physics-constrained-wear-nn: Tool Wear Prediction Using Physics-Constrained Neural Networks

physics-constrained-wear-nn is a physics-constrained neural network project for CNC turning scenarios, with the core goal of predicting the flank wear (VB) of cutting tools using force sensor signals. The project's innovations include: using architecture-level monotonicity guarantees (shifted Softplus activation function) to ensure the wear curve is non-decreasing, phase-aware loss functions adapted to the three stages of tool wear (initial, stable, and rapid), and a two-stage pipeline (signal processing + model training) to solve the data matching problem between sparse annotations and dense signals. This system ensures the physical consistency of prediction results without post-processing, providing a reliable solution for industrial predictive maintenance.

In CNC turning, the flank wear of cutting tools directly affects machining quality, efficiency, and cost. Traditional wear monitoring faces core challenges: wear measurement is mostly discrete manual operations with irregular time intervals, while force signals (Fx, Fy, Fz) during machining are continuously collected at high frequency, leading to a data density mismatch problem. How to learn wear patterns from dense force signals while relying only on sparse annotation data is the key problem this project aims to solve.

The project adopts a two-stage pipeline solution:

1. Signal Processing Stage: Identify effective cutting periods using algorithms like threshold segmentation and gap bridging, extract 9-dimensional statistical features (mean, peak value, standard deviation, etc.) from the three-axis force signals, resulting in a total of 27-dimensional features;
2. Model Training Stage: Use an MLP architecture, with input as 27-dimensional features + cumulative cutting length, and the output layer uses a shifted Softplus activation function to ensure non-negative wear increments (architecture-level monotonicity); the loss function is a composite form, including MSE, incremental supervision term, and shape constraint losses for the three wear stages (logarithmic for initial, linear for stable, exponential for rapid). Additionally, the project predicts wear increments (ΔVB) instead of absolute wear values to avoid cumulative errors and improve numerical stability.

The experiment compared the performance of multiple models:

Model	Test R²	Test MAE (µm)	Test RMSE (µm)	Monotonicity Guarantee
XGBoost (with cutting length)	0.922	12.89	17.49	Post-processing only
Monotonic NN (Softplus, no phase)	0.864	18.51	22.87	Architecture-level
Phase-aware NN (this project)	0.864	18.51	22.88	Architecture-level

Analysis: Tree models (e.g., XGBoost) have higher accuracy, but the model of this project ensures physical consistency through architecture-level design without post-processing; the phase-aware loss currently has limited effect but reserves space for future expansion.

Engineering Highlights:

• Modular design: signal_processing.py can be reused independently;
• Flexible data partitioning: supports partitioning by tool sample, coating system, and experiment batch;
• Interpretability: integrates SHAP analysis for feature importance;
• Experiment tracking: automatically generates prediction results, hyperparameters, training curves, etc.

Application Scenarios: Intelligent tool management (predicting remaining life), machining parameter optimization, quality control (abnormal wear warning). Expansion Directions: Adaptation to multiple tool types, online learning, uncertainty quantification, digital twin integration.

This project demonstrates the successful application of physics-informed machine learning in industrial predictive maintenance. The core methodological values include:

1. Embedding physical laws (monotonicity) into the model architecture to achieve end-to-end training;
2. Reconstructing the problem as incremental prediction to reduce learning difficulty and maintain interpretability;
3. Balancing accuracy and physical rationality to meet industrial application needs. In the context of intelligent manufacturing, such physics-aware models will play an important role in smart factories and provide a reference paradigm for time-series prediction problems.