SensorDBSCAN: Semi-Supervised Learning for Fault Diagnosis in Chemical Processes

This repository is based on code from SensorSCAN: Self-Supervised Learning and Deep Clustering for Fault Diagnosis in Chemical Processes. We've been accepted to IEEE Access with this method: Paper

Our modifications

• Changed second learning stage from Deep Clustering to semi-supervised encoder training with triplet loss (to make clusters separable) with DBSCAN applied later (so we don't have to specify number of clusters and could use triplet loss margin hyperparameter as epsilon for DBSCAN)
• Added active learning for semi-supervised learning (data sampling for labeling and effective triplets ranking for training). Made it effective: ~700 labeled samples is enough for a decent performance on a 2M dataset
• Added intermediate clusteing quality evaluation mechanism to estimate moment when we should sample more data
• Changed fixed positional embeddings (as is origianl Transformer paper) to RoPE embeddings
• Developed outlier handling technique (density-based clustering algorithms may produce outlier cluster)
• (Wasn't included in the IEEE Access paper) Developed a more complex and therefore more accurate sequence pooling mechanism

Requirements

To install dependencies, run

conda create -n sensordbscan --file environment.yml

Running experiments

Running experiments involves training and evaluating processes defined for each model in models folder. To run experiments, use the command line interface python main.py with the argument --config-name {model}_{dataset}. For example:

python main.py --config-name pca_kmeans_rieth_tep

Pretrained models are stored in saved_models folder. To evaluate a pretrained model, use path_to_model argument:

python main.py --config-name pca_kmeans_rieth_tep path_to_model=saved_models/pca_kmeans_rieth_tep.joblib

Experimental results are stored in results folder. All arguments are defined in configs folder.

Experimental results

Metrics in this section is fixed to date of IEEE Access paper submission. Some further experiments and improvements made our algorithm achieve better results

Results on rieth_tep

	ACC	ARI	NMI	Detection TPR	Detection FPR	CDR	ADD
PCA+KMeans	0.2745	0.1100	0.3634	0.3590	0.0000	0.7910	113.95
ST-CATGAN	0.1754	0.1135	0.2223	0.3044	0.0000	0.3238	102.63
ConvAE	0.1794	0.1565	0.2537	0.3631	0.0008	0.3664	164.76
SensorSCAN	0.5926	0.4747	0.6812	0.7316	0.0014	0.7351	57.15
SensorSCAN (paper)	0.7850	0.7030	0.8460	0.8400	0.0002	0.9200	5.21
SensorDBSCAN	0.7393	0.7124	0.8029	0.6877	0.0000	0.6913	121.42

Results on reinartz_tep

	ACC	ARI	NMI	Detection TPR	Detection FPR	CDR	ADD
PCA+KMeans	0.3513	0.1316	0.4484	0.3581	0.0000	0.9562	113.33
ST-CATGAN	0.3016	0.1287	0.3606	0.3627	0.0001	0.8882	135.04
ConvAE	0.4975	0.2381	0.5863	0.6023	0.0016	0.9402	155.16
SensorSCAN	0.5287	0.3336	0.7551	0.9013	0.0002	0.7219	30.98
SensorSCAN (paper)	0.7360	0.4810	0.8500	0.8700	0.0002	0.9600	28.47
SensorDBSCAN (paper)	0.7379	0.5410	0.8154	0.7808	0.0000	0.8433	164.91
SensorDBSCAN (best)	0.8297	0.6706	0.8800	0.8517	0.0009	0.9064	36.59