Intelligent Surface Roughness Detection Using Deep Learning for Deep Hole Drilling

Intelligent Surface Roughness Detection Using Deep Learning

Surface roughness measurement in deep hole drilling traditionally requires stopping production, withdrawing the tool, and running a profilometer or comparator down the bore — a time-consuming process that only samples a fraction of production. Recent advances in deep learning have made it possible to detect surface roughness in real-time during drilling by analyzing sensor signals.

A landmark 2025 study published in Measurement (ScienceDirect) introduced a Pyramid Adaptive Transformer (PAT) model that fuses multi-sensor vibration data to detect surface roughness in BTA deep hole drilling with 99.43% accuracy — a significant advance in in-process quality monitoring.

The Challenge: Measuring Roughness in Deep Holes

Why It’s Hard

Traditional Methods

The Deep Learning Solution

Approach

The researchers mounted three vibration sensors on the BTA machine — at the workpiece holder, the drill tube support, and the machine base. The vibration signals contain information about the cutting process that correlates with surface roughness: higher frequency vibrations indicate better surface finish (fine chip formation); lower frequency vibrations indicate rougher surfaces (tool chatter, chip breakage).

Multi-Sensor Fusion Architecture

The Pyramid Adaptive Transformer (PAT) model processes the data in stages:

Sensor 1 (workpiece)  ┐
Sensor 2 (tube support)┤→ MSST (Multi-Scale Synchrosqueezing Transform)
Sensor 3 (machine base)┘    → Time-frequency representation
                              → Pyramid Adaptive Transformer
                                → Multi-scale feature extraction
                                  → Attention-weighted fusion
                                    → Roughness prediction (Ra)

MSST: Multi-Scale Synchrosqueezing Transform

MSST converts the raw vibration signals into time-frequency images — similar to how a spectrogram represents audio. It provides higher resolution than conventional Fourier or wavelet transforms by “squeezing” the energy in each frequency band around the instantaneous frequency.

Pyramid Adaptive Transformer

The PAT architecture processes the MSST time-frequency images at multiple scales — similar to how a human eye examines a scene at different resolutions. It then applies self-attention mechanisms to identify which frequency bands and time segments are most predictive of surface roughness.

Performance Results

Comparison with Other Models

Parameter Efficiency

PAT achieves higher accuracy than a standard Transformer with 60% fewer parameters — critical for deployment on edge computing hardware near the machine, rather than requiring cloud processing.

Practical Implementation

Hardware Requirements

Installation

1. Mount accelerometers at workpiece holder, drill tube support, machine base
2. Connect to data acquisition module
3. Deploy trained PAT model on edge computer
4. Connect to machine control for alarm output

Data Requirements for Model Training

Applications

Best Use Cases

Limitations

Summary

Deep learning-based surface roughness detection using the Pyramid Adaptive Transformer with MSST multi-sensor fusion achieves 99.43% accuracy in BTA deep hole drilling — the first application of deep learning to in-process roughness monitoring for deep hole drilling. The model uses 62% fewer parameters than standard Transformers, making it suitable for edge deployment. With inference time under 50 ms, PAT enables real-time quality monitoring during drilling, potentially replacing post-process inspection for high-volume production. For AI monitoring concepts, see AI-based process monitoring for single-part production. For traditional measurement methods, see deep hole measurement methods guide.