[Paper Review] Estimating the electrical power output of industrial devices with end-to-end time-series classification in the presence of label noise

Andrea Castellani, Sebastian Schmitt+1 more|arXiv (Cornell University)|May 1, 2021

Time Series Analysis and ForecastingComputer Science65 references23 citations

TL;DR

This paper proposes Self-Re-Labeling with Embedding Analysis (SREA), a multi-task deep learning framework that jointly trains a time-series classifier and an autoencoder with shared representations to correct label noise in industrial power output estimation. SREA effectively detects and re-labels misclassified samples from faulty sensors in a self-supervised manner, significantly outperforming state-of-the-art methods on both synthetic and real-world noisy time-series data, including Combined Heat and Power (CHP) machine power prediction.

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.

Motivation & Objective

• To address the challenge of label noise in time-series classification for industrial energy monitoring, particularly when training labels are automatically generated from faulty sensors.
• To develop a self-supervised deep learning method that can detect and correct mislabeled samples without prior knowledge of noise levels.
• To improve the robustness of time-series classification models in real-world industrial settings where sensor failures lead to unreliable training data.
• To evaluate the method on both synthetic noise-corrupted benchmarks and a real-world CHP power output estimation use case with actual sensor failures.
• To analyze the sensitivity of the method to hyper-parameters and identify optimal architectural configurations.

Proposed method

• SREA employs a multi-task learning framework with a shared encoder that produces a shared embedding representation for both a time-series classifier and an autoencoder.
• The classifier predicts the CHP power output class from the shared embedding, while the autoencoder reconstructs the input time series from the same embedding.
• Label noise correction is achieved through a self-relabeling mechanism that iteratively reassigns labels based on the model’s confidence and the embedding space structure.
• The method uses a dynamic loss weighting strategy that starts with a high weight on the classification loss and gradually shifts to prioritize reconstruction and consistency in the embedding space.
• The training process includes a warm-up phase where the classifier is initially trained on original labels, followed by a joint optimization phase where both tasks are updated in tandem.
• The framework is designed to be robust to both symmetric and asymmetric label noise, as well as flip noise, without requiring prior knowledge of noise ratios.

Experimental results

Research questions

• RQ1 Can a self-supervised multi-task learning framework effectively detect and correct mislabeled time-series samples caused by sensor failures in industrial settings?
• RQ2 How does the proposed SREA method compare to state-of-the-art time-series classification models under various types of label noise (symmetric, asymmetric, flip)?
• RQ3 What is the impact of input signal selection (e.g., total power, ambient temperature, water temperature) on the performance of the model in the presence of label noise?
• RQ4 How sensitive is the SREA model to its hyper-parameters, particularly the initial loss weight and the schedule for adjusting the reconstruction loss?
• RQ5 Can SREA generalize to real-world industrial data with unstructured and unpredictable sensor failures, beyond controlled synthetic noise?

Key findings

• On synthetic noise-corrupted benchmarks, SREA achieved state-of-the-art performance, outperforming methods like MixUp-BMM, Co-teaching, and SIGUA, especially under high noise ratios.
• With 30% asymmetric label noise, SREA achieved an F1-score of 0.921, significantly outperforming the next-best method (Co-teaching at 0.871) on the CHP dataset.
• In real-world evaluation, SREA correctly re-labeled the period of PCHP sensor failure (19th Sept.) and detected other active CHP phases, demonstrating robustness to real sensor faults.
• The ablation study showed that including all three input signals (Ptot, Tamb, Twater) yielded the highest performance (F1-score of 0.978 under symmetric noise), confirming the importance of multimodal input.
• Hyper-parameter analysis revealed that λinit = 40 and ∆end = 30 provided the most stable performance across noise types, with minimal sensitivity to ∆start.
• The confusion matrix after correction showed a near-perfect class distribution, with 97.8% of samples correctly classified, indicating effective noise correction.