krishx009/mechsense-engine-whisperer

MechSense AI — Engine Whisperer

Bearing fault detection from smartphone microphone audio — 4 seconds, 8 fault classes.

A seasoned mechanic takes 20 minutes with a stethoscope. This model does it in 4 seconds from outside the bonnet.

Model Description

Engine Whisperer is a convolutional neural network that classifies bearing faults from mel-spectrogram representations of engine audio. It was designed to run on-device via ONNX Runtime — the primary deployment target is a smartphone microphone, making professional-grade bearing diagnostics accessible without any specialist hardware.

The model is the core of MechSense AI, a final-year engineering capstone project combining bearing fault detection with driving style analysis.

Model Architecture

Input: (1, 1, 128, 87) mel-spectrogram ↓ 4× Conv Block (Conv2D → BatchNorm → ReLU → MaxPool → Dropout2D) Channels: 1 → 32 → 64 → 128 → 256 ↓ Adaptive Global Average Pool → (256,) ↓ FC(256 → 128) → ReLU → Dropout(0.4) → FC(128 → 8) ↓ Output: (1, 8) logits

• Parameters: 1,206,568
• Input: 2-second audio window at 22,050 Hz → 128-band mel-spectrogram
• Framework: PyTorch 2.6, exported to ONNX opset 17
• Inference: ONNX Runtime CPU (< 5ms per 2s window)

Training Strategy

The model was trained in two phases:

Phase 1 — Base training (CWRU + MFPT + IMS) Trained from scratch on three standard bearing fault benchmark datasets using:

• File-level GroupShuffleSplit to prevent data leakage
• WeightedRandomSampler for class balance
• SpecAugment (time + frequency masking) + Gaussian noise augmentation
• Cosine annealing LR schedule
• Result: 99.1% validation accuracy (file-level split, no leakage)

Phase 2 — Cross-domain fine-tuning (HUST + Paderborn) Fine-tuned on two additional datasets from independent test rigs with:

• Frozen first 2 encoder blocks (preserve low-level frequency features)
• Differential learning rates (backbone: 5×10⁻⁵, head: 5×10⁻⁴)
• Extra augmentation on minority classes (intra-class mixup + pitch shift)
• Result: 88.6% cross-domain validation accuracy

Fault Classes

ID	Class	Description	Recommended Action
0	healthy	No bearing fault detected	Regular servicing
1	inner_race_fault	Inner race defect	Inspect within 1,000 km
2	inner_race_fault	Inner race defect (severe)	Workshop within 200 km
3	outer_race_fault	Outer race defect	Reduce speed, inspect soon
4	outer_race_fault	Outer race defect (severe)	Immediate inspection
5	ball_fault	Ball element defect	Monitor, next service
6	ball_fault	Ball element defect (severe)	Workshop within 300 km
7	degradation	Run-to-failure progressive wear	Plan bearing replacement

Training Datasets

Dataset	Source	Bearings	SR	Fault Types
CWRU	Case Western Reserve University	6204	12 kHz	IR, OR, Ball (3 diameters × 4 loads)
MFPT	Machinery Failure Prevention Technology	Multiple	97.6 kHz	IR, OR (varying loads)
IMS	NASA Ames PCOE	6203	20 kHz	Run-to-failure (natural degradation)
HUST	Hanoi University of Science & Technology	6204–6208 (5 types)	51.2 kHz	IR, OR, Ball (3 load levels)
Paderborn	University of Paderborn KAt-DataCenter	6203	64 kHz	IR, OR, real + artificial damage

All datasets are publicly available for academic use. Links in the Dataset Sources section.

Performance

Cross-domain Validation (HUST + Paderborn, 956 windows, file-level split)

Class	Precision	Recall	F1	Support
healthy	0.924	1.000	0.961	171
inner_race_fault	0.833	0.906	0.868	315
outer_race_fault	0.936	0.996	0.965	304
ball_fault	0.812	0.963	0.881	27
degradation	1.000	1.000	1.000	60
Overall			0.886	956

Note: Confidence level serves as a severity indicator — predictions above 85% confidence on a fault class suggest severe-stage faults requiring urgent attention.

Usage

Quick Start (Python + ONNX Runtime)

import numpy as np
import librosa
import onnxruntime as ort

CLASS_NAMES = [
    "healthy", "inner_race_fault", "inner_race_fault",
    "outer_race_fault", "outer_race_fault",
    "ball_fault", "ball_fault", "degradation"
]

def preprocess_audio(audio_path: str, target_sr: int = 22050):
    """Load audio and convert to mel-spectrogram windows."""
    audio, sr = librosa.load(audio_path, sr=target_sr, mono=True)
    
    # Normalize
    peak = np.max(np.abs(audio))
    if peak > 0:
        audio = audio / peak
    
    # 2-second windows with 0.5s hop
    window_samples = int(target_sr * 2.0)
    hop_samples    = int(target_sr * 0.5)
    windows = []
    for start in range(0, len(audio) - window_samples + 1, hop_samples):
        w = audio[start:start + window_samples]
        mel = librosa.feature.melspectrogram(
            y=w, sr=target_sr, n_mels=128,
            n_fft=2048, hop_length=512
        )
        mel_db = librosa.power_to_db(mel, ref=np.max)
        mn, mx = mel_db.min(), mel_db.max()
        if mx - mn > 0:
            mel_db = (mel_db - mn) / (mx - mn)
        windows.append(mel_db.astype(np.float32))
    
    return np.array(windows)[:, np.newaxis, :, :]   # (N, 1, 128, 87)


def predict(audio_path: str, model_path: str = "engine_whisperer_v2.onnx"):
    sess    = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
    specs   = preprocess_audio(audio_path)
    logits  = sess.run(None, {"spectrogram": specs})[0]          # (N, 8)
    
    # Softmax + average over windows
    exp     = np.exp(logits - logits.max(axis=-1, keepdims=True))
    probs   = (exp / exp.sum(axis=-1, keepdims=True)).mean(axis=0)  # (8,)
    
    pred    = CLASS_NAMES[probs.argmax()]
    conf    = float(probs.max())
    return pred, conf, {CLASS_NAMES[i]: float(probs[i]) for i in range(8)}


fault_class, confidence, all_probs = predict("engine_audio.wav")
print(f"Fault: {fault_class}  |  Confidence: {confidence:.1%}")

Using from HuggingFace Hub

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="YOUR_USERNAME/mechsense-engine-whisperer",
    filename="engine_whisperer_v2.onnx"
)
fault_class, confidence, all_probs = predict(model_path)

Files

File	Description	Size
engine_whisperer_v2.onnx	Main inference model (ONNX opset 17)	~5 MB
engine_whisperer_v2_best.pt	PyTorch checkpoint (for fine-tuning)	~5 MB
config.json	Class names, input specs, thresholds	< 1 KB

Dataset Sources

• CWRU: engineering.case.edu/bearingdatacenter
• MFPT: mfpt.org/fault-data-sets
• IMS/NASA: ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository
• HUST: data.mendeley.com/datasets/cbv7jyx4p9/3
• Paderborn: mb.uni-paderborn.de/kat/forschung/bearing-datacenter

Citation

If you use this model in your research, please cite:

@misc{mechsense2026,
  author    = {Krishna Chandana Giri},
  title     = {MechSense AI: Smartphone-based Bearing Fault Detection and Driving Style Analysis},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/YOUR_USERNAME/mechsense-engine-whisperer}
}

Project Links

• GitHub: github.com/YOUR_USERNAME/mechsense-ai
• Driving DNA Model: huggingface.co/YOUR_USERNAME/mechsense-driving-dna
• Live Demo: huggingface.co/spaces/YOUR_USERNAME/mechsense-demo

License

MIT License. Training datasets retain their original licenses — please refer to each dataset's terms before commercial use.