An Explainable Neural Network for Fault Diagnosis With a Frequency Activation Map

Min Su Kim,Jong Pil Yun,Poogyeon Park

doi:10.1109/access.2021.3095565

Abstract

This paper constructs an explainable neural network model for fault diagnosis with a 1D vibration signal of equipment and proposes an explainable method with a frequency activation map of the proposed model. The frequency activation map visualizes the classification criteria of the time-domain-based learned model in the frequency domain. Since the 1D vibration signal for monitoring the normal and faulty states of equipment is easy to interpret in the frequency domain, the frequency activation map provides the user with a specific frequency of vibration signal where the proposed model focuses on for the classification of normal and faulty states. To generate the frequency activation map, the proposed model structure for learning the 1D vibration signals is designed to filter the frequency components of the 1D vibration signals using a 1D convolutional filter with a norm constraint. Simulation results with two open datasets demonstrate that the proposed model and explainable method can visualize the classification criteria of the model learned with vibration signals through a frequency activation map. Based on the frequency activation map, characteristic frequencies of normal and faulty states are identified.

Highlights

Fault detection and diagnosis can be divided into model-based and data-based methods [1], [2]
Inspired by [20], this paper proposes a novel visualization method to visualize the classification criteria of the learned end-to-end model for fault diagnosis
The proposed explainable method, Frequency Activation Map (FAM), visualizes the frequency-domain-based classification criteria based on the high-level features of the model learned in the time domain

Summary

INTRODUCTION

Fault detection and diagnosis can be divided into model-based and data-based methods [1], [2]. The proposed explainable method, FAM, visualizes the frequency-domain-based classification criteria based on the high-level features (expressed as time-series signal) of the model learned in the time domain. The deep learning model determines the state of the equipment based on specific frequencies of the input signal This is to design a learning principle for visualization so that the feature extractor of the proposed model converts the vibration signal into time-series signals with band-passed frequency and the classifier of the proposed model classifies normal or faulty states based on these (Fig. 3 (c)). PROPOSED STRUCTURE 1) 1D CONVOLUTIONAL LAYER AND ACTIVATION LAYER The proposed model is designed to create the FAM It means that the proposed model must learn the frequency band of the vibration signal, which is a time series input signal. The FAM can be derived with a smaller amount of computation than performing frequency transformation on each of the N frequency maps (Eq (12))

IMPLEMENTATION DETAILS

CASE STUDY

CASE STUDY 1

CASE STUDY 2

Findings

CONCLUSION