1 Introduction
1.1 Drawbacks of time-series analysis methods used for artificial lift systems
1.2 Limitations of time-series clustering methods
1.3 A practical approach for streaming time-series analysis of artificial lift systems
2 Overview of Coal Seam Gas production
2.1 Progressive Cavity Pumps
2.2 Data gathering from CSG wells
3 Related work
3.1 Neural net-based anomaly detection
3.2 Neural net-based time-series clustering
3.3 Converting time-series data into performance heatmap images
3.3.1 Expanding window technique
3.3.2 Symbolic aggregation approximation (SAX)-based performance heatmaps for PCPs
3.3.3 Majority and anomaly heatmap images
3.4 Advantages of a human-in-the-loop approach for data labelling
4 Methodology
4.1 I. Auto-encoder-based dimensionality reduction
4.1.1 i Deep auto-encoder (DAE)
Parameter | Settings |
---|---|
Layer 1 | [16, 32] |
Layer 2 | [4, 8, 16] |
Train images | 134, 346 |
Test images | 33, 587 |
Loss function | Binary cross entropy |
Epochs | 100 |
Name | layer1 | layer2 | Loss | val_loss |
---|---|---|---|---|
sweep-6 | 16 | 4 | 0.043735 | 0.057785 |
sweep-5 | 16 | 8 | 0.021559 | 0.023302 |
sweep-4 | 16 | 16 | 0.021424 | 0.02292 |
sweep-3 | 32 | 4 | 0.064341 | 0.093155 |
sweep-2 | 32 | 8 | 0.032088 | 0.032394 |
sweep-1 | 32 | 16 | 0.021083 | 0.023502 |
Parameter | Settings |
---|---|
Layer 1 | [16] |
Layer 2 | [16] |
Layer 3 | [2, 4, 8] |
Train images | 134, 346 |
Test images | 33, 587 |
Loss function | Binary cross entropy |
Epochs | 100 |
Name | layer1 | layer2 | layer3 | loss | val_loss |
---|---|---|---|---|---|
sweep-3 | 16 | 16 | 2 | 0.067277 | 0.074296 |
sweep-2 | 16 | 16 | 4 | 0.057862 | 0.084062 |
sweep-1 | 16 | 16 | 8 | 0.022665 | 0.022379 |
Parameter | Settings |
---|---|
Layer 1 | [16] |
Layer 2 | [16] |
Layer 3 | [8] |
Layer 4 | [2, 4, 8] |
Train images | 134,346 |
Test images | 33,587 |
Loss function | Binary cross entropy |
Epochs | 100 |
Name | layer1 | layer2 | layer3 | layer4 | loss | val_loss |
---|---|---|---|---|---|---|
sweep-3 | 16 | 16 | 8 | 2 | 0.05247 | 0.068188 |
sweep-2 | 16 | 16 | 8 | 4 | 0.056899 | 0.069199 |
sweep-1 | 16 | 16 | 8 | 8 | 0.020568 | 0.022079 |
4.1.2 ii. Convolutional auto-encoder
Parameter | Settings |
---|---|
Layer 1 | [16] |
Layer 2 | [16] |
Layer 3 | [8] |
Layer 4 | [2, 4, 8] |
Train Images | 134, 346 |
Test images | 33, 587 |
Loss function | Binary cross entropy |
Epochs | 100 |
Encoder type | layer1 | layer2 | layer3 | layer4 | loss | val_loss |
---|---|---|---|---|---|---|
DAE | 16 | 16 | 8 | 8 | 0.020568 | 0.022079 |
CAE | 16 | 16 | 8 | 8 | 0.0215 | 0.02042 |
4.2 II. High-density dimensionality reduction
4.2.1 i t-Distributed stochastic neighbour embedding (t-SNE)
4.2.2 ii. Uniform manifold approximation and projection (UMAP)
4.2.3 iii. Minimum-distortion embedding (MDE)
4.3 III. Hierarchical density-based spatial clustering (HDBSCAN)
Parameter | Settings |
---|---|
Wells | [10, 20, 30, 40, 50, 60, 70] |
Cluster size | [5] |
Sample size | [200] |
Dimensionality Reduction | [t-SNE, UMAP, MDE] |
4.3.1 i. Clustering analysis
Parameter | Settings |
---|---|
Wells | [70] |
Cluster size | [2, 5, 10, 15, 25] |
Sample size | [5, 10, 25, 50, 100, 200] |
Dimensionality reduction | [UMAP, MDE] |
Parameter | Settings |
---|---|
Wells | [70] |
Cluster size | [2, 5, 10, 15, 25] |
Sample size | [5, 10, 25, 50, 100, 200] |
Dimensionality reduction | [UMAP] |
Cluster size | Sample size | Number of clusters | Number of outliers |
---|---|---|---|
2 | 5 | 996 | 0 |
15 | 5 | 995 | 0 |
25 | 5 | 994 | 0 |
5 | 5 | 996 | 0 |
10 | 5 | 996 | 0 |
5 | 10 | 996 | 0 |
2 | 10 | 996 | 0 |
10 | 10 | 996 | 0 |
25 | 10 | 994 | 0 |
15 | 10 | 995 | 0 |
10 | 25 | 994 | 0 |
25 | 25 | 993 | 0 |
5 | 25 | 996 | 18 |
15 | 25 | 993 | 0 |
2 | 25 | 996 | 18 |
5 | 50 | 996 | 110 |
2 | 50 | 999 | 116 |
25 | 50 | 989 | 128 |
15 | 50 | 993 | 69 |
10 | 50 | 995 | 115 |
15 | 100 | 991 | 1179 |
25 | 100 | 977 | 1002 |
5 | 100 | 1005 | 1118 |
10 | 100 | 998 | 1096 |
2 | 100 | 1016 | 1136 |
15 | 200 | 1006 | 4774 |
10 | 200 | 1031 | 4541 |
5 | 200 | 1061 | 4414 |
25 | 200 | 972 | 4976 |
2 | 200 | 1078 | 4402 |
4.3.2 ii. Analysing the UMAP and HDBSCAN clusters for Performance Heatmap grouping
4.4 IV Cluster labelling
4.4.1 i. Cluster labelling tool
5 Results
5.1 I. Grouping cluster labels
Major heatmap groups | Anomaly heatmap groups |
---|---|
High torque | High flow |
High high torque | Low flow |
Erratic torque | High torque |
Low flow, low torque | Low torque |
Low low flow | Flow and torque |
Low low flow, low low torque | |
High high flow, high high torque | |
Erratic flow | |
Ideal | |
Shutdown |