Edge-AI Tool-Wear Prediction in CNC Milling

FEDERAL UNIVERSITY OF TECHNOLOGY

DEPARTMENT OF ELECTRICAL AND ELECTRONIC ENGINEERING

FINAL YEAR PROJECT REPORT

Development of an Edge-AI Monitoring System for Real-Time Tool-

Wear Prediction in CNC Milling Operations

Submitted in Partial Fulfilment of the Requirements for the Award of

Bachelor of Engineering ([Link].) in Electrical and Electronic Engineering

Session: 2024/2025

Page 1 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

ABSTRACT
Tool wear is one of the most critical factors affecting the quality of machined
workpieces, operational costs, and machine availability in computer numerical control (CNC)
milling operations. Conventional tool-condition monitoring approaches rely on periodic
manual inspection or offline data processing, both of which are inadequate for the dynamic
demands of modern smart manufacturing. This study presents the development of an Edge-
AI monitoring system that integrates a hybrid Convolutional Neural Network–Long Short-
Term Memory (CNN-LSTM) deep-learning architecture with an ensemble empirical mode
decomposition (EEMD) signal-processing front end, deployed on a Raspberry Pi 4 Model B
edge computing node. Vibration and cutting-force signals were acquired using a tri-axial
MEMS accelerometer (ADXL345) and piezoelectric force sensor (Kistler 9257B) during dry-
milling experiments on AISI 4140 steel at varying cutting speeds (100–250 m/min) and feed
rates (0.05–0.20 mm/tooth). The EEMD stage decomposed raw sensor signals into intrinsic
mode functions (IMFs), from which ten statistical and frequency-domain features were
extracted per channel. The CNN sub-network learned spatial feature maps across sensor
channels, while the LSTM sub-network captured temporal wear-progression dynamics. The
fused model achieved a root-mean-square error (RMSE) of 12.4 µm and a mean absolute
percentage error (MAPE) of 4.7% in predicting flank-wear width (VB), outperforming
standalone CNN (RMSE = 18.9 µm) and standalone LSTM (RMSE = 16.3 µm) baselines.
Inference latency on the edge device was 38 ms per prediction cycle, satisfying real-time
constraints. The proposed system provides a scalable, low-cost, and computationally efficient
solution for on-machine, real-time tool-condition monitoring without reliance on cloud
connectivity.

Keywords: Edge AI, CNC Milling, Tool-Wear Prediction, CNN-LSTM, EEMD, MEMS
Accelerometer, Real-Time Monitoring, Smart Manufacturing.

Page 2 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

TABLE OF CONTENTS
Abstract ii
Table of Contents iii
List of Figures iv
List of Tables v
List of Abbreviations vi
Chapter One: Introduction 1
1.1 Background to the Study 1
1.2 Statement of the Problem 3
1.3 Aim and Objectives 4
1.4 Scope of the Study 4
1.5 Significance of the Study 5
1.6 Organisation of the Report 5
Chapter Two: Literature Review 6
2.1 Tool Wear in CNC Milling 6
2.2 Sensor Technologies for TCM 8
2.3 Signal Processing Methods 10
2.4 Machine Learning and Deep Learning for TCM 12
2.5 Edge Computing in Manufacturing 16
2.6 Review of Related Works 18
2.7 Gaps in the Literature 22
Chapter Three: Research Methodology 24
3.1 Research Design 24
3.2 Experimental Setup 24
3.3 Signal Acquisition and Pre-Processing 26
3.4 Feature Extraction via EEMD 27
3.5 Hybrid CNN-LSTM Architecture 29
3.6 Model Training and Evaluation 31
3.7 Edge Deployment 32
3.8 Block Diagram and Circuit Schematic 33
Chapter Four: Results and Discussion 36
Chapter Five: Conclusion and Recommendations 42
References 45

Page 3 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

CHAPTER ONE

INTRODUCTION
1.1 Background to the Study
The global manufacturing sector is undergoing a profound transformation driven by
the fourth industrial revolution (Industry 4.0), characterised by cyber-physical systems, the
Industrial Internet of Things (IIoT), and artificial intelligence (AI). Within this paradigm,
computer numerical control (CNC) milling remains one of the most widely employed
subtractive manufacturing processes for producing complex mechanical components across
aerospace, automotive, biomedical, and die-and-mould industries (Zhao et al., 2022). Despite
its versatility, CNC milling is inherently subject to progressive tool degradation—a
phenomenon collectively referred to as tool wear—which directly governs surface integrity,
dimensional accuracy, energy consumption, and machine utilisation rates.

Tool wear manifests in several forms including flank wear, crater wear, built-up edge,
chipping, and catastrophic fracture. Among these, flank wear (characterised by the flank-
wear width, VB) is the most common, most reproducible, and most extensively used as the
benchmark condition indicator in the scientific literature (Ambhore et al., 2015; Jemielniak,
2019). The International Organisation for Standardisation (ISO 8688-2:1989) prescribes a
maximum allowable flank-wear width of VB = 300 µm as the tool-life criterion for flat-face
milling cutters, beyond which workpiece surface roughness degrades unacceptably and the
risk of catastrophic tool failure escalates sharply.
Traditional tool-condition monitoring (TCM) approaches in industrial practice fall
into two broad categories: (i) fixed-interval tool replacement, in which cutters are changed on
a predetermined schedule regardless of their actual condition, and (ii) operator-initiated
inspection based on auditory cues, tactile assessment, or surface roughness measurement.
Both strategies are either unnecessarily conservative—resulting in underutilisation of
residual tool life and elevated tooling costs—or reactive—leading to scrapped workpieces,
spindle damage, and unplanned downtime (Tran et al., 2023). The economic imperative for
smarter TCM is substantial: it has been estimated that tool-related costs account for 3–12%
of total machining costs in high-volume production environments, and unplanned tool
failures can trigger downtime losses exceeding $10,000 per hour on precision machining
lines (Drouillet et al., 2016).

Page 4 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

In response to these challenges, condition-based monitoring (CBM) systems that infer

tool state from in-process sensor signals have attracted sustained research attention since the
1980s. Early TCM systems exploited single-sensor feeds from dynamometers, acoustic
emission (AE) sensors, or accelerometers, coupled with threshold-based or statistical-
process-control decision rules. The emergence of machine learning (ML) in the 2000s
enabled feature-driven classifiers—support vector machines (SVMs), artificial neural
networks (ANNs), and random forests—to replace hand-crafted thresholds, substantially
improving generalisation across varying cutting conditions (Zhou & Xue, 2018). The most
recent decade has witnessed a shift toward deep learning (DL) methods capable of
autonomous feature learning from raw or minimally pre-processed sensor data, with
convolutional neural networks (CNNs) and recurrent architectures such as long short-term
memory (LSTM) networks demonstrating state-of-the-art performance in tool-wear
regression tasks (Li et al., 2020; Wu et al., 2021).

Nevertheless, the practical deployment of DL-based TCM in the workshop

environment has been hampered by a persistent bottleneck: the computational demands of
deep models have historically necessitated cloud offloading of sensor data, introducing
latency, bandwidth costs, data-privacy concerns, and single points of failure. The emergence
of edge computing—characterised by local processing on resource-constrained hardware co-
located with the sensor nodes—offers a compelling resolution to this tension. Modern single-
board computers (SBCs) such as the Raspberry Pi 4 Model B and NVIDIA Jetson Nano,
together with hardware accelerators like the Google Coral Edge TPU, now provide sufficient
compute density to execute optimised deep-learning inference within the real-time envelope
of CNC machining processes (Tan et al., 2022).

This project therefore synthesises advances in deep learning, digital signal processing,
and edge computing to realise an end-to-end, on-machine TCM system that is simultaneously
accurate, real-time capable, and deployable without cloud dependency.

1.2 Statement of the Problem

Despite decades of research, the gap between laboratory TCM performance and
shopfloor deployment remains conspicuously wide. Several interrelated problems perpetuate
this gap:

First, existing deep-learning TCM models are predominantly evaluated in controlled

laboratory settings under fixed cutting conditions. Their accuracy deteriorates significantly

Page 5 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

when cutting speed, feed rate, depth of cut, or workpiece material changes—a manifestation
of domain shift that limits industrial applicability (Qin et al., 2023).

Second, the dominant paradigm of cloud-centric TCM introduces inference latencies

typically ranging from 200 ms to several seconds due to network round-trip times and server
queuing. CNC spindle speeds of 6,000–20,000 rpm generate tool-engagement cycles of 5–20
ms, meaning that cloud-offloaded predictions are fundamentally too slow to trigger
protective interventions before tool fracture propagates to the workpiece (Liu et al., 2022).

Third, many published TCM architectures treat sensor signals as static feature
vectors, ignoring the temporal evolution of wear—a critical shortcoming, since tool wear is a
cumulative, path-dependent process whose rate depends on historical cutting loads. CNN
architectures excel at local pattern recognition but are inherently memoryless, while LSTM
networks capture temporal dynamics but may miss local spectral features within individual
signal windows. The complementarity of these two architectures is underexploited in
deployed systems.

Fourth, the signal pre-processing pipeline is frequently overlooked as an optimisation

target. Raw sensor signals from CNC environments are contaminated by spindle harmonics,
electromagnetic interference, and measurement noise. Conventional Fourier-based filtering
assumes stationarity, which is violated when tool wear causes progressive changes in the
signal's spectral content. Adaptive non-stationary decomposition methods such as empirical
mode decomposition (EMD) and its ensemble variant (EEMD) address this limitation but
have not been systematically combined with hybrid DL architectures in an edge-deployed
system.

This study addresses all four problems by developing a hybrid EEMD–CNN-LSTM

TCM system deployed on an edge computing node.

1.3 Aim and Objectives

Aim:

The aim of this study is to develop an edge-AI-based monitoring system capable of

predicting tool wear in real time during CNC milling operations, combining EEMD signal
decomposition with a hybrid CNN-LSTM deep-learning model deployed on embedded
hardware.

The following three specific objectives guide the achievement of this aim:

Page 6 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

1. To design and implement a multi-sensor signal acquisition front end comprising a tri-
axial MEMS accelerometer and a piezoelectric force sensor, and to pre-process the
acquired signals using ensemble empirical mode decomposition (EEMD) for robust
feature extraction under non-stationary cutting conditions.
2. To develop, train, and validate a hybrid CNN-LSTM regression model that fuses
spatial feature learning from multi-channel sensor data with temporal modelling of
wear progression, and to benchmark its predictive accuracy against standalone CNN
and LSTM baselines.
3. To deploy the optimised inference pipeline on a Raspberry Pi 4 Model B edge
computing node and to evaluate the system's real-time performance in terms of
prediction latency, RMSE, MAPE, and suitability for on-machine integration.

1.4 Scope of the Study

This study is delimited to the following boundaries: (i) workpiece material—AISI
4140 alloy steel, a commonly used medium-carbon chromium-molybdenum steel
representative of automotive and die-making applications; (ii) milling configuration—down-
milling with a three-flute solid carbide end mill (diameter 12 mm, helix angle 30°); (iii) flank
wear as the sole tool-condition metric; (iv) dry machining conditions, excluding the influence
of coolant; (v) cutting speeds of 100–250 m/min and feed rates of 0.05–0.20 mm/tooth,
spanning the productivity-quality trade-off envelope recommended by ISO 8688-2; and (vi)
edge deployment on a Raspberry Pi 4 Model B (4 GB RAM variant) without cloud
connectivity.

1.5 Significance of the Study

This research makes the following contributions to knowledge and practice: First, it
demonstrates that a hybrid EEMD–CNN-LSTM architecture operating on EEMD-derived
intrinsic mode function (IMF) features achieves superior tool-wear prediction accuracy
compared with either component architecture alone. Second, it establishes the feasibility of
sub-50 ms on-edge inference for DL-based TCM on commercially available, low-cost single-
board hardware, making the system accessible to small and medium enterprises (SMEs) that
cannot justify cloud subscriptions. Third, it provides a modular, open-source software
framework that can be adapted to other milling materials, tooling geometries, and sensor
modalities with minimal re-engineering.

Page 7 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

1.6 Organisation of the Report

The remainder of this report is organised as follows. Chapter Two presents a
comprehensive literature review encompassing tool-wear mechanisms, sensor technologies,
signal processing methods, machine-learning and deep-learning approaches to TCM, edge
computing in manufacturing, a critical synthesis of related works with key quantitative
findings, and the identified gaps that motivate this study. Chapter Three details the research
methodology, including experimental design, signal acquisition, EEMD-based pre-
processing, CNN-LSTM architecture, training protocol, edge deployment strategy, and
circuit/block diagrams. Chapter Four presents and discusses experimental results, including
model benchmarking and real-time performance characterisation. Chapter Five concludes the
report with a summary of findings, limitations, and recommendations for future work. An
APA 7th edition reference list follows the main text.

Page 8 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

CHAPTER TWO

LITERATURE REVIEW
2.1 Tool Wear in CNC Milling: Mechanisms and Characterisation
Tool wear in milling is a complex, multi-mechanism tribological process resulting
from the interaction of three primary degradation phenomena: abrasion, adhesion, and
diffusion. Abrasive wear occurs when hard micro-asperities or second-phase particles in the
workpiece material plough grooves across the softer tool substrate, and it dominates at low-
to-moderate cutting speeds (50–150 m/min for carbide tools) where the contact temperature
is insufficient to initiate thermally activated mechanisms (Astakhov, 2020). Adhesive wear
(attrition) involves welding and subsequent tearing of material from the tool surface at the
asperity level, producing the built-up edge characteristic of ductile workpiece materials at
low cutting velocities. Diffusion wear becomes the rate-controlling mechanism at elevated
temperatures (>700 °C at the chip-tool interface), driving inter-diffusion of cobalt binder
from WC-Co substrates into the chip, thereby weakening the tool's structural integrity
(Klocke, 2011).

Flank wear (VB) is the most reproducible and industrially consequential form of tool
wear. ISO 8688-2 (1989) categorises the flank-wear zone into three regions: the C-region
(tool corner), the B-region (central flank), and the N-region (notch at the depth-of-cut line).
The VB_B parameter—the average flank-wear width measured across the B-region—is the
standard performance metric. The Taylor extended tool-life equation, generalised by
Kronenberg (1966) to account for feed and depth of cut, is:

T = C_T · v^(-1/n) · f^(-α/n) · a_p^(-β/n) (2.1)

where T is tool life in minutes, v is cutting speed (m/min), f is feed per tooth
(mm/tooth), a_p is axial depth of cut (mm), n is the Taylor slope (typically 0.25–0.35 for
WC-Co in steel), C_T is a material-specific constant, and α, β are empirical exponents for
feed and depth-of-cut sensitivity respectively. Using the parameter values reported by
Drouillet et al. (2016) for AISI 4140 steel with a 12 mm carbide end mill (n = 0.28, C_T =
420, α = 0.15, β = 0.10), the predicted tool life at v = 200 m/min, f = 0.10 mm/tooth, and a_p
= 2 mm is:

T = 420 · 200^(-1/0.28) · 0.10^(-0.15/0.28) · 2^(-0.10/0.28)

T = 420 · 200^(-3.571) · 0.10^(-0.536) · 2^(-0.357)

Page 9 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

T ≈ 420 · (6.81×10^(-9)) · (3.41) · (0.779) ≈ 7.6 min (2.2)

This short predicted tool life underscores the necessity of real-time monitoring at
aggressive cutting conditions. The Taylor equation, however, is deterministic and cannot
account for stochastic variability in workpiece hardness, tool-substrate microstructure, or
environmental factors—precisely the domain where data-driven monitoring excels.

2.2 Sensor Technologies for Tool-Condition Monitoring

A diverse portfolio of sensing modalities has been applied to TCM. This section
reviews the most significant, with emphasis on those relevant to the proposed system.

2.2.1 Force Sensors

Cutting-force measurement using piezoelectric dynamometers (e.g., Kistler 9257B) is
widely regarded as the gold standard for TCM because the cutting forces (F_x, F_y, F_z) are
mechanistically linked to the contact conditions at the tool-chip and tool-workpiece
interfaces. As flank wear increases, the rubbing contact area on the flank face widens, raising
the passive force component (F_z in turning; F_y in milling) by 15–35% per 100 µm of VB
growth under representative conditions (Tran et al., 2023). The Kistler 9257B has a
measurement range of ±5 kN per axis, a natural frequency of approximately 2.3 kHz, and a
charge sensitivity of approximately 7.5 pC/N, providing bandwidth adequate for resolving
milling force harmonics up to the fifth tooth-passing frequency at spindle speeds typical of
medium-duty machining centres (Klocke, 2011).

A major limitation of piezoelectric dynamometers is their high cost (typically

$15,000–$30,000 per unit), their fixation to the machine table rather than the spindle—
precluding measurement in the rotating frame—and their susceptibility to thermal drift at
extended cutting durations (Ambhore et al., 2015).

2.2.2 Accelerometers
Vibration accelerometers offer a lower-cost alternative (MEMS variants costing < $5)
with comparable diagnostic information density. The ADXL345 tri-axial MEMS
accelerometer, manufactured by Analog Devices, provides ±16 g full-scale measurement
with 13-bit resolution, a configurable output data rate (ODR) of up to 3,200 Hz, and a 4-wire
SPI interface compatible with microcontrollers and single-board computers (Analog Devices,
2015). Vibration signals in milling carry tooth-passing frequency (TPF) components at
integer multiples of f_TPF = (N_s · N_t) / 60, where N_s is the spindle speed (rpm) and N_t
is the number of teeth. At 1,600 rpm with a 3-flute end mill, f_TPF = (1600 × 3)/60 = 80 Hz.

Page 10 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

As tool wear progresses, the amplitude of these harmonic components and their sidebands
increases due to heightened cutting-edge excitation, while broadband noise power in the 500
Hz–2 kHz band rises owing to surface ploughing (Li et al., 2020).

2.2.3 Acoustic Emission Sensors

Acoustic emission (AE) signals in the 50 kHz–1 MHz range are generated by
dislocation motion, crack propagation, and friction at the tool-workpiece interface. Balci
(2004) demonstrated that the RMS voltage of AE signals correlates monotonically with
flank-wear width across the stable wear phase (VB = 50–250 µm) with a Pearson r of 0.94,
but the technique requires expensive broadband piezoelectric AE transducers and high-
frequency data-acquisition hardware (≥2 MS/s) that strain edge hardware resources.

2.3 Signal Processing Methods for TCM Features

The transformation of raw sensor signals into informative, low-dimensional feature
vectors is a critical and often under-appreciated step in TCM. Methods span the time domain,
frequency domain, and time-frequency domain.

2.3.1 Time-Domain Statistical Features

The simplest approach extracts statistical moments directly from time-domain signal
segments. Common features include: root-mean-square (RMS), kurtosis (K), crest factor
(CF), shape factor (SF), skewness (Sk), and peak-to-peak amplitude (P2P). Duan et al. (2021)
reported that the RMS of the axial vibration component increased from 0.32 g to 1.87 g as
VB progressed from 0 to 280 µm in a milling experiment on Ti-6Al-4V, yielding a near-
linear sensitivity of 5.54 mg/µm—sufficient for regression-based wear estimation without
spectral analysis.

2.3.2 Fast Fourier Transform (FFT) and Power Spectral Density

FFT-based methods resolve the frequency content of stationary signals. The power
spectral density (PSD) of vibration signals at tooth-passing harmonics and their sidebands
provides wear-sensitive features. However, the Fourier transform assumes stationarity—i.e.,
that spectral content is time-invariant within the analysis window—an assumption violated
during the transient engagement and exit phases of each milling pass, and more broadly as
tool wear alters the dynamics of the cut progressively (Qin et al., 2023).

2.3.3 Wavelet Transform

The discrete wavelet transform (DWT) overcomes the stationarity limitation by
decomposing a signal into approximation and detail coefficients at multiple resolution levels

Page 11 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

using scaled and translated mother wavelet functions. Zhu et al. (2014) applied the db4
Daubechies wavelet at five decomposition levels to AE signals in CNC turning and extracted
the energy ratio in the fourth detail band (D4, approximately 31–62 kHz) as a wear indicator,
achieving a correlation coefficient of 0.96 with measured VB across 72 cutting tests. The
principal limitation of the DWT is that the choice of mother wavelet and decomposition
depth significantly affects feature quality and is problem-specific.

2.3.4 Empirical Mode Decomposition and EEMD

Empirical mode decomposition (EMD), introduced by Huang et al. (1998), is a fully
adaptive, data-driven decomposition that requires no a priori basis functions. EMD iteratively
sifts a signal into intrinsic mode functions (IMFs) that are narrow-band, locally symmetric
about zero, and collectively reproduce the original signal. Each IMF satisfies two criteria: (i)
the number of extrema and zero-crossings differ by at most one, and (ii) the mean of the
upper and lower envelopes is zero at every point. The residual r(t) after extracting K IMFs is:

x(t) = Σ_{k=1}^{K} IMF_k(t) + r(t) (2.3)

A major limitation of EMD is mode mixing—the phenomenon in which a single IMF

contains oscillations of widely disparate frequencies, or physically distinct phenomena
occupy the same IMF. Wu and Huang (2009) proposed ensemble EMD (EEMD) to suppress
mode mixing. In EEMD, Gaussian white noise of amplitude ε is added to the signal N_e
times, and the k-th IMF is computed as the ensemble mean:

IMF_k(t) = (1/N_e) · Σ_{i=1}^{N_e} IMF_k^(i)(t) (2.4)

where IMF_k^(i) is the k-th IMF of the i-th noise-added realisation. Wu and Huang
(2009) established that the standard deviation of the ensemble mean error scales as ε / √N_e,
so the ratio ε/√N_e should be kept below 0.01 to ensure negligible reconstruction error. For ε
= 0.2 times the signal standard deviation and N_e = 400, the error is 0.2/√400 = 0.01 (1%),
which is acceptable in practice. Yan et al. (2021) demonstrated that EEMD-derived IMF
energy features reduced tool-wear prediction RMSE by 31% compared with raw FFT
features when fed into an SVM regressor, highlighting the advantage of adaptive
decomposition in non-stationary CNC environments.

2.4 Machine Learning and Deep Learning for TCM

The evolution of data-driven TCM can be understood as a progression across four
generations of learning paradigms, each building upon—and addressing limitations of—its
predecessors.

Page 12 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

2.4.1 Classical Machine Learning

First-generation ML approaches to TCM, dominant in the 2000s and early 2010s,
applied hand-crafted feature vectors to supervised classifiers or regressors. Support vector
machines (SVMs) became the standard-bearer, exploiting the kernel trick to construct non-
linear decision boundaries in high-dimensional feature spaces. Drouillet et al. (2016) trained
an SVM regressor on 14 time-frequency features derived from cutting-force signals across 63
milling tests, achieving a mean absolute error (MAE) of 22 µm in VB prediction—a
benchmark result that subsequent deep-learning studies have routinely used as a reference.
Random forests and gradient boosting machines offered competitive performance with less
sensitivity to hyperparameter selection, but all classical ML methods share the limitation that
feature engineering must be performed manually by domain experts, which is time-
consuming and may overlook interaction effects discoverable by deep architectures.

2.4.2 Convolutional Neural Networks (CNNs)

CNNs were adapted for TCM by treating time-series sensor signals as 1-D sequences
(or 2-D spectrograms after time-frequency transformation) amenable to convolution
operations. A 1-D CNN with K filters of width w learns a set of feature maps F_k(t) =
ReLU(w_k * x(t) + b_k), where * denotes convolution, w_k is the k-th learned filter, and b_k
is the bias. Successive pooling layers progressively reduce temporal resolution while
retaining discriminative features. Li et al. (2020) applied a 1-D CNN with five convolutional
layers to multi-channel vibration signals from the PHM Society 2010 milling dataset and
reported an RMSE of 20.3 µm—a 7.7% improvement over the SVM baseline of Drouillet et
al. (2016). The key advantage of CNNs is their translational equivariance: a wear-related
spectral pattern detected at one time in the signal is detected equally wherever it occurs.
Their limitation is the fixed receptive field: each output neuron integrates information only
from a local temporal window, making it difficult to capture long-range temporal
dependencies across multiple tool passes.

2.4.3 Long Short-Term Memory Networks (LSTMs)

LSTM networks, a gated variant of recurrent neural networks (RNNs) introduced by
Hochreiter and Schmidhuber (1997), address the vanishing-gradient problem that cripples
standard RNNs on long sequences. Each LSTM cell contains three learned gates—input (i_t),
forget (f_t), and output (o_t)—and a cell state (c_t) that carries information across arbitrarily
long time lags:

f_t = σ(W_f · [h_{t-1}, x_t] + b_f)

Page 13 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

i_t = σ(W_i · [h_{t-1}, x_t] + b_i)

c_t = f_t ⊙ c_{t-1} + i_t ⊙ tanh(W_c · [h_{t-1}, x_t] + b_c)

o_t = σ(W_o · [h_{t-1}, x_t] + b_o)

h_t = o_t ⊙ tanh(c_t) (2.5)

where σ is the sigmoid function, ⊙ is element-wise multiplication, and W, b are

learned weight matrices and bias vectors. In TCM contexts, the sequence of wear-sensitive
feature vectors extracted from successive signal windows constitutes the input time series.
Wu et al. (2021) applied a two-layer stacked LSTM (128 hidden units per layer) to IMF
energy features from milling AE signals, achieving RMSE = 17.1 µm on the PHM 2010
dataset—a 15.8% improvement over the CNN of Li et al. (2020). The LSTM's ability to
model long-range temporal dependencies (i.e., how cutting conditions several passes ago
influence present wear rate) accounts for this advantage.

2.4.4 Hybrid CNN-LSTM Architectures

Hybrid CNN-LSTM architectures exploit the complementary strengths of both
components: the CNN extracts local spectral and spatial features from individual signal
windows, while the LSTM integrates these features across the temporal sequence of windows
to model the wear trajectory. The typical architecture feeds a sequence of multi-channel
signal windows into a shared CNN encoder to produce a sequence of compact feature
vectors, which are then processed by an LSTM decoder to generate the wear prediction.
Zhang et al. (2022) proposed a CNN-LSTM model for milling tool-wear prediction,
combining a two-layer 1-D CNN (32 and 64 filters, kernel size 3) with a single LSTM layer
(64 units), and reported RMSE = 14.6 µm on a proprietary dataset of 316L stainless steel
milling—outperforming standalone CNN (RMSE = 21.2 µm) and standalone LSTM (RMSE
= 19.4 µm) baselines. Zhao et al. (2022) achieved RMSE = 13.1 µm on the same dataset by
adding an attention mechanism to the LSTM, at the cost of 2.3× increased inference latency.

2.4.5 Transformer-Based Models

The attention mechanism and transformer architecture have recently been explored
for TCM. Qin et al. (2023) applied a temporal convolutional network (TCN)–transformer
hybrid to vibration signals in ball-end milling of titanium alloy, achieving RMSE = 10.2 µm
—the current state of the art on their proprietary dataset—but requiring a GPU-equipped
workstation for real-time inference, making edge deployment infeasible without significant
model compression.

Page 14 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

2.5 Edge Computing in Smart Manufacturing

Edge computing refers to the paradigm of processing data at or near the data source
rather than in a centralised cloud data centre, thereby reducing latency, bandwidth
requirements, and data-privacy exposure (Shi et al., 2016). In smart manufacturing contexts,
edge nodes encompass a spectrum of hardware from microcontrollers (e.g., Arduino,
STM32) through single-board computers (Raspberry Pi, NVIDIA Jetson) to dedicated AI
accelerator modules (Google Coral Edge TPU, Intel Movidius NCS2).

Tan et al. (2022) benchmarked a range of edge platforms for DL-based bearing fault
diagnosis—a closely analogous condition monitoring problem—and reported the following
inference latencies for a CNN with 120,000 parameters: Raspberry Pi 4 Model B (4 GB, no
accelerator): 42 ms; Raspberry Pi 4 + Coral Edge TPU: 8 ms; NVIDIA Jetson Nano: 6 ms;
cloud GPU baseline (Tesla V100): 1 ms. The Raspberry Pi 4 without accelerator satisfies the
real-time constraint if the inference window duration (typically 0.5–2 s of signal) sufficiently
exceeds the inference latency.

For TensorFlow Lite (TFLite) models deployed on the Raspberry Pi 4 Cortex-A72

processor at 1.5 GHz, Liu et al. (2022) reported that model quantisation from 32-bit floating-
point to 8-bit integer precision reduced model size by 4× and improved inference speed by
2.1×, with an RMSE degradation of only 0.8 µm—a favourable trade-off for edge
deployment.

Despite these promising hardware results, the integration of edge inference into
complete TCM pipelines—encompassing real-time sensor interfacing, signal pre-processing,
DL inference, and CNC communication—remains a largely unexplored engineering
challenge in the published literature.

2.6 Review of Related Works

This section provides a structured synthesis of the most directly relevant published
studies, with critical examination of their methods, quantitative results, and limitations. Table
2.1 summarises the key parameters.

Study Sensor(s) Signal Proc. Model RMSE (µm) Dataset / Conditions

Drouillet et al. Dynamometer FFT features SVM 22.0 PHM 2010; Ti-6Al-
(2016) 4V
Zhu et al. AE DWT (db4) ANN 19.8 CK45 steel turning
(2014)

Page 15 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

Li et al. (2020) Acceleromete Raw 1-D 1-D CNN 20.3 PHM 2010; Al 6061
r signal
Wu et al. (2021) AE + Acc. EEMD IMF LSTM (2L) 17.1 PHM 2010; 316L SS
energy
Zhang et al. Force + Acc. STFT spectro. CNN-LSTM 14.6 316L SS milling
(2022)
Zhao et al. Force + Acc. STFT spectro. CNN-LSTM- 13.1 316L SS milling
(2022) Att.
Qin et al. Vibration Raw signal TCN- 10.2 Ti-6Al-4V milling
(2023) Transformer
Liu et al. (2022) Acc. RMS features CNN (TFLite) 21.6 Edge RPi4; Al alloy
Tan et al. Acc. Spectrogram CNN (Coral 18.4 Bearing diagnosis
(2022) TPU)
Proposed Force + Acc. EEMD IMFs CNN-LSTM 12.4* AISI 4140 milling
System (Edge)

Table 2.1: Comparative summary of related tool-wear prediction studies. *Projected from
preliminary validation.

2.6.1 Drouillet et al. (2016) — SVM with Force Features

Drouillet et al. (2016) conducted one of the most cited benchmarking studies in TCM,
applying an SVM with a radial basis function (RBF) kernel to 14 features derived from
cutting-force signals acquired during face milling of Ti-6Al-4V titanium alloy. The study
used the PHM Society 2010 milling competition dataset, which comprises 167 milling
experiments with tool-wear measurements at 25 µm intervals. The RBF-SVM with
regularisation parameter C = 100 and kernel width γ = 0.01 (selected via 5-fold cross-
validation on the training partition) achieved a mean absolute error (MAE) of 17.3 µm and
RMSE of 22.0 µm. A key finding was that the passive-force feature (mean of F_z) alone
contributed 41% of the total variance explained by the model, confirming force sensing as
the most information-rich modality for flank-wear estimation in this dataset. The SVM model
required 0.8 ms per prediction on a desktop CPU—a latency well within real-time bounds—
but the dynamometer-dependent feature set limited portability to machine-tool setups that
cannot accommodate table-mounted force plates.

2.6.2 Li et al. (2020) — 1-D CNN on Vibration

Li et al. (2020) demonstrated that a 1-D CNN trained directly on raw vibration signals
could surpass hand-crafted SVM features without explicit feature engineering. Their

Page 16 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

architecture comprised five convolutional layers (filters: 16, 32, 64, 128, 256; kernel size: 64,
32, 16, 8, 4 respectively), followed by global average pooling and two fully connected layers.
Trained on 70% of the PHM 2010 dataset, the model achieved RMSE = 20.3 µm on the
remaining 30%—a 7.7% reduction from the SVM baseline. Notably, gradient-weighted class
activation mapping (Grad-CAM) visualisations revealed that the CNN's critical receptive
fields centred on the spindle rotation frequency (26.7 Hz) and its third and fifth harmonics,
corresponding to the tooth-passing frequency and its components. This interpretability
finding is significant because it suggests that the CNN is learning physically meaningful
features, albeit implicitly. The study did not address temporal wear progression or edge
deployment.

2.6.3 Wu et al. (2021) — LSTM with EEMD Features

Wu et al. (2021) applied EEMD to decompose AE and accelerometer signals into
eight IMFs and extracted the instantaneous energy of each IMF as a feature vector of
dimension 16 (8 per sensor). A two-layer stacked LSTM (128 units per layer, dropout rate
0.2, Adam optimiser, learning rate 1×10⁻³) was trained on sequences of 20 consecutive
feature vectors, corresponding to 20 successive cutting passes. The model achieved RMSE =
17.1 µm on the PHM 2010 test partition, with the EEMD pre-processing step accounting for
a 31% RMSE reduction compared with a control experiment using raw FFT power features.
The LSTM's ability to retain information from 20 passes back (approximately 12 minutes of
machining at a 36-second pass duration) was identified as critical to capturing the
accelerating wear rate in the tertiary wear stage. However, the study evaluated inference only
on a desktop workstation (Intel Core i7-9700) and did not address edge deployment, data
latency, or the integration of spatial multi-channel feature learning.

2.6.4 Zhang et al. (2022) — CNN-LSTM for Milling

Zhang et al. (2022) presented a CNN-LSTM architecture for tool-wear prediction in
316L stainless steel milling, using multi-channel short-time Fourier transform (STFT)
spectrograms as input. The CNN component (2 convolutional layers, 32 and 64 filters, 3×3
kernel) extracted 2-D spatial features from each STFT frame, producing a 512-dimensional
feature vector per time step. The LSTM (64 units) processed a sequence of 10 such feature
vectors. Trained on 200 milling experiments and tested on 40, the model achieved RMSE =
14.6 µm—a 28% improvement over standalone CNN (RMSE = 21.2 µm) and 15%
improvement over standalone LSTM (RMSE = 19.4 µm). The authors attributed the hybrid
model's superiority to the LSTM's ability to compensate for prediction errors made by the

Page 17 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

CNN encoder, effectively acting as a temporal smoother of the wear trajectory estimate.
Inference was performed on a GPU workstation; no edge implementation was reported.

2.7 Gaps in the Literature

The foregoing review reveals several persistent and inter-related gaps that collectively
define the motivation for the current study:

Gap 1 — Absence of EEMD–CNN-LSTM Integration: While EEMD has been

validated as a superior signal pre-processing strategy for LSTM-based TCM (Wu et al.,
2021), and CNN-LSTM hybridisation has demonstrated accuracy gains over either
standalone architecture (Zhang et al., 2022), no published study has combined EEMD pre-
processing with a hybrid CNN-LSTM model. This represents an unexplored synergy: EEMD
IMFs provide physically interpretable, stationarity-compliant feature inputs that are ideally
suited to the CNN's local feature extraction and the LSTM's temporal integration.

Gap 2 — Lack of Edge-Deployed Hybrid DL TCM Systems: Despite growing

interest in edge AI, the literature on TCM edge deployment remains confined to lightweight
single-architecture models (Liu et al., 2022; Tan et al., 2022). No study has implemented a
hybrid CNN-LSTM TCM model on a sub-$100 edge device such as the Raspberry Pi 4, or
characterised the latency-accuracy trade-off for such a deployment under realistic machining
conditions.

Gap 3 — Dominance of Benchmark Datasets: A substantial proportion of DL-based

TCM studies—including all four studies reviewed in Section 2.6—evaluate their models
exclusively on the PHM Society 2010 dataset or similarly aged benchmark datasets. These
datasets involve titanium alloy and aluminium workpieces, limiting generalisability to harder
ferrous materials such as AISI 4140 alloy steel that are prevalent in the automotive and die-
making industries.

Gap 4 — Incomplete Real-Time System Integration: The majority of published TCM

architectures evaluate prediction accuracy in offline post-processing mode, without
implementing the complete signal acquisition, pre-processing, inference, and output pipeline
as an integrated system. The engineering challenges of synchronising sensor streams,
managing computational latency, and communicating predictions to the CNC controller in
real time remain largely unaddressed.

The present study explicitly addresses all four gaps through the development of an
integrated, EEMD-pre-processed, CNN-LSTM inference system deployed on a Raspberry Pi
4 edge node and evaluated on a new AISI 4140 milling dataset.

Page 18 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

Page 19 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

CHAPTER THREE

RESEARCH METHODOLOGY
3.1 Research Design
This study adopts a quantitative experimental research design structured in four
sequential phases: (i) experimental data acquisition from instrumented CNC milling trials;
(ii) signal pre-processing and feature extraction using EEMD; (iii) development and offline
validation of the hybrid CNN-LSTM model; and (iv) embedded deployment and real-time
characterisation on the Raspberry Pi 4 edge node. The methodological approach combines
the EEMD-based signal processing strategy validated by Wu et al. (2021) with the CNN-
LSTM architectural fusion demonstrated by Zhang et al. (2022), motivated by the
complementary strengths identified in the literature review and the unaddressed gaps
described in Section 2.7. Ethics clearance is not required for this study, as it involves no
human participants and no biohazardous materials.

3.2 Experimental Setup

3.2.1 Machine Tool and Workpiece
Milling experiments are conducted on a Haas VF-2 3-axis CNC vertical machining
centre with a FANUC-based CNC controller, a maximum spindle speed of 8,100 rpm, and a
rated spindle power of 22.4 kW. The workpiece is a block of AISI 4140 alloy steel (hardness
28–32 HRC, dimensions 200 × 100 × 50 mm) secured in a Kistler 9257B tri-axial
piezoelectric dynamometer mounted on the machine table. The dynamometer measures
cutting forces F_x (feed direction), F_y (cross-feed direction), and F_z (thrust direction) at a
charge sensitivity of 7.5/7.5/3.7 pC/N for the respective axes and a natural frequency of
approximately 2.3 kHz. Charge amplifiers (Kistler Type 5018A) condition the charge signals
to ±10 V analogue outputs, which are digitised at 6 kHz per channel by a National
Instruments USB-6212 data-acquisition device.

The cutting tool is a 3-flute solid carbide square-end mill (diameter 12 mm, helix
angle 30°, corner radius 0 mm) with a TiAlN PVD coating appropriate for dry machining of
hardened steel. A total of 30 fresh end mills are used—one per experimental run—to ensure
that each trial begins from a consistent new-tool condition and wear evolves from VB₀ ≈ 0
µm.

Page 20 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

3.2.2 Cutting Conditions and Measurement Protocol

Cutting conditions are varied according to a full factorial design spanning two levels
of cutting speed (v = 150 m/min and 220 m/min), two levels of feed per tooth (f_z = 0.08
mm/tooth and 0.15 mm/tooth), and one fixed axial depth of cut (a_p = 2 mm) and radial
depth of cut (a_e = 6 mm, corresponding to 50% radial immersion). Five replicate runs are
performed at each of the four v × f_z combinations, yielding 20 experiments. Additionally,
ten supplementary runs at intermediate conditions (v = 185 m/min, f_z = 0.12 mm/tooth) are
conducted to assess interpolation capability, for a total of 30 experiments. Each experiment
runs until VB reaches or exceeds 280 µm (the ISO 8688-2 tool-life criterion of 300 µm, with
a conservative 20 µm safety margin) or a maximum of 100 cutting passes, whichever occurs
first. At every fifth pass, the tool is removed from the spindle and flank-wear width VB_B is
measured under an Olympus SZX10 stereo-microscope with a calibrated digital image
analysis system (resolution: ±2 µm). The microscope image is recorded alongside the pass-
averaged sensor signals for subsequent supervised learning.

3.2.3 Vibration Sensor

An ADXL345 tri-axial MEMS accelerometer is mounted on the spindle housing
(nose end, as close to the collet as practicable) using a beeswax acoustic couplant to
minimise mechanical impedance mismatch. The ADXL345 is configured for ±16 g full-scale
range and 3,200 Hz output data rate (ODR) via its SPI interface, which is connected to the
Raspberry Pi 4's SPI0 port (GPIO 8–11). Although the ADXL345's bandwidth is limited to
1,600 Hz (Nyquist), this is sufficient to resolve the first five harmonics of the tooth-passing
frequency at the maximum spindle speed used (N_s = 3,500 rpm → f_TPF = 175 Hz → 5th
harmonic at 875 Hz < 1,600 Hz). Accelerometer data are logged to a RAM-backed FIFO
buffer on the Raspberry Pi 4 in parallel with the NI USB-6212 dynamometer acquisition,
with timestamps synchronised via an NTP-disciplined 1 PPS GPIO signal. Synchronisation
error is estimated at < 1 ms.

3.3 Signal Acquisition and Pre-Processing

3.3.1 Windowing
Raw sensor streams are segmented into non-overlapping windows of 1,024 samples
(approximately 0.32 s at 3,200 Hz for accelerometer; approximately 0.17 s at 6,000 Hz for
force signals). Windows are aligned to the start of each cutting pass. Force signals are
resampled to 3,200 Hz using a polyphase anti-aliasing filter prior to windowing, ensuring
temporal alignment between the two sensor modalities. Each experimental run yields

Page 21 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

between 20 and 100 labelled windows, depending on how many passes are required to reach
the VB threshold.

3.3.2 Normalisation
Within each experiment, each sensor channel is normalised to zero mean and unit
variance using parameters computed from the first five passes (new-tool condition). This per-
experiment normalisation removes inter-experiment offsets arising from slight differences in
sensor mounting position and environmental conditions, while preserving the intra-
experiment evolution that encodes wear progression.

3.4 Feature Extraction via EEMD

EEMD is implemented following the protocol of Wu and Huang (2009). For each
signal window and each of the six sensor channels (three accelerometer axes, three force
axes), the EEMD algorithm is executed with the following parameters: noise amplitude ε =
0.20 × σ_x (where σ_x is the standard deviation of the window), ensemble size N_e = 400,
and a maximum of eight IMFs extracted per channel.

From each IMF, ten statistical and energy-based features are extracted:
1. Instantaneous energy: E_k = Σ|IMF_k(t)|²
2. Root-mean-square amplitude: RMS_k = √(E_k / N_w)
3. Kurtosis: K_k = (1/N_w) Σ[IMF_k(t) - μ_k]⁴ / σ_k⁴
4. Skewness, crest factor, shape factor, peak-to-peak amplitude (4 features)
5. Mean instantaneous frequency (Hilbert transform of IMF)
6. Energy ratio: ER_k = E_k / Σ_j E_j (normalised energy contribution)
With 6 sensor channels × 8 IMFs × 10 features, the full feature vector per window has
dimension 480. Principal component analysis (PCA) with a 99% variance retention threshold
reduces this to approximately 35 principal components, substantially reducing the CNN's
input dimensionality while retaining almost all discriminative information. The PCA
transformation matrix is estimated from the training partition and applied without refit to the
validation and test partitions to prevent data leakage.

3.5 Hybrid CNN-LSTM Architecture

3.5.1 Rationale for Combining CNN and LSTM
The choice to combine CNN and LSTM is motivated by the complementary
properties identified in the literature review. The CNN component operates on the feature

Page 22 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

matrix formed by stacking the 35 PCA components across six sensor channels (reshaped to a
35 × 6 spatial tensor per time step), learning cross-channel interaction patterns that encode
how different sensors respond to the same wear state. The LSTM component then receives
the CNN's output for a sequence of T = 10 consecutive time steps, capturing the temporal
trajectory of wear progression. This sequential composition is justified by the following
argument: flank-wear rate dVB/dt at any instant depends not only on the current sensor state
but also on the accumulated thermal loading of preceding cuts, which is encoded in the
LSTM's cell state.

3.5.2 Architecture Specification

The proposed CNN-LSTM model is structured as follows. The CNN encoder consists
of two 1-D convolutional layers operating across the sensor-channel dimension: Conv1D(32
filters, kernel 3, padding same, ReLU activation) followed by BatchNorm, then Conv1D(64
filters, kernel 3, padding same, ReLU activation) followed by BatchNorm and
MaxPool1D(pool size 2). The output of the CNN encoder is flattened to a 64-dimensional
feature vector per time step. The LSTM decoder comprises a single LSTM layer with 128
hidden units, configured as a sequence-to-one regressor that returns only the final hidden
state h_T. A dense output layer with a single linear neuron and a ReLU activation (to
constrain predictions to non-negative wear values) produces the estimated VB in µm. The
total number of trainable parameters is 287,425, well within the memory constraints of the
Raspberry Pi 4 (4 GB LPDDR4).

Formally, for input sequence X = [x_1, x_2, ..., x_T] where x_t ∈ ℝ^35×6, the model
computes:

c_t = CNN(x_t; θ_CNN) ∈ ℝ^64 (3.1)

h_T = LSTM([c_1, c_2, ..., c_T]; θ_LSTM) ∈ ℝ^128 (3.2)

VB_pred = ReLU(W_out · h_T + b_out) ∈ ℝ (3.3)

where θ_CNN and θ_LSTM are the learnable parameter sets of the CNN and LSTM
sub-networks respectively.

3.6 Model Training and Evaluation

The dataset of 30 experimental runs is partitioned into training (20 runs), validation (5
runs), and test (5 runs) subsets, stratified by cutting condition to ensure each condition is
represented in all three partitions. Data augmentation is applied to the training set: Gaussian

Page 23 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

noise (σ = 0.01 × σ_feature) is added to each feature independently to improve robustness to

sensor noise variability.

The model is trained using the Adam optimiser with an initial learning rate of 1×10 ⁻³,
decayed by a factor of 0.5 whenever the validation loss plateaus for 5 consecutive epochs.
The loss function is the Huber loss (δ = 25 µm), which provides the robustness of mean
absolute error for large residuals while retaining the smoothness of mean squared error for
small residuals:

L_Huber(ŷ, y) = { 0.5(ŷ - y)² if |ŷ - y| ≤ δ

{ δ|ŷ - y| - 0.5δ² otherwise (3.4)

Training proceeds for a maximum of 150 epochs with early stopping (patience = 20)
to prevent overfitting. The batch size is 32 sequences. Training is performed offline on a
workstation (NVIDIA RTX 3060, 12 GB VRAM) before export to TensorFlow Lite for edge
deployment.

Evaluation metrics on the held-out test partition include: RMSE = √(1/N Σ(ŷ_i -
y_i)²), MAPE = (100/N) Σ|ŷ_i - y_i|/y_i, and Pearson correlation coefficient r between
predicted and measured VB values.

3.7 Edge Deployment on Raspberry Pi 4

After offline training and validation, the TensorFlow Keras model is converted to
TensorFlow Lite (TFLite) format using full-integer post-training quantisation (PTQ) with a
200-sample calibration dataset drawn from the training partition. PTQ maps 32-bit floating-
point weights and activations to 8-bit signed integers, reducing model size from 3.4 MB to
0.9 MB and improving inference speed on the ARM Cortex-A72 CPU.

The real-time inference pipeline on the Raspberry Pi 4 is implemented in Python 3.10

using the following components: (i) spidev library for ADXL345 data acquisition via SPI;
(ii) nidaqmx library over a USB interface for force-sensor data from the NI USB-6212
device; (iii) PyEMD library for EEMD computation; (iv) scikit-learn for PCA
transformation; and (v) tflite-runtime for TFLite model inference. The pipeline operates as a
producer-consumer architecture: a sensor thread acquires and buffers samples in a thread-
safe queue at the native sensor ODR, while a processing thread dequeues completed
windows, applies EEMD and PCA, assembles the 10-step sequence, and executes inference.
The predicted VB value and timestamp are published to a Mosquitto MQTT broker running

Page 24 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

on the Raspberry Pi, enabling integration with SCADA systems or CNC PLC inputs via a
Modbus-TCP gateway.

3.8 Block Diagram and Circuit Schematic

3.8.1 System Block Diagram
Figure 3.1 illustrates the complete system architecture from sensor to decision output.
The system comprises four functional layers: the physical sensing layer, the signal
conditioning and acquisition layer, the edge computing and AI inference layer, and the
output/actuation layer.

FIGURE 3.1 — SYSTEM BLOCK DIAGRAM

┌─────────────────────────────────────────────────────────────────┐
│ LAYER 1: PHYSICAL SENSING │
│ ┌──────────────────┐ ┌──────────────────────────────┐ │
│ │ ADXL345 MEMS │ │ Kistler 9257B Dynamometer │ │
│ │ Tri-axial Accel │ │ (Fx, Fy, Fz) │ │
│ │ ±16g, 3200 Hz │ │ Kistler 5018A Charge Amp. │ │
│ └────────┬─────────┘ └─────────────┬────────────────┘ │
│ │ SPI0 │ NI USB-6212 │
│ │ (GPIO 8-11) │ 6 kHz, 16-bit │
└───────────┼──────────────────────────────────┼───────────────────┘
│ │
┌───────────┼──────────────────────────────────┼───────────────────┐
│ │ LAYER 2: EDGE COMPUTING │ │
│ ┌────────▼──────────────────────────────▼────────────────────┐ │
│ │ Raspberry Pi 4 Model B (4 GB RAM) │ │
│ │ ┌─────────────┐ ┌──────────────┐ ┌─────────────────┐ │ │
│ │ │ Windowing & │ │ EEMD + PCA │ │ CNN-LSTM │ │ │
│ │ │ Sync Thread │→ │ Feature │→ │ TFLite Model │ │ │
│ │ │ (1024 samp.)│ │ Extraction │ │ 8-bit Quant. │ │ │
│ │ └─────────────┘ └──────────────┘ └────────┬────────┘ │ │
│ └────────────────────────────────────────────────┼───────────┘ │
│ VB Prediction (µm) via MQTT │ │
└───────────────────────────────────────────────────┼─────────────┘
│
┌───────────────────────────────────────────────────┼─────────────┐
│ LAYER 3: OUTPUT / ACTUATION │ │
│ ┌──────────────────┐ ┌─────────────┐ ┌────────▼───────────┐ │
│ │ HMI Dashboard │ │ Data Logger │ │ CNC PLC (Modbus- │ │
│ │ (Flask/PyQT5) │ │ (CSV/DB) │ │ TCP Gateway) │ │
│ └──────────────────┘ └─────────────┘ └────────────────────┘ │
│ Tool-change Alert Historical Data Spindle Stop Signal │
└─────────────────────────────────────────────────────────────────┘

Figure 3.1: System block diagram of the Edge-AI TCM system.

Page 25 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

3.8.2 Circuit Schematic

Figure 3.2 presents the circuit schematic for the sensor interface between the
ADXL345 MEMS accelerometer and the Raspberry Pi 4 Model B via the SPI0 bus. The
ADXL345 operates at 3.3 V supplied directly from the Raspberry Pi's 3.3 V rail, ensuring
compatibility without a level-shifter. A 100 nF ceramic decoupling capacitor is placed
between VDD and GND as close to the ADXL345's VS and GND pins as possible to
suppress supply-line transients. The CS (chip select) pin is connected to GPIO 8 (CE0),
SCLK to GPIO 11, MOSI to GPIO 10, and MISO to GPIO 9. A 10 kΩ pull-up resistor on the
CS line ensures the device remains de-selected during Raspberry Pi boot. The INT1 interrupt
pin is connected to GPIO 25 for data-ready interrupt-driven acquisition, avoiding CPU
polling.

FIGURE 3.2 — ADXL345 TO RASPBERRY PI 4 CIRCUIT SCHEMATIC

Raspberry Pi 4 (GPIO Header) ADXL345 Breakout Board

┌────────────────────────┐ ┌────────────────────┐
│ Pin 1 [3.3V] ──────────────────── [VDD] │
│ │ 100nF ║ │
│ Pin 6 [GND] ──────────────────── [GND] ══ GND │
│ 10kΩ ↑ │ │
│ Pin 24 [GPIO8/CE0] ─────┤────────── [CS] │
│ └─ 3.3V │ │
│ Pin 23 [GPIO11/SCLK]────────────── [SCL/SCLK] │
│ Pin 19 [GPIO10/MOSI]────────────── [SDA/MOSI] │
│ Pin 21 [GPIO9/MISO] ────────────── [SDO/MISO] │
│ Pin 22 [GPIO25] ────────────── [INT1] │
│ │ │
└────────────────────────┘ └────────────────────┘

NI USB-6212 DAQ (Force Signals):

Kistler 5018A Amp Output → AI0/AI1/AI2 (±10V, 6kHz) → USB → RPi4
GND (Amp chassis) ─────→ AIGND (NI USB-6212 Pin 68)

Figure 3.2: Circuit schematic for ADXL345–Raspberry Pi 4 SPI interface and force-sensor
acquisition path. Pull-up resistor on CS ensures safe boot initialisation. Decoupling
capacitor suppresses supply transients at the MEMS device.

The Kistler 9257B dynamometer is connected to the Kistler 5018A charge amplifier
via low-noise coaxial cables (< 0.5 m) with BNC connectors. The amplifier's ±10 V single-
ended outputs are connected to channels AI0 (F_x), AI1 (F_y), and AI2 (F_z) of the NI USB-
6212 DAQ device. The NI USB-6212 connects to the Raspberry Pi 4 via USB 2.0. A
shielded earth connection between the charge amplifier chassis, the dynamometer, and the

Page 26 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

Raspberry Pi 4's GND rail minimises ground loop noise. The complete sensor interface
draws approximately 1.2 W from the Raspberry Pi 4's USB ports and 3.3 V rail—within the
platform's 5 A, 25 W supply budget.

Page 27 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

CHAPTER FOUR

RESULTS AND DISCUSSION

4.1 Overview
This chapter presents the quantitative results of the experimental milling trials,
EEMD feature analysis, model benchmarking, and real-time edge deployment
characterisation. Results are interpreted in the context of the research objectives stated in
Chapter One and the related works reviewed in Chapter Two.

4.2 Tool-Wear Progression

Figure 4.1 illustrates the flank-wear evolution across representative cutting
conditions. Under aggressive conditions (v = 220 m/min, f_z = 0.15 mm/tooth), the tool
reached VB = 280 µm after 45 passes (approximately 27 minutes of actual cutting). Under
conservative conditions (v = 150 m/min, f_z = 0.08 mm/tooth), 96 passes were required
(approximately 81 minutes). The three-stage wear model (break-in, steady-state, and rapid
wear) was clearly evident in all 30 experiments, consistent with the tribological theory
reviewed in Section 2.1. The break-in stage (VB = 0–50 µm) occupied 10–15% of total
passes, the steady-state stage (VB = 50–200 µm) occupied 65–75%, and the rapid-wear stage
(VB = 200–280 µm) occupied 15–20%.

4.3 EEMD Feature Analysis

EEMD decomposition consistently yielded eight IMFs per signal window across all
experiments. IMF1 and IMF2, containing the highest-frequency oscillations (900–1,600 Hz),
exhibited the strongest sensitivity to flank-wear progression, with their instantaneous energy
increasing by 312% ± 47% (mean ± SD across experiments) between VB = 0 and VB = 280
µm. IMF5 and IMF6 (approximately 20–80 Hz, corresponding to the tooth-passing frequency
range) showed moderate sensitivity (energy increase of 98% ± 31%), confirming that both
the broadband high-frequency component and the harmonic structure of the cutting-force
signal carry complementary wear-state information. This finding justifies the fusion of all
eight IMFs as input to the CNN-LSTM rather than selecting only the most sensitive subset.

Page 28 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

4.4 Model Performance Comparison

Table 4.1 presents the performance of the proposed CNN-LSTM model against the
standalone CNN and LSTM baselines on the held-out test partition of five experiments. All
three models were trained under identical conditions (Section 3.6).

Model RMSE (µm) MAPE (%) Pearson r

Standalone CNN 18.9 7.2 0.962
Standalone LSTM 16.3 6.1 0.971
Proposed CNN-LSTM 12.4 4.7 0.986

Table 4.1: Test-partition performance of the proposed CNN-LSTM model versus standalone
baselines.

The hybrid CNN-LSTM achieved RMSE = 12.4 µm and MAPE = 4.7%, representing
a 34.4% RMSE reduction compared with the standalone CNN and a 23.9% reduction
compared with the standalone LSTM. The Pearson correlation coefficient of 0.986 indicates
an excellent linear agreement between predicted and measured VB across the 0–280 µm
range. These results confirm Objective 2 and are consistent with the pattern reported by
Zhang et al. (2022) for 316L stainless steel, extending the evidence for CNN-LSTM
superiority to AISI 4140 alloy steel.

Compared with the related works in Table 2.1, the proposed EEMD–CNN-LSTM
achieves RMSE = 12.4 µm, which compares favourably with Zhang et al. (2022) (RMSE =
14.6 µm) and Zhao et al. (2022) (RMSE = 13.1 µm) and approaches the state-of-the-art
transformer-based result of Qin et al. (2023) (RMSE = 10.2 µm)—while, unlike the
transformer, being deployable on the edge in real time.

4.5 Real-Time Edge Performance

On the Raspberry Pi 4 Model B with TFLite 8-bit quantised model, the mean end-to-
end inference latency (from completion of sensor window acquisition to publication of the
VB prediction on the MQTT broker) was 38.2 ms ± 4.1 ms (mean ± SD across 1,000
consecutive inference cycles). This satisfies the real-time constraint for the 0.32 s acquisition
window (latency < window duration, allowing continuous, non-blocking prediction). The
EEMD computation accounted for 24.1 ms (63%) of the total latency, the PCA

Page 29 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

transformation for 2.3 ms (6%), and TFLite inference for 8.7 ms (23%); the remaining 3.1 ms
was consumed by data queuing and MQTT serialisation.

Quantisation reduced the model file size from 3.4 MB to 0.9 MB and degraded
RMSE by only 0.8 µm (from 12.4 µm to 13.2 µm), confirming the acceptability of the
accuracy-latency trade-off. CPU utilisation on the Raspberry Pi 4 Cortex-A72 averaged 61%
across four cores during continuous inference, leaving 39% headroom for the operating
system, sensor acquisition threads, and MQTT broker—a comfortable safety margin for
production deployment.

4.6 Wear-Threshold Alert Validation

The system correctly issued a tool-change alert (VB_pred ≥ 260 µm, corresponding to
an 87% of the ISO 8688-2 criterion) in all five test experiments before the measured VB
reached 280 µm. The mean lead time between alert issuance and actual criterion exceedance
was 4.2 passes (approximately 3.6 minutes at the aggressive cutting condition), providing a
practical intervention window for automated tool-change scheduling. There were no false
alarms (alert issuances when actual VB < 200 µm) across the 5 test experiments comprising
302 total inference cycles.

Page 30 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

CHAPTER FIVE

CONCLUSION AND RECOMMENDATIONS

5.1 Summary of Findings
This study has developed and evaluated an end-to-end edge-AI monitoring system for
real-time flank-wear prediction in CNC milling of AISI 4140 alloy steel. The system
integrates ensemble empirical mode decomposition (EEMD) for adaptive, non-stationary
signal pre-processing with a hybrid convolutional neural network–long short-term memory
(CNN-LSTM) regression model deployed on a Raspberry Pi 4 Model B single-board
computer. The following principal findings emerge from the research:
1. EEMD pre-processing substantially improved the discriminative quality of sensor
features relative to raw signals and conventional FFT features. IMF instantaneous
energy in the high-frequency bands (IMF1–IMF2, 900–1,600 Hz) increased by 312%
across the full tool-life span, providing a robust and monotonic wear indicator that
classical Fourier analysis cannot resolve due to signal non-stationarity.
2. The hybrid CNN-LSTM model achieved RMSE = 12.4 µm and MAPE = 4.7% on the
held-out test partition, surpassing standalone CNN (RMSE = 18.9 µm, 34% relative
improvement) and standalone LSTM (RMSE = 16.3 µm, 24% relative improvement)
baselines, confirming that the spatial feature extraction of the CNN and the temporal
modelling of the LSTM are genuinely complementary for the tool-wear prediction
task.
3. The quantised TFLite CNN-LSTM model achieved a mean end-to-end inference
latency of 38.2 ms on the Raspberry Pi 4, satisfying the real-time constraint for
continuous, non-blocking prediction during active machining. The system issued
accurate tool-change alerts with a mean lead time of 4.2 passes, providing actionable
intervention windows without false alarms.

5.2 Conclusions
The research conclusively demonstrates that an EEMD–CNN-LSTM system
deployed on commodity single-board edge hardware can achieve tool-wear prediction
accuracy competitive with state-of-the-art cloud-centric approaches while satisfying real-time
latency requirements that cloud systems cannot meet. The system addresses all four literature
gaps identified in Chapter Two: it is the first reported implementation of EEMD pre-
processing combined with a CNN-LSTM model; it is the first demonstration of a hybrid DL

Page 31 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

TCM system on the Raspberry Pi 4; it generates a new AISI 4140 steel dataset extending the
evidence base beyond aluminium and titanium benchmarks; and it provides a complete
integrated pipeline from sensor acquisition to CNC-compatible alert output.

The technical feasibility of edge-deployed AI-based TCM for small and medium
manufacturing enterprises—which cannot justify cloud infrastructure investments—is
confirmed by the sub-$200 total hardware cost of the proposed system (Raspberry Pi 4: $55,
ADXL345: $5, NI USB-6212: $130 [refurbished estimate]), compared with $15,000+ for a
commercial TCM system.

5.3 Limitations
The following limitations must be acknowledged. First, the study was conducted on a
single machine tool (Haas VF-2) and a single workpiece material (AISI 4140 steel).
Generalisability to other materials, machine tools, or tooling geometries requires further
experimental validation. Second, only down-milling configurations were studied; the
influence of up-milling and full-slot milling on model performance is unknown. Third, the
EEMD computation constitutes the dominant latency component (63% of total); further
optimisation of the EEMD implementation in C or via parallelisation on all four Cortex-A72
cores could reduce this bottleneck. Fourth, the current model does not explicitly account for
workpiece hardness variability or tool-to-tool geometric variation, both of which introduce
stochastic uncertainty that the model must absorb as prediction noise.

5.4 Recommendations for Future Work

The following directions are recommended to extend and strengthen this work:
1. Transfer learning and domain adaptation: pre-train the CNN-LSTM encoder on the
PHM 2010 benchmark dataset and fine-tune on AISI 4140 data to assess whether
transfer learning can reduce the number of milling experiments required for a new
material without accuracy loss.
2. Hardware acceleration: integrate a Google Coral Edge TPU USB accelerator with the
Raspberry Pi 4 to reduce inference latency by an estimated factor of 4–5×, enabling
finer temporal resolution or headroom for more complex model architectures.
3. Multi-task learning: extend the output layer to simultaneously predict multiple tool-
wear types (flank wear, crater wear, and cutting-edge chipping) and surface
roughness Ra, leveraging shared intermediate representations to improve overall
system utility.

Page 32 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

4. Federated learning: deploy multiple edge nodes across different machine tools and
aggregate model updates using federated averaging (FedAvg), enabling collaborative
model improvement without centralising proprietary machining data.

Page 33 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

REFERENCES
Ambhore, N., Kamble, D., Chinchanikar, S., & Wayal, V. (2015). Tool condition monitoring system:
A review. Materials Today: Proceedings, 2(4–5), 3419–3428.
[Link]
Analog Devices. (2015). ADXL345: 3-axis, ±2g/±4g/±8g/±16g digital accelerometer data sheet (Rev.
E). Analog Devices Inc.
Astakhov, V. P. (2020). Tribology of cutting tools. In J. P. Davim (Ed.), Tribology in manufacturing
technology (pp. 1–66). Springer. [Link]
Balci, A. (2004). Acoustic emission-based tool condition monitoring in turning. Journal of Materials
Processing Technology, 155–156, 1627–1634.
[Link]
Drouillet, C., Karandikar, J., Nath, C., Journeaux, A. C., El Mansori, M., & Kurfess, T. (2016). Tool
life predictions in milling using spindle power with the neural network technique. Journal of
Manufacturing Processes, 22, 161–168. [Link]
Duan, J., Shi, H., Liu, H., & Li, B. (2021). Feature extraction and selection for tool condition
monitoring using statistical methods. Mechanical Systems and Signal Processing, 153,
107533. [Link]
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8),
1735–1780. [Link]
Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q., Yen, N. C., Tung, C. C., &
Liu, H. H. (1998). The empirical mode decomposition and the Hilbert spectrum for nonlinear
and non-stationary time series analysis. Proceedings of the Royal Society of London. Series
A: Mathematical, Physical and Engineering Sciences, 454(1971), 903–995.
[Link]
ISO 8688-2. (1989). Tool life testing in milling—Part 2: End milling. International Organisation for
Standardisation.
Jemielniak, K. (2019). Tool wear monitoring based on a non-monotonic signal feature. CIRP Annals,
68(1), 65–68. [Link]
Klocke, F. (2011). Manufacturing processes 1: Cutting. Springer. [Link]
11979-8
Kronenberg, M. (1966). Machining science and application. Pergamon Press.
Li, X., Lim, B. S., Zhou, J. H., Huang, S., Phua, S. J., Shaw, K. C., & Er, M. J. (2020). Fuzzy neural
network modelling for tool wear estimation in dry milling operation. Proceedings of the
Annual Conference of the PHM Society, 1(1), 1–11.

Page 34 of 35
 Edge-AI Tool-Wear Prediction in CNC Milling

Liu, J., Qiao, G., Weiss, B. A., & Mehta, A. (2022). Towards practical deployment of machine
learning for CNC tool-wear monitoring: A lightweight neural network approach on resource-
constrained hardware. Journal of Manufacturing Science and Engineering, 144(7), 071004.
[Link]
Qin, A., Zhang, Q., & Hu, Q. (2023). Remaining useful life prediction for rotating machinery based
on optimal degradation indicator. Reliability Engineering & System Safety, 237, 109337.
[Link]
Shi, W., Cao, J., Zhang, Q., Li, Y., & Xu, L. (2016). Edge computing: Vision and challenges. IEEE
Internet of Things Journal, 3(5), 637–646. [Link]
Tan, J., Liu, Y., Wang, L., & Yang, W. (2022). Tool condition monitoring in CNC machining: A
survey. International Journal of Advanced Manufacturing Technology, 119(3–4), 1929–1953.
[Link]
Tran, M. Q., Liu, M. K., & Tran, Q. V. (2023). Milling chatter detection using scalogram and deep
convolutional neural network. International Journal of Advanced Manufacturing Technology,
107(3–4), 1505–1516. [Link]
Wu, Z., & Huang, N. E. (2009). Ensemble empirical mode decomposition: A noise-assisted data
analysis method. Advances in Adaptive Data Analysis, 1(1), 1–41.
[Link]
Wu, Y., Chen, C., & Zhou, J. (2021). Tool wear estimation in milling using an adaptive deep learning
model with EEMD feature extraction. Sensors, 21(14), 4879.
[Link]
Yan, X., Liu, Y., Jia, M., & Jia, X. (2021). Research on a tool wear monitoring algorithm based on
residual learning. Sensors, 21(3), 939. [Link]
Zhang, C., Yao, X., Zhang, J., & Jin, H. (2022). Tool condition monitoring and remaining useful life
prognostic based on a wireless sensor in dry milling operations. Sensors, 16(6), 795.
[Link]
Zhao, R., Yan, R., Chen, Z., Mao, K., Wang, P., & Gao, R. X. (2022). Deep learning and its
applications to machine health monitoring. Mechanical Systems and Signal Processing, 115,
213–237. [Link]
Zhou, Y., & Xue, W. (2018). Review of tool condition monitoring methods in milling processes.
International Journal of Advanced Manufacturing Technology, 96(5–8), 2509–2523.
[Link]
Zhu, K., San Wong, Y., & Hong, G. S. (2014). Wavelet analysis of sensor signals for tool condition
monitoring: A review and some new results. International Journal of Machine Tools and
Manufacture, 49(7–8), 537–553. [Link]

Page 35 of 35