Drilling in the dark? The AI that 'feels' the rock while drilling

Imagine digging a tunnel hundreds of meters underground, in total darkness. How do you know if the drill bit is chewing through soft coal, hard sandstone, or crumbly mudstone without seeing it? Until recently, we relied on human experience or slow lab samples. Today, a study published in *Applied Sciences* changes the game: just "listen" to the drill's vibrations and pressure.

Researchers took raw data from anchor cable drilling — the steel cables holding up tunnel roofs — and cleaned it with a "mathematical sponge" called wavelet transform, which removes background noise. Then they trained four artificial "brains" (LSTM and BiLSTM neural networks) to recognize the rock. The real champion? A hybrid model boosted by the **DBO** (Dung Beetle Optimizer) algorithm, inspired by the dung beetle rolling dung balls with geometric efficiency.

Result: on coal, easy to distinguish, all models score 100%. But on sandstone and mudstone, which look very similar, **DBO-BiLSTM** pulls ahead: +9% accuracy, +12% recall, +13% F1 score. Basically, the system "feels" rock hardness in real-time while drilling (MWD technology). This means safer tunnels, less waste, and precise engineering even in complex geology. A win for miner safety and smart engineering.

The study addresses a classic *Measurement While Drilling* (MWD) challenge: the raw drilling signal is inherently noisy, non-stationary, and non-linear. The methodological pipeline consists of three phases. First, **data cleaning** to remove outliers and missing values caused by spurious mechanical vibrations. Second, **Wavelet Transform (WT)** for denoising: multi-resolution decomposition isolates high-frequency components (instrumental noise) from low-frequency ones (geological signal), applying thresholding on detail coefficients to reconstruct the "clean" signal of torque, pressure, rotation speed, and feed rate.

Third, **Deep Learning modeling**. Recurrent architectures (LSTM/BiLSTM) are chosen for their ability to capture long temporal dependencies in sequential drilling data. The Bidirectional LSTM processes the sequence forward and backward, capturing past and future context. The innovation lies in hyper-optimization via **Dung Beetle Optimizer (DBO)**, a *swarm-based* metaheuristic algorithm mimicking the dung beetle's ball-rolling navigation, breeding dance, and foraging. DBO avoids local minima better than PSO or GA, finding the optimal combination of learning rate, hidden units, batch size, dropout, and window length.

Validation on field datasets (strata: coal, sandstone, mudstone, limestone) uses standard metrics. The confusion matrix reveals that "difficult" classes (sandstone/mudstone) benefit most from bidirectionality combined with global optimization. Implementation in embedded MWD systems now requires model quantization (e.g., TensorRT, ONNX) for real-time inference on edge hardware (industrial FPGA/GPU), ensuring latency < 100ms per drilled meter.