sensor reliability is imperfect. By integrating self-supervised reconstruction and geometric clustering, the model effectively filters out structured noise. Future work will explore incorporating temporal dependencies between data windows using Recurrent Neural Networks or Transformers, and extending the framework to real-time anomaly detection for identifying sensor failures as they occur.
Summary updated