This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON RELIABILITY 1

Two-Dimensional Optimization Framework of

Online Interpretable Time-Frequency Feature
Learning for Practical Machine Health Monitoring
Tongtong Yan , Graduate Student Member, IEEE, Dong Wang , Member, IEEE,
Tangbin Xia , Member, IEEE, Lifeng Xi , and Min Xia

Abstract—Data-driven feature extraction for machine health nearing failure. Consequently, it is essential to keep assessing the
monitoring has garnered significant attention, yet two key limi- operational performance of machines on a regular basis. Ongo-
tations remain unaddressed: lack of interpretability and the need
ing monitoring of machine health helps in preventing unexpected
for extensive historical fault data. To overcome these problems,
an online two-dimensional optimization framework is proposed breakdowns and devising cost-effective maintenance strategies.
that enables interpretable time-frequency feature extraction and The availability of extensive monitoring data has spurred
health index (HI) construction without requiring faulty samples advancements in data-driven continuous health monitoring tech-
for model training. Our approach introduces a convex hull-based nologies. For instance, Xu et al. [4] combined a stacked
closest point optimization model for estimating time-frequency
auto-encoder with an exponential function to produce highly
instances and learning interpretable time-frequency features. By
leveraging a small set of baseline vibration samples and recent on- monotonic health indices (HIs). Ma et al. [5] applied wavelet
line data, rapid fault diagnosis can be achieved based on optimized packages, statistical features, and Grassmann manifold tech-
interpretable time-frequency features. This method also facilitates niques for machine performance evaluation through locally lin-
long-term degradation tracking by constructing and updating an ear embedding. Wu et al. [6] incorporated an inverse hyperbolic
HI from collected time-frequency spectrograms. Once machine
tangent function in a multiscale convolutional neural network
faults appear, updated time-frequency features can show apparent
and interpretable fault signatures for prompt fault alarming. More- to construct HIs for trend prediction, using frequency signals
over, the proposed framework allows continuous HI updates for as input. Duong et al. [7] used frequency amplitudes around
incipient fault detection and degradation tracking. The proposed fault characteristic frequencies in a deep convolutional neural
framework is validated by using two run-to-failure datasets and network for bearing fault detection. Qin et al. [8] implemented
ablation experiments are conducted to demonstrate its superiority.
a knowledge-based loss function in a variational autoencoder to
Index Terms—Closest point optimization model, convex hull, create HIs for continuous monitoring. To overcome the challenge
health index (HI), interpretable time-frequency feature, machine of manual feature extraction, several end-to-end training frame-
health monitoring. works have been introduced for continuous health monitoring.
For example, Wen et al. [9] used multilevel fusion and genetic
programming to create an end-to-end system for HI development
I. INTRODUCTION
and trend prediction. Shi and Chehade [10] implemented a two-
ERFORMANCE deterioration can happen at any time dur-
P ing the operational phase [1], [2]. Given that the decline
in machine performance typically unfolds over an extended
stage continuous monitoring approach using long short-term
memory networks.
While purely data-driven approaches offer significant poten-
period, it is crucial to implement ongoing health monitoring of tial to improve model performance and reduce reliance on expert
machinery [3]. This process involves gathering condition data knowledge, they often lack transparency and reliability in prac-
from machines ranging from those in optimal condition to those tical applications. Consequently, integrating data-driven meth-
ods with signal-processing techniques is frequently necessary.
Signal processing features are more interpretable and practical.
Received 22 November 2023; revised 27 May 2024 and 27 August 2024; However, traditional time-domain and frequency-domain data
accepted 27 October 2024. This work was supported by the National Natural processing methods fall short in handling nonstationary and
Science Foundation of China under Grant 52475112. Associate Editor: E. Pohl.
(Corresponding author: Dong Wang.) nonlinear signals for weak fault signature enhancement [11].
Tongtong Yan and Min Xia are with the Department of Mechanical and Time-frequency methods, such as short-time Fourier transform
Materials Engineering, The University of Western Ontario, London, ON N6A (STFT) and wavelet transform (WT), are more effective for these
3K7, Canada (e-mail: [email protected]; [email protected]).
Dong Wang, Tangbin Xia, and Lifeng Xi are with the Department of Industrial types of signals. Numerous studies have combined these tech-
Engineering and Management, School of Mechanical Engineering, and the niques with data-driven models for machine health monitoring.
State Key Laboratory of Mechanical System and Vibration, Shanghai Jiao For instance, Cao et al. [12] used time-frequency representations
Tong University, Shanghai 200240, China (e-mail: [email protected];
[email protected]; [email protected]). of vibration signals with marginal spectra and applied them
Digital Object Identifier 10.1109/TR.2024.3489589 to a temporal broad learning system for health assessment.

1558-1721 © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON RELIABILITY

Du et al. [13] converted data into two-dimensional images using framework in detecting incipient machine faults and tracking
continuous WT and employed these images in a convolutional degradation by identifying informative frequency bands and
neural network with global attention for trend evaluation. Hu providing progressively updated HIs.
et al. [14] fed STFT-derived grey images into a deep convolu- This article is structured as follows. Section II explores the
tional network for assessing performance degradation. Time- signatures and characteristics of faulty samples using simulation
frequency techniques have shown significant advantages in signals and introduces convex hulls to analyze time-frequency
machine health monitoring across various applications. How- maps. Section III presents a convex hull-based closest point
ever, practical implementation faces several challenges. Data- optimization model and proposes an online two-dimensional
driven methods often lack interpretability, leaving engineers optimization framework for practical health monitoring. Sec-
unable to fully understand how these models make decisions. tion IV demonstrates the application of this framework using two
In addition, these models typically require faulty samples for run-to-failure datasets, illustrating how it continuously extracts
training, which are rarely available, particularly for new ma- physics-informed features from the time-frequency domain and
chines still in their early stages. Although time-frequency maps constructs HIs for real-time monitoring. In addition, perfor-
are widely used, their integration into opaque models limits the mance comparisons and ablation studies are provided to evaluate
understanding of critical frequency bands and fault characteris- the proposed optimization model and HIs.
tics. Consequently, these approaches remain largely inscrutable.
The convex hull is a fundamental geometric concept, defined II. FAULT CHARACTERISTICS IN THE TIME-FREQUENCY
as the smallest convex set enclosing a dataset. Recently, opti- DOMAIN AND ITS GEOMETRIC REPRESENTATION BASED
mization models based on convex hulls have found applications ON CONVEX HULLS
in machine health monitoring, primarily for fault diagnosis, This section introduces the mathematical basis of STFT, a
where they serve as classification tools. These models estimate famous time-frequency analysis technique. It then discusses the
the distribution of different fault classes based on convex hulls time-frequency domain characteristics of fault signals based on
and use distance metrics to classify new samples. For instance, STFT. Following this, convex hulls are utilized to represent
Zeng et al. [15] proposed a flexible convex hull based classifi- time-frequency maps in matrix form, allowing for a compact
cation model and applied this model to machine fault diagnosis expression and characterization of geometric features across
[16]. Li et al. [17] enhanced this approach by integrating the various health states.
Riemannian manifold into the model, while He et al. [18] intro-
duced a kernel-based, flexible convex hull for fault classification. A. Fault Signatures in Time-Frequency Representation
Another study [19] proposed a convex hull-based degradation
model for analyzing spectral lines, though it was limited to Vibration signals from machines often exhibit nonlinear
frequency amplitudes and could not handle time-varying signals. and nonstationary characteristics, especially in harsh operat-
In addition, this approach requires faulty samples for training, ing conditions. Traditional spectral analysis methods such as
which are often not available. the Fourier transform are inadequate for such signals, whereas
In summary, existing data-driven models for continuous time-frequency techniques, including STFT and WT, are more
health monitoring face two key problems. First, despite the suitable. This study emphasizes fault features within the time-
integration of signal processing algorithms, these models still frequency domain, using STFT as an example to convert
suffer from interpretability problems. Second, they rely on fault time-domain vibration signals into time-frequency spectro-
data for model training. To overcome these limitations, this grams. STFT operates by applying a windowed Fourier trans-
study proposes a two-dimensional optimization framework for form to capture local spectral information. For a continuous
online, interpretable time-frequency feature extraction and HI signal s(t), the STFT of the s(t) can be mathematically written
construction, removing the reliance on faulty data for model as follows [20]:
 ∞
training. The key contributions of this article are summarized
as follows. First, it thoroughly examines the time-frequency TFs (t, f ) = [s (τ ) g ∗ (τ − t)] e−j2πf τ dτ (1)
−∞
characteristics of fault vibration signals using simulation data. A
convex hull-based closest point optimization model is proposed where TFs means the time-frequency map of s(t); g(τ − t) is a
to extract interpretable time-frequency features by approximat- window function with a short time length; ∗ means the operation
ing two-dimensional maps with convex hulls. Second, a novel of complex conjugate. It is observed from (1) that the STFT
online two-dimensional optimization framework is introduced of the s(t) at time t can be regarded as the Fourier transform
for practical health monitoring. This framework can identify performed by multiplying the signal s with the moving window
informative time-frequency bands without requiring faulty sam- g; for a discrete signal s(k), the STFT of the s(k) can be obtained
ples, enabling rapid fault diagnosis. Long-term monitoring data by discretizing the STFT in (1) as follows:
∞

of time-frequency spectra and their optimized features are used
to update an HI for incipient fault detection and degradation TFs (μ, υ) = s (k) g ∗ (kT − μT )e−j2π(υF )k (2)
tracking. Finally, the study demonstrates the proposed method- k=−∞

ology with two run-to-failure datasets, showcasing continuous where T > 0 and F > 0 are respectively sampling periods for
health monitoring without faulty samples for model training. time and frequency; μ and υ are integers and they means
Validation results confirm the effectiveness of the proposed that TFs (t, f ) is sampled at equidistant time-frequency grid

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YAN et al.: TWO-DIMENSIONAL OPTIMIZATION FRAMEWORK OF ONLINE INTERPRETABLE TIME-FREQUENCY FEATURE LEARNING 3

representation for time-frequency maps under varying condi-

tions, suitable optimization models are needed for effective early
fault detection. Meanwhile, the constructed optimization model
based on convex hulls should be sensitive to anomalies caused
by incipient faults rather than to noise.

III. PROPOSED FRAMEWORK FOR ONLINE CONTINUOUS

HEALTH MONITORING WITHOUT FAULTY SAMPLES
This section presents a practical two-dimensional optimiza-
tion framework for online continuous health monitoring with-
out requiring faulty samples. The framework utilizes a convex
hull-based closest point optimization model to update physics-
informed time-frequency features and corresponding HIs in
real time. As a machine approaches the onset of a fault, the
Fig. 1. Fault characteristics of repetitive transients in the time-frequency
domain transformed from STFT. time-frequency features help identify key frequency bands for
rapid fault diagnosis, while updated HIs reveal degradation
trends and indicate fault initiation. Finally, the application of this
method for ongoing health monitoring without faulty samples is
points (μT, υF ). Based on STFT, vibration samples in the time-
summarized.
frequency domain can be converted to time-frequency maps. A
fault signal from a simulation model [21] along with its STFT is
displayed in Fig. 1. The time-frequency map reveals distinct fault A. Closest Point Optimization Model for Convex Hull
signatures, with transient pulses appearing over time, indicating Separation
a sparse nature. These transients are concentrated in specific
This section presents a closest point optimization model
frequency bands, highlighting their aggregation and grouping
that accounts for both the separation of time-frequency maps
characteristics.
represented by convex hulls under different conditions and the
sparsity of extracted features in the time-frequency domain.
B. Convex Hulls for Time-Frequency Map Characterization
First, it is assumed that sets of healthy and faulty time-
To identify fault characteristics for interpretable continuous frequency maps are available, such as those derived from
health monitoring, convex hulls are used to geometrically rep- raw vibration samples using STFT. For example, raw vibra-
resent time-frequency maps in two dimensions. A convex hull tion samples can be immediately transformed from STFT.
is defined as the smallest convex set that encloses a given set Healthy time-frequency maps are C normal = {TFH i ∈R
q×w
,
of samples. Assuming a sample set C includes time-frequency i = 1, 2, . . . , m} and faulty time-frequency maps are C fault
=
maps TFi ∈ Rq×w (i = 1, 2, . . . , n). The time-frequency maps {TFF j ∈R
q×w
, j = 1, 2, . . . , l}. Herein, m and l are respec-
can be represented as matrix data. q and w respectively represent tively the number of healthy and faulty time-frequency maps.
the number of rows and columns of the time-frequency maps. The convex hull of healthy time-frequency maps C normal =
C can be called a convex set if its contained samples satisfy the {TFH i ∈R
q×w
, i = 1, 2, . . . , m} can be mathematically written
following equation: as follows:
θTFi + (1 − θ) TFj ∈ C (3)
covC normal =
where TFi ∈ C and TFj ∈ C. The value of θ is ranged from 0  m  m 
 
to 1; the convex hull of the C can be mathematically defined as H H
ai TFi  ai = 1, 0 ≤ ai ≤ 1, TFi ∈ C normal
. (5)
follows: 
i=1 i=1
 n  n 
 

covC = δi TFi  δi = 1, 0 ≤ δi ≤ 1, TFi ∈ C Similarly, the convex hull of faulty time-frequency maps

i=1 i=1 C fault = {TFF
j ∈R
q×w
, j = 1, 2, . . . , l} can be mathematically
(4) written as follows:
where covC is the convex representation of the dataset C; δi
is the coefficients of the covC. It is observed from (4) that covC fault =
the covC has a compact expression, in which the sum of the ⎧ ⎫
⎨ l  l ⎬
coefficients is equal to 1 and these coefficients are ranged from  
0 to 1. Due to its compact geometric representation, the convex bj TFF 
j  bj = 1, 0 ≤ bj ≤ 1, TFF
j ∈C
fault
. (6)
⎩ ⎭
hull may exhibit sensitivity to outliers [22]. Early-stage faults in j=1 j=1
machines produce weak signal responses, which can be further
diminished by noise due to signal transmission through multiple The closest point optimization model that aims to maximize
paths. Thus, while convex hulls provide a precise geometric the margin between these two convex hulls covC normal and

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON RELIABILITY

covC fault can be formulated as follows: further investigate the impact of these two objectives on the
m l
physics-informed features. Combining these objectives helps in
1  
effectively identifying fault characteristics in the time-frequency
min ai TFH
i − bj TFF
j
ai ,bj 2 i=1 j=1
domain. Specifically, the key differences between ideal healthy
2
and faulty time-frequency maps often involve recurring fault
m
 transients within a relevant frequency range, exhibiting spar-
s.t. ai = 1, 0 ≤ ai ≤ 1 sity. By optimizing contrastively, the proposed features should
i=1
enhance the detection of important frequency components and
l
 reduce noise in the time-frequency domain.
bj = 1, 0 ≤ bj ≤ 1 In essence, If the proposed physics-informed features φ
j=1 have the same coordinate with TFF H
j and TFi in the time-
i = 1, 2, . . . , m; j = 1, 2, . . . , l. (7) frequency domain, then φ can be understood as an optimized
time-frequency feature and it can be capable of filtering out
Next, considering the group sparsity of fault signatures in the noise automatically. To mitigate reliance on faulty samples, an
time-frequency domain, a norm regulation can be added to (7) online two-dimensional optimization framework is presented.
to enhance the sparsity of extracted time-frequency features as This framework enables continuous and interpretable learning
follows: of time-frequency features and HI construction. The updated
m
 l
 time-frequency features are derived from baseline and recent
min ai TFH
i − bj TFF
j samples, facilitating prompt fault diagnosis through a convex
ai ,bj
i=1 j=1
2
optimization model in (8). All collected data and their cor-
responding optimized features are recorded, allowing for the
m
 l
 continuous updating of HI for monitoring emerging faults and
+λ ai TFH
i − bj TFF
j tracking degradation trends.
i=1 j=1
1
m
 l

s.t. ai = 1, 0 ≤ ai ≤ 1, bj = 1, 0 ≤ bj ≤ 1 B. Online Two-Dimensional Optimization Framework Without
i=1 j=1 Needing Faulty Samples
i = 1, 2, . . . , m; j = 1, 2, . . . , l (8) In practical scenarios, especially with newly introduced ma-
chines, there is often a shortage of faulty samples and extensive
where λ is a hyperparameter to balance two terms in the objective
historical data. Addressing machine health monitoring under
functions. It can be observed from (8) that the optimization
these conditions remains a significant challenge. This section
model is convex, as each term in the objective function and con-
proposes a practical two-dimensional optimization framework
straints is convex. If a feasible solution exists, global optimal so-
for ongoing health monitoring, leveraging the optimization
lutions a∗ and b∗ are guaranteed. Herein, a∗ = [a∗1 , a∗2 , . . . , a∗m ]
model in (8). When a new machine Ω is first monitored, healthy
and b∗ = [b∗1 , b∗2 , . . . , b∗l ] . Based on this, the proposed inter-
vibration data is collected. To conserve storage space, high-
pretable time-frequency features φ are defined as follows:
frequency data is not recorded continuously, as performance
l
 m
 degradation typically occurs gradually. Instead, data is gathered
φ= b∗j TFF
j − a∗i TFH
i (9) in fixed-length segments at regular intervals. These intervals
j=1 i=1 can be adjusted according to specific conditions and needs,
where φ ∈ Rq×w is the proposed physics-informed fea- providing insight into changes in machine conditions. Assume
tures in the time-frequency that the time interval for data collection is set to T so that online
m domain. Based on (9), the first
l
objective function 12 i=1 ai Hi − j=1 bj Nj 2 in (8) is
collection vibration data can be denoted as vT , v2T , . . . ,vsT .
equivalent to 12 φ2 while the second objective function Once they are acquired, they are immediately transformed into
m l 1 time-frequency maps TFT , TF2T , . . . ,TFsT based on STFT.
i=1 ai Hi − j=1 bj Nj 1 is equivalent to 2 φ1 . Geomet- Since the machine Ω can generally be regarded as being in a
rically, the first objective function is to maximize the sep-
completely normal state when it is first used, the first collected
aration between time-frequency maps of healthy and faulty
small number of time-frequency maps TFmT i=T (m  s) are used
states estimated by convex hulls and 2 norm is used to
as healthy time-frequency maps to establish a baseline convex
penalize large elements in the proposed physics-informed
hull. Herein, TFmTi=T (m  s) means the time-frequency maps
features, aiming to reduce overfitting. The second objective
of collected vibration signals from T to mT . For subsequent
function employs a 1 norm to promote sparsity in these 2mT
obtained time-frequency maps T Fi=(m+1)T (m  s), they are
features, which improvesm their robustness
l to outliers. There-
subjectively classified as “faulty time-frequency maps,” allow-
fore, the first term i=1 ai Hi − j=1 bj Nj 2 in the ob- ing the optimization model in (8) to be applied. The optimization
jective function aims to highlight the significant differences
model in (8) can be simplified as follows:
between healthy and 
faulty time-frequency maps. The second
m l
term i=1 ai Hi − j=1 bj Nj 1 seeks to enforce sparsity  
in these differences. Ablation experiments in case studies will (a∗ , b∗ , φ) = Φ C normal , C fault (10)

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YAN et al.: TWO-DIMENSIONAL OPTIMIZATION FRAMEWORK OF ONLINE INTERPRETABLE TIME-FREQUENCY FEATURE LEARNING 5

where C normal and C fault are the inputs of the optimization model. Algorithm 1: Iterative Optimization Framework for Online
a∗ and b∗ are the outputs of the optimization model and they are Machine Health Monitoring.
the coefficients of convex hulls corresponding to C normal and
Input: baseline normal dataset {T F mT i=T } with STFT,
C fault . φ is the interpretable time-frequency features. Therefore,
online collected dataset {T F sT i=(m+1)T } with STFT,
the interpretable time-frequency features can be updated for
sample of baseline normal data and online data for the
quick fault monitoring as follows:
optimization model m, λ
⎧   mT  fault  2mT 
⎪
⎪ Φ C normal
= TF , C = TF Output: a, b, φ
⎪
⎪ i=T i=(m+1)T
1. For k = 1, 2, . . . , (s − 2m + 1) do
⎪
⎪
⎪t ∈ (0, 2mT ) 2. baseline normal dataset T F H mT
⎪
⎪
⎪   fault  (2m+1)T  i = {T F i=T }
⎪
⎨Φ C normal = TFmT i=T , C = TFi=(m+2)T F (k+2m−1)T
3. online update dataset T F j = {T F j=(k+m)T }
φ = t ∈ ((m + 2) T, (2m + 1)T ) . 4. #optimization formulation for CVX solver
⎪
⎪
⎪
⎪ . 5. cvx begin
⎪
⎪.. 
⎪
⎪   fault  sT  6. variablesa(m)b(m)
⎪
⎪Φ C normal = TFmT = TFi=sT −mT +1
⎪ i=T , C
⎩ 7.
t ∈ (sT − mT + 1, sT )
m
 l

(11)
It can be observed from (11) that initial interpretable time- objective f unction = ai T F H
i − bj T F F
j
mT i=1 j=1
frequency features are obtained based on C normal = {Pi=T } 2
2mT
and C fault
= {Pi=(m+1)T }. Subsequently, the first updated m
 l

interpretable time-frequency features are obtained based on +λ ai T F H
i − bj T F F
j
mT (2m+1)T
C normal = {Pi=T } and C fault = {Pi=(m+2)T }. The assumed i=1 j=1 1
fault time-frequency maps are truncated from (m + 1)T to 8. minimize(objectivef unction)
sT through a window with a length of mT and the window
moves step by step based on collection time T . With this subject to
approach, interpretable time-frequency features can be updated m
 l

online with each data collection, enabling swift fault diagnosis. ai = 1, ai ≥ 0, bj = 1, bj ≥ 0
In addition, the collected time-frequency maps and updated i=1 j=1
features facilitate long-term monitoring of degradation. Online 9. cvx end
updated interpretable time-frequency features are denoted as 10. calculate the time-frequency feature φ based on (9)
φ2mT , φ(2m+1)T , φ(2m+2)T , . . . ,φsT . The update frequency of 11. end for
interpretable time-frequency features can match the data collec-
tion interval. These features allow for real-time fault diagnosis
and identification of informative frequency bands. While these
features indicate the current health status, an HI is necessary maps with transparent construction processes, and it can identify
for tracking long-term degradation. Combining interpretable incipient degradation and reflect ongoing deterioration. Besides,
time-frequency features with an HI provides a comprehen- the interpretable features enhance the HI’s sensitivity to early
sive approach to monitoring machine health over its lifecycle. faults and improve its effectiveness in monitoring degradation
The HI is updated based on the latest interpretable features to trends.
offer a clear method for detecting incipient faults and tracking To calibrate HI scales across different collection times, each
degradation and it is defined as follows: HI is normalized, allowing for the detection of incipient faults
⎧ based on trending changes. This study introduces an adaptive
⎪
⎪ sum (TF  i ◦ φ2mT ) , i ∈ (0, 2mT ) threshold method, using a three-sigma rule [23], to identify fault
⎪
⎨sum TFi ◦ φ(2m+1)T , i ∈ (0, (2m + 1) T ) initiation. The threshold is updated online with each new HI.
HIi = .. (12) To address noise and outliers, a fault alert is triggered only if
⎪
⎪ .
⎪
⎩ multiple consecutive HI values exceed the adaptive threshold.
sum (TFi ◦ φsT ) , i ∈ (0, sT )
This approach ensures precise detection of incipient faults and
where ◦ is the operation of Hadamard product; Since matrices enables the assessment of degradation trends and future predic-
TFi and φ have the same dimension, the result of TFi ◦ φ is tions.
a matrix with the same dimension with TFi and φ. Its element The practical two-dimensional optimization framework for
is equal to the multiplication between the elements of TFi and online machine health monitoring is summarized in Fig. 2
φ in the corresponding place; sum(·) is the operation of sum and detailed in Algorithm 1. Initially, a small set of baselined
and its result is to add all elements in a matrix. As shown in and online time-frequency matrices are collected. The convex
(12), the HI at each collection time integrates time-frequency optimization model in (8) can be addressed by using some
maps of vibration signals with updated interpretable features in established algorithms such as the interior point method and
a data-level fusion way. This HI can be continuously updated for tools such as CVX. This study uses CVX based on a MATLAB
tracking degradation. The validity of the HI is supported by two environment [24] as a demonstration to solve the optimization
factors. It is built from physically interpretable time-frequency model in (8). Solutions provide updated interpretable features for

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON RELIABILITY

Fig. 3. Life cycle endurance platform for NASA bearings [11].

Fig. 2. Proposed practical two-dimensional optimization for online machine

health monitoring based on time-frequency maps and closet point optimization
models.

real-time fault diagnosis. Long-term degradation maps are then

collected to update HIs for ongoing monitoring and degradation
tracking. This framework offers both ante-hoc and post-hoc
interpretability. Its objective functions and constraints have clear
geometric and physical meanings and yield global optimal so-
lutions. Time-frequency features derived are physics-informed Fig. 4. Raw vibration signals in the time domain at different collection times.
and interpretable, allowing engineers to further analyze them. (a) Vibration data at collection time 1. (b) Vibration data at collection time 100.
(c) Vibration data at collection time 533. (d) Vibration data at collection time
In addition, an integrated HI can detect early faults and support 534. (e) Vibration data at collection time 703. (f) Vibration data at collection
effective health monitoring and degradation tracking, enhancing time 900.
the framework’s practical utility.

IV. REAL EXPERIMENTAL CASE STUDY collection point, 20 480 vibration samples were recorded to
assess the bearings’ health at that particular time. In total, the
This section illustrates the application of the proposed
NASA bearings underwent 984 data collection intervals before
methodology for online continuous health monitoring through
one of them completely failed. Fig. 4 illustrates raw vibration
two endurance tests of bearings, demonstrating their perfor-
data from various collection points, showing that the ampli-
mance from initial new conditions to eventual failure.
tude of vibration samples increases with time. However, the
vibration energy recorded at collection time 900 is lower than
A. Developed Practical Methodology to Realize Online
that at collection time 703, suggesting that despite worsening
Incipient Fault Detection and Degradation Tracking of NASA bearing performance, the vibration responses might diminish.
Bearings
Furthermore, fault characteristics observed at the early stages of
The experimental setup for assessing bearing life cycles failure are notably weak, as demonstrated in Fig. 4(c) and (d),
is illustrated in Fig. 3. As depicted, the setup includes four which display patterns similar to those in Fig. 4(a) and (b). This
bearings, labeled bearings 1–4, mounted on a single shaft. A issue presents significant challenges for detecting early faults
motor positioned on the shaft’s left side maintains a constant and accurately assessing the degradation process.
rotational speed of 2000 r/min. To monitor vibrations during the The proposed online approach employs a convex hull opti-
endurance test of the bearings, accelerometers were affixed to mization model combined with time-frequency maps to enable
the bearing housings. The test was terminated once the bearings continuous health monitoring of the NASA bearings without
surpassed their designed life span. The specifics of the vibration needing prior fault sample knowledge. In this section, param-
data collection process are outlined as follows: a sampling rate eters λ and m are defined as 1 and 10, respectively. This
of 20 000 Hz was employed, with data being gathered every implies that ten healthy time-frequency maps and ten faulty
10 min. Consequently, T is 10 min for this study. At each time-frequency maps are utilized to extract interpretable features

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YAN et al.: TWO-DIMENSIONAL OPTIMIZATION FRAMEWORK OF ONLINE INTERPRETABLE TIME-FREQUENCY FEATURE LEARNING 7

Fig. 5. Initial and updated interpretable time-frequency features at different

collection times. (a) Initial time-frequency feature at collection time 20. (b) Up-
dated time-frequency feature at collection time 138. (c) Updated time-frequency Fig. 6. Updated HIs for incipient fault detection during continuous health
feature at collection time 338. (d) Updated time-frequency feature at collection monitoring of the NASA bearing at different collection times. (a) HI during
time 519. (e) Updated time-frequency feature at collection time 638. (f) Updated collection times 11–109. (b) HI during collection times 11–209. (c) HI dur-
time-frequency feature at collection time 738. (g) Updated time-frequency ing collection times 11–309. (d) HI during collection times 11–409. (e) HI
feature at collection time 838. (h) Updated time-frequency feature at collection during collection times 11–509. (f) HI during collection times 11–609.
time 919. (i) Updated time-frequency feature at collection time 969. (g) HI during collection times 11–709. (h) HI during collection times 11–809.
(i) HI during collection times 11–909.

from the time-frequency data. As vibration data are collected at online monitoring of machine health. Furthermore, this approach
each time point, they are promptly converted into time-frequency allows for adaptive updates to pinpoint informative frequency
maps using STFT. Consequently, a new time-frequency map is bands accurately.
produced at each collection interval. Thus, initial interpretable Subsequently, the HIs derived from the proposed method are
features are derived from time-frequency maps collected be- used to assess fault signatures within the interpretable time-
tween times 1 and 20. Specifically, the maps from times 1 frequency features, enhancing the precision of early fault de-
through 10 are considered healthy, whereas those from times tection. Once optimized time-frequency features are obtained at
11 through 20 are treated as faulty. These features are then each collection interval, the HI trajectory is updated accordingly.
updated using time-frequency maps from collection periods Fig. 6 displays the HIs at various collection times for online
1–10 and 12–21. In this case, maps from times 1–10 are still early fault detection, based on (12). Fig. 6 demonstrates that the
used as healthy references, and maps from times 12–21 are proposed HIs effectively monitor the degradation process of the
assumed to represent faults. Given that 984 collection points NASA bearing. During the healthy state, the HIs derived from
were recorded for NASA bearings, this method allows for the the interpretable time-frequency features show random fluctu-
update of 965 time-frequency features. Fig. 5 illustrates some ations and follow a Gaussian distribution, with no observable
selectively chosen interpretable time-frequency features from degradation trend, as shown in Fig. 6(a)–(e). However, when an
various collection times, demonstrating how the proposed fea- incipient fault develops, the HIs begin to exhibit a noticeable
tures evolve as the collection times progress. increasing trend, signaling the onset of the fault, as depicted in
During the healthy phase of the bearing, neither the ini- Fig. 6(f).
tial nor the updated time-frequency features before collec- According to the threshold determination method outlined
tion time 519 display any distinct signatures, with their in [23], an early fault in the NASA bearing was identified at
energy distributions appearing relatively uniform, as illustrated collection time 533. Typically, because degradation is usually
in Fig. 5(a)–(d). Following collection time 519, the NASA an irreversible process, one would expect the HI to show a
bearing transitions into a faulty phase, with a gradual decline consistent monotonic trend throughout a bearing’s lifecycle.
in performance. Concurrently, the energy within the proposed Nonetheless, after collection time 533, some fluctuations are ob-
interpretable time-frequency features begins to focus on specific served in the proposed HIs. This behavior has been examined in
informative frequency bands, as depicted in Fig. 5(e)–(i). This various studies, which have applied the concept of self-healing to
indicates that the proposed features are informed by physical account for these fluctuations [25]. Despite this, such variations
principles and effectively highlight the informative frequency do not affect the precise detection of early faults. The HI values
bands crucial for subsequent fault diagnosis. The methodology consistently surpass the threshold during the lifecycle, indicating
applied here employs a data-driven approach to derive inter- its sensitivity to early faults and its capability for online updates
pretable features within the time-frequency domain, facilitating to monitor the degradation process. In addition, the efficacy

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON RELIABILITY

Fig. 7. Degradation trends of RMS and kurtosis for continuous health moni-
toring of the NASA bearing. (a) RMS. (b) Kurtosis.

Fig. 9. Comparison between a HI based on the sum of original time-frequency

maps and the proposed HI for NASA bearing. (a) HI based on the sum of
original time-frequency maps. (b) Proposed HI based on the sum of weighted
time-frequency maps.

to early bearing faults and demonstrates a steeper degradation

trend. This improvement is attributed to the use of interpretable
time-frequency features as effective weights, which are ap-
plied to the original time-frequency maps to generate the HI.
Fig. 8. Degradation trends of negative entropy, Gini index, and smoothness
The results in Fig. 9 suggest that these features signifi-
index for continuous health monitoring of the NASA bearing. (a) Negative cantly enhance the HI’s capability for early fault detection.
entropy. (b) Gini index. (c) Smoothness index. Finally, a widely recognized technique known as Fast Kur-
togram, designed to identify informative frequency bands for
early fault diagnosis, is compared with the frequency bands
of the proposed HIs is well-supported by the time-frequency highlighted by the proposed interpretable time-frequency fea-
tures. Fig. 10 shows the informative frequency bands iden-
features it extracts. Since these features promptly highlight fault
characteristics in the time-frequency domain, they are highly tified by the Fast Kurtogram at collection times 533, 534,
effective for detecting incipient faults. and 535. Once these bands are determined, they are used
To ensure a fair evaluation, several statistical measures are to filter raw vibration samples, with the squared envelope
used to quantify vibration samples collected at each interval. spectra of these filtered signals shown in Fig. 11. Based on
Initially, two commonly utilized HIs, including root mean square the proposed interpretable time-frequency features in Fig. 5,
(RMS) [26], [27] and kurtosis [28], [29] are employed for com- the informative frequency band is located between 4000 and
parison. RMS provides an average measure of data amplitude 5000 Hz. The corresponding squared envelope spectra for col-
at each collection point, whereas kurtosis gauges the sharpness lection times 533, 534, and 535 are displayed in Fig. 12.
of distributions. Fig. 7 illustrates the performance of both RMS Figs. 11 and 12 demonstrate that the proposed frequency
and kurtosis for the NASA bearing, revealing that neither effec- band reveals more fault characteristic frequencies. Therefore,
tively indicates degradation trends associated with early fault the proposed interpretable time-frequency features not only
detection. Subsequently, various sparsity measures are applied serve as informative weights for constructing the HIs but also
to assess the sparsity of vibration signals in monitoring the prove more effective in identifying significant frequency com-
conditions of NASA bearing. Negative entropy, Gini index, and ponents and filtering out weak fault signals for early fault
smoothness index are presented in Fig. 8, but these measures diagnosis.
do not exhibit sensitivity to early fault occurrences. In addition,
B. Developed Practical Methodology to Realize Online
these statistical HIs lack flexibility and are heavily influenced
Incipient Fault Detection and Degradation Tracking of
by specific conditions and the nature of the vibration data. In
contrast, the proposed HIs, which are based on data-level fusion, XJTU-SY Bearings
leverage extracted interpretable time-frequency features to offer Next, a bearing endurance test established by Xi’an Jiaotong
valuable insights for fault diagnosis. University and the Changxing Sumyoung Technology Co., Ltd.
Furthermore, to highlight the effectiveness and significance is used to verify the proposed online continuous health mon-
of the optimized interpretable time-frequency features shown itoring methodology. As seen in Fig. 13, two unidirectional
in Fig. 5, a comparison is made between the HI calculated acceleration sensors were fixed on the horizontal and vertical
from the sum of the original time-frequency maps and the directions of testing bearings through magnetic seats to obtain
proposed HI derived from the weighted time-frequency maps, bearing full life cycle vibration signals. During the life cycle
as depicted in Fig. 9. It is evident from Fig. 9 that the proposed experiments, a sampling frequency and a sampling interval
HI, utilizing weighted time-frequency maps, is more responsive were respectively set to 25.6 kHz and 1 min. Therefore, T

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YAN et al.: TWO-DIMENSIONAL OPTIMIZATION FRAMEWORK OF ONLINE INTERPRETABLE TIME-FREQUENCY FEATURE LEARNING 9

Fig. 12. Square envelope spectra at incipient faults times 533, 534, and 535 by
using the indicated informative frequency band based on proposed interpretable
time-frequency features. (a) Collection time 533. (b) Collection time 534.
(c) Collection time 535.

Fig. 10. Indicated informative frequency bands at incipient faults times 533,
534, and 535 by using the fast kurtogram. (a) Collection time 533. (b) Collection
time 534. (b) Collection time 535.

Fig. 13. Bearing endurance test for life cycle vibration data collection [30].

Fig. 11. Square envelope spectra at incipient faults times 533, 534, and 535
by using the fast kurtogram. (a) Collection time 533. (b) Collection time 534.
(c) Collection time 535.

is equal to 1 min. In total, vibration data were sampled 123

times until one bearing completely failed. At each collection
time, 32 767 samples were collected. First, raw vibration data Fig. 14. Raw vibration signals in the time domain at different collection
times of the XJTU-SY bearing. (a) Vibration data in collection time 1.
at different collection times are given in Fig. 14. Like the first (b) Vibration data in collection time 30. (c) Vibration data in collection time 78.
case study, vibration amplitudes show an overall increasing trend (d) Vibration data in collection time 80. (e) Vibration data in collection time 100.
with collection times. Nevertheless, energy changes at incipient (f) Vibration data in collection time 120.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON RELIABILITY

Fig. 15. Proposed interpretable time-frequency features at different collection

Fig. 16. Updated HIs for incipient fault detection during continuous health
times for continuous health monitoring of the XJTU-SY bearing. (a) Initial
monitoring of the XJTU-SY bearing at different collection times. (a) HI during
time-frequency feature at collection time 20. (b) Updated time-frequency feature
collection time 11–59. (b) HI during collection time 11–64. (c) HI during col-
at collection time 49. (c) Updated time-frequency feature in collection time 69.
lection time 11–69. (d) HI during collection time 11–74. (e) HI during collection
(d) Updated time-frequency feature in collection time 90. (e) Updated time-
time 11–79. (f) HI during collection time 11–84. (g) HI during collection time
frequency feature in collection time 101. (f) Updated time-frequency feature
11–89. (h) HI during collection time 11–94. (i) HI during collection time 11–99.
in collection time 109. (g) Updated time-frequency feature in collection time
114. (h) Updated time-frequency feature in collection time 119. (i) Updated
time-frequency feature in collection time 123.

fault times at collection times 78 and 80 are not obvious while

their energy may weaken in a bearing late life stage.
The proposed methodology is then employed to extract in-
terpretable features from the time-frequency domain for moni-
toring the health of XJTU-SY bearings online. Data from the
first ten collection points are used to establish a baseline of
healthy samples. Data from subsequent collections, such as
times 11–20, 12–21, 13–22, etc., are assumed to reflect faults and Fig. 17. Degradation trends of RMS and kurtosis for continuous health mon-
are fed into the convex optimization model detailed in (8) as they itoring of the XJTU-SY bearing. (a) RMS. (b) Kurtosis.
become available. The updated interpretable time-frequency
features at various collection points are shown in Fig. 15, demon-
strating how these features progressively emphasize informa-
tive frequency bands as bearing degradation advances. During
the healthy phase, the energy distribution of the optimized
time-frequency features is consistent across different frequency
regions. Upon the onset of an incipient fault, these features
increasingly concentrate on specific frequency bands that aid in
fault localization. Fig. 16 illustrates the updated HI trajectories,
which, although fluctuating in the healthy phase, do not reveal a
Fig. 18. Degradation trends of negative entropy, Gini index, and smoothness
clear degradation trend. However, during the degradation phase, index for continuous health monitoring of the XJTU-SY bearing. (a) Negative
the HI trajectories confirm that the incipient fault occurs around entropy. (b) Gini index. (c) Smoothness index.
collection time 78, with a subsequent monotonic degradation
trend useful for assessing and predicting future degradation.
A comparison of various HIs for continuous health monitor- Negative entropy, Gini index, and smoothness index exhibit ini-
ing of the XJTU-SY bearing is presented in Figs. 17 and 18. tial increasing trends but show abrupt changes at the fault onset,
Fig. 17 shows that while RMS displays a slight upward trend which are unsuitable for describing the lifecycle process. In
during normal conditions, it is insufficient for pinpointing the contrast, the proposed methodology for feature extraction and HI
onset of faults. Kurtosis, on the other hand, fluctuates unpre- construction demonstrates reliable and consistent performance
dictably in the healthy stage, making fault detection challenging. across different datasets.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YAN et al.: TWO-DIMENSIONAL OPTIMIZATION FRAMEWORK OF ONLINE INTERPRETABLE TIME-FREQUENCY FEATURE LEARNING 11

Fig. 19. Comparison between a HI based on the sum of original time-frequency

maps and the proposed HI for XJTU-SY bearing. (a) HI based on the sum of
original time-frequency maps. (b) Proposed HI based on the sum of weighted
time-frequency maps.

Fig. 19 compares the HI derived from the sum of original

time-frequency maps with the proposed HI for XJTU-SY bear-
ing condition monitoring. Fig. 19(a) reveals that the HI based
on the original time-frequency maps shows instability during
normal conditions, hindering early fault detection. Conversely,
the proposed HI maintains stability under normal conditions
and exhibits a clear increasing trend as early faults develop,
proving more effective for bearing health monitoring. Figs. 20
and 21 summarize the performance of informative frequency Fig. 20. Indicated informative frequency bands of the XJTU-SY bearing at
band localization for fault diagnosis using the Fast Kurtogram. incipient fault times 78, 79, and 80 by using the Fast Kurtogram. (a) Collection
The squared envelope spectra based on the proposed inter- time 78. (b) Collection time 79. (c) Collection time 80.
pretable time-frequency features for diagnosing incipient faults
in the XJTU-SY bearings are displayed in Fig. 22. Although
the Fast Kurtogram identifies fault characteristic frequencies
at collection times 79 and 80, it also captures numerous ex-
traneous components. The proposed squared envelope spectra
offer a clearer representation of fault characteristics, leading to
more accurate fault detection. Consequently, the extracted time-
frequency features and integrated HI demonstrate robustness and
consistent performance across various datasets.

C. Ablation Experiments
In this section, ablation experiments of the proposed two-
dimensional optimization model in (8) are conducted to show the
functions of different objective functions. The results between
the first objective function based solely on 2 norm, the second
objective function based solely on 1 norm, and their combi-
nation are compared. This section uses the XJTU-SY bearing
dataset as an illustrative example.
By using the same online optimization strategy, the optimized
interpretable time-frequency features and their fused HI at dif- Fig. 21. Square envelope spectra of the XJTU-SY bearing at incipient fault
ferent collection times based solely on 2 norm are respectively times 78, 79, and 80 by using the Fast Kurtogram. (a) Collection time 78.
(b) Collection time 79. (c) Collection time 80.
depicted in Figs. 23 and 24. And the optimized interpretable

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON RELIABILITY

Fig. 24. Updated HIs for incipient fault detection during continuous health
monitoring of the XJTU-SY bearing at different collection times by solely using
2 norm. (a) HI during collection time 11–59. (b) HI during collection time
Fig. 22. Square envelope spectra of the XJTU-SY bearing at incipient fault
11–64. (c) HI during collection time 11–69. (d) HI during collection time 11–74.
times 78, 79, and 80 by using the indicated informative frequency band
(e) HI during collection time 11–79. (f) HI during collection time 11–84. (g) HI
of proposed interpretable time-frequency features. (a) Collection time 78.
during collection time 11–89. (h) HI during collection time 11–94. (i) HI during
(b) Collection time 79. (c) Collection time 80.
collection time 11–99.

Fig. 23. Proposed interpretable time-frequency features at different collection

times for continuous health monitoring of the XJTU-SY bearing by solely using Fig. 25. Proposed interpretable time-frequency features at different collection
2 norm. (a) Initial time-frequency feature at collection time 20. (b) Updated times for continuous health monitoring of the XJTU-SY bearing by solely using
time-frequency feature at collection time 49. (c) Updated time-frequency feature 1 norm. (a) Initial time-frequency feature at collection time 20. (b) Updated
in collection time 69. (d) Updated time-frequency feature in collection time 90. time-frequency feature at collection time 49. (c) Updated time-frequency feature
(e) Updated time-frequency feature in collection time 101. (f) Updated time- in collection time 69. (d) Updated time-frequency feature in collection time 90.
frequency feature in collection time 109. (g) Updated time-frequency feature in (e) Updated time-frequency feature in collection time 101. (f) Updated time-
collection time 114. (h) Updated time-frequency feature in collection time 119. frequency feature in collection time 109. (g) Updated time-frequency feature in
(i) Updated time-frequency feature in collection time 123. collection time 114. (h) Updated time-frequency feature in collection time 119.
(i) Updated time-frequency feature in collection time 123.

time-frequency features and their fused HI at different collection

times based solely on 1 norm are respectively depicted in
Figs. 25 and 26. For a fair comparison, it is worth mentioning Fig. 28. More specifically, the energies of proposed interpretable
that Figs. 23 and 25 are calculated under the same samples as time-frequency features in Fig. 28(c) are more concentrated
Fig. 15. Figs. 24 and 26 are fused based on the same interpretable and sparser distributed in the informative frequency band than
time-frequency features as Fig. 16. interpretable time-frequency features based on a single norm in
As shown in Fig. 27, the design of the optimization func- Fig. 28(a) and (b). Concentrated and aggregated energies in the
tion by only using 2 norm or 1 norm has similar results in time-frequency domain are more conducive to distinguishing
fused HIs compared to their combination. Nevertheless, their information frequency bands. Therefore, it is better to combine
performances in the optimized interpretable time-frequency fea- 2 norm and 1 norm as the objective function for interpretable
tures have slight differences and their comparisons are given in time-frequency feature optimization.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YAN et al.: TWO-DIMENSIONAL OPTIMIZATION FRAMEWORK OF ONLINE INTERPRETABLE TIME-FREQUENCY FEATURE LEARNING 13

Fig. 26. Updated HIs for incipient fault detection during continuous health Fig. 29. Proposed interpretable time-frequency features at different collection
monitoring of the XJTU-SY bearing at different collection times by solely using times for continuous health monitoring of the XJTU-SY bearing (λ = 0.01). (a)
1 norm. (a) HI during collection time 11–59. (b) HI during collection time Initial time-frequency feature at collection time 20. (b) Updated time-frequency
11–64. (c) HI during collection time 11–69. (d) HI during collection time 11–74. feature at collection time 49. (c) Updated time-frequency feature in collection
(e) HI during collection time 11–79. (f) HI during collection time 11–84. (g) HI time 69. (d) Updated time-frequency feature in collection time 90. (e) Updated
during collection time 11–89. (h) HI during collection time 11–94. (i) HI during time-frequency feature in collection time 101. (f) Updated time-frequency
collection time 11–99. feature in collection time 109. (g) Updated time-frequency feature in collection
time 114. (h) Updated time-frequency feature in collection time 119. (i) Updated
time-frequency feature in collection time 123.

Fig. 27. Comparison between fused HIs for the XJTU-SY bearing by using
different objective functions. (a) 2 norm. (b) 1 norm. (c) Combination of 2
norm and1 norm.

Fig. 30. Updated HIs for incipient fault detection during continuous health
monitoring of the XJTU-SY bearing at different collection times (λ = 0.01).
(a) HI during collection time 11–59. (b) HI during collection time 11–64. (c)
HI during collection time 11–69. (d) HI during collection time 11–74. (e)
HI during collection time 11–79. (f) HI during collection time 11–84. (g) HI
during collection time 11–89. (h) HI during collection time 11–94. (i) HI during
collection time 11–99.

D. Discussion of Parameter λ
This section presents additional experiments to examine the
effect of the model parameter λ on time-frequency features and
their associated HIs. Due to space constraints, the XJTU-SY
bearing is also chosen as a case study to illustrate the influence
of λ. Various representative values for λ, including λ = 0.01,
Fig. 28. Comparison between interpretable time-frequency features at collec-
tion time 78 for continuous health monitoring of the XJTU-SY bearing by using λ = 0.1, and λ = 100, are evaluated. The time-varying features
different objective functions. (a) 2 norm. (b) 1 norm. (c) Combination of 2 and corresponding HIs for different values of λ are depicted in
norm and1 norm. Figs. 29–34. It can be observed that λ does not significantly affect
the performance of the proposed interpretable time-frequency

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON RELIABILITY

Fig. 31. Proposed interpretable time-frequency features at different collection Fig. 33. Proposed interpretable time-frequency features at different collection
times for continuous health monitoring of the XJTU-SY bearing (λ = 0.1). (a) times for continuous health monitoring of the XJTU-SY bearing (λ = 100) (a)
Initial time-frequency feature at collection time 20. (b) Updated time-frequency Initial time-frequency feature at collection time 20. (b) Updated time-frequency
feature at collection time 49. (c) Updated time-frequency feature in collection feature at collection time 49. (c) Updated time-frequency feature in collection
time 69. (d) Updated time-frequency feature in collection time 90. (e) Updated time 69. (d) Updated time-frequency feature in collection time 90. (e) Updated
time-frequency feature in collection time 101. (f) Updated time-frequency time-frequency feature in collection time 101. (f) Updated time-frequency
feature in collection time 109. (g) Updated time-frequency feature in collection feature in collection time 109. (g) Updated time-frequency feature in collection
time 114. (h) Updated time-frequency feature in collection time 119. (i) Updated time 114. (h) Updated time-frequency feature in collection time 119. (i) Updated
time-frequency feature in collection time 123. time-frequency feature in collection time 123..

Fig. 32. Updated HIs for incipient fault detection during continuous health
monitoring of the XJTU-SY bearing at different collection times (λ = 0.1). Fig. 34. Updated HIs for incipient fault detection during continuous health
(a) HI during collection time 11–59. (b) HI during collection time 11–64. monitoring of the XJTU-SY bearing at different collection times (λ = 100).
(c) HI during collection time 11–69. (d) HI during collection time 11–74. (e) (a) HI during collection time 11–59. (b) HI during collection time 11–64.
HI during collection time 11–79. (f) HI during collection time 11–84. (g) HI (c) HI during collection time 11–69. (d) HI during collection time 11–74. (e)
during collection time 11–89. (h) HI during collection time 11–94. (i) HI during HI during collection time 11–79. (f) HI during collection time 11–84. (g) HI
collection time 11–99. during collection time 11–89. (h) HI during collection time 11–94. (i) HI during
collection time 11–99.

features and their related HIs. The proposed method demon-

faulty samples. Herein, interpretable time-frequency features
strates consistent and reliable performance in continuous bearing
were real-time updated by using a small amount of baselined and
health monitoring under different settings of λ.
latest samples for quick fault diagnosis. The tasks of incipient
fault detection and degradation trajectory tracking were realized
V. CONCLUSION by continuous updates of HIs. Based on case studies, some
To solve the dilemma of small sample sizes and zero faulty insights into the closest point optimization model based on
samples, a convex hulls based closest point optimization model convex hulls are given as follows. First, the combination of 2
was first presented to learn interpretable time-frequency fea- norm and 1 norm in a closest point optimization model can
tures. Further, a practical online optimization framework of enhance energy aggregation and sparsity, which is useful for
the model was proposed to realize quick fault diagnosis and learning the significant difference between input instances, such
long-term degradation tracking based on an HI without needing as informative frequency bands in the time-frequency domain.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YAN et al.: TWO-DIMENSIONAL OPTIMIZATION FRAMEWORK OF ONLINE INTERPRETABLE TIME-FREQUENCY FEATURE LEARNING 15

Second, if a feature can reflect fault signatures, a data-level [12] Y. Cao, M. Jia, X. Zhao, X. Yan, and Z. Liu, “Semi-supervised machinery
fusion scheme can be used to develop an intuitive and phys- health assessment framework via temporal broad learning system em-
bedding manifold regularization with unlabeled data,” Expert Syst. Appl.,
ical HI for incipient fault detection. Third, the parameter λ vol. 222, 2023, Art. no. 119824, doi: 10.1016/j.eswa.2023.119824.
does not have a significant impact on the optimization of the [13] X. Du, W. Jia, P. Yu, “RUL prediction based on GAM–CNN for rotating
model. For online deployment, according to our tests, due to machinery,” J. Braz. Soc. Mech. Sci. Eng., vol. 45, no. 3, Mar. 2023,
doi: 10.1007/s40430-023-04062-8.
the use of matrices as model inputs and norm optimization in [14] M. Hu, G. Wang, K. Ma, Z. Cao, and S. Yang, “Bearing
the proposed optimization model, the inference time of inter- performance degradation assessment based on optimized EWT
pretable time-frequency features requires over ten minutes while and CNN,” Measurement, vol. 172, 2021, Art. no. 108868,
doi: 10.1016/j.measurement.2020.108868.
they show slight differences in a degradation stage and they [15] M. Zeng, Y. Yang, J. Zheng, and J. Cheng, “Maximum margin classification
mainly concentrate on some informative frequency bands. The based on flexible convex hulls,” Neurocomputing, vol. 149, pp. 957–965,
computation of a HI took within one second. Therefore, the 2015, doi: 10.1016/j.neucom.2014.07.038.
[16] M. Zeng, Y. Yang, J. Zheng, and J. Cheng, “Maximum margin clas-
interpretable time-frequency features can be updated according sification based on flexible convex hulls for fault diagnosis of roller
to an application demand, such as 15 min while the HI can be bearings,” Mech. Syst. Signal Process., vol. 66, pp. 533–545, Jan. 2016,
updated once new data are collected. Finally, some limitations doi: 10.1016/j.ymssp.2015.06.006.
[17] X. Li, Y. Yang, W. Ping, W. Jian, and J. Cheng, “A bearing fault diagno-
of the proposed methodology are discussed as follows. First, sis scheme with statistical-enhanced covariance matrix and Riemannian
more advanced and efficient algorithms are needed to shorten the maximum margin flexible convex hull classifier,” ISA Trans., vol. 111,
inference time of interpretable time-frequency features, which pp. 323–336, May 2021, doi: 10.1016/j.isatra.2020.11.018.
[18] Z. He, H. Shao, J. Cheng, Y. Yang, and J. Xiang, “Kernel flexible and
will be investigated in the future. Second, the applications of displaceable convex hull based tensor machine for gearbox fault intelli-
the proposed online optimization framework mainly focus on gent diagnosis with multi-source signals,” Measurement, vol. 163, 2020,
bearings in this study, which can be extended to other mechanical Art. no. 107965, doi: 10.1016/j.measurement.2020.107965.
[19] T. Yan, Y. Wang, T. Xia, B. Hou, L. Xi, and D. Wang, “Sparse
components in future studies. Finally, effective strategies for and flexible convex-hull representation for machine degradation mod-
updating the baseline normal dataset and further enhancing eling,” IEEE Trans. Rel., vol. 72, no. 1, pp. 27–36, Mar. 2023,
method performances will be considered. doi: 10.1109/tr.2022.3164976.
[20] J. Allen, “Short term spectral analysis, synthesis, and modification by
discrete Fourier transform,” IEEE Trans. Acoust., Speech Signal Process.,
vol. 25, no. 3, pp. 235–238, Jun. 1977.
REFERENCES [21] M. Zhao, J. Lin, X. Xu, and Y. Lei, “Tacholess envelope order analysis
and its application to fault detection of rolling element bearings with
[1] A. K. S. Jardine, D. Lin, and D. Banjevic, “A review on machin-
varying speeds,” Sensors, vol. 13, no. 8, pp. 10856–10875, Aug. 2013,
ery diagnostics and prognostics implementing condition-based mainte-
doi: 10.3390/s130810856.
nance,” Mech. Syst. Signal Process., vol. 20, no. 7, pp. 1483–1510, 2006,
[22] M. Zeng, Y. Yang, S. Luo, and J. Cheng, “One-class clas-
doi: 10.1016/j.ymssp.2005.09.012.
sification based on the convex hull for bearing fault detec-
[2] H. Zhu, C. Shen, J. Wang, B. Chen, D. Wang, and Z. Zhu, “Few-shot class-
tion,” Mech. Syst. Signal Process., vol. 81, pp. 274–293, 2016,
incremental learning with adjustable pseudo-incremental sessions for bear-
doi: 10.1016/j.ymssp.2016.04.001.
ing fault diagnosis,” IEEE Sensors J., vol. 24, no. 12, pp. 19543–19552,
[23] N. Li, Y. Lei, J. Lin, and S. X. Ding, “An improved exponential
Jun. 2024, doi: 10.1109/jsen.2024.3395515.
model for predicting remaining useful life of rolling element bearings,”
[3] T. Yan, D. Wang, T. Xia, E. Pan, Z. Peng, and L. Xi, “Online piecewise
IEEE Trans. Ind. Electron., vol. 62, no. 12, pp. 7762–7773, Dec. 2015,
convex-optimization interpretable weight learning for machine life cycle
doi: 10.1109/tie.2015.2455055.
performance assessment,” IEEE Trans. Neural Netw. Learn. Syst., vol. 35,
[24] M. Grant and S. Boyd, “CVX: MATLAB software for disciplined convex
no. 5, pp. 6048–6060, May 2024, doi: 10.1109/TNNLS.2022.3183123.
programming,” version 2.0 beta, Sep. 2013. [Online]. Available: http://
[4] F. Xu, Z. Huang, F. Yang, D. Wang, and K. L. Tsui, “Constructing a
cvxr.com/cvx
health indicator for roller bearings by using a stacked auto-encoder with an
[25] N. Yang, Z. Wang, W. Cai, and Y. Li, “Data regeneration based
exponential function to eliminate concussion,” Appl. Soft Comput., vol. 89,
on multiple degradation processes for remaining useful life esti-
2020, Art. no. 106119, doi: 10.1016/j.asoc.2020.106119.
mation,” Rel. Eng. Syst. Saf., vol. 229, 2023, Art. no. 108867,
[5] M. Ma, X. Chen, X. Zhang, B. Ding, and S. Wang, “Locally linear em-
doi: 10.1016/j.ress.2022.108867.
bedding on grassmann manifold for performance degradation assessment
[26] A. Klausen, H. van Khang, and K. G. Robbersmyr, “RMS based health
of bearings,” IEEE Trans. Rel., vol. 66, no. 2, pp. 467–477, Jun. 2017,
indicators for remaining useful lifetime estimation of bearings,” Model.,
doi: 10.1109/tr.2017.2691730.
Identification Control, Norwegian Res. Bull., vol. 43, no. 1, pp. 21–38,
[6] C. Wu, F. Feng, S. Wu, P. Jiang, and J. Wang, “A method for constructing
2022, doi: 10.4173/mic.2022.1.3.
rolling bearing lifetime health indicator based on multi-scale convolutional
[27] J. Meng, C. Yan, G. Chen, Y. Liu, and L. Wu, “Health indicator of
neural networks,” J. Braz. Soc. Mech. Sci. Eng., vol. 41, no. 11, 2019,
bearing constructed by rms-CUMSUM and GRRMD-CUMSUM with
doi: 10.1007/s40430-019-2010-6.
multifeatures of envelope spectrum,” IEEE Trans. Instrum. Meas., vol. 70,
[7] B. Duong et al., “A reliable health indicator for fault prognosis of bearings,”
Jan. 2021, Art. no. 3510216, doi: 10.1109/tim.2021.3054000.
Sensors, vol. 18, no. 11, 2018, Art. no. 3740, doi: 10.3390/s18113740.
[28] L. Wang, J. Pan, Y. Shao, Q. Zeng, and X. Ding, “Two new kurtosis-
[8] Y. Qin, J. Zhou, and D. Chen, “Unsupervised health Indicator construc-
based similarity evaluation indicators for grinding chatter diagnosis un-
tion by a novel degradation-trend-constrained variational autoencoder
der non-stationary working conditions,” Measurement, vol. 176, 2021,
and its applications,” IEEE/ASME Trans. Mechatron., vol. 27, no. 3,
Art. no. 109215, doi: 10.1016/j.measurement.2021.109215.
pp. 1447–1456, Jun. 2022, doi: 10.1109/tmech.2021.3098737.
[29] D. Wang, Z. Peng, and L. Xi, “The sum of weighted normal-
[9] P. Wen, Y. Li, S. Chen, and S. Zhao, “Remaining useful life prediction of
ized square envelope: A unified framework for kurtosis, negative en-
IIoT-enabled complex industrial systems with hybrid fusion of multiple in-
tropy, Gini index and smoothness index for machine health monitor-
formation sources,” IEEE Internet Things J., vol. 8, no. 11, pp. 9045–9058,
ing,” Mech. Syst. Signal Process., vol. 140, 2020, Art. no. 106725,
Jun. 2021, doi: 10.1109/jiot.2021.3055977.
doi: 10.1016/j.ymssp.2020.106725.
[10] Z. Shi and A. Chehade, “A dual-LSTM framework combining change
[30] Y. Jun, Y. A. N. Tao, L. I. Naipeng, W. Biao, H. A. N. Tianyu, and
point detection and remaining useful life prediction,” Rel. Eng. Syst. Saf.,
L. E. I. Yaguo, “XJTU-SY rolling element bearing accelerated life test
vol. 205, 2021, Art. no. 107257, doi: 10.1016/j.ress.2020.107257.
datasets: A tutorial,” J. Mech. Eng., vol. 55, no. 16, pp. 1–6, 2019,
[11] H. Qiu, J. Lee, J. Lin, and G. Yu, “Wavelet filter-based weak signa-
doi: 10.3901/jme.2019.16.001.
ture detection method and its application on rolling element bearing
prognostics,” J. Sound Vib., vol. 289, no. 4/5, pp. 1066–1090, 2006,
doi: 10.1016/j.jsv.2005.03.007.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on January 12,2025 at 12:57:43 UTC from IEEE Xplore. Restrictions apply.