Working-face gas concentration shifts in nonlinear, time-varying ways—threshold alarms alone rarely anticipate trends. Liang et al. (2020) propose bidirectional GRU (BiGRU) with Adamax optimization. Below is a concise summary of the problem, method, and reported results.
Scenario and goals
- Signal: mine gas concentration time series.
- Goal: improve forecast accuracy when data quality is controlled.
- Constraint: models should fit deployment and monitoring/alert pipelines.
Method highlights
1. Data cleaning first
The paper uses a 3σ (Laida) rule for gross errors and Lagrange interpolation for gaps—treating preprocessing as part of modeling, not an afterthought.
2. BiGRU
BiGRU combines forward and backward context; the authors report gains over unidirectional RNN, LSTM, and GRU on their test set.
3. Adamax-BiGRU
MSE loss; Adamax chosen after optimizer comparison; architecture details (e.g. two hidden layers, 32 units) are given in the original study.
4. Reported test-set improvements (their data)
| Baseline | Approx. error reduction | |----------|-------------------------| | RNN | 25.58% | | LSTM | 12.53% | | Unidirectional GRU | 3.01% |
Results are dataset-specific.
Practical takeaways
- Clean series before training.
- Benchmark BiGRU and optimizers on local data.
- Connect forecasts to live monitoring and alerts.
Reference
Liang, R.; Chang, X.; Jia, P.; Xu, C. Mine Gas Concentration Forecasting Model Based on an Optimized BiGRU Network. ACS Omega 2020, 5 (44), 28579–28586.