Deep Learning for Bearing Fault Detection

Project: Bearing Fault Detection and Classification Using Temporal Convolutions and LSTM Networks in Induction Machine Systems.

This project applies deep learning to detect and classify faults in bearings of induction machines. The model leverages Temporal Convolutional and LSTM layers to effectively identify various fault types.

---

📄 Publication Reference

For a comprehensive exploration of our methodology and findings, consult our published paper:

Title: Classification of Fault Severity in Induction Machine Systems Using Temporal Convolutions and Recurrent Networks
Authors: V. Mashayekhi, S. Hasani Borzadaran, M. Hoseintabar Marzebali
Published: February 16, 2022
Access: DOI

This paper outlines the advanced deep learning techniques developed to detect and classify fault severity in induction machine systems.

---

Requirements

To reproduce the results, install the dependencies in a Python 3.7 environment.

• Python: 3.7
• Required Packages: Listed in requirements.txt

Dataset Summary

The dataset consists of bearing fault signals, pre-processed for experiments with optional downsampling. The table below summarizes the data structure:

Downsampling	Sequence Length	Training Samples	Test Samples	Classes
Yes	899	3095	890	16
No	1000	-	-	16

---

Model Evaluation

The following table shows the performance of GRU-based models across various configurations, using different downsampling and architecture options to optimize classification accuracy.

Run	Epochs	Batch Size	Architecture	Weights	Downsampling	Accuracy
1	600	128	#1	1	Yes	71.46%
2	1200	128	#1	2	Yes	84.26%
3	1200	128	#2	3	Yes	84.94%
4	1200	128	#3	4	Yes	87.52%
5	900	64	#1	5	Yes	83.37%
_	_	_	_	_	_	_
6	200	128	#4	-	No	71.35%
7	500	128	#4	-	No	94.65%
8	650	128	#4	-	No	95.37%
9	700	128	#4	-	No	95.48%
10	900	128	#4	-	No	95.62%
11	1050	128	#4	-	No	95.77%
_	_	_	_	_	_	_
12	600	-	#5	-	No	91.91%
13	850	-	#5	-	No	93.70%
14	1000	-	#5	-	No	92.43%
15	1050	-	#5	-	No	92.43%

---

Model Architectures

This section details the model architectures used for bearing fault detection, optimized with different downsampling strategies to enhance efficiency and accuracy. Each architecture utilizes 1D convolutional layers with varying configurations.

---

Architectures with Downsampling (Yes)

These architectures employ downsampling to reduce sequence length, facilitating faster training while retaining essential temporal features.

• #1 Architecture: A four-layer convolutional model that maintains consistent dimensionality across the first three layers, with a reduction in the final layer.

  Conv1D(128) => Conv1D(128) => Conv1D(128) => Conv1D(64)

• #2 Architecture: An enhanced version of #1 with increased dimensionality in the middle layers for improved feature extraction.

  Conv1D(128) => Conv1D(256) => Conv1D(256) => Conv1D(64)

• #3 Architecture: This configuration introduces a more gradual increase in filter sizes, potentially capturing finer-grained features before reduction.

  Conv1D(64) => Conv1D(128) => Conv1D(256) => Conv1D(64)

---

Architectures without Downsampling (No)

These architectures retain the full sequence length, designed for more detailed temporal feature extraction without losing data granularity.

• #4 Architecture: A stable model with identical filter sizes in the first three layers, enabling consistent feature maps, followed by a final reduction layer.

  Conv1D(128) => Conv1D(128) => Conv1D(128) => Conv1D(64)

Figure 1: Results for Architecture #4

---

• #5 Architecture: A deeper configuration with an initial smaller filter size, followed by an increase, allowing the model to learn a hierarchical representation of features.

  Conv1D(64) => Conv1D(64) => Conv1D(64) => Conv1D(128) => Conv1D(64)

Figure 2: Results for Architecture #5

---

Each architecture has been tested with specific training conditions to measure its accuracy and performance, providing insights into the effectiveness of downsampling for fault detection in induction machine systems.