Federated Learning for Distributed CNC Tool Wear Prediction

This paper investigates federated learning for CNC tool wear prediction using the MATWI dataset, which contains sensor recordings and images from 17 cutting-tool sets. The authors formulate tool wear prediction as a distributed learning problem where each client represents an autonomous industrial site that keeps raw data local while participating in collaborative model training. The federated learning setup uses horizontal federated learning with FedAvg aggregation, where a central server coordinates training rounds and clients perform local computation without sharing raw sensor recordings or images.

The methodology includes preprocessing steps to identify active cutting intervals using discrete wavelet decomposition with the Daubechies 4 wavelet, removing air-cut portions that contain limited information about tool condition. An ablation experiment showed that air-cut removal reduces total MAE from 26.24 µm to 16.40 µm, mainly driven by improvements in flank wear prediction. Two pipelines are evaluated: a sensor-based pipeline using a 1D-CNN with sliding windows, and an image-based pipeline using ResNet50 regression. The training data is partitioned across three simulated clients, with validation and test sets remaining unchanged.

The results show that federated learning achieves performance close to centralized learning for both modalities. For the sensor modality, the federated model achieves a total MAE of 18.73 µm compared to 17.90 µm for the centralized model, with comparable performance on flank wear. For the image modality, the federated model achieves a total MAE of 29.97 µm compared to 26.73 µm for the centralized model. The federated model outperforms local client baselines, achieving lower total error than Clients B and C in the sensor modality and lower total error than all three local client models in the image modality.

The authors conclude that federated learning provides a practical compromise between centralized and local client models, supporting collaborative tool wear prediction in distributed CNC manufacturing environments. The approach is privacy-preserving by design since raw operational data remain at client sites. Future research directions include applying more involved aggregation methods like FedProx or personalized federated learning, and exploring multimodal federated learning to combine the complementary strengths of sensor and image data.

Improvements to AI Systems:

1. Domain-Aware Preprocessing Module: Integrate a wavelet-based (Daubechies 4) air-cut detection and removal step into the AI pipeline. This reduces noise from non-informative data segments, improving prediction accuracy (MAE drop from 26.24 µm to 16.40 µm) and enabling the model to focus on active cutting phases.

2. Federated Learning with Modality-Specific Architectures: Deploy a horizontal federated learning framework (FedAvg) that supports both 1D-CNN (for sensor time-series) and ResNet50 (for image regression). This allows collaborative training across distributed industrial sites without centralizing raw data, preserving privacy while achieving near-centralized performance.

3. Adaptive Client Weighting for Heterogeneous Data: Enhance FedAvg by dynamically weighting client contributions based on local data distribution (e.g., cutting-tool wear stages). This reduces the performance gap between federated and centralized models (e.g., sensor MAE: 18.73 µm vs 17.90 µm) and prevents domination by clients with more uniform wear patterns.

4. Cross-Modal Federated Fusion (Future-Enabled): Design a multimodal federated aggregation layer that combines latent features from sensor and image pipelines at the server side (e.g., via knowledge distillation or feature alignment). This leverages complementary strengths—sensor data for temporal dynamics, images for visual wear—without sharing raw data, improving robustness for tools with complex wear modes.

5. Personalized Federated Fine-Tuning: Replace global model averaging with a two-stage approach: train a global model via FedAvg, then fine-tune per-client with local data (e.g., using FedProx or local adaptation layers). This improves performance for clients with unique tool geometries or operating conditions, as seen in the image modality where federated (29.97 µm) still lags centralized (26.73 µm).

What the Improved AI System Can Do:

• Predict CNC tool wear (flank and crater) in real-time with high accuracy (MAE < 19 µm for sensor data) across multiple factories, while keeping proprietary sensor recordings and images on-premises.

• Automatically ignore air-cut segments, reducing computational load and false alarms during idle machine periods.

• Adapt to new clients (e.g., a new CNC machine) with minimal data transfer, using personalized fine-tuning to maintain performance without retraining from scratch.

• Provide a unified dashboard that fuses sensor and image predictions, flagging tools for maintenance earlier than single-modality systems, especially when one modality is noisy or unavailable.

• Operate under bandwidth constraints by transmitting only model updates (gradients/weights), not raw data, enabling deployment in low-connectivity industrial environments.

Tool wear prediction is an important task in CNC machining, where accurate monitoring of tool condition supports product quality and process reliability. Machine learning methods have shown potential for this task, but their use in industrial environments is limited by the distributed nature of machining data and by restrictions on data sharing between machines, sites, or organizations. Federated learning offers a suitable framework for this setting by enabling collaborative model training without transferring raw operational data. This paper investigates federated learning for CNC tool wear prediction. Tool trajectories are distributed across simulated clients to represent a federated learning scenario. The federated models are compared against centralized references and local client baselines. Results show that federated learning achieves performance close to centralized learning and improves significantly over local client models. These findings indicate that federated learning can support collaborative tool wear prediction in distributed CNC manufacturing environments.