pytorch implementation for "Anomalous Sound Detection Using Spectral-Temporal Information Fusion"

The paper is available in link.

Installation

---

sh run.sh or

$ conda create -n stgram_mfn python=3.7
$ conda activate stgram_mfn
$ pip install -r requirements.txt
$ python run.py

Dataset

---

DCASE2020 Task2 Dataset:

• development dataset
• additional training dataset
• Evaluation dataset

data path can be set in config.yaml

Model Weights File

---

Our trained model weights file for loading can be get in https://zenodo.org/record/7194640#.Y0t1WXZBxD8

Result on development dataset

---

machine Type	AUC(%)	pAUC(%)	mAUC(%)
Fan	94.04	88.97	81.39
Pump	91.94	81.75	83.48
Slider	99.55	97.61	98.22
Valve	99.64	98.44	98.83
ToyCar	94.44	87.68	83.07
ToyConveyor	74.57	63.60	64.16
Average	92.36	86.34	84.86

ToyCar		
id	AUC	pAUC
1	0.830719697	0.652198679
2	0.951617251	0.874308413
3	0.995218329	0.981046957
4	0.99993531	0.999659526
Average	0.944372647	0.876803394
ToyConveyor		
id	AUC	pAUC
1	0.869696875	0.766776316
2	0.641580986	0.547164566
3	0.725856264	0.594070052
Average	0.745711375	0.636003645
fan		
id	AUC	pAUC
0	0.950638821	0.894607526
2	0.996852368	0.983433514
4	0.813936782	0.681034483
6	0.999972299	0.999854206
Average	0.940350067	0.889732432
pump		
id	AUC	pAUC
0	0.892342657	0.744939271
2	0.83481982	0.693219535
4	0.9999	0.999473684
6	0.950490196	0.832301342
Average	0.919388168	0.817483458
slider		
id	AUC	pAUC
0	1	1
2	0.982209738	0.906367041
4	0.99988764	0.999408634
6	0.999775281	0.998817268
Average	0.995468165	0.976148236
valve		
id	AUC	pAUC
0	1	1
2	0.988333333	0.952192982
4	1	1
6	0.99725	0.985526316
Average	0.996395833	0.984429825
Total Average	0.923614376	0.863433498

Cite

---

If you think this work is useful to you, please cite:

@INPROCEEDINGS{9747868,
 author={Liu, Youde and Guan, Jian and Zhu, Qiaoxi and Wang, Wenwu},
 booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
 title={Anomalous Sound Detection Using Spectral-Temporal Information Fusion}, 
 year={2022},
 volume={},
 number={},
 pages={816-820},
 doi={10.1109/ICASSP43922.2022.9747868}}