deep-performance-rul

Code accompanying the paper “A Deep Learning Approach for Performance-Based Prediction of Remaining Useful Life”.

This repository implements a performance-based RUL prediction framework that preserves an explicit estimation–prediction structure, combining latent health index estimation with particle filter–based uncertainty-aware degradation modeling.

Qualitative Results

Example RUL prediction

Installation

This project uses uv for dependency management.

1. Install uv

macOS (recommended if you use Homebrew):

brew install uv

macOS or Linux (if you do not use Homebrew):

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows (PowerShell):

powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Verify installation:

uv --version

---

2. Clone the repository

git clone https://github.com/4belito/deep-performance-rul.git
cd deep-performance-rul

---

3. Install dependencies

uv sync

---

Running Experiments

The following steps reproduce the results reported in the paper.

---

0. Experiment Configuration

The main experiment settings are defined in experiment_config.py.

The default values correspond to those used in the paper.
To reproduce the reported results, select the desired dataset in this file and keep all other parameters unchanged.

---

1. Data Preparation

Run the notebook:

1-data_preparation.ipynb

Execute it for both data splits:

• data_split = "dev"
• data_split = "test"

Make sure the deep-performance-rul kernel (uv environment) is selected.

---

2. Train Operation Condition Normalization Network

Run:

2-train_ocnorm.ipynb

This trains the operation condition normalization model.

---

3. Apply Operation Condition Normalization (Estimation Step)

Run:

3-estimation_data.ipynb

Execute it for both:

• data_split = "dev"
• data_split = "test"

This step:

• Applies operation condition normalization
• Normalizes performance metrics to the interval [0, 1]
• Removes metrics that do not exhibit a valid degradation pattern
• Removes training units for which none of the performance metrics approach their EOL

The selected performance metrics are displayed before saving the processed data.

---

4. Train Degradation Models (Offline Units)

Run:

4-degmodel_train.ipynb

This trains one degradation model per development unit. These models are used:

• To initialize the particle filter state
• As priors in the particle filter correction step

Optional visualization notebooks:

• 4p-init_states_plots.ipynb — Plots the learned stochastic degradation processes
• 4p-mixture_model_plot.ipynb — Plots the learned mixture model (initial PF prior)

---

5. Train Particle Filter Controller Network

Run:

5-pf_net_train.ipynb

Train the controller network separately for each performance metric selected in Step 3.

Execute the notebook once for each performance metric, for example:

• perform_name = "T48"

Optional:

• 5v-pf_test_video.ipynb — Generates a video of the predictive distribution of the performance metric over time on test data.

---

6. RUL Prediction

Run:

6-rul_test_rep.ipynb

This predicts RUL on the test set for N_REP repetitions
(default: N_REP = 10, configurable in experiment_config.py).

Optional:

• 6p-rul_test_rep_plot.ipynb — RUL prediction plots
• 6v-rul_test_video.ipynb — RUL prediction video

---

7. Evaluation

Run:

7-results.ipynb

This notebook:

• Averages predictions across the N_REP runs
• Computes evaluation metrics in three regions:

1. Full lifetime
2. Last 65 time steps (RUL ≤ 65), for comparison with clipped methods
3. Degradation region, defined by the health-state (hs) variable