EarthQuake Seismic Classification

Implementation of paper Research on Seismic Signal Analysis Based on Machine Learning

installation 安装

pip install -r requirements.txt

data splitting 数据划分

train:valid:test 训练:验证:测试 0.63: 0.07:0.3

(13883, 3, 50, 40) (1543, 3, 50, 40) (6612, 3, 50, 40)

2D shape data

选项 arg --shape 0: 6000, 1: 2000, 2: 1000, 3: 500, 4:1600

shape 0: 6000 -> [3, 50, 40]
shape 1: 2000 -> [1, 50, 40]
shape 2: 1000 -> [1, 40, 25]
shape 3: 500 -> [1, 25, 20]
shape 4: 1600 -> [1, 40, 40]

resnet

python resnet_test.py   or python main_all.py --model resnet --shape 0

precision recall F1 score in micro: (0.9774652147610405, 0.9774652147610405, 0.9774652147610405, None)
precision recall F1 score in macro: (0.9742350942479397, 0.9784838179920086, 0.9762684751952175, None)
precision recall F1 score in weighted: (0.9777493925150258, 0.9774652147610405, 0.9775278578108789, None)

30 epochs takes 210.65927863121033 seconds 0:03:30.659279 by GPU

Note: specify 指定 GPU id CUDA_VISIBLE_DEVICES=4 4 is the GPU id you want to use or --gpu-id 4

Note: about the micro and macro, weighted modes: https://scikit-learn.org/stable/modules/generated/sklearn.metrics.precision_recall_fscore_support.html

python main_all.py --gpu-id 0

vgg

python main_all.py --model vgg  --shape 0

precision recall F1 score in micro: (0.9807924984875983, 0.9807924984875983, 0.9807924984875983, None)
precision recall F1 score in macro: (0.9807680111517341, 0.979085444315043, 0.9799133389389202, None)
precision recall F1 score in weighted: (0.9808440645335728, 0.9807924984875983, 0.9808039652414201, None)

30 epoch takes 228 seconds, 0:03:48.908452

inception

python main_all.py --model inception  --shape 0

precision recall F1 score in micro: (0.9640048396854205, 0.9640048396854205, 0.9640048396854205, None)
precision recall F1 score in macro: (0.9675209970511508, 0.9558780226812885, 0.9610584015381325, None)
precision recall F1 score in weighted: (0.9643438850153745, 0.9640048396854205, 0.9636412281604345, None)

30 epoch takes 155 seconds, 0:02:35.260036 by Titan RTX GPU

MiniRocket

python minirocket_pytorch.py --shape 0

precision recall F1 score in micro: (0.9765577737447066, 0.9765577737447066, 0.9765577737447066, None)
precision recall F1 score in macro: (0.9740317458166179, 0.9760858799525116, 0.975037360831401, None)
precision recall F1 score in weighted: (0.9766728008981629, 0.9765577737447066, 0.9765968085618346, None)

30 epoch training time takes: 0:00:32.0839

total time(feature init + 30 epochs training + evaluation) takes 62 seconds, 0:01:02.968496

MiniRocket + KMeans

python minirocket_kmeans.py

Valid (Train data)

precision recall F1 score in micro: (0.72650071308181, 0.72650071308181, 0.72650071308181, None)
precision recall F1 score in macro: (0.768166980467447, 0.7778092485870426, 0.7269018362922858, None)
precision recall F1 score in weighted: (0.8061755802259496, 0.72650071308181, 0.7177761482591168, None)

Test (unseen data)

precision recall F1 score in micro: (0.7280701754385965, 0.7280701754385965, 0.7280701754385965, None)
precision recall F1 score in macro: (0.7688228818383352, 0.7821433192910333, 0.7306720810775338, None)
precision recall F1 score in weighted: (0.8059558883062758, 0.7280701754385965, 0.7192791769292779, None)

total time(feature init + KMeans training + evaluate cluster accuracy) takes 100 seconds, 0:01:40.92

Much slower, because sklearn KMeans uses multi-cores CPUs. So please use servers, instead of your own PC/laptop 不要用自己的电脑，容易卡死

MiniRocket + Gaussian Mixture

python minirocket_gmm.py

Valid(Train data)

precision recall F1 score in micro: (0.7281861791780111, 0.7281861791780111, 0.7281861791780111, None)
precision recall F1 score in macro: (0.772654993160231, 0.7804282970534618, 0.7288347746132154, None)
precision recall F1 score in weighted: (0.8115435180379689, 0.7281861791780111, 0.7195912425526907, None)

Test (unseen data)

precision recall F1 score in micro: (0.6677253478523896, 0.6677253478523896, 0.6677253478523896, None)
precision recall F1 score in macro: (0.4913682236442349, 0.5983473703037866, 0.5340606981630641, None)
precision recall F1 score in weighted: (0.5391397590979354, 0.6677253478523896, 0.5924183550697695, None)

total time(feature init + GM training + evaluate cluster accuracy) takes 341 seconds, 0:05:41.382244

MiniRocket + cluster head

python minirocket_cluster_head.py

precision recall F1 score in micro: (0.8114035087719298, 0.8114035087719298, 0.8114035087719298, None)
precision recall F1 score in macro: (0.8178379533964444, 0.8394204262915687, 0.8132728614856477, None)
precision recall F1 score in weighted: (0.8449325470490134, 0.8114035087719298, 0.813164746070741, None)

Best test clustering accuracy 0.8373 data shape (13883, 1, 6000)(1543, 1, 6000)(6612, 1, 6000) total time(feature init + Cluster Head training + evaluate cluster accuracy) takes 246 seconds, 0:04:06.971629

KMeans only

python kmeans.py

Valid

precision recall F1 score in micro: (0.41443018280824584, 0.41443018280824584, 0.41443018280824584, None)
precision recall F1 score in macro: (0.36466642369961283, 0.35768726352641655, 0.3509105103790023, None)
precision recall F1 score in weighted: (0.39853839805663566, 0.41443018280824584, 0.3952050271484101, None)

Test

precision recall F1 score in micro: (0.40955837870538414, 0.40955837870538414, 0.4095583787053841, None)
precision recall F1 score in macro: (0.3614725644567485, 0.3519464691407293, 0.34555335489579786, None)
precision recall F1 score in weighted: (0.3934958204550063, 0.40955837870538414, 0.3894361854535951, None)

total time(KMeans training + evaluate cluster accuracy) takes 43 seconds, 0:00:43.213827

Gaussian Mixture only

Valid

precision recall F1 score in micro: (0.6155840788279529, 0.6155840788279529, 0.6155840788279529, None)
precision recall F1 score in macro: (0.5325245187452413, 0.5746818320085421, 0.5426358216476126, None)
precision recall F1 score in weighted: (0.5645584996493351, 0.6155840788279529, 0.5809346168689176, None)

Test

precision recall F1 score in micro: (0.7265577737447066, 0.7265577737447066, 0.7265577737447066, None)
precision recall F1 score in macro: (0.5178998258431012, 0.6411639141669475, 0.5700647733106167, None)
precision recall F1 score in weighted: (0.5717435580617377, 0.7265577737447066, 0.6366422396828016, None)

total time(GM training + evaluate cluster accuracy) takes 526 seconds, 0:08:46

python gmm.py

Observations

Visualization

Please refer to minirocket_tsne.ipynb

Normalization

--norm 1  with data normalization
--norm 0  without data normalization

Normalization (in data_util.py) is crucial for resnet_test.py/main_all.py --model (98% -> 50%).

But MiniRocket works well without normalization(supervised: 97.8% -> 97.0%, kmeans: 72.7% -> 68.7%, cluster head: 85.1% -> 80.5%)

Citation

If you found this work useful and used it on your own research, please concider citing this paper.


@Article{xiulongyang20220506,
    author = {Yin, Xinxin and Liu, Feng and Cai, Run and Yang, Xiulong and Zhang, Xiaoyue and Ning, Meiling and Shen, Siyuan},
    title = {Research on Seismic Signal Analysis Based on Machine Learning},
    journal = {Applied Sciences},
    volume = {12},
    year = {2022},
    number = {16},
    article-NUMBER = {8389},
    url = {https://www.mdpi.com/2076-3417/12/16/8389},
    issn = {2076-3417},
    doi = {10.3390/app12168389}
}

sndnyang/LCL-SSS