rz0718/deftcnn

Explainable Machine Health Monitoring Systems

Decoupled Feature-Temporal CNN: Explaining Deep Learning-Based Machine Health Monitoring

This repo contains the code used in our ICSMD paper. The published version will be updated soon.

Motivation

Our work is trying to make "machine doctors" act as the real doctors who not only make diagnosis, but also describe patients' symptoms. Our proposal algorithms will generate the saliency map to explain which parts of our input features contribute to the algorithms' decisions.

Table of Contents

• Decoupled Feature-Temporal CNN: Explaining Deep Learning-Based Machine Health Monitoring
  • Table of Contents
  • Data
  • Code
    • Data Preprocessing
    • Feature Extraction
    • DEFTCNN

Data

This folder contains the features and labels in a clean format, which will be used as the input data for the main notebook: DEFTCNN.

Code

This folder contains codes for data preprocessing, feature extraction and the implementation of DEFTCNN.

Data Preprocessing

Sample the cwru data.

Feature Extraction

Based on the sampled data, extract different domain features from the time series data. And the features extractions are applied on time windows.

DEFTCNN

deftcnn's Issues

Can Grad-CAM work if we don't know the true label?

To compute cam, we should know the true label of the input data to get the gradient. Can Grad-CAM work if we don't know the true label? Intuitively explaination should not depend on the true label. If it can could you tell me how to implement, and if not are there any methods to solve the problem?