SyedaArisha/predictive-maintenance-rag-system

Predictive Maintenance RAG System

Author: Syeda Arisha Hassan Domain: FMCG & Industrial Manufacturing Stack: XGBoost · Random Forest · PyTorch LSTM · 1D-CNN-LSTM · FAISS · SmolLM2-135M

Overview

An end-to-end, closed-loop predictive maintenance platform built for FMCG and continuous industrial manufacturing. The system integrates four ML/DL pipelines with a Retrieval-Augmented Generation (RAG) layer to deliver actionable, plain-language diagnostics and automated production rescheduling.

Unexpected machine breakdowns in FMCG plants cost $10,000 to $36,000 per hour (McKinsey). This system addresses that by predicting failures before they happen and automatically reallocating production jobs to healthy machines.

Repository Contents

models/
├── classification/
│   ├── ai4i_ensemble_model.pkl        # XGBoost + RF soft-voting ensemble (AI4I 2020)
│   └── pump_ensemble_model.pkl        # XGBoost + RF soft-voting ensemble (Pump Sensor)
├── regression/
│   ├── best_cmapss_lstm_checkpoint.pt     # 2-Layer PyTorch LSTM (NASA CMAPSS FD001)
│   └── best_cmapss_cnn-lstm_checkpoint.pt # 1D-CNN-LSTM Regressor (NASA CMAPSS FD001)
└── rag/
    ├── faiss_index.bin                # FAISS IndexFlatL2 — 7,966 vectors × 384-d
    └── faiss_texts.pkl                # Synthesized maintenance log corpus

System Architecture

[ Sensor Telemetry ] → [ ML Classifiers ] → [ Risk Score ]
                     → [ DL RUL Regressor ] → [ Hours Remaining ]
                     → [ FAISS Retrieval ] → [ Historical Context ]
                     → [ LLM Explainer (SmolLM2) ] → [ Diagnostic Report ]
                     → [ Production Scheduler ] → [ Adjusted Job Schedule ]

Pipelines & Datasets

1. AI4I 2020 — Milling Machine Classification

• —Dataset: 10,000 records, 14 features (UCI Machine Learning Repository)
• —Task: Binary failure classification (3.4% class imbalance)
• —Approach: Optuna-tuned XGBoost + Random Forest, Soft Voting Ensemble, SMOTE oversampling
• —Results:
• —ROC-AUC: 0.9726
• —Recall: 0.87 (87% of true failures detected)
• —Confusion Matrix: [[1892, 40], [9, 59]]

2. Pump Sensor — Temporal Early Warning

• —Dataset: 220,320 minute-interval readings, 52 sensors (April–August 2018)
• —Task: 60-minute lookahead pre-failure warning classification
• —Approach: Strict temporal 50/50 train/test split, TimeSeriesSplit Optuna tuning, RECOVERING rows dropped to eliminate target leakage
• —Model: models/classification/pump_ensemble_model.pkl

3. NASA CMAPSS FD001 — Turbofan Engine RUL Regression

• —Dataset: 100 training engines, 100 test engines, 26 sensor channels
• —Task: Remaining Useful Life (RUL) regression in cycles
• —Approach: 3D sliding window sequences (N × 30 × 24), MinMaxScaler, piecewise RUL capping at 125 cycles
• —Architectures:
• —2-Layer LSTM: Hidden sizes 64→32, LayerNorm, early stopping (patience=10)
• —1D-CNN-LSTM: Conv1D spatial extraction + LSTM temporal recurrence
• —Results: MAE ~14.2–16.8 cycles, RMSE ~18.9–22.4 cycles
• —Checkpoints: models/regression/

4. Microsoft Azure PdM — Fleet RAG Database

• —Dataset: 876k hourly telemetry rows across 100 machines + error/maintenance/failure logs
• —Approach: 7,966 natural language maintenance records synthesized, embedded with all-MiniLM-L6-v2 (384-d), indexed in FAISS
• —Index: models/rag/faiss_index.bin (12.2 MB, exact L2 search)
• —LLM Explainer: HuggingFaceTB/SmolLM2-135M-Instruct running on CPU with deterministic anti-parroting post-processor

Risk Status Thresholds

How to Load the Models

import joblib
import torch

# Load classification ensemble
ensemble = joblib.load("models/classification/ai4i_ensemble_model.pkl")
scaler   = joblib.load("data/processed/ai4i_scaler.pkl")

# Load LSTM checkpoint
checkpoint = torch.load("models/regression/best_cmapss_lstm_checkpoint.pt",
                        map_location="cpu")

# Load FAISS index
import faiss, pickle
index = faiss.read_index("models/rag/faiss_index.bin")
with open("models/rag/faiss_texts.pkl", "rb") as f:
    texts = pickle.load(f)

Serving Dashboard

The full system is served via a FastAPI backend with an interactive web dashboard featuring:

• —Real-time failure probability gauge and RUL countdown per machine
• —Fleet health overview across all 100 machines
• —LLM-generated Markdown diagnostic reports (RAG + SmolLM2)
• —Automated production rescheduling with job reallocation audit log

pip install -r requirements.txt
python -m uvicorn app:app --host 127.0.0.1 --port 8000

Citation

If you use these models or datasets in your research, please credit the original dataset sources:

• —AI4I 2020: UCI Machine Learning Repository
• —NASA CMAPSS: NASA Prognostics Center of Excellence
• —Microsoft Azure PdM: Microsoft Azure AI Gallery