Sonus – Acoustic Machine Fault Detection

Hackathon project · 4th place out of 28 teams

Sonus listens to an industrial fan and decides whether it sounds healthy or faulty. A small CNN classifies Mel spectrograms of the fan's sound. A desktop dashboard plays the audio, shows the result, and stops the motor through an Arduino when it detects a fault. It then asks Gemini to write a short technical report.

How it works

fan sound (.wav) ─► first channel, 16 kHz, 3 s ─► Mel spectrogram (128 mels, dB)
        ─► resize to 128×128, min-max normalize ─► CNN ─► P(fault)
                                                            │
                         P(fault) > 0.5 ─► stop motor (Arduino "M0") + Gemini report

Model. The CNN has two Conv–ReLU–MaxPool blocks (32 and 64 filters) and two fully connected layers (128 → 1). A sigmoid output gives P(fault). The model has about 8.4M parameters, most of them in the first dense layer. It was trained with BCE loss and the Adam optimizer (lr = 0.001) for 20 epochs, with batch size 16.

Data. Training used a mini subset of the MIMII dataset: the fan machine type at the noisiest setting (−6 dB SNR). The subset has 400 clips, 200 normal and 200 abnormal. Training ran on a Kaggle GPU notebook.

Dashboard. The dashboard (CustomTkinter + Matplotlib) runs a two-step demo:

1. It plays a normal clip, draws its waveform with a moving cursor, and shows the model's score.
2. It plays an abnormal clip. If the model flags it, the app sends M0 over serial to stop the motor and shows a "CRITICAL" state. It then asks Gemini for a short technical analysis.

If no Arduino is connected, the app keeps running in simulation mode.

Results

Metric	Value	Note
Training accuracy (epoch 20)	96.5%	measured on the training data
Training loss (epoch 20)	0.112	BCE
Demo check	6 / 6 correct	3 normal, 3 abnormal clips

Limitation: These numbers are not held-out test results. The training notebook trains on all 400 clips without a separate test split. The demo clips come from the same dataset, so they may also have been seen during training. A proper evaluation is listed under next steps.

Repository structure

File	Purpose
Sonus/EgitimOrtami.ipynb	Training notebook (Kaggle): dataset class, preprocessing, model, training loop
Sonus/Uygulama.ipynb	Desktop dashboard: model inference, Arduino control, Gemini report
Sonus/fan_model_minus6db.pth	Trained model weights (PyTorch state_dict)
Sonus/normal.wav, Sonus/anormal.wav	Example clips from MIMII (fan, −6 dB) used by the demo

Run the dashboard

pip install torch torchaudio librosa numpy matplotlib customtkinter sounddevice soundfile pyserial google-genai

1. Open Sonus/Uygulama.ipynb and set the values in Config:
   • ARDUINO_PORT (default COM10, 9600 baud)
   • API_KEY for Gemini. Keep this out of git, for example by reading it from an environment variable.
2. Run the cell from the Sonus/ folder so that the model and .wav files are found.
3. Press ▶ SENARYOYU BAŞLAT (start the scenario).

Arduino protocol: M1\n turns the motor on, and M0\n turns it off.

Retrain the model

Open Sonus/EgitimOrtami.ipynb on Kaggle and attach the "MIMII mini" dataset (fan, −6 dB). Then run the cells. The notebook saves the model as fan_model_minus6db.pth.

Tech stack

Python · PyTorch · torchaudio · librosa · CustomTkinter · Matplotlib · sounddevice · pySerial · Arduino · Google Gemini API · Kaggle

Next steps

• Split the data into train, validation and test sets, then report test accuracy, precision, recall and AUC.
• Test the model at the other MIMII noise levels and on other machine types.
• Classify live microphone input instead of stored clips.
• Move the dashboard from a notebook into an app.py file and add a requirements.txt.

Dataset and acknowledgements

This project uses audio from the MIMII dataset (Hitachi, Ltd.):

H. Purohit, R. Tanabe, K. Ichige, T. Endo, Y. Nikaido, K. Suefusa, Y. Kawaguchi, "MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection," DCASE 2019 Workshop. https://zenodo.org/record/3384388

Check the dataset license on the Zenodo page before you reuse or redistribute the audio files.