Skip to content

Bearing Fault Diagnosis with a 1D CNN

A 1D CNN that classifies bearing faults from raw vibration signals (CWRU dataset), tested on an unseen motor load and under added noise, with a Streamlit app for uploading a signal.
  • Deep Learning
  • Signal Processing
  • Predictive Maintenance
Date
June 2025
Result
100% on an unseen motor load; 87% with 50% added noise
  • 1D CNN on raw vibration windows, PyTorch
  • Tested on an unseen motor load (2 HP)
  • Noise test: holds at 25% noise, fails past 50%

Can a small 1D CNN tell a healthy bearing from three kinds of fault using only the raw vibration signal, and how far can it be trusted?

Watch the video demo · Open the app

Setup

  • Data: the public Case Western Reserve University (CWRU) bearing dataset, four classes: normal, inner-race fault (IRF), outer-race fault (ORF) and ball fault (BF).
  • Windows: each recording is cut into 2,048-sample windows with 50% overlap.
  • Split by motor load, not at random: training and validation use the 0 HP and 1 HP recordings; the test set is the 2 HP recordings (824 windows), which the model never sees. A random split would put overlapping windows of the same recording on both sides and inflate the score.
  • Imbalance: normal windows outnumber each fault class about four to one. Two models were trained for 25 epochs each: one on the original data, one on data balanced with SMOTE.
  • Model: three Conv1d + BatchNorm + pooling blocks and a small classifier, in PyTorch.

Results

Test (2 HP, 824 windows)Accuracy
Original data100%
SMOTE-balanced data100%
+ noise at 25% of signal std100%
+ noise at 50%87.0%
+ noise at 75%28.8%
+ noise at 100%23.7%

What the numbers actually say

  • CWRU is easy for a CNN. Both models were perfect on the unseen load, so this test cannot show whether SMOTE helps. A harder dataset, or a test on different bearings, would.
  • Noise is where it breaks. At 50% noise, ball faults are mostly mistaken for inner-race faults (BF recall 9%). At 75%, almost everything is called an inner-race fault. A real sensor in a noisy plant would need noise augmentation in training, or denoising before the model.
  • The app handles real-length signals. Inputs shorter than a window are extended by reflection padding; longer ones are cut into overlapping windows, and the class probabilities are averaged across them.

The Streamlit app

Built with

Python, PyTorch, scikit-learn and imbalanced-learn (scaling, SMOTE), pandas and NumPy, and Streamlit for the app. The four notebooks go from data exploration to the trained model and the app.

Deep Learning · Self-Supervised LearningDistill-SSLSelf-Supervised Teachers, Small StudentsDeep Learning · Medical ImagingMRI Brain Tumor Segmentation