The International Journal of Advanced Manufacturing Technology

https://doi.org/10.1007/s00170-024-13590-z

ORIGINAL ARTICLE

Chatter detection in simulated machining data: a simple refined

approach to vibration data
Matthew Alberts1 · Sam St. John1 · Bradley Jared3 · Jaydeep Karandikar2 · Anahita Khojandi1 · Tony Schmitz2,3 ·
Jamie Coble4

Received: 31 January 2024 / Accepted: 3 April 2024

© The Author(s) 2024

Abstract
Vibration monitoring is a critical aspect of assessing the health and performance of machinery and industrial processes.
This study explores the application of machine learning techniques, specifically the Random Forest (RF) classification
model, to predict and classify chatter—a detrimental self-excited vibration phenomenon—during machining operations. While
sophisticated methods have been employed to address chatter, this research investigates the efficacy of a novel approach to an
RF model. The study leverages simulated vibration data, bypassing resource-intensive real-world data collection, to develop a
versatile chatter detection model applicable across diverse machining configurations. The feature extraction process combines
time-series features and Fast Fourier Transform (FFT) data features, streamlining the model while addressing challenges
posed by feature selection. By focusing on the RF model’s simplicity and efficiency, this research advances chatter detection
techniques, offering a practical tool with improved generalizability, computational efficiency, and ease of interpretation. The
study demonstrates that innovation can reside in simplicity, opening avenues for wider applicability and accelerated progress
in the machining industry.

Keywords Random forest · Machine learning · Chatter · Stability · Recursive feature elimination simulation ·
Additive manufacturing

1 Introduction plex methodologies and advanced algorithms, it is crucial

not to overlook the simplicity and efficiency embodied in the
Vibration monitoring has been an integral practice in assess- Random Forest (RF) model.
ing the health of rotating machinery and industrial processes In recent years, the machining industry has increasingly
for decades. Among various applications of vibration-based turned to vibration monitoring to evaluate machine health and
damage detection, monitoring rotating machinery has stood optimize processes [2]. The vibrations emitted by machinery
out as one of the most mature and successful endeavors [1]. serve as early indicators, empowering operators to make crit-
While the field of anomaly detection boasts an array of com- ical decisions during machining operations or even before
their commencement [3]. Yet, the challenge of chatter, a dis-
This manuscript has been authored in part by UT-Battelle, LLC, under ruptive self-excited vibration occurring between the tool and
contract DE-AC05-00OR22725 with the US Department of Energy the workpiece, has persisted in machining for over a century
(DOE). The US government retains and the publisher, by accepting [4]. Chatter results in suboptimal surface finishes, reduced
the article for publication, acknowledges that the US government
material removal rates, and potential damage to both machine
retains a nonexclusive, paid-up, irrevocable, worldwide license to
publish or reproduce the published form of this manuscript, or allow tools and workpieces [5].
others to do so, for US government purposes. DOE will provide public Our research offers a fresh perspective on this age-old
access to these results of federally sponsored research in accordance problem: the power of a meticulously designed RF clas-
with the DOE Public Access Plan (https://www.energy.gov/doe-
sification model, emphasizing simplicity, as an effective
public-access-plan).
alternative. This model can be easily comprehended and
B Jamie Coble swiftly deployed on the shop floor.
[email protected] This study makes a distinct contribution to the ongo-
Extended author information available on the last page of the article ing research in machining by exploring the potential of

123
 The International Journal of Advanced Manufacturing Technology

vibration-based features combined with time-series features This integration not only enhances the precision of chatter
as classifiers for the identification and prediction of chatter in detection but also paves the way for a model that is intuitively
machining processes. In a landscape where intricate methods interpretable and rapidly deployable in industrial settings.
dominate, our research boldly champions the idea of harness-
ing the straightforward RF classification model. Building on
1.1.1 Traditional feature extraction methods
the foundation laid by Shevchik et al. in their predictive appli-
cation of random forest models [6], this study acknowledges
Traditional feature extraction methods such as Autocorrela-
the existence of more advanced methodologies. However,
tion (AC), Power Spectral Density (PSD), and Fast Fourier
our emphasis lies in demonstrating that the dependable RF
Transform (FFT) have been the cornerstone of vibration data
model can yield remarkable results, especially in industrial
analysis. Despite their widespread application, these meth-
settings where accessibility, speed, and interoperability are
ods encounter challenges in handling the intricate nature of
paramount. In an era that prizes computational speed and
machining vibrations. Our study builds upon these conven-
practicality, our research sets out to prove that innovation
tional techniques by refining the feature selection process,
does not always necessitate complexity.
thus enabling the detection of chatter with heightened sensi-
This perspective aligns chatter, a fault detectable through
tivity and specificity. Previous researchers have also explored
vibration data, with conventional fault diagnosis techniques,
combining multiple traditional feature extraction methods.
which typically entail feature extraction and fault identifica-
For instance, Yesilli et al. utilized coordinated peaks in AC,
tion [7]. Although a range of signal processing techniques,
PSD, and FFT plots as features for chatter classification in
including statistical analysis [8], Fast Fourier Transforms
turning [13]. However, determining which peaks are mean-
(FFT) [9], wavelet packet transform (WT) [10], empirical
ingful in FFT can be challenging due to signal variability
mode decomposition (EMD) [11], and sparse representa-
across different machine setups and machining parameters.
tion methods [12], have been proposed for feature extraction,
This study address the limitations inherent in traditional
these methods often require substantial prior knowledge and
methods, such as their inability to process non-stationary
may have limitations in complex, dynamic working environ-
signals, by implementing a tailored RF model adept at navi-
ments.
gating the complexities of machining data.
This research seeks to bridge the gap between advanced
theory and practical application. By demonstrating that a sim-
plified RF model can effectively classify chatter and aspire 1.1.2 Advanced feature extraction techniques
to make this valuable tool more accessible to machinists
and industry professionals. This approach retains the merits In recent years, a plethora of feature extraction methods for
of precision and reliability while streamlining the process, the prediction of chatter in boring processes has been intro-
reducing the reliance on extensive computational resources duced. Techniques such as Empirical Mode Decomposition
and specialized expertise. (EMD) and Support Vector Machine (SVM) classifiers have
The subsequent sections will delve into the methodology, shown promise. Li et al. proposed an EMD-based approach,
experimental results, and implications of our findings. While breaking down vibration signals into intrinsic mode func-
the chatter detection field has seen the advent of more com- tions (IMFs) and creating feature vectors for these IMFs
plex techniques, this research will show that simplicity, as [14]. Similarly, Chen and Zheng utilized SVM in tandem with
exemplified in the “keep it simple” approach, can be a form of Recursive Feature Extraction to detect chatter in milling oper-
sophistication that catalyzes broader applicability and accel- ations [15]. Despite their effectiveness, these methods often
erates progress in the machining industry. require extensive manual pre-processing and suffer from
computational inefficiencies, leading to potential drawbacks
such as the need for identifying informative decompositions
1.1 Feature extraction approaches manually [16][17].
Moreover, additional research has embraced sophisticated
Despite the plethora of feature extraction methods available techniques like stacked denoising autoencoders, entropy, and
for chatter prediction in machining, our research introduces coarse-grained information for feature extraction in various
a nuanced methodology that elevates the efficacy of these machining and grinding scenarios [18, 19]. Notably, Li et al.
techniques. Traditional feature extraction methods have laid combined multiscale power spectral entropy (MPSE), mul-
a strong foundation, yet they often fall short when address- tiscale permutation entropy (MPE), and Laplacian scores
ing the dynamic and non-linear characteristics of chatter. In to forecast chatter in milling operations [20]. While these
contrast, our approach synergistically integrates time-series advanced methods are promising, they are not without chal-
analysis with RF classifiers, presenting an innovative solution lenges, including intricate parameter selection and high
that simplifies the complex landscape of anomaly detection. computational demands.

123
The International Journal of Advanced Manufacturing Technology

To address these challenges, our method introduces the intricacies of parameter selection and the sensitivity to
a streamlined RF model that integrates the strengths of noise in the data, to the interpretation of results and the scal-
advanced feature extraction techniques, such as EMD and ability issues that emerge when dealing with large datasets
SVM, while bypassing their complexities and extensive harboring numerous features. Such complexities often hin-
pre-processing requirements. This innovative approach sig- der the effectiveness and applicability of these methods in
nificantly simplifies the process, eliminating the need for real-world scenarios, particularly in the vibration analysis of
manual intervention and making it a more viable and prac- milling operations.
tical tool for real-time machining diagnostics. Our proposed Recognizing these challenges, our research delves into
model stands out for its elegant simplicity, offering a more a systematic approach aimed at mitigating these obstacles.
efficient solution for on-the-floor machining diagnostics. The subsequent sections will introduce and detail our pro-
posed methodology, which seeks to capitalize on the inherent
strengths of the RF model. This model is renowned for its
1.1.3 Specific application cases
robustness against noise and its proficient ability to evaluate
the importance of features that are instrumental in enhancing
In the domain of monitoring rotating assets for opera-
the generalizability of the analysis [23]. By harnessing these
tional anomalies, the detection of issues such as chatter is
capabilities, our method aims to transcend the traditional lim-
paramount. Aslan and Altinas notably proposed the utiliza-
itations associated with feature extraction techniques.
tion of FFT and spectral data to facilitate online chatter
Our approach does not merely stop at identifying the
detection from spindle drive motor data [21]. This approach,
challenges; it proactively addresses them by integrating the
while innovative, is not without its limitations. Specifi-
simplicity and efficiency of the RF model with advanced ana-
cally, FFT’s inherent assumptions regarding the stationarity
lytical strategies. This integration results in a sophisticated
and linearity of signals pose significant challenges, as real-
yet practical tool, one that is adept at managing the nuances
world vibration signals from such machinery often exhibit
of parameter selection, reducing the impact of noise on the
non-stationary and transient characteristics, potentially com-
data, and interpreting complex results in a meaningful way.
promising the accuracy of FFT-based methods. Another
Furthermore, our method significantly improves scalability,
method that is focused on the spindle is that of Spindle
enabling the analysis of extensive datasets without compro-
Speed Variation (SSV). SSV is meant to vary the speed of
mising on accuracy or computational efficiency. Through
the spindle should chatter be detected; this does require a
these enhancements, our proposed approach promises to
machine with a SSV setup and is not found on most milling
revolutionize the vibration analysis of milling operations,
machines in additive manufacturing environments. Li et al.
offering a model that is not only grounded in solid theo-
also mentioned the possible adverse effects of speed varia-
retical foundations but also distinguished by its applicability
tion on tool wear, power fluctuation, and harmful overheating
and effectiveness in real-world industrial settings.
[22]. This infers that modeling milling data and predicting
chatter before machining begins would be optimal.
1.2 Approach and research contributions
Recognizing these limitations, our research pivots towards
adapting the RF model to better align with the dynamic nature
This comprehensive study introduces a streamlined and inno-
of machine vibrations. By accounting for the inherently tran-
vative approach for the prediction and detection of chatter
sient and non-stationary characteristics of these signals, our
in milling operations. Central to our methodology is the
adapted RF model offers a more accurate and reliable mech-
integration of simulated data, advanced time-series feature
anism for real-time chatter detection, particularly in critical
extraction using FFT data features, and the robust classifica-
applications like spindle drive monitoring. This enhanced
tion capabilities of a RF model. This multifaceted approach
adaptability not only improves the detection accuracy but
is designed not only to enhance the precision of chatter detec-
also ensures seamless integration with existing monitoring
tion but also to address and overcome the common challenges
systems. By doing so, our model demonstrates the consider-
associated with limited real-world datasets.
able potential to transform chatter detection across various
Our primary contribution is the strategic utilization of
machining operations, setting a new standard for precision
simulated vibration data, which serves as a powerful tool to
and reliability in the field.
circumvent the constraints of small-scale real-world datasets.
Feng et al. showcased how similar grinding topography sim-
1.1.4 Challenges and limitations ulated data could be to real-world data supporting this study’s
use of simulated matching data to be an actual representa-
Feature extraction methods are fundamental in the analysis tion of real-world data [24]. This innovative use of simulated
of complex data sets, yet each method inherently carries its data significantly expands the versatility and applicability of
own set of distinct challenges. These challenges range from our chatter detection model, making it suitable for a wide

123
 The International Journal of Advanced Manufacturing Technology

array of machining scenarios. By employing FFT for vibra- gated. This aligns with Dutta et al. who saw RF have a more
tion data feature extraction, a well-established method that than 30% reduction in computation time and power to other
is recognized and valued in the industry for its reliability and methods [26]. This reduction aligns with our goal of achiev-
efficacy. ing model generalizability, a principle emphasized by Jia et
Furthermore, our RF classification model stands out for its al. [27].
unique ability to harness the potential of this simulated data,
leading to a solution that not only reduces the dependency on 1.2.3 Experimental validation
extensive data collection from physical experiments but also
exhibits exceptional adaptability to various machining condi- This section outlines the experimental setup and the valida-
tions. The incorporation of FFT and time-series features into tion process for assessing the efficacy of the RF model in
our model ensures computational efficiency and produces classifying chatter in machining processes. The simulated
results that are readily interpretable by industry profession- data generation process will be described, and how this syn-
als, thus bridging the gap between theoretical research and thetic dataset emulates real-world machining conditions will
practical application. be explained. Also, the details of the metrics and criteria will
The experimental validation of our approach, which is be employed to evaluate the performance of our RF model,
detailed in the subsequent sections of this study, provides a including accuracy, precision, recall, and F1 score.
clear demonstration of our model’s capability to accurately In summary, this study presents an efficient and versa-
emulate real-world machining scenarios. This validation tile approach to chatter prediction, empowered by simulated
firmly establishes the practicality and applicability of our data, FFT feature extraction, and the RF classification model.
model in industrial settings, offering a promising solution By emphasizing simplicity and feature selection precision,
to the pervasive challenge of chatter prediction in milling this study not only advances the field of chatter detection but
operations. By marrying the strengths of simulated data, FFT also offers a practical tool with broad applicability, simplified
feature extraction, and the RF classification model, this study interpretability, and enhanced generalizability, thus address-
presents a holistic approach that is not only innovative in its ing the current limitations and complexities inherent in the
design but also exemplary in its performance and relevance domain.
to the industry.

1.2.1 Leveraging simulated vibration data 2 Methodology

Our approach distinguishes itself by harnessing the power This investigation unfolds an innovative methodology desig-
of simulated vibration data. This novel strategy obviates ned to predict chatter in milling operations with a pioneer-
the need for resource-intensive and time-consuming real- ing integration of simulated vibration data and sophisticated
world data collection, a common hurdle in previous studies. time-series feature extraction techniques, with a particular
Moreover, our model’s adaptability extends beyond specific emphasis on FFT data features. Our methodological inno-
machining configurations, offering the capability to extrapo- vation lies in the utilization of a RF classification model,
late to diverse setups. This work builds upon the foundation conceived to circumvent the constraints typically posed by
laid by prior research, such as the investigations conducted by limited real-world datasets. Such constraints often hinder
Yesilli and Khasawneh, which employed traditional feature the scalability and flexibility required for a comprehensive
extraction techniques like FFT, autocorrelation, and power chatter detection model. This study demonstrates the unique
spectral density [25]. potential of simulated datasets to replicate a diverse array of
milling scenarios, thus broadening the applicability of the RF
1.2.2 Streamlined feature selection for enhanced model model in practical settings.
efficiency
2.1 Cross-validation techniques and dataset size
While earlier studies employed a range of features, including importance
FFT, MPSE, MPE, or Laplacian scores, this study stream-
lined the approach by focusing solely on FFT and time-series To bolster the robustness and reliability of our predic-
data features. This strategic simplification aims to address the tive model, our research meticulously adopts k-fold cross-
challenges associated with feature selection in chatter clas- validation techniques, a method celebrated for its effec-
sification models. By reducing the feature set, the model’s tiveness in evaluating model performance. This approach
comprehensibility is enhanced, computational demands are provides a rigorous framework for assessing the model’s
reduced, and the risk of overfitting the training data is miti- robustness, significantly mitigating risks of overfitting by

123
The International Journal of Advanced Manufacturing Technology

Fig. 1 Example milling

machine setup

partitioning the data into training and testing subsets multi- 2.2 Data description and dataset division
ple times. Such a meticulous evaluation allows for a thorough
understanding of the model’s predictive capabilities, ensur- This comprehensive study delves into the intricacies of
ing that its performance is not just a result of chance or machining dynamics by employing a state-of-the-art sim-
overfitting to a particular set of data. ulated vibration data source. This data is meticulously
Furthermore, this study recognizes and emphasizes the generated from a sophisticated machining simulation model,
pivotal role of dataset size and diversity in the performance a brainchild of Schmitz and Smith as detailed in their sem-
of machine learning models. Large and diverse datasets are inal work, “Machining Dynamics: Frequency Response to
instrumental in enhancing the model’s ability to generalize Improved Productivity” [5]. This model is not just a mere
and make accurate predictions across various scenarios. In representation but a close emulation of the conditions expe-
this study, the substantial and carefully curated dataset, espe- rienced by a single milling machine, as visually encapsulated
cially with the deliberate exclusion of extrapolated data from in Fig. 1.
the training set, plays a crucial role. This strategic dataset The depth of this simulation is further exemplified through
composition reinforces the model’s proficiency, enabling it to Figs. 5 and 6, which provide a stark visual contrast between
operate effectively and consistently across a broad spectrum the raw vibration acceleration data for datasets characterized
of machining conditions. Through this combination of rigor- by the presence and absence of chatter. A notable observation
ous validation techniques and strategic dataset management, from this comparative analysis is the marked tenfold decrease
our approach ensures the utmost robustness and reliability of in acceleration values during instances of chatter, a testament
the predictive model in practical, real-world applications. to the precision and authenticity of the simulation.

Fig. 2 Simulated data curve

123
 The International Journal of Advanced Manufacturing Technology

Moreover, the subtle intricacies of this simulation are 2.3 Data sets A and A0
brought to the forefront in Figs. 7 and 8. These figures meticu-
lously illustrate data runs that teeter on the delicate threshold Data sets A and A0 are derived from the same machining tool
between stability and instability, a domain notoriously chal- setup, with variations in cutting depth and rotational speed
lenging for chatter prediction. This can best be seen in Fig. 2, of the machine tool head. The cutting parameters remain
where the instability zone is between the chatter and no chat- consistent, with a cutting force of 700 × 106 N /m 2 , feed per
ter areas. It can also be noted that in this dataset how stability tooth of 0.1 mm/tooth, and a tool diameter of 10 mm equipped
changes based on inputs like RPM and cutting depth, and with four cutting teeth. Set A consists of 4000 data files,
is not a one to one relationship. It is within this complex covering RPM settings ranging from 6000 to 10,000 rpm in
landscape that the prowess of FFT and time-series features, 50 rpm increments, with the cutting depth varying from 0.2
particularly those extracted via TSFresh, becomes indispens- to 10 mm in 0.2 mm increments. Set A0 comprises 2500 data
able. The FFT, with its ability to spotlight peak values at files, representing RPM settings from 10,050 to 12,000 rpm
specific frequencies, emerges as a critical tool in identifying in 50 rpm increments, with the cutting depth varying from
potential machine issues. When synergized with the nuanced 10.2 to 15 mm in 0.2 mm increments. Notably, Set A0 serves
insights provided by time-series features, a comprehensive as an extrapolated data set and is excluded from the training
framework for effective chatter classification is established, set for the RF classifier model. In Fig. 2, the example dataset
extending its applicability not only within but also beyond will first see chatter at a cutting depth of around 6 mm; this is
the confines of the training datasets A and A0. Doing this will in conjunction with the specific machine setup and will vary
allow the model to include the use of nonlinear parameters to based on inputs like cutting teeth, cutting force, material,
evaluate the degree of randomness, complexity, or similarity RPM, machine dampening devices, etc.
of the signals [28].
This simulation, with its detailed visual representations 2.4 Feature extraction and selection
and the integration of FFT and time-series features, under-
scores the nuanced differences in vibration acceleration In our goal to refine the predictive capabilities of our model
between stable operations and those prone to chatter. The while ensuring its operational efficiency, the method of
precision and depth of this simulated data echo the opera- recursive feature elimination (RFE) has been strategically
tional realities of a milling machine under varied conditions, incorporated into this model. This technique stands at the
offering an unprecedented level of detail and accuracy. Con- forefront of our feature selection process, iteratively pruning
sequently, this study not only leverages but also significantly the less significant features from our initial comprehen-
enriches the foundational work of Schmitz and Smith, offer- sive set. The essence of RFE lies in its ability to distill
ing novel insights and methodologies that hold the potential the core, most impactful features, thereby optimizing the
to transform the field of machining dynamics. model’s interpretability and significantly reducing compu-

Table 1 Set A0 and A0 FFT

Feature Description
features extracted
Acceleration peak (g) peak=max(|a|)
 n
Acceleration RMS (g) R M S = (1/n) i=1 ai 2
Crest factor Cr est = X peak /X R M S
 N
Standard deviation (g) std = N 1−1 i=1 (xi − x)2

Velocity RMS (in/s) V R M S = A2 /2

Displacement RMS (in) D R M S = (a/(2 ∗ pi ∗ f )2 )2
Peak frequency (Hz) Highest peak frequency (Hz)
 65
RMS (g) from 1 to 65hz R M S1−65 = (1/n) i=1 ai 2
 300
RMS (g) from 65 to 300hz R M S65−300 = (1/n) i=65 ai 2 r
 1000
RMS (g) from 300 to 6000hz R M S300−1000 = (1/n) i=300 ai 2 r

Note: FFT features are common in many industries using vibration data to determine the failure on rotating
components

123
The International Journal of Advanced Manufacturing Technology

tational demands [29, 30]. This rigorous approach not only 2.5 Data preprocessing
sharpens the model’s focus on the most influential attributes
but also serves as a preventive measure against the risk of In the data preprocessing phase of this study, a struc-
overfitting, ensuring robust and reliable predictions. tured approach was adopted for feature extraction from the
The application of RFE in our methodology has been collected raw vibration data during machining operations.
instrumental in achieving a streamlined feature set, intri- Custom Python scripts were meticulously developed for the
cately documented in our detailed tables. Table 1 is a extraction of FFT features. Concurrently, the TSFresh library
testament to this, presenting the FFT features meticulously was utilized for the extraction of time-series features, reflect-
extracted from each dataset, capturing the frequency-based ing a systematic integration of computational techniques
nuances integral to our analysis. Complementing this, Table 2 [31]. The transformation of raw data sets A and A0 into fea-
offers an exhaustive overview of the time-series features ture data sets B and B0 was methodically executed through
extracted from each dataset, illustrating the rich temporal the application of TSfresh and FFT methodologies. The chal-
patterns and dynamics encapsulated in our data. These tables lenges associated with discerning stability and instability
not only serve as a repository of the refined features but also from FFT data alone are highlighted in Figs. 9, 10, 11, and
underscore the thoroughness of our feature selection process. 12.
The result of this feature selection process is a refined and Table 2 details the time-series features extracted from each
optimized set of FFT and time-series features, which form data set, culminating in the generation of 846 features. Rec-
the foundational inputs for our RF classification model. By ognizing the potential risk of model overfitting and loss of
harnessing RFE, it ensures that the RF model is fed with generalizability due to the extensive number of features, a
only the most crucial features, significantly enhancing the process of RFE was employed. This process effectively dis-
model’s interpretability and computational efficiency. This tilled the feature set to 17 most significant features for the
strategic distillation of features sets the stage for a stream- RF model, enhancing the model’s predictive accuracy while
lined and effective RF classification model, poised to deliver maintaining its computational efficiency. These 17 features
precise and actionable insights. In essence, the integration of were categorized into two distinct groups: 10 FFT features
RFE in our feature selection process reflects our unwavering and 7 time-series features, each set contributing uniquely to
commitment to precision, efficiency, and clarity, driving the the model’s performance. The refined feature data sets are
development of a robust and interpretable predictive model designated to be utilized as input for the RF classification
for our analytical endeavors. model.

Table 2 Set A0 and A0

Feature Description
TSFresh features extracted
Ratio value number to time series length The number of unique values versus the
total number of values
Benford correlation How often a value starts with a certain
number, in analytics overwhelmingly a
value is most likely to start with a 1
Change quant f-agg "var" False qh 1.0 ql 0.4 Aggregator function of the differences
taken over a specific range of upper and
lower quartiles
FFT coefficient attr "imag" coeff 55 Fast Fourier Transform of the imaginary
part of the data with a coefficient of 55
FFT coefficient attr "imag" coeff 77 Fast Fourier Transform of the imaginary
part of the data with a coefficient of 77
Agg linear trend "stderr" len 10 f agg "min" Linear least squares regression for cer-
tain attributes for a certain number of time
series data points
Permutation entropy dimension 4 tau 1 Counts the frequency of permutation and
returns the appropriate entropy, this is a
complexity measure for time series data

123
 The International Journal of Advanced Manufacturing Technology

The transformation of raw datasets into precisely refined 2.7 Evaluation metrics
feature sets was achieved through the diligent application of
custom Python scripts alongside the TSFresh library, ensur- The framework of this study delineates the suite of eval-
ing a meticulous analysis of each feature’s significance in uation metrics that have been systematically employed to
the context of machine vibration analysis. The preprocessing assess the performance of our RF model in the prediction of
phase is critically underscored as instrumental in accurately chatter within milling operations. The explication of these
identifying the markers of stability and instability within the metrics is paramount for a comprehensive understanding
datasets, as depicted in our figures. This systematic approach of the model’s effectiveness and its connection to practical
to feature extraction and selection forms the cornerstone implications within the operational environment.
of the RF classification model, ensuring a robust analytical Central to our evaluative strategy is the employment of the
framework for the prediction of chatter in milling operations. Area Under the Curve (AUC) metric. The AUC metric stands
as a pivotal and robust indicator of the model’s classification
accuracy, offering an integrative assessment that considers
the entirety of possible classification thresholds. AUC val-
2.6 Model development ues range from 0.5 to 1.0. An AUC of 0.5 suggests that the
model performs no better than random chance, while an AUC
In the initial phase of feature engineering, a comprehensive of 1.0 indicates perfect classification. Values greater than 0.5
correlation analysis was undertaken to discern the predic- demonstrate the model’s ability to distinguish chatter from
tive value of the generated features in the context of the no chatter with better-than-random accuracy. Specifically, it
vibration data acquired during machining processes. Notable quantifies the percentage at which the model correctly pre-
among these features were the aggregated FFT outputs, along dicts chatter occurrences. For instance, a model with a 97%
with statistical measures such as the spectral centroid, mean, AUC score signifies a 97% accuracy rate in predicting chatter.
variance, skewness, and kurtosis of the absolute Fourier This assessment allows us to ascertain the model’s generality
spectrum. Despite preliminary indications suggesting their and its potential to be effectively deployed across a broader
potential utility, further analysis relegated these features to spectrum of machining speeds and cutting depths. The simu-
the non-predictive category. lated data contains a chatter metric that indicates whether or
Proceeding to the feature elimination phase, the RFE strat- not the dataset experienced chatter. This can be seen in Fig. 2
egy was employed with methodical precision. This strategy below. From this, you will see the line between chatter and
facilitated the exclusion of non-contributory features, thereby no chatter. It will be in this zone that will be the most difficult
refining the feature set to a succinct ensemble that is computa- to predict chatter or no chatter, given the minute changes in
tionally less demanding and more aligned with the predictive variables that can influence whether or not chatter is experi-
exigencies of the model. enced.
The development of the model, as systematically outlined In concert with the AUC, the robustness of the model’s
in Fig. 13, underscores our exacting approach to feature selec- predictive validity is further reinforced through the imple-
tion. This process was instrumental in ensuring the retention of mentation of k-fold cross-validation. This methodological
only those features that exhibited substantial predictive merit. approach enables a comprehensive validation of the model’s
The RF analysis was characterized by judicious employment performance by systematically partitioning the data into sub-
of various training-to-test data splits, encompassing ratios sets, thus ensuring that the evaluation is reflective of the
of 70-30, 50-50, and 40-60, to validate the model’s predictive model’s ability to generalize across various subsets of data.
robustness across a spectrum of scenarios. Moreover, the significance of dataset size in the context
In essence, this study’s methodology exemplifies a para- of model evaluation is highlighted, recognizing that the vol-
digm of structured feature selection that is critical to the ume and diversity of data are critical to the model’s learning
construction of an accurate and computationally efficient and subsequent predictive performance. By incorporating a
RF classification model. The approach delineates a clear substantial and diverse dataset, it ensures that the model’s
trajectory from the identification through to the elimina- utility is not confined to theoretical or controlled scenarios
tion of non-essential features, thus sharpening the model’s but extends to practical, real-world machining environments.
predictive acuity and enhancing its operational expediency. In summary, the evaluation framework adopted herein is
This structured approach not only highlights the paramount both rigorous and multifaceted, designed to thoroughly scru-
importance of feature relevance in RF analysis but also tinize the model’s efficacy. It is through this comprehensive
exemplifies the adaptability and analytical rigor requisite for battery of metrics, with the AUC as the cornerstone of our
handling complex machining data. evaluation, supported by k-fold cross-validation and a sub-

123
The International Journal of Advanced Manufacturing Technology

stantive dataset, that this study affirms the robustness and Study 3: Generalizability of RF models in extrapolated
practical utility of our RF model in the realm of chatter pre- data
diction for milling operations. Objective: This segment explores the capability of RF
models to sustain predictive accuracy and reliability when
applied to extrapolated data, thereby testing the models’
3 Investigation adaptability in broader application scenarios.

This study undertakes a methodological investigation through Step 1: Model development for extrapolated data: RF mod-
a series of three distinct analyses designed to rigorously els are adeptly developed for sets i ∈ {A, A0} and
assess the robustness of key vibration and time-series fea- j ∈ {B, B0}, catering to both original and extrapo-
tures, evaluate the predictive acumen of RF models within lated data contexts.
diverse datasets, and scrutinize the generalizability of the Step 2: Performance appraisal for extrapolated data: The
models with respect to extrapolated data. models’ performance within each dataset is scrupu-
Study 1: Consistency of key features within Set i ∈ lously measured using the AUC metric, represented
{A, A0} as AU C B,B , for dataset j ∈ {B, B0}.
Objective: This analytical segment is devoted to clarifying Step 3: Comprehensive evaluation of generalizability: An
the consistency of salient features in machining vibration data extensive assessment of the models’ performance
across identical tool configurations. is conducted across all datasets i ∈ A, A0 and
sets j ∈ B, B0 utilizing the AUC metric, denoted
Step 1: Feature identification and evaluation: The study as AU C A,B , AU C A0,B , AU C B,A , and AU CcB,A0 ,
commences with the identification and subsequent thus affirming the models’ adaptability and forecast-
evaluation of paramount features, denoted by the set ing potential.
t Fi „ within each dataset i ∈ {A, A0}
Step 2: Intersection of feature sets: The investigation pro- This systematic and rigorous investigation ensures an in-
ceeds to discern the intersection of feature sets, depth examination of feature consistency, model predictive-
represented by FA ∩ FA0 , to gauge the uniformity ness, and generalizability, contributing invaluable insights
of feature significance across datasets. into the potential of RF models to advance chatter prediction
Step 3: Quantification of common features: The cardinal- in machining operations.
ity, denoted as n(FA ∩ FA0 ), quantifies the common
features, thereby providing a metric for consistency
across the datasets. 4 Results

Study 2: Generalizability of RF models within Set i ∈ This section delineates the findings of our comprehensive
{A, A0} investigation, each study contributing nuanced insights into
Objective: This phase aims to assess the generalizability the effectiveness, precision, and adaptability of our RF model
of the RF models in predicting chatter across identical tool in the context of chatter prediction.
configurations, thereby establishing the models’ predictive
stability. 4.1 Features and descriptive statistics

Step 1: RF model construction: Individual RF models are The initial phase of feature extraction, facilitated by the
meticulously constructed for each dataset i ∈ TSFresh Python package, yielded an exhaustive array of
{A, A0} 846 features, spanning both time-series and FFT-based vibra-
Step 2: Intra-dataset performance assessment: The predic- tion attributes. A methodical feature reduction process was
tive performance of each model within its respective undertaken to distill this extensive set down to 17 paramount
dataset is rigorously evaluated using the AUC met- features, thereby ensuring the model’s efficiency and predic-
ric, denoted as AU C cA,A and AU C A0,A0 . tive strength.
Step 3: Cross-dataset performance evaluation: The mod-
els’ predictive acumen is further scrutinized across 4.2 Investigation results
datasets i ∈ A, A0 employing the AUC metric, rep-
resented as AU Cc A,A0 and AU C A0,A , ensuring a The investigation unfolded through a series of structured
comprehensive assessment of the models’ robust- studies, each clarifying distinct facets of the model’s per-
ness. formance and the viability of the implemented features.

123
 The International Journal of Advanced Manufacturing Technology

Fig. 3 Set A and A0 ten

vibration features

4.2.1 Study 1 results The AUC metrics for Set A serve as a testament to the
predictability of different feature groups within the datasets.
The first analytical study illuminated the internal consis- Specifically, the predictability of vibration features within Set
tency of key features across sets A and A0, corroborating the A, denoted as AU C A,Avibe , reached the ceiling value of 1.000.

model’s acuity in identifying features indicative of chatter. In contrast, the predictability of time-series features, denoted
Figures 3 and 4 elegantly illustrate this congruence, pre- time , was slightly lower at 0.995. However, the
as AU C A,A
senting the top features in both vibration and time-series top_15
combined feature predictability, represented by AU C A,A ,
categories for the respective sets. achieved a score of 1.000, confirming that vibration features
marginally outperform time-series features in terms of impor-
tance for this specific dataset.
4.2.2 Study 2 results Similarly, the results for Set A0 align closely with those
of Set A, manifesting high classification accuracy across all
The analyses conducted in Study 2 furnish compelling evi- examined feature groups and reinforcing the model’s capa-
dence regarding the high predictability and generalizability bility to generalize effectively. The predictability of vibration
features within Set A0, as indicated by AU C A0,A0 vibe , was
of the RF models across the datasets in classifying chatter.
The empirical results for Set A were remarkably robust, with reported at 0.992, while the predictability of time-series
features within the same set, as indicated by AU C A0,A0 time ,
all AUC values surpassing the threshold of 0.997, indica-
tive of the model’s profound predictive precision, as detailed attained the optimal value of 1.000. When considering the
in Table 3. Table 4 details the results for set A0. This level combined feature predictability for Set A0, as represented
top_15
of accuracy establishes the model as exceptionally reliable by AU C A0,A0 , the model still exhibits an impressive AUC
within the dataset population under study. value of 0.997.

Fig. 4 Set A and A0 time-series

features

123
The International Journal of Advanced Manufacturing Technology

Table 3 Set A Random Forest

Model features AUC Sensitivity Specificity
model predictability
Combinds features 1.000 ± 0.000 1.000 ± 0.000 1.000 ± 0.000
Vibration features 0.995 ± 0.005 0.989 ± 0.013 0.972 ± 0.033
Time-series features 1.000 ± 0.000 1.000 ± 0.000 1.000 ± 0.000

In summary, the RF model’s exemplary performance is analysis. The initial challenge, presented by the daunting
underlined by the AUC values, which serve as robust indi- quantity of features generated by the TSFresh package, was
cators of the model’s ability to classify chatter with high surmounted by the judicious application of RFE. This tool
accuracy. The consistency of these results across different proved instrumental in distilling the essence of the dataset,
feature sets and dataset variations underscores the model’s selectively sifting through the multitude of features to iden-
remarkable generalizability. The predictive precision of the tify those with paramount significance for chatter prediction.
model, as illustrated by near-perfect AUC values, instills a The integration of selected features, encompassing both
high degree of confidence in its deployment for practical time-series data and traditional FFT-based vibration analysis,
applications within the domain of chatter classification in has substantially augmented the model’s predictive accuracy
milling operations. and generalizability. This strategic amalgamation has not only
enhanced the model’s capabilities in stability classification
4.2.3 Study 3 results but has also achieved this enhancement without delving into
more intricate and computationally intensive methodologies.
Study 3 reveals favorable predictability of the RF models Such a balance between simplicity and effectiveness is seldom
within dataset B, as outlined in Table 5. The predictability of achieved, underscoring the innovative edge of our approach.
vibration features within Set B, represented by AU C B,Bvibe , is The empirical findings of our research outline a pro-
0.824, while the predictability of time-series features within nounced synergy between feature importance and classi-
time is 0.869. The combined feature predictability
Set B AU C B,B fication accuracy. This synergy was consistently observed
top 17 across datasets A and A0, affirming the adeptness of RFE
for Set B AU C B,B stands at 0.879.
in unearthing the most predictive features. Furthermore, the
Study 3 also demonstrates the RF model’s ability to extrap-
application of our RF model to extrapolated data in set B
olate data beyond the original training dataset’s range. This
demonstrated notable generalizability. Achieving an impres-
extrapolation includes machine tool rotations per minute and
sive classification accuracy of 87% on data that extends
cutting depth values outside the training dataset’s range.
beyond the initial training scope, the model exhibited a mod-
Notably, the combination of vibration and time-series fea-
est decline in performance, a testament to its robustness and
tures yields the best generalizability results.
adaptability across diverse operational conditions.
The practical implications of our findings are profound.
The RF model, harnessed with FFT and time-series features,
5 Discussion epitomizes efficiency and deployability. Its capacity for swift
integration into the shop floor, coupled with its accessibility
This section provides a comprehensive interpretation of the to professionals across a spectrum of expertise levels, addresses
outcomes of our investigation, underscoring the significance the pressing demands of the contemporary industrial milieu,
of the methodological simplicity and operational efficiency which places a premium on agility and practical utility.
inherent in our approach. These attributes not only contribute
to the robustness of our model but also render it a pragmatic
asset within the realm of industrial machining applications. 6 Conclusion
Our study leverages the capabilities of RFE within the
framework of RF models, a methodological choice that This investigation concludes with a reflective analysis of the
substantially streamlines the complexity of vibration data strides made through our methodological approach, charac-

Table 4 Set A0 Random Forest

Model features AUC Sensitivity Specificity
model predictability
Combined features 0.997 ± 0.003 0.994 ± 0.002 0.990 ± 0.008
Vibration features 0.992 ± 0.008 0.972 ± 0.032 0.957 ± 0.048
Time-series features 1.000 ± 0.000 1.000 ± 0.000 1.000 ± 0.000

123
 The International Journal of Advanced Manufacturing Technology

Table 5 Set B Random Forest

Model features AUC Sensitivity Specificity
model predictability
Combined features 0.879 ± 0.121 0.872 ± 0.151 0.869 ± 0.115
Vibration features 0.824 ± 0.076 0.812 ± 0.180 0.808 ± 0.178
Time-series features 0.869 ± 0.131 0.861 ± 0.142 0.852 ± 0.109

terized by its intrinsic simplicity and operational efficacy. The ultimate ambition of these future explorations is to
The study has demonstrated the competence of a well-crafted facilitate a transition from retrospective to real-time pre-
RF classification model in accurately categorizing simulated dictive modeling. By harnessing the synergistic potential
vibration data for the detection of chatter in machining pro- of voluminous, heterogeneous datasets, and real-time data
cesses. The dataset, albeit comprehensive within its current integration, the goal is to evolve a model that not only
scope, encompasses less than 10,000 data files. In light of the prognosticates chatter occurrences with heightened preci-
potential for augmented model robustness and generalizabil- sion but also does so in an online, proactive manner. Such a
ity, future research endeavors will be directed toward expanding leap towards real-time predictive modeling is poised to rev-
the dataset substantially, aiming to surpass the threshold of olutionize the domain of chatter detection, heralding new
100,000 datasets. This expansion is anticipated to substan- possibilities for predictive maintenance and operational effi-
tially refine the model’s predictive precision and its adeptness ciency in the manufacturing sector.
in navigating the multifaceted landscape of operational con- This research substantiates the viability of a streamlined
ditions prevalent in additive manufacturing settings. novel yet insightful approach to technological innovation in
Furthermore, the union of empirical data with simu- the realm of industrial machining. It signals the industry to
lated data presents a promising avenue for exploration. The recognize and adopt this paradigm, which underscores that
reliance on purely simulated data, while instrumental in practical, impactful solutions can indeed emerge from sim-
establishing the foundational robustness of the model, may plicity and strategic ingenuity. As research progresses, our
not fully encapsulate the stochastic nature of real-world oper- endeavor will remain steadfastly focused on broadening the
ational settings. To surmount this limitation, forthcoming empirical base of our model, enriching its predictive faculties,
research initiatives will concentrate on integrating empirical and ultimately, endowing the industrial machining domain
data with simulated datasets. This integrative approach seeks with a tool of unparalleled value and utility.
to leverage the granularity and unpredictability inherent in
real-world data, thereby enhancing the model’s predictive
acumen and its adaptability to diverse operational paradigms. Appendix A Figures

Fig. 5 Example vibration acceleration, stable

123
The International Journal of Advanced Manufacturing Technology

Fig. 6 Example vibration acceleration, unstable

Fig. 7 Example vibration acceleration slightly below chatter line, stable

Fig. 8 Example vibration acceleration slightly above chatter line, unstable

123
 The International Journal of Advanced Manufacturing Technology

Fig. 9 Example FFT data

indicating clear instability

Fig. 10 Example FFT data indicating clear stability

Fig. 11 Example FFT data slightly above chatter line, unstable

123
The International Journal of Advanced Manufacturing Technology

Fig. 12 Example FFT data slightly below chatter line, stable

Fig. 13 The research approach

includes feature pruning,
recursive feature elimination,
and RF analysis. Steps in this
approach are provided with a
generalized flow from top to
bottom

123
 The International Journal of Advanced Manufacturing Technology

Author Contributions Conceptualization, creation of data pipeline, and 4. Quintana G, Ciurana J (2011) Chatter in machining processes: a
experimentation within data pipeline done by MA. Article writing is review. Int J Mach Tools Manuf 51(5):363–376. https://doi.org/10.
done by MA with key contributions for machining from JK, BJ, and 1016/j.ijmachtools.2011.01.001
TS. Data acquisition for simulated data performed by MA. Review of 5. Schmitz T, Smith K (2019) Machining dynamics: frequency
paper done by JC, BJ, JK, AK, TS, and SS. response to improved productivity, 2nd edn. Springer, New York,
NY
Funding The research leading to these results received funding in part 6. Shevchik SA, Saeidi F, Meylan B, Wasmer K (2017) Prediction of
from UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the failure in lubricated surfaces using acoustic time-frequency fea-
US Department of Energy (DOE). The US government retains and the tures and random forest algorithm. IEEE Trans Industr Inform
publisher, by accepting the article for publication, acknowledges that the 13(4):1541–1553. https://doi.org/10.1109/tii.2016.2635082
US government retains a nonexclusive, paid-up, irrevocable, worldwide 7. Wu SD, Wu PH, Wu CW, JJD, Chun Chieh Wang Chun-Chieh
license to publish or reproduce the published form of this manuscript, or Wang (2012) Scilit Preprints.orgGoogle Scholar View Publications
allow others to do so, for US government purposes. DOE will provide 3, Bearing fault diagnosis based on multiscale permutation entropy
public access to these results of federally sponsored research in accor- and support vector machine. Entropy 14(8):1343–1356. https://doi.
dance with the DOE Public Access Plan (http://energy.gov/downloads/ org/10.3390/e14081343
doe-public-access-plan). 8. Heng RB, Nor MJM (1998) Statistical analysis of sound and
The authors also gratefully acknowledge seed funding from the Univer- vibration signals for monitoring rolling element bearing con-
sity of Tennessee-Oak Ridge Innovation Institute (UT-ORII) to partially dition. Appl Acoust 23(1–3):211–226. https://doi.org/10.1016/
support this research. s0003-682x(97)00018-2
9. Rao KR, Kim DN, Hwang JJ (2010) Fast Fourier Transform - algo-
Data availability Machining files and derived data are not publicly rithms and applications. Springer, New York, NY
available. 10. Huo Z, Zhang Y, Francq P, Shu L, Huang J (2017) Incipient fault
diagnosis of roller bearing using optimized wavelet transform based
Code availability Code written in this study is not publicly available. multi-speed vibration signatures. IEEE Access 25:19,442–19,456
https://doi.org/10.1109/access.2017.2661967
11. Ma X, Hu J, Zhang L (2017) EMD-based online filtering of pro-
Declarations
cess data. Control Eng Pract 62:79–91. https://doi.org/10.1016/j.
conengprac.2017.03.008
12. Yang J, Wright J, Huang TS, Ma Y (2010) Image super-resolution
Ethics approval The authors claim that there are no ethical issues via sparse representation. IEEE Trans Image Process 19(11):2861–
involved in this research. 2873. https://doi.org/10.1109/TIP.2010.2050625
13. Yesilli MC, Khasawneh FA (2021) On transfer learning of tradi-
Consent to participate All the authors consent to participate in this tional frequency and time domain features in turning. Electr Eng
research and contribute to the research. Syst Sci doi 10(1115/1):0005358v
14. LiDe X, Chen QMC (2010) An effective EMD-based feature
Consent for publication All the authors consent to publish the research. extraction method for boring chatter recognition. Appl Mech
There are no potential copyright/plagiarism issues involved in this Mater 34–35:1058–1063. https://doi.org/10.4028/www.scientific.
research. net/amm.34-35.1058
15. Chen GS, Zheng QZ (2017) Online chatter detection of the end
Conflict of interest The authors declare no competing interests. milling based on wavelet packet transform and support vector
machine recursive feature elimination. J Adv Manuf Technol
Open Access This article is licensed under a Creative Commons 95:775–784. https://doi.org/10.1007/s00170-017-1242-9
Attribution 4.0 International License, which permits use, sharing, adap- 16. Konar P, Chattopadhyay P (2011) Bearing fault detection of induc-
tation, distribution and reproduction in any medium or format, as tion motor using wavelet and support vector machines (SVMS).
long as you give appropriate credit to the original author(s) and the Appl Soft Comput 11(6):4203–4211. https://doi.org/10.1016/j.
source, provide a link to the Creative Commons licence, and indi- asoc.2011.03.014
cate if changes were made. The images or other third party material 17. Yesilli MC, Khasawneh FA, Mann BP (2022) Transfer learning
in this article are included in the article’s Creative Commons licence, for autonomous chatter detection in machining. Int J Mach Tools
unless indicated otherwise in a credit line to the material. If material Manuf 80:1–27. https://doi.org/10.1016/j.jmapro.2022.05.037
is not included in the article’s Creative Commons licence and your 18. Wan S, Li X, Yin Y, Hong J (2021) Milling chatter detection by
intended use is not permitted by statutory regulation or exceeds the multi-feature fusion and Adaboost-SVM. Mech Syst Signal Pro-
permitted use, you will need to obtain permission directly from the copy- cess 156(107):671. https://doi.org/10.1016/j.ymssp.2021.107671
right holder. To view a copy of this licence, visit http://creativecomm 19. Gradišiek J, Baus A, Govekar E, Klocke F, Grabec I, (2003)
ons.org/licenses/by/4.0/. Automatic chatter detection in grinding. Mech Syst Sig-
nal Process 43(14):1397–1403. https://doi.org/10.1016/s0890-
6955(03)00184-6
20. Li K, He S, Li B, Liu H, Mao X, Shi C (2020) A novel online
References chatter detection method in milling process based on multiscale
entropy and gradient tree boosting. Mech Syst Signal Process
1. Farrar C, Doebling S (1999) Damage detection II: field applications 135(14):106,385 https://doi.org/10.1016/j.ymssp.2019.106385
to large(Kluwer Academic Publishers) 21. Aslan D, Altintas Y (2018) On-line chatter detection in milling
2. Goyal D, Pabla B (2016) The vibration monitoring methods and using drive motor current commands extracted from CNC.
signal processing techniques for structural health monitoring: a Int J Mach Tools Manuf 132:64–80. https://doi.org/10.1016/j.
review. Arch Comput Methods Eng 23(4):585–594 ijmachtools.2018.04.007
3. Carden EP, Fanning P (2004) Vibration-based condition monitor-
ing: a review. Struct Health Monit 3(4):355–377

123
The International Journal of Advanced Manufacturing Technology

22. Li Z, Song Q, Jin P, Liu Z, Wang B, Ma H (2023) Chatter suppres- 28. Navarro-Devia JH, Chen Y, Dao DV, Li H (2023) Chatter detection
sion techniques in milling processes: astate of the art review. Chin in milling processes-a review on signal processing and condition
J Aeronaut. https://doi.org/10.1016/j.cja.2023.10.001 classification. Int J Adv Manuf Technol 125(9–10):3943–3980.
23. Nadi A, Moradi H (2019) Increasing the views and reducing the https://doi.org/10.1007/s00170-023-10969-2
depth in random forest. Expert Syst Appl 138(112):801. https:// 29. Xia S, Yang Y (2023) A model-free feature selection technique of
doi.org/10.1016/j.eswa.2019.07.018 feature screening and random forest-based recursive feature elim-
24. Feng Z, Yi H, Shu A, Tang L (2023) Simulation of grinding surface ination. Int J Intell Syst pp. 1–16 https://doi.org/10.1155/2023/
topography considering wheel wear and wheel vibration. Int J Adv 2400194
Manuf Tech 130(1–2):475–490. https://doi.org/10.1007/s00170- 30. Wang Q, Song L, Zhao J, Wang H, Dong L, Wang X, Yang Q
023-12675-5 (2022) Application of the gradient boosting decision tree in the
25. Yesilli MC, Khasawneh FA (2021) On transfer learning of tradi- online prediction of rolling force in hot rolling. Int J Adv Manuf
tional frequency and time domain features in turning. Am Soc Mech Technol 125(1–2):387–397. https://doi.org/10.1007/s00170-022-
Eng doi 10(1115/1):0005358v 10716-z
26. Dutta C, Nagarajan N, Adhoni ZA, Dhandapany SK, Alagarsamy 31. Christ M, Braun N, Neuffer J, Kempa-Liehr AW (2018) Time series
A, Chakravarthy AK, Praveen N, Pushpanathan G, Kannan S feature extraction on basis of scalable hypothesis tests (tsfresh - a
(2023) Outlier detection in additive manufacturing using novel python package). Neurocomputing 307:72–77. https://doi.org/10.
machine learning algorithm. Int J Adv Manuf Tech. https://doi. 1016/j.neucom.2018.03.067
org/10.1007/s00170-023-12798-9
27. Jia W, Sun M, Lian J, Hou S (2022) Feature dimensionality reduc-
tion: a review. Complex Intell Syst 8(3):2663–2693. https://doi.
Publisher’s Note Springer Nature remains neutral with regard to juris-
org/10.1007/s40747-021-00637-x
dictional claims in published maps and institutional affiliations.

Authors and Affiliations

Matthew Alberts1 · Sam St. John1 · Bradley Jared3 · Jaydeep Karandikar2 · Anahita Khojandi1 ·
Tony Schmitz2,3 · Jamie Coble4

Matthew Alberts Tony Schmitz

[email protected] [email protected]
Sam St. John 1 Department of Industrial and Systems Engineering, University
[email protected]
of Tennessee, 851 Neyland Drive, Knoxville, TN 37996, USA
Bradley Jared 2 Manufacturing Science Division, Oak Ridge National
[email protected]
Laboratory, 1 Bethel Valley Road, Oak Ridge, TN 37830, USA
Jaydeep Karandikar 3 Department of Mechanical, Aerospace, and Biomedical
[email protected]
Engineering, University of Tennessee, 1512 Middle Drive,
Anahita Khojandi Knoxville, TN 37996, USA
[email protected] 4 Department of Nuclear Engineering, University of Tennessee,
863 Neyland Drive, Knoxville, TN 37996, USA

123