Machine Learning based Analytical Framework for Automatic Hyperspectral Raman Analysis of Lithium-ion Battery Electrodes

Ankur Baliyan,Hideto Imai

doi:10.1038/s41598-019-54770-2

Abstract

The intelligence to synchronously identify multiple spectral signatures in a lithium-ion battery electrode (LIB) would facilitate the usage of analytical technique for inline quality control and product development. Here, we present an analytical framework (AF) to automatically identify the existing spectral signatures in the hyperspectral Raman dataset of LIB electrodes. The AF is entirely automated and requires fewer or almost no human assistance. The end-to-end pipeline of AF own the following features; (i) intelligently pre-processing the hyperspectral Raman dataset to eliminate the cosmic noise and baseline, (ii) extract all the reliable spectral signatures from the hyperspectral dataset and assign the class labels, (iii) training a neural network (NN) on to the precisely “labelled” spectral signature, and finally, examined the interoperability/reusability of already trained NN on to the newly measured dataset taken from the same LIB specimen or completely different LIB specimen for inline real-time analytics. Furthermore, we demonstrate that it is possible to quantitatively assess the capacity degradation of LIB via a capacity retention coefficient that can be calculated by comparing the LMO signatures extracted by the analytical framework (AF). The present approach is suited for real-time vibrational spectroscopy based industrial applications; multicomponent chemical reactions, chromatographic, spectroscopic mixtures, and environmental monitoring.

Highlights

Hyperspectral Raman imaging caters the ability to image multiple chemical signatures simultaneously, and the spatial information is collected in the X–Y plane, and the spectral information is represented in the Z-direction, hyperspectral images are represented in the form of data cubes (Fig. 1)
The capacity retention inside the lithium-ion battery (LIB) cell decrease with an increase in the charging/discharging cycles and, until 300 cycles, the drop in capacity retention curve is proportional to the square-root of cycle number
To investigate the effect of side reaction on to cathode electrode before and after the charging/discharging of the of LIB cell, three LIB samples were subjected to Raman spectral mapping; (1) Pristine, (2) 500_IN sample (after 500 cycle of charge/discharge form interior outer region, and (3) 500_Out sample

Summary

Introduction

Hyperspectral Raman imaging caters the ability to image multiple chemical signatures simultaneously, and the spatial information is collected in the X–Y plane, and the spectral information is represented in the Z-direction, hyperspectral images are represented in the form of data cubes (Fig. 1). For any chosen dataset processed via MCR-ALS, the end-results using such methods are not consistent and resulting in a different number of spectral and concentration profiles. This inconsistency curtails the reliability of the MCR-ALS analysis[2,6]. The AF pipeline to make an analytical model involve; intelligently pre-processing the hyperspectral Raman data with fewer or no human assistance, accurately identifying the reliable spectral signature from the hyperspectral dataset and assign the class labels, training a neural network (NN) on to the accurately “labelled” spectral signature, and testing the reusability of already trained NN to evaluate other test samples in real-time (Fig. 2 for schematic diagram). By averaging all the concentration images and corresponding spectral profiles, within an individual retained cluster, provides a trusted singleton concentration image and spectrum profile

Methods

Results

Conclusion