CNNIRPLASTICS

This repository now only contains multiple FTIR spectral datasets used for polymer classification and machine learning experiments.

The datasets are primarily focused on the following six commodity plastics:

• PET — Polyethylene Terephthalate
• HDPE — High-Density Polyethylene
• LDPE — Low-Density Polyethylene
• PP — Polypropylene
• PS — Polystyrene
• PVC — Polyvinyl Chloride

---

Dataset Overview

Dataset	Samples	Description
openspecy_polymer_dataset.csv	8390	Extracted from the OpenSpecy spectral library
FTIR_PLASTIC_c4.csv	3000	Balanced FTIR plastics dataset
FTIR_PLASTIC_c8.csv	3000	Balanced FTIR plastics dataset

---

1. OpenSpecy Dataset

File

openspecy_polymer_dataset.csv

Description

This dataset was generated from the OpenSpecy spectral library using custom building scripts.

The dataset contains:

• FTIR and related infrared spectra
• 1983 spectral features per sample
• 6 polymer classes

Each row corresponds to one spectrum.

---

Dataset Shape

8390 rows × 1984 columns

• Columns V1 → V1983: spectral intensity values
• Column label: polymer class label

---

Class Distribution

Polymer	Samples
HDPE	1057
LDPE	867
PET	2267
PP	2451
PS	1491
PVC	257

---

Spectral Features

The 1983 feature columns correspond to spectral intensity values measured across different wavenumbers.

Additional wavenumber metadata is stored separately in:

openspecy_wavenumbers.csv

Each feature index corresponds to a specific FTIR wavenumber.

---

2. FTIR Plastics Dataset

Files

FTIR_PLASTIC_c4.csv
FTIR_PLASTIC_c8.csv

---

Description

These datasets are balanced FTIR polymer datasets containing 1000 samples for each plastic class.

Each sample stores:

• polymer metadata
• spectral wavelength values
• spectral intensity values

---

Dataset Shape

Each dataset contains:

3000 samples

with balanced polymer classes.

---

Columns

Column	Description
IDE	Sample identifier
Polymer	Polymer class label
Technic	Spectroscopy technique
Sample	Sample description
BR	Metadata field
RST	Metadata field
Data(x)	Wavelength / wavenumber values
Data(y)	Spectral intensity values

---

Important Note About Spectral Data

In FTIR spectroscopy:

• Data(x) usually represents the wavenumber axis (cm⁻¹)
• Data(y) represents absorbance, transmittance, or intensity values

The exact interpretation depends on the original acquisition setup.

---

3. Functional Requirements

3.1 Spectral Input Processing

The system shall:

• Accept FTIR spectral data as input.
• Preprocess spectra into a standardized format suitable for machine learning.
• Normalize absorbance values before classification.

---

3.2 Plastic Classification

The system shall:

• Predict the plastic category from a given infrared spectrum.
• Support classification into the six defined plastic classes.
• Output the predicted class and confidence score.

---

3.3 Machine Learning Models

The project shall implement and evaluate:

• A 1D Convolutional Neural Network (1D-CNN)
• A Random Forest classifier

The models shall be compared using:

• Classification accuracy
• Inference speed
• Robustness
• Suitability for real-time deployment

---

3.4 Model Persistence

The system shall:

• Save trained models to disk.
• Reload trained models for future inference without retraining.

---

3.5 Evaluation

The system shall:

• Evaluate models using validation and test datasets.
• Produce accuracy metrics and classification reports.
• Support single-sample testing for inference verification.

---

4. Non-Functional Requirements

4.1 Accuracy

The classifier should achieve high classification accuracy across all supported plastic categories.

(Done) Current experimental results using the 1D-CNN model have achieved near-perfect classification performance on the available dataset.

---

4.2 Performance

The system should support fast inference suitable for real-time or near real-time recycling machine operation.

(Need improvement) Inference time should remain sufficiently low to support conveyor-belt-based plastic sorting systems.

---

4.3 Scalability (maybe)

The architecture should support:

• Addition of new plastic classes
• Integration of future datasets
• Deployment to embedded or edge-computing systems

---

4.4 Reliability (more tests needed)

The system should maintain stable predictions under repeated testing and varying spectra conditions.

---

Suggested Workflow

Recommended pipeline:

1. Train models on:

   • FTIR_PLASTIC_c4.csv
   • FTIR_PLASTIC_c8.csv

2. Evaluate generalization on:

   • openspecy_polymer_dataset.csv

This helps test model robustness across:

• instruments
• preprocessing pipelines
• spectral variability
• real-world samples

---

Notes

• Spectral preprocessing may still be required before training.

• Recommended preprocessing:

  • normalization
  • baseline correction
  • interpolation
  • smoothing
  • noise filtering

• OpenSpecy data contains spectra from multiple sources and instruments, which may introduce distribution shifts.

---

License

Please refer to the original dataset and OpenSpecy licenses for redistribution and usage terms.