Research on Unstructured Text Data Mining and Fault Classification Based on RNN-LSTM with Malfunction Inspection Report

Daqian Wei, Bo Wang, Gang Lin, Dichen Liu, Zhaoyang Dong, Hesen Liu, Yilu LiuEnergies 2017 · DOI

TL;DR

This paper tackles the challenge of extracting valuable insights from unstructured malfunction inspection reports for power transformers. They propose a deep learning approach using Recurrent Neural Networks with Long Short-Term Memory (RNN-LSTM) to automatically classify faults, offering a more efficient alternative to manual analysis.

What Problem Does This Paper Address?

The operation and maintenance of power grids increasingly rely on condition-based maintenance (CBM) for power transformers. While structured data (numeric, categorical) has been extensively used for fault diagnosis, a significant portion (around 80%) of crucial information resides in unstructured text, specifically in malfunction inspection reports. Traditional methods for analyzing these reports are time-consuming and often fail to capture the full spectrum of insights. This paper aims to bridge this gap by developing an automated method to process and classify faults from these unstructured reports, thereby improving grid safety and efficiency.

Methods

The core of this research lies in applying deep learning, specifically a combination of Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks, to process natural language text from malfunction inspection reports.

  • RNNs: These networks are well-suited for sequential data like text, as they can maintain a "memory" of previous inputs to inform current predictions. However, standard RNNs struggle with capturing long-term dependencies due to the vanishing gradient problem.
  • LSTMs: To overcome RNN limitations, LSTMs were developed. They introduce a "cell state" and "gates" (forget, input, and output gates) that allow the network to selectively remember or forget information over extended sequences. This makes them particularly effective for tasks involving long-term dependencies in text.
  • Text Representation: Words are converted into vectors, where each dimension can represent a word.
  • Training Strategy: A straightforward training strategy is employed where targets are replicated at each sequence step. Fault labels are provided in the dataset to calculate classification accuracy.

Datasets

The study utilizes a dataset of malfunction inspection reports from the China Southern Power Grid, collected over approximately 10 years. This dataset comprises six main information bodies: "DeviceInfo," "TripInfo," "Faultinfo," "DigitalStatus," "DigitalEvent," and "SettingValue." The faults are categorized into five types: mechanical fault, electrical fault, secondary equipment fault, fault caused by external environment, and fault caused by human factors. The paper also analyzes semantic relationships within the data, categorizing them into nine types, including Cause-Effect (CE) and Entity-Origin (EO).

Key Insights and Ideas

  • Unlocking Unstructured Data: The paper highlights that a vast amount of critical information for power transformer maintenance is locked within unstructured text reports, which are often overlooked.
  • RNN-LSTM for Textual Fault Analysis: The research demonstrates that RNN-LSTM, a powerful deep learning architecture, is effective in capturing the nuances of textual fault descriptions and classifying them accurately.
  • Parameter Tuning is Crucial: The study emphasizes that careful selection and tuning of key parameters (number of LSTM units, activation function, batch size, traversal times) are essential for achieving optimal fault recognition accuracy.
  • Beyond Simple Keywords: The RNN-LSTM approach goes beyond simple keyword matching by understanding the context and semantic relationships within the reports, leading to more robust fault classification.

Key Results

The experimental results show that the proposed RNN-LSTM method can successfully classify labeled malfunction inspection reports.

  • Accuracy Improvement: The accuracy of fault recognition generally increases with the number of traversals (epochs) and the number of LSTM units, up to a certain point.
  • Optimal Parameters: The study identifies that an LSTM unit count of 128 and a batch size of 20, along with sigmoid or Sofamax activation functions, tend to yield better results. For instance, with 128 LSTM units and 50 traversal times, an accuracy of approximately 0.64587 was achieved with a batch size of 20.
  • ROC Analysis: Receiver Operating Characteristic (ROC) curve analysis further supports the effectiveness of the RNN-LSTM model, with AUC values indicating good discriminative ability. For example, with 256 LSTM units, the AUC reached 0.8319.

Limitations and Future Work

  • Timestamped Data: The current study uses fault labels without timestamps. The authors suggest that incorporating timestamped diagnoses would enable models to perform early fault diagnosis by predicting future conditions, which would be a significant advancement.
  • Parameter Optimization: While key parameters were explored, further optimization and exploration of hyperparameter tuning could potentially lead to even higher accuracy.
  • Generalizability: While the dataset is from a specific power grid, future work could explore the generalizability of the model to reports from different regions or with varying reporting styles.

Summary updated

Abstract

This paper documents the condition-based maintenance (CBM) of power transformers, the analysis of which relies on two basic data groups: structured (e.g., numeric and categorical) and unstructured (e.g., natural language text narratives) which accounts for 80% of data required. However, unstructured data comprised of malfunction inspection reports, as recorded by operation and maintenance of the power grid, constitutes an abundant untapped source of power insights. This paper proposes a method for malfunction inspection report processing by deep learning, which combines the text data mining–oriented recurrent neural networks (RNN) with long short-term memory (LSTM). In this paper, the effectiveness of the RNN-LSTM network for modeling inspection data is established with a straightforward training strategy in which we replicate targets at each sequence step. Then, the corresponding fault labels are given in datasets, in order to calculate the accuracy of fault classification by comparison with the original data labels and output samples. Experimental results can reflect how key parameters may be selected in the configuration of the key variables to achieve optimal results. The accuracy of the fault recognition demonstrates that the method we proposed can provide a more effective way for grid inspection personnel to deal with unstructured data.
BibTeX
@inproceedings{wei2017research,
  author = {Daqian Wei and Bo Wang and Gang Lin and Dichen Liu and Zhaoyang Dong and Hesen Liu and Yilu Liu},
  title = {{Research on Unstructured Text Data Mining and Fault Classification Based on RNN-LSTM with Malfunction Inspection Report}},
  booktitle = {Energies},
  year = {2017},
  doi = {10.3390/en10030406},
}
Figure