Federated Learning for Distributed CNC Tool Wear Prediction

# Federated Learning for Distributed CNC Tool Wear Prediction
Source: [https://arxiv.org/html/2608.11281](https://arxiv.org/html/2608.11281)
11institutetext:Maastricht University, Maastricht, Netherlands
11email:a\.khan@maastrichtuniversity\.nlMorris StallmannMarcin PietrasikCharis KouzinopoulosAnna Wilbik

###### Abstract

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\.

## 1Introduction

Cutting tools are crucial in machining processes such as milling, drilling, or sawing to guarantee product quality\. Even though significant effort is put into engineering high\-quality tools\[[11](https://arxiv.org/html/2608.11281#bib.bib15)\], even the best ones are affected by wear due to abrasion, adhesion, and diffusion during the machining process and eventually break or reach a state where the produced parts become scrap\[[14](https://arxiv.org/html/2608.11281#bib.bib14)\]\. Therefore, it is important to recognize the wear state to plan a timely replacement of the tool\.

Several recent studies have shown that machine learning \(ML\) methods achieve accurate wear state predictions\[[18](https://arxiv.org/html/2608.11281#bib.bib3)\],\[[17](https://arxiv.org/html/2608.11281#bib.bib2)\]\. While training reliable and generalizableMLmodels often depend on the availability of high\-quality and sufficiently diverse data, collecting such a dataset may be challenging for a single manufacturer or production cell, but it may become feasible if multiple parties collaborate and combine their data\. In a realistic CNC manufacturing environment, machining data may be distributed across machines, production cells, factories, or organizations\. Centralizing such data can be costly because sensor recordings and images are large, continuously generated, and tied to operational processes\. It can also be undesirable because tool wear data can reveal information about production conditions, machining strategies, material properties, or process quality\.Federated Learning \(FL\)is a machine learning paradigm that allows collaborators to jointly train a machine learning model without having to share raw data instances and while keeping local data private\. As such, it is a natural fit for environments with distributed data and data sharing constraints such as CNC manufacturing, as several recent works have noticed \(Section[2](https://arxiv.org/html/2608.11281#S2)\)\.

Although these works show promising results, they leave open whetherFLactually provides benefits over siloed local client training under the assumption of realistic data processing challenges in distributed systems \(Section[2](https://arxiv.org/html/2608.11281#S2)\)\. In this work, we propose to use the MATWI dataset\[[3](https://arxiv.org/html/2608.11281#bib.bib16)\]in combination with two novelFLsystems to study this question\. Each system is composed of modules for data preprocessing to harmonize disparate local datasets, learning algorithms to identify local patterns, and an averaging mechanism to derive the global model \(Section[3](https://arxiv.org/html/2608.11281#S3)\), and is benchmarked against local client training\. The MATWI dataset contains images and measurements of accelerometer and acoustic sensors along with the wear labels of 17 sets of cutting tools \(Section[3\.1](https://arxiv.org/html/2608.11281#S3.SS1)\)\. The data is unprocessed, sufficiently large to simulate a federated setting, contains multiple modalities, and is naturally distributed across different machines\.

In Section[2](https://arxiv.org/html/2608.11281#S2), we introduce the federated learning framework and related works that apply it to learn tool wear prediction models\. Section[3](https://arxiv.org/html/2608.11281#S3)formulates the distributed learning problem of this work and describes how to solve it usingFL\. In Section[4](https://arxiv.org/html/2608.11281#S4), Section[5](https://arxiv.org/html/2608.11281#S5)Section[6](https://arxiv.org/html/2608.11281#S6), the evaluation on the MATWI dataset is described\. Limitations of this work and potential future research directions are discussed in Section[7](https://arxiv.org/html/2608.11281#S7)\. Lastly, the work is concluded in Section[8](https://arxiv.org/html/2608.11281#S8)\.

## 2Related Work

### 2\.1Federated Learning

InFL, multiple*clients*collaboratively train aMLmodel to utilize distributed data while preserving privacy\. Each client holds a local dataset that must remain private, that is, it cannot be shared with any other entity\. In centralizedFL, the training protocol can be described in terms of five steps\[[5](https://arxiv.org/html/2608.11281#bib.bib13)\]:

1. 1\.The orchestrator \(central server\) selects the clients that participate in the next round of training\.
2. 2\.The central server shares the global model with the selected clients\.
3. 3\.The clients optimize a local learning objective on their local data and share the local updates with the central server\.
4. 4\.The central server aggregates all local updates to a global update and applies it to derive the new global model\.
5. 5\.The updated global model is shared with the clients\. If a convergence criterion is met, the training ends\. If not, the training process continues from the first step\.

FLsettings differ in how data is partitioned \(vertically or horizontally\) and how the learning process is orchestrated \(centralized or decentralized\)\. Horizontally split data is partitioned by sample, and vertically split data by feature\. In centralizedFL, the model training is orchestrated by a central server, whereas clients communicate directly with each other in the decentralized case\. We consider centralized, horizontalFederated Learningin this work\.

Common challenges inFederated Learningstem from the distributed nature of the data and include communication overhead, system heterogeneity, and data heterogeneity\[[16](https://arxiv.org/html/2608.11281#bib.bib6)\]\. Data heterogeneity, or non\-IIDness, inFLoccurs when the clients’ local data follow different distributions, which can impact training convergence and model performance\. Approaches addressing the key challenge of data heterogeneity often apply variants of the standard aggregation algorithm FedAvg \(see Section[3](https://arxiv.org/html/2608.11281#S3)\), group compatible clients for focused model updates, or apply data augmentation techniques\[[20](https://arxiv.org/html/2608.11281#bib.bib1)\]\. Despite its challenges, the academic literature demonstrates the applicability ofFLto various application domains, including health care\[[13](https://arxiv.org/html/2608.11281#bib.bib7)\], financial service security\[[7](https://arxiv.org/html/2608.11281#bib.bib4)\], and predictive maintenance\[[15](https://arxiv.org/html/2608.11281#bib.bib5)\]\.

### 2\.2Tool Wear Prediction with Federated Learning

FLhas attracted increasing research attention for tool wear prediction in industrial environments\. Using sensory information, including accelerometers, vibration, or acoustic emission, the authors of\[[6](https://arxiv.org/html/2608.11281#bib.bib9)\]formulate tool wear as a forecasting problem\. They propose a combination of feature calculation techniques and a federated \(Bi\)LSTM architecture to estimate the target value and demonstrate promising results on the PHM dataset\[[8](https://arxiv.org/html/2608.11281#bib.bib10)\]\. However, it remains open whether the federated approach is beneficial compared to local\-only training since the work focused on proposing a novel federated method\. In\[[4](https://arxiv.org/html/2608.11281#bib.bib12)\], a new federated aggregation algorithm addressing client data heterogeneity is introduced and evaluated on the tool wear dataset introduced in\[[9](https://arxiv.org/html/2608.11281#bib.bib11)\]\. Since the focus of the work is on the aggregation algorithm, no emphasis is put on data processing challenges in distributed environments\. In fact,\[[9](https://arxiv.org/html/2608.11281#bib.bib11)\]reports that data normalization is applied to raw data, a step that may be hard to replicate in a distributed, privacy\-preserving setting\. A tool wear prediction approach using images is introduced in\[[2](https://arxiv.org/html/2608.11281#bib.bib8)\]\. The authors observe thatFLmodels can produce better results than per\-client training, thereby motivating its application in tool wear prediction\. However, they also acknowledge the limited generality of their results given the small scale of their experimental data\.

In summary, several recent works demonstrate the applicability ofFederated Learningto tool wear prediction in the machining industry\. However, these works either neglect data preprocessing challenges inherent to distributed data settings or do not prove the benefits of training aFLmodel justify the increased complexity of such an implementation\.

## 3Methodology

This paper formulates tool wear prediction as a distributed learning problem in which industrial data are generated and retained by separate data owners\. The proposed setting therefore treats each client as an autonomous industrial site that keeps its raw data local while participating in collaborative model training\.

The considered setting corresponds to horizontal federated learning\. In horizontal federated learning, clients share the same feature space and learning task but hold different samples\. This matches the tool wear scenario because each client trains the same type of prediction model for a given modality, while the samples are distributed across different tools or sites\.

Federated learning provides the distributed coordination mechanism\. Instead of moving raw data to a central location, the model is sent to the data\. A central server coordinates training rounds, while clients perform local computation using their own data\. The server never observes raw sensor recordings or images\. It only receives model parameters from the clients and aggregates them into a new global model\. This design reduces the need for data centralization and supports data locality by construction\. The approach is privacy\-preserving by design, in that raw operational data remain at the client sites\.

LetKKdenote the number of clients\. Each clientkkowns a local dataset𝒟k\\mathcal\{D\}\_\{k\}containingnkn\_\{k\}samples, and letn=∑k=1Knkn=\\sum\_\{k=1\}^\{K\}n\_\{k\}be the total number of samples across all clients\. For a sample\(x,y\)∈𝒟k\(x,y\)\\in\\mathcal\{D\}\_\{k\}, wherexxis the model input andyyis the corresponding tool wear label, letℓ​\(θ;x,y\)\\ell\(\\theta;x,y\)denote the loss for model parametersθ\\theta\. The local objective for clientkkis

ℒk​\(θ\)=1nk​∑\(x,y\)∈𝒟kℓ​\(θ;x,y\)\.\\mathcal\{L\}\_\{k\}\(\\theta\)=\\frac\{1\}\{n\_\{k\}\}\\sum\_\{\(x,y\)\\in\\mathcal\{D\}\_\{k\}\}\\ell\(\\theta;x,y\)\.The global objective is to learn a single model that minimizes the sample\-weighted average of the local client objectives:

ℒ​\(θ\)=∑k=1Knkn​ℒk​\(θ\),\\mathcal\{L\}\(\\theta\)=\\sum\_\{k=1\}^\{K\}\\frac\{n\_\{k\}\}\{n\}\\mathcal\{L\}\_\{k\}\(\\theta\),and the target model is

θ∗=arg⁡minθ⁡ℒ​\(θ\)\.\\theta^\{\*\}=\\arg\\min\_\{\\theta\}\\mathcal\{L\}\(\\theta\)\.
Training follows a client\-server protocol based on synchronous communication rounds\. At roundtt, the server broadcasts the current global parametersθt\\theta\_\{t\}to all participating clients\. Each client initializes its local model withθt\\theta\_\{t\}, performs local optimization on𝒟k\\mathcal\{D\}\_\{k\}, and returns updated parametersθtk\\theta\_\{t\}^\{k\}to the server\. The server then applies FedAvg aggregation\[[12](https://arxiv.org/html/2608.11281#bib.bib20)\]:

θt\+1=∑k=1Knkn​θtk\.\\theta\_\{t\+1\}=\\sum\_\{k=1\}^\{K\}\\frac\{n\_\{k\}\}\{n\}\\theta\_\{t\}^\{k\}\.
This aggregation provides clients with more local samples with proportionally greater influence while maintaining a single shared global model\. From a distributed systems perspective, each round consists of global synchronization, parallel local computation, communication of model updates, and server\-side aggregation\. The cost of collaboration is therefore expressed through communication rounds and exchanged model parameters rather than through raw data transfer\.

The distributed training protocol is independent of modality\. The same client\-server coordination, local training, communication, and aggregation procedure can be applied to different data modalities, provided that the clients use a compatible model architecture for the selected modality\. In this paper, the protocol is instantiated for sensor data and image data\. The model architecture differs by modality, but the system\-level protocol remains unchanged\.

![Refer to caption](https://arxiv.org/html/2608.11281v1/architecture.png)Figure 1:Federated Learning Architecture for Tool Wear Prediction### 3\.1The MATWI Dataset

The Multimodal Automatic Tool Wear Inspection \(MATWI\) dataset was introduced by De Pauw et al\.\[[3](https://arxiv.org/html/2608.11281#bib.bib16)\]to support the development of automated tool\-wear monitoring methods for CNC milling\. It contains data from 17 cutting\-tool sets, with each set representing the wear progression of one physical milling insert over its operational lifetime\. For each machining cycle, MATWI provides labeled wear observations together with synchronized process\-sensor recordings and an image of the cutting edge\. The sensor data include cutting\-force measurements along three axes, accelerometer signals, and acoustic measurements, while the visual modality consists of images captured after machining to document the condition of the insert\.

Each sample is associated with quantitative wear labels and wear\-type annotations\. The wear categories include flank wear, adhesive wear, and combined flank\-and\-adhesive wear\. In\[[3](https://arxiv.org/html/2608.11281#bib.bib16)\], sets 1–13 were used for the main baseline image\-based experiments\. Sets 14–17 were excluded because they were collected while machining a different workpiece material, which resulted in increased adhesive wear and a visibly different tool appearance\. This created a domain shift relative to the earlier sets\. Following the same protocol, Sets 1–13 are used for the benchmarking experiments\. Each sample is associated with quantitative wear labels and wear\-type annotations\. The wear categories include flank wear, adhesive wear, and combined flank\-and\-adhesive wear\.

The structure of MATWI is suitable for federated learning because each set represents the complete wear trajectory of a single physical tool\. In practical manufacturing environments, comparable data may be generated and retained locally by different machines, production cells, or sites rather than centrally pooled\. To emulate this setting, tool sets are allocated across federated clients, while all sensor recordings and images associated with a particular tool remain at the same client\. Clients collaboratively train a shared model through model updates without exchanging raw machining data\. This partitioning also prevents tool\-level data leakage and preserves natural heterogeneity across tools, including differences in wear progression and wear type\.

### 3\.2Preprocessing

The sensor recordings contain both air\-cut intervals and active\-cutting intervals\. Air cuts occur when the tool follows the machining path without engaging the workpiece and therefore does not remove material\. These portions contain limited information about tool condition, whereas active cutting intervals capture the interaction between the tool and the workpiece and contain patterns relevant to wear prediction\.

To identify the active cutting interval, a level four discrete wavelet decomposition\[[10](https://arxiv.org/html/2608.11281#bib.bib17)\]with the Daubechies 4 wavelet\[[1](https://arxiv.org/html/2608.11281#bib.bib18)\]is applied to the selected sensor channels\. For a sensor signalx​\[n\]x\[n\], the decomposition is represented as

DWT​\(x​\[n\]\)=A4,D4,D3,D2,D1,\\mathrm\{DWT\}\(x\[n\]\)=\{A\_\{4\},D\_\{4\},D\_\{3\},D\_\{2\},D\_\{1\}\},whereA4A\_\{4\}denotes the approximation coefficients andDjD\_\{j\}denotes the detail coefficients at leveljj\. The magnitude of the finest scale detail coefficients is used as an activity measure:

E​\[k\]=\|D1​\[k\]\|\.E\[k\]=\|D\_\{1\}\[k\]\|\.
A baseline is estimated from the first 1000 values ofE​\[k\]E\[k\]\. LetμE\\mu\_\{E\}andσE\\sigma\_\{E\}denote the mean and standard deviation of this baseline\. The normalized activity score is calculated as:

z​\[k\]=E​\[k\]−μEσE\+10−9\.z\[k\]=\\frac\{E\[k\]\-\\mu\_\{E\}\}\{\\sigma\_\{E\}\+10^\{\-9\}\}\.
After the scores are aligned with the original signal length, samples satisfyingz​\[n\]\>10z\[n\]\>10are considered active\. The first and last active samples define the preliminary cutting boundaries\. The detected interval is expanded by 5

The procedure is performed independently for the two selected sensor channels\. Given detected intervals\[s1,e1\]\[s\_\{1\},e\_\{1\}\]and\[s2,e2\]\[s\_\{2\},e\_\{2\}\], the final retained interval is defined as:

s=max⁡\(s1,s2\),e=min⁡\(e1,e2\)\.s=\\max\(s\_\{1\},s\_\{2\}\),\\qquad e=\\min\(e\_\{1\},e\_\{2\}\)\.
When fewer than 1000 active samples are detected for a channel, the complete recording is retained\. Only the resulting active cutting segments are used in the experiments \(Figure[2](https://arxiv.org/html/2608.11281#S3.F2)\)\.

![Refer to caption](https://arxiv.org/html/2608.11281v1/remove_aircut.png)Figure 2:Sensor recordings for machine cycle on Set 1 before and after air cut removal\. The left column shows the complete machining cycle, with the retained active cutting interval highlighted\. The right column shows the corresponding retained signals for the force, accelerometer, and acoustic channels\.To evaluate the effect of air\-cut removal, an ablation experiment was conducted for the model trained on sensor data\. The model was compared under two conditions: using the complete sensor signals and using only the detected active cutting intervals\. As shown in Figure[3](https://arxiv.org/html/2608.11281#S3.F3), air\-cut removal reduces the total MAE from26\.24​μ​m26\.24~\\mu mto16\.40​μ​m16\.40~\\mu m\. The improvement is mainly driven by flank wear, where the MAE decreases from26\.02​μ​m26\.02~\\mu mto11\.36​μ​m11\.36~\\mu m\. Since flank wear is the dominant wear type in the test set, this improvement has a strong effect on the total result\. Adhesion and combined flank wear & adhesion do not show the same improvement, which may be due to their smaller number of samples present in the dataset and the fact that adhesion is a surface material sticking phenomenon that is less consistently reflected in sensor signals\. Overall, the ablation supports the use of air\-cut removal because it focuses the sensor model on the active cutting region and improves the total sensor modality performance\.

![Refer to caption](https://arxiv.org/html/2608.11281v1/aircut_ablation_sensor.png)Figure 3:Effect of air\-cut removal on performance of the model trained on sensor data\. The model architecture is a 1D\-CNN, and the reported metric is MAE inμ\\mum for per wear type and in total\.

## 4Experimental Setup

A centralized baseline is used as the reference setting\. In this setting, all training data are pooled at one location, and a single model is trained\. Following the MATWI dataset paper protocol\[[3](https://arxiv.org/html/2608.11281#bib.bib16)\], Sets 1, 2, 5, 7, 8, 10, and 11 are used for training, Sets 3, 6, and 12 for validation, and Sets 4, 9, and 13 for testing\. The split is performed at the tool set level, so no cuts from the same physical tool appear in more than one split\.

### 4\.1Prediction Task

The task is supervised regression, where the model predicts the continuous tool wear value for each machining cycle\. Two pipelines are evaluated: a sensor\-based pipeline using process signals recorded during machining, and an image\-based pipeline using cutting edge images captured after machining\. Both pipelines are evaluated under centralized, local, and federated training\.

All models are trained with Huber loss, which is used to reduce sensitivity to occasional large prediction errors while retaining smooth regression behaviour for small errors:

ℒδ​\(r\)=\{12​r2,\|r\|≤δ,δ​\(\|r\|−12​δ\),\|r\|\>δ\.\\mathcal\{L\}\_\{\\delta\}\(r\)=\\begin\{cases\}\\frac\{1\}\{2\}r^\{2\},&\|r\|\\leq\\delta,\\\\ \\delta\\left\(\|r\|\-\\frac\{1\}\{2\}\\delta\\right\),&\|r\|\>\\delta\.\\end\{cases\}wherer=y−y^r=y\-\\hat\{y\}, andδ=1\.0\\delta=1\.0\.

### 4\.2Sensor Pipeline

The full active cutting signal is not used as a single model input because it is long and contains time\-varying patterns across the machining cycle\. Yang et al\.\[[19](https://arxiv.org/html/2608.11281#bib.bib19)\]used local segmentation of tool wear sensor signals to capture local signal characteristics before modeling the overall wear condition\. Following the same motivation, each active cutting signal is divided into sliding windows\. The window size is 2048 samples, and the stride is 1024 samples\. Each window inherits the cut\-level wear label\. During inference, window predictions from the same cut are averaged to obtain one cut\-level prediction\.

The sensor model is a 1D\-CNN\. Training settings are Adam optimizer, learning rate10−410^\{\-4\}, batch size 32, and Huber loss withδ=1\.0\\delta=1\.0\. In addition to the windowed sensor input, the model receives a context vector formed by concatenating the previous wear value with a one\-hot encoding of the wear type\. Since the wear type has three categories, flank wear, adhesive wear, and combined flank and adhesive wear, the resulting context vector has dimension 4\. Early stopping with patience 10 is used for centralized, local, and federated training\. In federated sensor training, the maximum number of communication rounds is 100, with 5 local epochs per round\.

### 4\.3Image Pipeline

The image pipeline uses a ResNet50 regression model trained on cutting edge images\. Training settings follow the original MATWI baseline: Adam optimizer, learning rate3×10−43\\times 10^\{\-4\}, batch size 16, and Huber loss withδ=1\.0\\delta=1\.0\. Images are cropped using the per set crop coordinates from the dataset metadata, resized to224×224224\\times 224pixels using Lanczos resampling, and normalized with ImageNet statistics: mean\[0\.485,0\.456,0\.406\]\[0\.485,0\.456,0\.406\], standard deviation\[0\.229,0\.224,0\.225\]\[0\.229,0\.224,0\.225\]\.

Early stopping with patience 10 is used for centralized, local, and federated image training\. In federated image training, the maximum number of communication rounds is 30, with 5 local epochs per round\.

### 4\.4Federated Training and Evaluation

To simulate a distributed industrial environment, the training sets are partitioned across three clients, each representing an independent manufacturing site\. Client A receives Sets 1, 5, and 7; Client B receives Sets 2 and 10; and Client C receives Sets 8 and 11\. The validation and test sets remain unchanged across all experiments\. In addition to centralized training, a local client baseline is evaluated, where each client trains a separate model using only its own assigned data\. The federated setting uses the same client partitions, but the clients collaboratively train a shared model by exchanging model updates\.

The comparison is designed to assess whether federated training improves over the local baseline while approaching the centralized baseline\. Federated learning is implemented with Flower 1\.30\.0 using FedAvg\. In each round, all three clients participate\. The server sends the global model to the clients, clients train locally, and the server aggregates the returned parameters using a sample\-weighted average\.

Final evaluation is performed on test Sets 4, 9, and 13\. Results are reported using MAE and the standard deviation of absolute errors per wear category, following the original MATWI evaluation format\. All models are implemented in PyTorch and trained on the Snellius HPC cluster using a single NVIDIA H100 GPU\. A fixed seed of 777 is used for Python, NumPy, PyTorch, CUDA, and DataLoader shuffling\.

## 5Results

This section presents the performance of the centralized, federated, and local client models on the held\-out test sets, Sets 4, 9, and 13\. Performance is reported using mean absolute error \(MAE\) measured in micrometers\. The centralized model represents the pooled\-data reference, the local client models represent isolated training without collaboration, and the federated model represents collaborative training without raw data sharing\.

### 5\.1Sensor Modality

![Refer to caption](https://arxiv.org/html/2608.11281v1/centralised_sensor_eval.png)\(a\)Centralized model
![Refer to caption](https://arxiv.org/html/2608.11281v1/federated_sensor_eval.png)\(b\)Federated model

Figure 4:Performance of centralised and federated models \(Sensor Modality\)Figure[4](https://arxiv.org/html/2608.11281#S5.F4)shows the performance of the models trained on the sensor modality\. For flank wear, which is the most represented and practically important wear type in this evaluation, the two models show comparable performance\. For adhesion, both models show larger deviations from the reference trend, with the federated model having a higher error than the centralized model\. For combined flank wear and adhesion, the federated model performs slightly better than the centralized model according to the MAE reported in Table[2](https://arxiv.org/html/2608.11281#S5.T2)\. The local client models provide the isolated training baselines for the sensor modality\. Compared with these baselines \(Table[2](https://arxiv.org/html/2608.11281#S5.T2)\), the federated model achieves lower total error than Clients B and C and remains close to Client A\. This shows that federated training provides a competitive distributed model while allowing the clients to keep their raw operational data local\.

### 5\.2Image Modality

![Refer to caption](https://arxiv.org/html/2608.11281v1/centralised_image.png)\(a\)Centralized model
![Refer to caption](https://arxiv.org/html/2608.11281v1/federated_image.png)\(b\)Federated model

Figure 5:Performance of centralised and federated models \(Image Modality\)Figure[5](https://arxiv.org/html/2608.11281#S5.F5)shows the performance of the models trained on the image modality\. The federated model has performance close to the centralized model as well, indicating that the distributed training protocol produces results close to the pooled\-data reference for this modality\.

From Table[2](https://arxiv.org/html/2608.11281#S5.T2)we observe that the federated model remains close to the centralized reference in total MAE\. The flank wear type shows similar performance between the two models\. The federated model achieves lower error for adhesion, while the centralized model performs better for combined flank wear and adhesion\. The local client models provide the isolated training baselines for the image modality \(Table[2](https://arxiv.org/html/2608.11281#S5.T2)\)\. The federated model achieves lower total error than all three local client models\. This shows that collaborative learning across clients is more effective than training each client independently\. This improvement is achieved without exchanging raw operational data between clients or with the server\.

Table 1:Tool wear prediction performance of centralized, federated, and local client models using sensor modality\.Wear TypeCentralizedModelMAE \(μ\\mum\)FederatedModelMAE \(μ\\mum\)Client AModeltextbfMAE \(μ\\mum\)Client BModeltextbfMAE \(μ\\mum\)Client CModeltextbfMAE \(μ\\mum\)Adhesion58\.36±55\.7458\.36\\pm 55\.7470\.75±52\.0570\.75\\pm 52\.0556\.26±63\.40\\mathbf\{56\.26\\pm 63\.40\}57\.62±40\.1457\.62\\pm 40\.1462\.33±55\.7862\.33\\pm 55\.78Flank Wear11\.36±12\.04\\mathbf\{11\.36\\pm 12\.04\}12\.50±12\.5812\.50\\pm 12\.5813\.67±13\.3913\.67\\pm 13\.3927\.56±27\.8427\.56\\pm 27\.8416\.63±20\.8416\.63\\pm 20\.84Flank Wear & Adhesion23\.66±18\.3723\.66\\pm 18\.3722\.96±9\.70\\mathbf\{22\.96\\pm 9\.70\}30\.27±19\.0330\.27\\pm 19\.0349\.45±25\.6449\.45\\pm 25\.6462\.52±26\.3262\.52\\pm 26\.32Total16\.40±22\.24\\mathbf\{16\.40\\pm 22\.24\}17\.90±22\.3217\.90\\pm 22\.3219\.21±23\.9519\.21\\pm 23\.9533\.28±30\.2133\.28\\pm 30\.2127\.58±32\.0227\.58\\pm 32\.02
Table 2:Tool wear prediction performance of centralized, federated, and local client models using image modality\.Wear TypeCentralizedModelMAE \(μ\\mum\)FederatedModelMAE \(μ\\mum\)Client AModeltextbfMAE \(μ\\mum\)Client BModeltextbfMAE \(μ\\mum\)Client CModeltextbfMAE \(μ\\mum\)Adhesion35\.19±40\.4335\.19\\pm 40\.4326\.73±44\.21\\mathbf\{26\.73\\pm 44\.21\}39\.17±52\.7039\.17\\pm 52\.7036\.39±40\.0636\.39\\pm 40\.0639\.60±36\.8239\.60\\pm 36\.82Flank Wear13\.64±11\.20\\mathbf\{13\.64\\pm 11\.20\}14\.05±13\.9814\.05\\pm 13\.9828\.15±21\.7328\.15\\pm 21\.7331\.43±23\.7931\.43\\pm 23\.7940\.65±20\.4140\.65\\pm 20\.41Flank Wear & Adhesion31\.08±21\.42\\mathbf\{31\.08\\pm 21\.42\}36\.00±31\.3836\.00\\pm 31\.3834\.58±31\.8234\.58\\pm 31\.8253\.61±30\.3153\.61\\pm 30\.31108\.62±169\.62108\.62\\pm 169\.62Total18\.06±18\.42\\mathbf\{18\.06\\pm 18\.42\}18\.73±22\.7318\.73\\pm 22\.7329\.97±26\.8329\.97\\pm 26\.8335\.68±27\.6235\.68\\pm 27\.6252\.69±78\.7652\.69\\pm 78\.76

## 6Discussion

The results show that federated learning provides a practical compromise between centralized and local client models for CNC tool wear prediction\. The centralized model represents the pooled\-data reference, while the federated model allows clients to contribute to a shared model through parameter updates\. The comparison with local client models shows whether collaboration through federated learning is more effective than keeping each client model isolated\.

For the sensor modality, the federated model remains close to the centralized model\. This is mainly driven by the flank wear results, where the two models have very similar errors\. Since flank wear is the largest wear type in the test set, similar performance on this wear type has a strong effect on the overall result\. For combined flank wear and adhesion, the federated model slightly improves over the centralized model, while for adhesion the federated model has a higher error\. This difference is expected because adhesion can be more difficult to infer from process signals alone\. Sensor signals capture indirect effects of wear during cutting, whereas adhesion refers to material sticking to the tool surface and may not always produce a clear or consistent signal pattern\.

For the image modality, the federated model also remains close to the centralized model\. The difference between the two models is small, showing that the federated protocol also works effectively when the input modality is visual\. The federated model performs better for adhesion, while the centralized model performs better for flank wear and combined flank wear and adhesion\. The stronger adhesion result in the image modality is plausible because adhesion is a visual surface phenomenon, as it occurs when workpiece material sticks to the cutting edge, making it easier to observe directly in images than through sensor measurements\.

The comparison with local client models shows the benefit of federated collaboration\. A local client model is built only from the data available at one client, so it reflects a narrower subset of the available tool trajectories\. In the image modality, the federated model achieves lower total error than all three local client models\. In the sensor modality, the federated model achieves lower total error than Clients B and C and remains close to Client A\. These results show that federated learning provides a stronger alternative to isolated local models by allowing information from multiple clients to be combined through model aggregation\. This improvement is achieved without exchanging raw sensor recordings or images between clients or with the server\.

The remaining gap between federated and centralized models reflects the constraints of federated optimization\. The centralized model has access to pooled data and can optimize directly over the combined training set\. The federated model, in contrast, depends on client\-side updates followed by server\-side aggregation\. Differences in client data size, wear progression, and wear type composition can influence the updates received by the server\. As a result, the aggregated model may not exactly match the centralized reference, even when all clients participate in each communication round\.

The wear type results also show that performance is not equally stable across all labels\. Flank wear forms the largest part of the train and test set and gives the most consistent results across centralized and federated models\. Adhesion and combined flank wear and adhesion contain fewer samples, so their MAE values are more affected by individual prediction errors\. This is visible in both modalities, where the smaller wear types show larger variation than flank wear\. Therefore, differences for adhesion and combined wear should be interpreted together with the number of available test samples\.

From a federated systems perspective, the results support the use of data\-local collaborative learning for industrial monitoring\. Each client performs computation locally, while the server coordinates communication rounds and aggregates model parameters\. The server does not receive raw operational data, and the clients do not need to exchange datasets with one another\. This makes the approach relevant for industrial environments where machining data may be distributed across machines, production cells, factories, or organizations, and where direct data pooling may be costly, restricted, or undesirable\.

## 7Limitations and Future Directions

The federated learning setup supports privacy by design because raw operational data remain at the client sites and are not transferred to the server or other clients\. Stronger privacy protection would require additional mechanisms such as secure aggregation or differential privacy, since model updates may still contain information about local data\. The evaluation is based on a controlled federated simulation using a benchmark dataset\. In larger deployments with more clients, the data distribution may become more heterogeneous, especially when clients differ in machine type, cutting conditions, workpiece materials, tooling, data volume, or wear progression\. Such non\-IID behaviour can make federated optimization more difficult and may affect the stability of FedAvg\. In these cases, alternative aggregation and optimization strategies, such as FedProx or personalized federated learning, may become useful for handling stronger client heterogeneity\. The results also show that the two modalities contribute differently across wear types\. The sensor modality gives strong performance for flank wear, where process signals capture the cutting behaviour associated with progressive tool degradation\. The image modality is more effective for adhesion, which is a visual surface phenomenon caused by material sticking to the cutting edge\. This suggests that sensor and image data provide complementary information\. A promising direction is therefore multimodal federated learning, where both modalities are used jointly within the federated framework to combine their advantages while still keeping raw operational data local to each client\.

## 8Conclusion

In this work, we study the applicability of federated learning to tool wear prediction with real\-world, naturally distributed CNC milling data containing multiple modalities in scenarios where data cannot be centralized\. The problem is formulated to fit into theFLparadigm and solved by preprocessing to harmonize local data distributions and by a federated training protocol using FedAvg\. Through experimentation on the MATWI dataset, it is demonstrated that the federated method outperforms local\-only models, therefore justifying the increased complexity of federated learning compared to local\-only machine learning\. Future research can focus on improving prediction accuracy through the application of more involved aggregation methods or by applying multimodal federated learning to learn a single model from multiple, complementary modalities\.

## References

- \[1\]I\. Daubechies\(1990\)The wavelet transform, time\-frequency localization and signal analysis\.IEEE transactions on information theory36\(5\),pp\. 961–1005\.Cited by:[§3\.2](https://arxiv.org/html/2608.11281#S3.SS2.p2.1)\.
- \[2\]G\. L\. de Melo Rosa, A\. Gwose, P\. Ganser, and T\. Bergs\(2026\)Security\-enhanced cutting tool wear segmentation with federated learning\.Procedia CIRP138,pp\. 415–420\.Note:18th CIRP Conference on Intelligent Computation in Manufacturing EngineeringExternal Links:ISSN 2212\-8271,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.procir.2026.01.072),[Link](https://www.sciencedirect.com/science/article/pii/S2212827126000727)Cited by:[§2\.2](https://arxiv.org/html/2608.11281#S2.SS2.p1.1)\.
- \[3\]L\. De Pauw, T\. Jacobs, and T\. Goedemé\(2023\)MATWI: a multimodal automatic tool wear inspection dataset and baseline algorithms\.InInternational Conference on Computer Vision Systems,pp\. 255–269\.Cited by:[§1](https://arxiv.org/html/2608.11281#S1.p3.1),[§3\.1](https://arxiv.org/html/2608.11281#S3.SS1.p1.1),[§3\.1](https://arxiv.org/html/2608.11281#S3.SS1.p2.1),[§4](https://arxiv.org/html/2608.11281#S4.p1.1)\.
- \[4\]G\. Huang and C\. Lee\(2024\)Industrial federated learning algorithm \(p\-pfedsgd\) for tool wear estimation\.Future Generation Computer Systems158,pp\. 150–157\.External Links:ISSN 0167\-739X,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.future.2024.04.026),[Link](https://www.sciencedirect.com/science/article/pii/S0167739X24001547)Cited by:[§2\.2](https://arxiv.org/html/2608.11281#S2.SS2.p1.1)\.
- \[5\]P\. Kairouz, H\. B\. McMahan, B\. Avent, A\. Bellet, M\. Bennis, A\. Nitin Bhagoji, K\. Bonawitz, Z\. Charles, G\. Cormode, R\. Cummings, R\. G\. L\. D’Oliveira, H\. Eichner, S\. El Rouayheb, D\. Evans, J\. Gardner, Z\. Garrett, A\. Gascón, B\. Ghazi, P\. B\. Gibbons, M\. Gruteser, Z\. Harchaoui, C\. He, L\. He, Z\. Huo, B\. Hutchinson, J\. Hsu, M\. Jaggi, T\. Javidi, G\. Joshi, M\. Khodak, J\. Konecný, A\. Korolova, F\. Koushanfar, S\. Koyejo, T\. Lepoint, Y\. Liu, P\. Mittal, M\. Mohri, R\. Nock, A\. Özgür, R\. Pagh, H\. Qi, D\. Ramage, R\. Raskar, M\. Raykova, D\. Song, W\. Song, S\. U\. Stich, Z\. Sun, A\. T\. Suresh, F\. Tramèr, P\. Vepakomma, J\. Wang, L\. Xiong, Z\. Xu, Q\. Yang, F\. X\. Yu, H\. Yu, and S\. Zhao\(2021\-06\)Advances and open problems in federated learning\.Found\. Trends Mach\. Learn\.14\(1–2\),pp\. 1–210\.External Links:ISSN 1935\-8237,[Link](https://doi.org/10.1561/2200000083),[Document](https://dx.doi.org/10.1561/2200000083)Cited by:[§2\.1](https://arxiv.org/html/2608.11281#S2.SS1.p1.1)\.
- \[6\]I\. S\. Kaleli, P\. Unal, B\. U\. Deveci, O\. Albayrak, and A\. M\. Ozbayoglu\(2024\)A domain\-aware federated learning study for cnc tool wear estimation\.InMobile Web and Intelligent Information Systems,M\. Younas, I\. Awan, D\. Petcu, and B\. Feng \(Eds\.\),Cham,pp\. 250–265\.Cited by:[§2\.2](https://arxiv.org/html/2608.11281#S2.SS2.p1.1)\.
- \[7\]C\. H\. Kennedy, A\. Hilal, and M\. Momeni\(2025\)The role of federated learning in improving financial security: a survey\.In2025 IEEE Global Conference on Artificial Intelligence and Internet of Things \(GCAIoT\),Vol\.,pp\. 1–8\.External Links:[Document](https://dx.doi.org/10.1109/GCAIoT68269.2025.11275535)Cited by:[§2\.1](https://arxiv.org/html/2608.11281#S2.SS1.p3.1)\.
- \[8\]Cited by:[§2\.2](https://arxiv.org/html/2608.11281#S2.SS2.p1.1)\.
- \[9\]Y\. Lin, C\. Lee, and M\. Lu\(2022\)Robust tool wear monitoring system development by sensors and feature fusion\.Asian Journal of Control24\(3\),pp\. 1005–1021\.External Links:[Document](https://dx.doi.org/https%3A//doi.org/10.1002/asjc.2741),[Link](https://onlinelibrary.wiley.com/doi/abs/10.1002/asjc.2741),https://onlinelibrary\.wiley\.com/doi/pdf/10\.1002/asjc\.2741Cited by:[§2\.2](https://arxiv.org/html/2608.11281#S2.SS2.p1.1)\.
- \[10\]S\. G\. Mallat\(1989\)A theory for multiresolution signal decomposition: the wavelet representation\.IEEE transactions on pattern analysis and machine intelligence11\(7\),pp\. 674–693\.Cited by:[§3\.2](https://arxiv.org/html/2608.11281#S3.SS2.p2.1)\.
- \[11\]P\. Mativenga, J\. Schoop, I\.S\. Jawahir, D\. Biermann, M\. Kipp, Z\. M\. Kilic, T\. Özel, R\. Wertheim, P\. Arrazola, and D\. Boing\(2024\)Engineered design of cutting tool material, geometry, and coating for optimal performance and customized applications: a review\.CIRP Journal of Manufacturing Science and Technology52,pp\. 212–228\.External Links:ISSN 1755\-5817,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.cirpj.2024.06.001),[Link](https://www.sciencedirect.com/science/article/pii/S1755581724000798)Cited by:[§1](https://arxiv.org/html/2608.11281#S1.p1.1)\.
- \[12\]S\. Mehta and A\. Aneja\(2024\)Securing data privacy in machine learning: the fedavg of federated learning approach\.In2024 4th Asian Conference on Innovation in Technology \(ASIANCON\),pp\. 1–5\.Cited by:[§3](https://arxiv.org/html/2608.11281#S3.p8.5)\.
- \[13\]A\. Noor, U\. M\. Chaman, and M\. S\. Kabir\(2026\)Federated learning in healthcare: a comprehensive survey on privacy, scalability and clinical applications\.ICT Express\.External Links:ISSN 2405\-9595,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.icte.2026.05.011),[Link](https://www.sciencedirect.com/science/article/pii/S2405959526000834)Cited by:[§2\.1](https://arxiv.org/html/2608.11281#S2.SS1.p3.1)\.
- \[14\]D\. Patel and S\. Muthuswamy\(2026\-05\)Tool wear state recognition in CNC milling using reliable labels, robust domain features, and lightweight AI models\.J\. Intell\. Manuf\.37\(5\),pp\. 2129–2164\(en\)\.Cited by:[§1](https://arxiv.org/html/2608.11281#S1.p1.1)\.
- \[15\]A\. A\. Purkayastha and S\. Aggarwal\(2024\-08\)Federated Learning for Predictive Maintenance: A Survey of Methods, Applications, and Challenges\.In2024 IEEE 67th International Midwest Symposium on Circuits and Systems \(MWSCAS\),Springfield, MA, USA,pp\. 238–242\.External Links:ISBN 979\-8\-3503\-8717\-9,[Link](https://ieeexplore.ieee.org/document/10658900/),[Document](https://dx.doi.org/10.1109/MWSCAS60917.2024.10658900)Cited by:[§2\.1](https://arxiv.org/html/2608.11281#S2.SS1.p3.1)\.
- \[16\]T\. Z\. Sana, S\. Abdulla, A\. Nag, A\. Das, Md\. M\. Hassan, Z\. Z\. Fiza, A\. Karim, and S\. R\. R\. Kabir\(2025\)Advancing Federated Learning: A Systematic Literature Review of Methods, Challenges, and Applications\.IEEE Access13,pp\. 153817–153844\.External Links:ISSN 2169\-3536,[Link](https://ieeexplore.ieee.org/document/11146653/),[Document](https://dx.doi.org/10.1109/ACCESS.2025.3605165)Cited by:[§2\.1](https://arxiv.org/html/2608.11281#S2.SS1.p3.1)\.
- \[17\]R\. Shah, N\. Pai, G\. Thomas, S\. Jha, V\. Mittal, K\. Shirvni, and H\. Liang\(2024\-11\)Machine learning in wear prediction\.Journal of Tribology147\(4\),pp\. 040801\.Cited by:[§1](https://arxiv.org/html/2608.11281#S1.p2.1)\.
- \[18\]M\. Soori, B\. Arezoo, and R\. Dastres\(2023\)Machine learning and artificial intelligence in cnc machine tools, a review\.Sustainable Manufacturing and Service Economics2,pp\. 100009\.External Links:ISSN 2667\-3444,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.smse.2023.100009),[Link](https://www.sciencedirect.com/science/article/pii/S2667344423000014)Cited by:[§1](https://arxiv.org/html/2608.11281#S1.p2.1)\.
- \[19\]C\. Yang, J\. Zhou, E\. Li, M\. Wang, and T\. Jin\(2022\)Local\-feature and global\-dependency based tool wear prediction using deep learning\.Scientific Reports12\(1\),pp\. 14574\.Cited by:[§4\.2](https://arxiv.org/html/2608.11281#S4.SS2.p1.1)\.
- \[20\]H\. Zhu, J\. Xu, S\. Liu, and Y\. Jin\(2021\-11\)Federated learning on non\-IID data: A survey\.Neurocomputing465,pp\. 371–390\.External Links:ISSN 0925\-2312,[Link](https://www.sciencedirect.com/science/article/pii/S0925231221013254)Cited by:[§2\.1](https://arxiv.org/html/2608.11281#S2.SS1.p3.1)\.