Learning to Monitor Machine Health with Convolutional Bi-Directional LSTM Networks

Rui Zhao, Ruqiang Yan, Jinjiang Wang, Kezhi Mao

TL;DR

This paper introduces a novel deep learning architecture, Convolutional Bi-directional Long Short-Term Memory networks (CBLSTM), for machine health monitoring. By combining Convolutional Neural Networks (CNNs) for local feature extraction with Bi-directional Long Short-Term Memory (Bi-LSTM) networks for temporal information encoding, CBLSTM effectively predicts tool wear from raw sensor data, outperforming existing methods.

What Problem Does This Paper Address?

Modern manufacturing relies heavily on machine health monitoring to ensure efficiency and prevent failures. While data-driven approaches are promising, raw sensor data often suffers from noise, varying lengths, and irregular sampling, making it difficult to feed directly into traditional models. Previous methods often required extensive manual feature engineering and expert knowledge, which is time-consuming and can be a bottleneck. This paper aims to overcome these limitations by developing a deep learning model that can learn robust representations directly from raw, sequential sensor data for tasks like tool wear prediction.

Methods

The proposed Convolutional Bi-directional Long Short-Term Memory (CBLSTM) network is a hybrid deep learning architecture designed to process raw time-series sensor data. It consists of three main components:

  1. Local Feature Extractor (CNN): A single-layer Convolutional Neural Network (CNN) is used first. It slides filters over the raw input sequence to extract local, robust, and informative features. This step helps to reduce noise and compress the sequence length, making it more manageable for subsequent layers.
  2. Temporal Encoder (Bi-directional LSTM): Following the CNN, a two-layer Bi-directional Long Short-Term Memory (Bi-LSTM) network is employed. LSTMs are known for their ability to capture long-term dependencies in sequential data. The bi-directional nature of this component allows the model to consider both past and future contexts within the sequence, leading to a more comprehensive understanding of temporal patterns.
  3. Output Layers: The output from the Bi-LSTM is then passed through two fully-connected dense layers for higher-level representation learning, followed by a linear regression layer to predict the target value (tool wear depth).

The model also incorporates dropout as a regularization technique to prevent overfitting, which is crucial given the potential for limited training data in real-world machine monitoring scenarios.

Datasets

The study utilizes data from a high-speed CNC machine operating under dry milling conditions. Six different sensor signals (three-component cutting forces and three-axis vibrations) were collected. The raw data, sampled at 50 kHz, was pre-processed by dividing it into 100 sections and taking the max and mean values to form sequential data samples of length 100 with a dimensionality of 12 per time step.

Three tool life tests (C1, C4, and C6) were conducted and used as datasets. A three-fold cross-validation strategy was employed for training and testing, where two tests were used for training and the remaining one for testing. This approach ensures robust evaluation across different operational conditions.

Key Insights and Ideas

  • End-to-End Learning from Raw Data: The CBLSTM model demonstrates the power of deep learning to learn directly from raw sensor data, eliminating the need for manual feature engineering and expert knowledge.
  • Synergy of CNN and Bi-LSTM: The combination of CNNs for local feature extraction and Bi-LSTMs for temporal modeling proves highly effective. CNNs filter noise and extract salient local patterns, while Bi-LSTMs capture complex temporal dependencies, including future context.
  • Importance of Bi-directionality: The bi-directional nature of the LSTM component significantly improves performance by allowing the model to leverage information from both past and future time steps, which is crucial for understanding the evolving state of a machine.
  • Regularization is Key: Dropout is shown to be effective in mitigating overfitting, a common challenge in deep learning, especially when dealing with potentially limited datasets in industrial settings.

Key Results

The proposed CBLSTM model achieved superior performance compared to several baseline methods, including Linear Regression (LR), Support Vector Regression (SVR), Multi-layer Perceptron (MLP), Recurrent Neural Networks (RNNs), and various configurations of LSTMs and Bi-LSTMs.

  • MAE and RMSE: Across the three testing datasets (C1, C4, C6), CBLSTM consistently achieved the lowest Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). For instance, on dataset C6, CBLSTM achieved an MAE of 8.1 and an RMSE of 9.8, outperforming the next best model (Deep BLSTM) which had an MAE of 14.4 and RMSE of 18.9.
  • Qualitative Results: Visualizations of predicted tool wear against ground truth data showed that CBLSTM accurately follows the trend of actual tool wear, indicating its effectiveness in capturing the underlying dynamics.
  • Efficiency: The model demonstrated efficient training and testing times, with training for one epoch taking approximately 5 seconds and testing each sample taking only 0.027 seconds on a NVIDIA Tesla K40c GPU, making it suitable for real-time machine monitoring.

Limitations and Future Work

  • Limitations: The paper acknowledges that while dropout helps, the model's performance can still be affected by the availability of sufficient training data, which can be a challenge in industrial environments.
  • Future Work: The authors propose exploring the integration of wavelet transformation into their deep neural network models. They believe that combining shallow feature extraction (like wavelets) with deep feature extraction (CNN-BiLSTM) could lead to even more effective and efficient machine health monitoring solutions.

Summary updated

Abstract

In modern manufacturing systems and industries, more and more research efforts have been made in developing effective machine health monitoring systems. Among various machine health monitoring approaches, data-driven methods are gaining in popularity due to the development of advanced sensing and data analytic techniques. However, considering the noise, varying length and irregular sampling behind sensory data, this kind of sequential data cannot be fed into classification and regression models directly. Therefore, previous work focuses on feature extraction/fusion methods requiring expensive human labor and high quality expert knowledge. With the development of deep learning methods in the last few years, which redefine representation learning from raw data, a deep neural network structure named Convolutional Bi-directional Long Short-Term Memory networks (CBLSTM) has been designed here to address raw sensory data. CBLSTM firstly uses CNN to extract local features that are robust and informative from the sequential input. Then, bi-directional LSTM is introduced to encode temporal information. Long Short-Term Memory networks (LSTMs) are able to capture long-term dependencies and model sequential data, and the bi-directional structure enables the capture of past and future contexts. Stacked, fully-connected layers and the linear regression layer are built on top of bi-directional LSTMs to predict the target value. Here, a real-life tool wear test is introduced, and our proposed CBLSTM is able to predict the actual tool wear based on raw sensory data. The experimental results have shown that our model is able to outperform several state-of-the-art baseline methods.
BibTeX
@inproceedings{zhao2017learning,
  author = {Rui Zhao and Ruqiang Yan and Jinjiang Wang and Kezhi Mao},
  title = {{Learning to Monitor Machine Health with Convolutional Bi-Directional LSTM Networks}},
  booktitle = {Sensors},
  year = {2017},
  url = {https://openreview.net/forum?id=Z4HhSaYzn2},
}
Figure