Integrating Trend Monitoring and Change Point Detection for Wind Turbine Blade Diagnostics: A Physics-Driven Evaluation of Erosion and Twist Faults

Abstract

Robust condition monitoring of wind turbine blades is essential for reducing downtime and maintenance costs, particularly under variable operating conditions. While recent studies suggest that combining trend monitoring (TM) with change point detection (CPD) can improve diagnostic performance, it remains unclear whether such integration is beneficial for all fault types. This study experimentally evaluates the integration of TM and CPD using vibration data from a laboratory-scale wind turbine for two representative blade faults: leading-edge erosion and twist misalignment. For the erosion case, discrete wavelet transform (DWT) energy features exhibit a clear and persistent increase in mid-frequency content, with energy deviations of approximately 34–45% relative to the healthy state. However, Bayesian Online Change Point Detection (BOCPD) does not reveal distinct change points, indicating that CPD provides limited additional value for gradual, steady-state degradation. In contrast, for twist misalignment, the short-time Fast Fourier Transform (FFT) features reveal dynamic spectral redistribution, and CPD applied to spectral centroid trends produces a sharp, localized detection signature. These results demonstrate that integrating TM with CPD significantly enhances fault detectability for dynamic, instability-driven faults, while TM alone is sufficient for smooth, steady-state degradation. This study provides an evidence-based guideline for selectively integrating CPD into wind turbine blade condition monitoring systems based on fault physics.

1. Introduction

Wind energy has transitioned from a supplementary renewable source to a central pillar of global decarbonization efforts. As nations pursue climate neutrality and energy independence, wind capacity—especially offshore—continues to expand rapidly, with installations expected to quadruple by 2050 [1]. Alongside this growth, the industry’s priorities have shifted from simply increasing capacity to reducing costs. Operation and maintenance (O&M) activities, particularly in remote offshore environments, now represent 25–30% of the levelized cost of electricity (LCOE) [2]. Unplanned downtime and major failures not only halt energy production but also generate substantial logistical expenses. As a result, the long-term economic viability of wind power increasingly depends on replacing traditional “reactive” maintenance with predictive, condition-based strategies that identify faults at their earliest stages.

Within a wind turbine, the rotor blades are among the most critical and vulnerable components. Unlike internal mechanical elements (e.g., gearboxes or generators) that operate in controlled, lubricated environments, blades are continuously exposed to fluctuating aerodynamic loads and harsh weather conditions. Their aerodynamic performance directly determines the turbine’s power curve, and recent industry reports identify blade damage as one of the leading causes of turbine downtime [3]. Blade-related faults generally fall into two principal categories: (1) Surface degradation (leading-edge erosion)—a gradual, steady-state process in which rain, hail, and airborne particles deteriorate the blade surface. This erosion disrupts the laminar boundary layer, reduces the lift-to-drag ratio, and can decrease annual energy production (AEP) by 5–15% [4,5,6,7]. Because the damage accumulates slowly, it typically appears as a gradual drift in vibration features. (2) Structural and installation asymmetries (twist faults)—arising from manufacturing tolerances, installation errors, or pitch control malfunctions. These faults introduce geometric asymmetry, causing aerodynamic imbalance and periodic loading that stresses the drivetrain and tower [8,9]. In contrast to erosion, twist faults produce dynamic, harmonic instabilities that are clearly observable in the frequency spectrum.

The development of robust structural health monitoring (SHM) for these components has seen significant advancements through diverse methodological lenses. As highlighted in a comprehensive state-of-the-art review by Haseeb and Krawczuk [10], the field is increasingly moving toward integrated frameworks that combine multi-sensor data acquisition with sophisticated analysis methods like machine learning to handle the high complexity of modern rotor blades. For instance, Joshuva and Sugumaran [11] demonstrated the effectiveness of using vibration signals coupled with statistical histogram features and J48 decision tree algorithms to successfully classify distinct fault states, including surface erosion and pitch angle twist. However, the reliability of these diagnostic tools depends heavily on understanding the underlying physics of the fault. Recent experimental and theoretical research by Chen et al. [12] on stall-induced vibrations confirms that blade stability and aero-damping are extremely sensitive to pitch and twist configurations, where even minor aerodynamic perturbations can lead to divergent edge-wise oscillations. Furthermore, active monitoring systems, such as the machine-learning-based approach proposed by Milani et al. [13], have shown that pitch misalignments can be detected with high precision under time-varying operational loads.

Detecting such diverse blade faults early is challenging due to the inherently non-stationary nature of wind conditions. Although Supervisory Control and Data Acquisition (SCADA) systems are widely deployed across wind farms, their diagnostic capability is limited. SCADA typically records only 10 min averaged operational parameters (e.g., power, temperature, rotor speed) [14], which are sufficient for identifying major failures but lack the temporal resolution needed to capture subtle, early-stage blade anomalies. As a result, small fault signatures are often obscured by turbulence-induced variability, making incipient damage difficult to detect using SCADA data alone [15].

Consequently, high-frequency vibration analysis has become the gold standard for blade condition monitoring. Vibration signals capture rich, instantaneous information about both the structural and aerodynamic state of the turbine. However, these raw signals are inherently noisy and highly complex, making advanced signal-processing methods essential for extracting meaningful diagnostic features. Once such features are obtained, the central challenge becomes interpreting how—and to what extent—these features change over time. In practice, current diagnostic frameworks rely on two primary mathematical paradigms:

• Trend Monitoring (TM): TM tracks the long-term evolution of statistical features (e.g., energy, mean, kurtosis) to reveal gradual deviations from a healthy operating state [16]. It is effective for identifying slow, wear-related degradation but generally cannot pinpoint the exact moment a fault begins.
• Change Point Detection (CPD): CPD uses statistical methods, such as Cumulative Sum (CUSUM) [17] or Bayesian Online Change Point Detection (BOCPD) [18], to detect the precise time at which the underlying data generation process changes abruptly. These techniques excel at identifying sudden or dynamic fault events [19,20].

While machine learning (ML) and deep learning (DL) models have become increasingly popular for wind turbine diagnostics [21,22,23,24,25], their “black-box” nature limits interpretability and hinders industrial certification. In contrast, statistical approaches such as TM and CPD are inherently interpretable but often applied without considering the specific characteristics of the fault. In practice, researchers seldom evaluate whether a given fault truly requires a sophisticated CPD algorithm or whether simple TM is sufficient. As a result, mismatches frequently occur: applying CPD to a slowly developing fault such as erosion can produce false negatives, while relying solely on TM for highly dynamic faults such as twist misalignment may fail to pinpoint the moment of instability [24]. This highlights the need for a more fault-aware selection and integration of TM and CPD methods.

Rather than proposing new diagnostic algorithms, this study aims to experimentally clarify when the integration of trend monitoring (TM) and change point detection (CPD) is beneficial, redundant, or ineffective, depending on the physical nature and temporal evolution of wind turbine blade faults.

To address this gap, this paper introduces an integrated monitoring and diagnostic framework that aligns specific time–frequency signal processing techniques with appropriate statistical monitoring tools, selected according to the underlying fault physics. We argue that “one size does not fit all”: the diagnostic approach must be tailored to whether the fault exhibits primarily steady-state (stationary) behavior or dynamic, time-varying characteristics.

This study tests the hypothesis through two experimental case studies:

• Case 1—Leading-edge erosion: Discrete wavelet transform (DWT), suited for multi-resolution transient analysis, is paired with BOCPD to assess whether advanced CPD methods improve the detection of gradual, surface-level degradation.
• Case 2—Twist misalignment: The short-time Fast Fourier Transform (FFT), effective for capturing harmonic behavior, is combined with the CUSUM algorithm to identify the spectral redistribution and instability associated with aerodynamic imbalance.

The primary contributions of this work are as follows:

• Methodological integration: We present a unified diagnostic workflow that systematically links time–frequency feature extraction with statistical decision-making tools.
• Physics-informed tool selection: Through experimental evidence, we show that leading-edge erosion introduces a global energy offset—making it best identified through feature-based trend monitoring—whereas twist faults induce spectral redistribution, for which centroid-based CPD is more effective.
• Practical engineering guidelines: We provide a comparative, application-oriented roadmap to help practitioners select appropriate monitoring strategies based on the physical characteristics of the anticipated fault.

Accordingly, this research seeks to answer the following questions:

• For gradual erosion faults characterized by persistent spectral deviation, does incorporating advanced CPD (e.g., BOCPD) provide additional diagnostic value beyond feature-based trend monitoring?
• For twist-induced aerodynamic instabilities, can the combined use of frequency-domain analysis and the CUSUM algorithm effectively isolate and pinpoint the resulting spectral fluctuations?

The remainder of this paper is organized as follows. Section 2 describes the experimental setup, vibration dataset, and the trend monitoring and change point detection techniques employed in this study. Section 3 presents the diagnostic results for the two blade fault scenarios—leading-edge erosion and twist misalignment—using the proposed TM–CPD framework. Section 4 discusses the findings in the context of fault physics and signal characteristics, highlighting why different diagnostic strategies are appropriate for different fault types. Finally, Section 5 summarizes the main conclusions and outlines limitations and directions for future work.

2. Materials and Methods

2.1. Material

2.1.1. Dataset Description

The vibration data used in this study were sourced from the open access dataset “Wind Turbine Blades Fault Diagnosis Based on Vibration Dataset Analysis” by Ogaili et al. (2023), published in Data in Brief [25]. The dataset was generated using a controlled laboratory-scale wind turbine test rig at the Renewable Energy Laboratory, Department of Mechanical Engineering, University of Mustansiriyah, Baghdad, Iraq.

2.1.2. Experimental Setup

The measurements were conducted using a Computer-Controlled Wind Energy Unit (EEEC) provided by Edibon Equipment. The test setup includes a stainless-steel wind tunnel (2000 × 550 × 550 mm) with transparent viewing panels and a variable-speed axial fan capable of generating wind speeds from 1.3 m/s to 5.3 m/s. A schematic overview of the experimental setup is shown in Figure 1.

The laboratory-scale turbine consists of a 510 mm diameter rotor connected to a 60 W generator. Vibration data were collected using a PCB Piezotronics 352C65 uniaxial accelerometer mounted on the nacelle near the hub using an adhesive coupling. This sensor is commonly employed in vibration-based condition monitoring and its key specifications are summarized in

Table 1. Properties of the PCB Piezotronics 352C65 sensor.

Specification Parameter	Metric Value (SI)	Imperial/Standard Value
Broadband sensitivity	10.2 mV/(m/s2)	100 mV/g (±10%)
Dynamic range (peak)	±491 m/s2 pk	±50 g pk
Resolution (broadband)	0.0016 m/s2 RMS	16 μg RMS
Operational bandwidth	0.5 Hz–10 kHz	0.5–10,000 Hz

. It should be noted that the vibration measurements were not acquired directly from the blade surface. The accelerometer was mounted on the nacelle near the rotor hub; therefore, the recorded signals represent the global dynamic response of the blade–rotor–nacelle assembly rather than localized blade vibrations. As a result, distances along the blade as well as intrados or extrados distinctions are not applicable in the present study.

The accelerometer signal was conditioned through its built-in charge amplifier and digitized using a National Instruments USB-4431 data acquisition module, which provides 24-bit resolution across five analog input channels. Data were sampled at a fixed rate of 1000 Hz, fully satisfying the Nyquist criterion [26] for the vibration frequency range of interest. All measurements were recorded and visualized using LabVIEW version 2020.

2.1.3. Measurement Procedure and Fault Scenarios Investigated

The turbine blades used in the experiment were fiber-reinforced polymer (FRP) components with a solid core, each measuring 300 mm in length and operating at a 60° blade angle under healthy conditions. For every test, 500-time samples—approximately 0.5 s of vibration data—were collected at a constant wind speed. The mechanical setup and environmental conditions were kept identical across all experiments to ensure consistency and comparability among cases [25]. Two fault scenarios were investigated in this study, each assessed relative to its corresponding healthy baseline:

Case Study 1—Surface Erosion Fault: The surface erosion fault investigated in Case Study 1 corresponds to an early-stage, mild degradation of the blade’s leading-edge surface. Following the experimental procedure described by Ogaili et al. [25], erosion was intentionally introduced by uniformly abrading the blade’s outer surface using fine sandpaper to simulate aerodynamic wear caused by prolonged exposure to rain, dust, and airborne particles. This degradation primarily affects the leading-edge region, where aerodynamic sensitivity is highest, resulting in increased surface roughness and subtle but persistent changes in the vibration response. As the dataset is intended for vibration-based fault diagnosis rather than material characterization, erosion depth and geometric material loss were not quantified; instead, the erosion severity is defined qualitatively as early-stage surface degradation. The vibration response of the eroded blade at a wind speed of 5 m/s was compared against that of the healthy blade under identical operating conditions.

Case Study 2—Twist Fault: The twist fault investigated in Case Study 2 represents a blade angle misalignment leading to aerodynamic asymmetry. Under healthy conditions, all three blades operate at a uniform pitch angle of 60°. To simulate a twist-related fault, one of the three blades was deliberately set to a pitch angle of 50°, while the remaining blades retained the nominal 60° configuration, following the experimental procedure described by Ogaili et al. [25]. This 10° pitch-angle deviation introduces an imbalance in aerodynamic loading, as the twisted blade generates a different lift and drag force compared to the other blades. The resulting asymmetric aerodynamic forces produce periodic fluctuations in rotor loading and vibration response, which are reflected in increased spectral energy and redistribution of frequency content. The vibration signal recorded at a wind speed of 5.3 m/s was analyzed relative to the healthy blade under identical operating conditions.

2.1.4. Dataset Structure

The complete dataset consists of 35 CSV files, each corresponding to a specific combination of wind speed and blade condition. Vibration acceleration is recorded in gravitational units (1 g = 9.80665 m/s²). Thanks to its precisely controlled laboratory conditions, this dataset provides a reliable benchmark for evaluating the effects of distinct mechanical faults under realistic aerodynamic loading [25].

For this study, four specific subsets were selected to evaluate two distinct fault scenarios:

Case 1—Surface Erosion Fault: Comparison between a healthy blade and an eroded blade at 5 m/s. This subset was chosen to assess the DWT–BOCPD integration’s ability to capture transient, non-stationary energy changes caused by surface degradation.

Case 2—Twist Fault: Comparison between a healthy blade and a twisted blade at 5.3 m/s. This subset was selected to evaluate the FFT–CUSUM integration’s capability to detect gradual frequency-domain deviations resulting from aerodynamic imbalance.

These pairs allow a direct assessment of how surface erosion and blade-angle misalignment affect the turbine’s vibration response.

Although wind speed and rotational speed are known to influence the vibration spectrum of wind turbine blades, this study controls for operating-condition variability by performing fault diagnosis through within-condition comparisons. For each investigated case, vibration signals from the faulty blade are directly compared with healthy reference signals acquired at the same wind speed and under identical experimental conditions. As a result, the extracted features and change detection results primarily reflect fault-induced deviations rather than changes caused by operating condition variability. This baseline-referenced strategy is commonly adopted in vibration-based condition monitoring to isolate structural and aerodynamic fault effects when operating conditions cannot be fully decoupled.

2.2. Trend Monitoring and Change Point Detection Methods

2.2.1. Trend Monitoring

TM is a diagnostic approach that tracks the temporal evolution of operational parameters to identify gradual degradation and incipient faults before they escalate. Unlike simple threshold-based methods, TM analyzes long-term variations and emerging patterns, making it highly effective for detecting subtle fault signatures amid the noisy and stochastic operational environment of wind turbines. Continuous monitoring of variables such as torque, vibration, and temperature is essential for early fault detection and maintenance optimization [27]. While TM—particularly torque analysis—has proven valuable for diagnosing drivetrain and gearbox faults [28], its industrial adoption is limited by the high cost and complexity of torque sensor installation compared to mature vibration-based systems. Nevertheless, TM remains a vital complementary tool for enhancing fault detection reliability [29]. In this study, we employ TM tools based on FFT and DWT to extract and analyze relevant signal features.

(a)

Fast Fourier Transform (FFT)

The signal processing pipeline for wind turbine condition monitoring is fundamentally based on the Continuous Fourier Transform (CFT), which decomposes a time-domain signal into its constituent frequency components [30]. In practical digital implementations, this operation is approximated using the Discrete Fourier Transform (DFT) and efficiently computed using the FFT, reducing the computational cost to $O\left(NlogN\right)$. The resulting frequency spectrum X(k) is expressed in Equation (1).

$$X\left(k\right)={\sum }_{n=0}^{N-1}x\left(n\right)e^{-j2\pi kn/N}$$

(1)

Although the FFT is effective for analyzing stationary signals, its output represents a global spectral average and therefore loses temporal information. This poses a significant limitation in wind turbine applications, where varying aerodynamic loads produce non-stationary signals containing short-lived or evolving fault signatures. To overcome this, time–frequency analysis techniques, such as the CWT and the Short-Time Fourier Transform (STFT), are used to localize signal energy simultaneously in time and frequency. While the CWT offers multi-resolution capabilities (discussed later), the STFT remains widely used due to its straightforward extension of the Fourier framework and its computational efficiency.

The STFT applies a sliding window function ω(n) to the signal, enabling the computation of the DFT within each localized segment [31,32]. This approach preserves both temporal and spectral information, making it possible to track how fault-related frequencies evolve over time. The STFT is mathematically expressed in Equation (2).

$${STFT}_x\left(m, \omega \right)={\sum }_{n=-\infty }^{\infty }x\left(n\right)\omega (n-m)e^{-j\mathit{\omega }n}$$

(2)

By selecting an appropriate window function (e.g., Hamming or Hanning), the STFT produces a spectrogram that reveals how spectral content evolves over time. This time-resolved representation is particularly effective for highlighting intermittent or evolving mechanical anomalies—common in wind turbine drivetrain systems—that would otherwise be obscured in conventional FFT analysis.

(b)

Wavelet Transform (WT) and Discrete Wavelet Decomposition (DWT)

To analyze the non-stationary vibration signals characteristic of wind turbine operation, this study employs the wavelet transform (WT). In contrast to the short-time Fourier transform (STFT), which relies on a fixed window size, WT provides a multi-resolution analysis framework with adaptive time–frequency localization. This property enables high temporal resolution for high-frequency transients and high spectral resolution for low-frequency components, making WT well suited for vibration-based fault diagnosis in rotating machinery [33].

While the continuous wavelet transform (CWT) is commonly used for qualitative time–frequency visualization, the present study employs the discrete wavelet transform (DWT) for quantitative feature extraction and condition monitoring. DWT provides an orthogonal and computationally efficient multi-resolution decomposition, which is particularly suitable for energy-based feature analysis. Accordingly, all wavelet-based results presented in this work are obtained using DWT.

In the discrete wavelet transform, the vibration signal $x\left(n\right)$ is decomposed into approximation and detail components through dyadic scaling and translation of a selected mother wavelet. At a given decomposition level $j$, the approximation coefficients $Aj\left(k\right)$ and detail coefficients $Dj\left(k\right)$ are computed as

$$Aj\left(k\right)={\sum }_{n}x\left(n\right) {\mathsf{\varphi }}_{j,k}\left(n\right)\hspace{1em}\hspace{1em}Dj\left(k\right)={\sum }_{n}x\left(n\right) {\psi }_{j,k}\left(n\right)$$

(3)

where ${\mathsf{\varphi }}_{j,k}\left(n\right)$ and ${\psi }_{j,k}\left(n\right)$ denote the discrete scaling and wavelet functions, respectively. Each successive decomposition level isolates a lower frequency band of the original signal.

In this study, DWT decomposition using the Daubechies wavelet (db4) is applied, and statistical energy features are extracted from the detail coefficients at four decomposition levels. These coefficients, denoted as D1–D4, correspond to progressively lower frequency bands and are subsequently used for trend monitoring and change point detection.

2.2.2. Change Point Detection

CPD is a statistical methodology used to identify abrupt transitions in the underlying probability distribution of a time series—such as changes in mean, variance, or correlation structure. In contrast to TM, which targets gradual and long-term behavioral shifts, CPD is expressly designed to detect sudden structural changes that may signal emerging component faults or shifts in operating conditions [34]. The foundational principles of CPD were established by Basseville and Nikiforov, whose work on sequential statistical decision-making became central to industrial quality assurance and process control [35]. In recent years, CPD has been increasingly applied in the wind energy sector, particularly for analyzing SCADA data to diagnose gearbox degradation, power-curve anomalies, and other rapid-onset failures [15,36,37,38,39]. Despite challenges posed by environmental variability and the inherently non-stationary nature of turbine signals, CPD remains an essential diagnostic tool capable of detecting sudden degradation events that may be missed by trend-based approaches. The following subsections describe the specific CPD algorithms employed in this study.

(a)

Bayesian Online Change Point Detection (BOCPD)

BOCPD offers a principled probabilistic framework for detecting abrupt structural changes in time-series data [40]. In wind turbine condition monitoring, it plays a key role in identifying deviations from nominal behavior caused by emerging faults or physical deterioration [41]. Building on this foundation, Tsaknaki et al. [42] introduced a Bayesian extension that dynamically updates autoregressive model parameters, improving robustness to non-stationary operating conditions and enabling accurate, real-time adaptation. The central task of BOCPD is to estimate the run length $r_t$, defined as the time elapsed since the last change point. This inference is performed by recursively updating the posterior distribution of $r_t$, as expressed in Equation (4).

$$P(r_t\mid x_{1:t})\propto P(x_t\mid r_t,x_{t-r_{t:t-1}})P(r_t\mid r_{t-1})$$

(4)

where

• $r_t$ denotes the current run length at time t.
• $x_t$ represents the observed data (e.g., vibration amplitude, temperature).
• $P(r_t\mid r_{t-1})$ is the transition probability, governed by the hazard function.
• $P(x_t\mid r_t,x_{t-r_{t:t-1}})$ is the likelihood of the new observation given the current run length.

The transition dynamics are controlled by the Hazard function ${H(r}_t)$, which quantifies the probability of a change point occurring at the current step, as shown in Equation (5).

$${H(r}_{t-1})=P(r_t=0\left|r_{t-1}\right)$$

(5)

To assess anomalies, we compute the predictive distribution. Given a statistical model $f\left(x_t\right|\theta )$ with parameters $\theta$ (e.g., Gaussian parameters for vibration data), the predictive probability is obtained by marginalizing over $\theta$ using Equation (6):

$$P\left(x_t∣r_t,x_{t-r_{t:t-1}}\right)= \int f\left(x_t|\theta \right)\mathrm{P}\left(\theta |x_{t-r_{t:t-1}}\right)d\theta$$

(6)

This integration facilitates the robust detection of deviations from expected behavior by accounting for parameter uncertainty.

In this study, no fixed posterior probability threshold is imposed; instead, BOCPD outputs are interpreted qualitatively based on the presence or absence of pronounced posterior probability peaks.

(b)

Cumulative Sum (CUSUM)

While BOCPD is highly effective for detecting abrupt structural changes, the CUSUM algorithm is specifically designed to identify small but persistent shifts in system behavior [17]. In wind turbine condition monitoring, CUSUM plays a critical role in tracking gradual deviations in parameters such as vibration amplitude, power output, or component temperature, providing early warning of emerging faults before they escalate into major failures. CUSUM is computationally efficient and works by cumulatively summing deviations of the observed signal from a reference mean. This memory-based mechanism enables the detection of subtle drifts that would otherwise remain hidden within the noise of instantaneous thresholding methods. In this study, we employ the tabular (one-sided) CUSUM formulation to monitor both upward and downward trends using two statistics: the upper cumulative sum ($S_t^{-}$) and the lower cumulative sum ($S_t^{+}$). The update rules for these statistics are given in Equations (7) and (8) below:

$$S_t^{-}=\mathrm{m}\mathrm{a}\mathrm{x}(0,S_{t-1}^{+}+\left(x_t{-\mu }_0-k\right))$$

(7)

$$S_t^{+}=\mathrm{m}\mathrm{a}\mathrm{x}(0,S_{t-1}^{-}+\left({\mu }_0-x_t-k\right))$$

(8)

where

• $x_t$ is the observed value at time $t$.
• ${\mu }_0$ represents the baseline mean under normal operating conditions.
• $k$ is the drift reference value, typically set to $\delta /2$, where $\delta$ is the magnitude of the shift to be detected.
• The max (0, …) function resets the accumulation when the signal remains close to the target, preventing negative drift.

A fault alarm is conventionally triggered when either cumulative statistic exceeds a predefined decision threshold $h$, that is, when $S_t^{+}>h$ or $S_t^{-}>h$. In the present study, this formulation is used to illustrate the underlying detection mechanism; however, diagnostic interpretation is based on relative divergence from the healthy reference signal rather than on a fixed numerical threshold, enabling robust comparison across operating conditions.

It should be emphasized that the diagnostic framework adopted in this study does not rely on fixed or universal threshold values to separate serviceable and faulty states. Instead, fault detection is performed through baseline-referenced comparison with healthy vibration signals acquired under identical operating conditions. For the CUSUM-based analysis, the healthy signal defines the reference cumulative behavior, and fault presence is inferred from sustained divergence or pronounced localized deviations relative to this baseline rather than from an absolute decision threshold. For BOCPD, diagnostic interpretation is probabilistic, where the presence or absence of statistically meaningful change points is assessed based on the evolution of posterior probabilities instead of predefined threshold levels. This approach is well suited for comparative fault analysis under controlled experimental conditions and avoids overfitting to specific operating states.

2.3. Proposed TM–CPD Integration

To overcome the limitations of treating signal processing and statistical monitoring as independent tasks, this study introduces an integrated monitoring and diagnostic framework that systematically couples time–frequency analysis with appropriate statistical decision tools. The core principle underlying this framework is that diagnostic strategies must reflect the physics of the fault. More specifically, it is supposed that stationary, steady-state faults demand different analytical tools than dynamic, time-varying faults.

To evaluate this approach, two representative fault types are used to validate two specific combinations of TM and CPD methods:

• Leading-edge surface erosion detection using DWT–BOCPD integration: DWT, which provides multi-resolution insight into transient energy variations, is paired with BOCPD. This combination is used to determine whether sophisticated CPD techniques enhance the detection of gradual surface degradation.
• Twist misalignment detection using FFT–CUSUM integration: FFT, well suited for resolving harmonic behavior, is integrated with the CUSUM algorithm to capture the subtle spectral redistribution and instability caused by aerodynamic imbalance.

The main goal is to answer two research questions: (1) Does integrating advanced CPD methods such as BOCPD yield additional diagnostic value for slow, progressive erosion, or is feature-based trend monitoring alone sufficient? (2) Can combining frequency-domain analysis with the CUSUM algorithm effectively detect, isolate, and localize the spectral fluctuations introduced by twist-related aerodynamic instability?

3. Case Studies and Results

3.1. Case 1—Surface Erosion Fault

3.1.1. DWT Analysis

To characterize the vibration behavior of the wind turbine blade under healthy and erosion conditions, a DWT analysis was performed using the Daubechies-4 (db4) wavelet with four decomposition levels. This wavelet family is widely adopted in vibration-based condition monitoring for its capacity to capture both high-frequency transients and low-frequency structural responses within an interpretable multiresolution framework [43]. For each recorded signal, detail coefficients D1–D4 were extracted, and two standard metrics—energy and Shannon entropy—were calculated to quantify the intensity and distributional complexity of the vibration content [44]. The resulting global features are shown in

Table 2. Global DWT features for the healthy and eroded blade at a wind speed of 5 m/s using db4, level = 4, whole signal.

Feature	Healthy Blade	Eroded Blade	Absolute Difference	Percentage Difference (%)
Energy D1	0.001550	0.001700	+0.000150	+9.68%
Entropy D1	3.0841	2.7904	−0.2937	−9.52%
Energy D2	0.001065	0.001545	+0.000480	+45.07%
Entropy D2	3.5792	3.2736	−0.3056	−8.54%
Energy D3	0.002002	0.002686	+0.000684	+34.17%
Entropy D3	4.0809	4.1865	+0.1056	+2.59%
Energy D4	0.003380	0.004125	+0.000745	+22.04%
Entropy D4	4.6143	4.8221	+0.2078	+4.50%

, which indicates that erosion drives a consistent increase in wavelet energy, with the most pronounced deviations observed in the D2 (+45%) and D3 (+34%) bands. Conversely, entropy exhibits smaller, mixed fluctuations. Given the dominance of the energy changes in these bands, the Energy D2 and Energy D3 features were selected for short-time analysis. These windowed features serve as time-series inputs for subsequent BOCPD [18], enabling the precise detection of localized deviations associated with surface erosion.

Figure 2 compares the mid-frequency vibration content of the healthy and eroded blades across multiple overlapping time windows for the short-time DWT Energy D2 feature. The eroded blade consistently exhibits higher D2-level energy than the healthy blade throughout the signal duration, characterized by a pronounced peak around 0.13 s, followed by a gradual decline while remaining above the healthy trend. In contrast, the healthy blade shows a smoother and steadily increasing energy profile. This consistent elevation in Energy D2 feature reflects the intensified mid-frequency activity resulting from surface roughness and the associated aerodynamic disturbances [45]. The distinct separation between the two curves confirms that the Energy D2 feature is highly sensitive to erosion-induced vibration changes, identifying it as a robust candidate for the subsequent change-point analysis using BOCPD.

Figure 3 reveals distinct differences in the lower mid-frequency vibration behavior between the healthy and eroded blades for the short-time DWT Energy D3 feature. Across all time windows, the eroded blade displays substantially higher D3-level energy, characterized by a distinct peak near 0.13 s, followed by a moderate decline that nonetheless remains consistently above the healthy baseline. In contrast, the healthy signal shows a modest increase up to approximately 0.19 s before stabilizing at lower values. This persistent elevation in Energy D3 feature also indicates stronger low–mid-frequency fluctuations, a phenomenon linked to the modified aerodynamic loads and performance degradation caused by surface erosion [46]. The clear and stable energy gap between the two signals demonstrates that Energy D3 feature effectively captures erosion-related vibration characteristics, supporting its use as a complementary feature to D2 in the BOCPD method.

3.1.2. BOCPD Analysis

The dominance and stability of the energy changes in the D2 and D3 levels identify these components as the most sensitive indicators of erosion-related vibration variations. Therefore, to complete the proposed integrated framework, these two wavelet bands were selected as the time-series inputs for the subsequent BOCPD analysis, enabling a probabilistic assessment of fault onset and temporal evolution.

Figure 4 presents the BOCPD analysis of the short-time DWT Energy D2 feature. The upper subplot displays the Z-scored energy trajectories, revealing a clear separation between the healthy (orange) and eroded (blue) conditions. The eroded blade’s signal exhibits a distinct non-stationary pattern with a pronounced amplitude peak at t = 0.13 s, contrasting with the smoother and gradually increasing trend of the healthy blade. This consistent elevation in the eroded blade’s signal reflects the stronger mid-frequency vibration content generated by surface roughness and aerodynamic disturbance [46].

Despite these visible differences in signal morphology, the corresponding BOCPD analysis in the lower subplot did not produce distinct or localized change-points within either signal. Instead, the change-point probability remained approximately constant and low (≈0.02) across all time steps. No significant probability spikes were observed, indicating that the algorithm did not detect a definitive structural break. This behavior is expected given the short length of the time series and the gradual evolution of the Energy D2. Although the Energy D2 feature effectively differentiates the two operating conditions, the erosion-induced changes manifest as smooth, global shifts rather than abrupt temporal transitions. As a result, BOCPD does not detect strong intra-signal regime changes for this feature under the present operating speed of 5 m/s.

Figure 5 presents the BOCPD analysis of the short-time DWT Energy D3 feature. The upper subplot illustrates the temporal evolution of the Z-scored energy, revealing a distinct contrast between the two conditions. The eroded blade’s signal (blue) consistently exhibits higher energy values, characterized by a sharp, high-magnitude transient peak at t = 0.13 s (Z-score ≈ 2.0), followed by a rapid decline. In contrast, the healthy blade’s signal (orange) shows a smoother, wave-like progression with a modest increase up to a delayed maximum around t = 0.19 s before stabilizing. This confirms that D3-level wavelet energy effectively captures lower-mid-frequency variations associated with surface degradation and aerodynamic performance loss, serving as a sensitive indicator of erosion effects [46].

Despite the prominent localized spike observed in the eroded blade’s signal, the BOCPD results in the lower subplot show no prominent change-points. The posterior probability remains nearly uniform and negligible (≈0.02) across all time indices. This outcome indicates that the algorithm treated the transient peak as an outlier or part of the global variance rather than a sustained regime shift. The absence of detection is consistent with the smooth, slowly varying nature of the Energy D3 sequence and the limited number of windows available for analysis. Consequently, while the Energy D3 feature clearly separates the healthy and erosion conditions globally, the lack of abrupt internal fluctuations means BOCPD does not reveal time-localized change-points under these conditions.

3.2. Case 2—Twist Fault

Figure 6 presents the overlaid FFT amplitude spectra of the healthy and twist-faulted wind-turbine blades operating at 5.3 m/s. Both signals exhibit a dominant peak at the low-frequency region associated with the rotor’s fundamental rotational frequency. However, the twist-fault spectrum shows a noticeably broader and more irregular amplitude distribution across the full 0–500 Hz range. In the healthy case, the spectral amplitudes remain relatively smooth and concentrated, with lower energy content beyond the main harmonic. In contrast, the twist-fault condition produces consistently higher amplitudes at multiple frequencies, particularly in the mid-band and high-frequency regions (e.g., near 50–100 Hz, 150–200 Hz, 250–300 Hz, and beyond). This elevated and more scattered spectral pattern reflects the aerodynamic imbalance introduced by an unsynchronized blade angle [47]. Because the twist fault disturbs the periodic loading on the rotor, the resulting vibrations spread energy into a wider set of harmonics and sidebands, leading to a noisier and more broadband spectrum. The wider spectral footprint and multiple elevated peaks therefore serve as clear indicators of twist-induced instability relative to the stable, narrow-band behavior of the healthy blade.

3.2.1. FFT Features

To characterize the frequency-domain behavior of the healthy and twist-faulted blades at 5.3 m/s, the vibration signals were first examined using the full FFT amplitude spectrum. This initial comparison demonstrated that while both signals contain the same fundamental rotor frequency, the twist-faulted blade exhibits a broader and more irregular spectral distribution with elevated amplitudes across multiple higher-frequency bands. This indicates that twist deformation introduces additional aerodynamic perturbations that redistribute vibration energy across a wider range of frequencies [47]. To enable a time-resolved comparison and prepare the data for subsequent change point detection, a short-time FFT analysis was performed. Each vibration signal was segmented into six overlapping windows of 128 samples (50% overlap) using a Hann window to obtain the single-sided amplitude spectrum. From each window, four standard spectral descriptors were extracted: spectral energy, spectral centroid, spectral entropy, and dominant frequency. The resulting feature matrices as shown in

Table 3. Short-time FFT features for the healthy blade at a wind speed of 5.3 m/s.

Time Center (s)	Spectral Energy	Spectral Centroid	Spectral Entropy	Dominant Frequency (Hz)	Fs (Hz)
0.0635	0.037193	220.212941	4.698935	50.78125	1000
0.1275	0.036908	238.313912	4.712699	0.00000	1000
0.1915	0.042527	233.527067	4.704256	0.00000	1000
0.2555	0.044467	242.466711	4.680576	46.87500	1000
0.3195	0.034379	233.744984	4.624165	50.78125	1000
0.3835	0.047053	222.422595	4.732791	50.78125	1000

and

Table 4. Short-time FFT features for the twist-faulted blade at a wind speed of 5.3 m/s.

Time Center (s)	Spectral Energy	Spectral Centroid	Spectral Entropy	Dominant Frequency (Hz)	Fs (Hz)
0.0635	0.046868	238.201584	4.705423	46.87500	1000
0.1275	0.053766	240.903462	4.678903	50.78125	1000
0.1915	0.053763	221.446806	4.689938	50.78125	1000
0.2555	0.059613	248.396287	4.737508	50.78125	1000
0.3195	0.056793	212.019799	4.658341	50.78125	1000
0.3835	0.062206	225.994091	4.615763	46.87500	1000

reveal that the twist-fault condition produces systematically higher spectral energy in every time window (=0.048–0.094) and greater fluctuation in the spectral centroid compared to the stable, low-energy profile of the healthy blade. These consistent increases confirm that the twist fault produces a measurable spectral deviation that persists over time. Consequently, spectral energy—identified as the feature most sensitive to the fault—was selected as the input for a CUSUM-based detector [17] to identify the onset and persistence of twist-induced vibrational changes.

Figure 7 compares the short-time spectral energy of the healthy and twist-faulted blades across six consecutive analysis windows. The healthy blade exhibits a relatively stable energy profile, fluctuating mildly between approximately 0.034 and 0.047, consistent with a well-balanced rotor under steady aerodynamic loading. In contrast, the twist-faulted blade shows noticeably higher spectral energy in every window, with values rising progressively from 0.046 to more than 0.062. This consistent elevation in broadband energy reflects the aerodynamic asymmetry introduced by the twisted blade angle, which increases turbulent airflow interaction and injects additional vibration energy into the system [47]. The clear separation between the two curves demonstrates that spectral energy is a highly sensitive indicator of twist-induced structural-aerodynamic changes.

Figure 8 shows the short-time spectral centroid for both operating conditions. The healthy blade presents a smooth, gradually varying centroid trajectory between approximately 220 and 242 Hz, representing a stable distribution of spectral mass around the rotor’s dominant frequency components. The twist-faulted blade shows a more irregular pattern, with centroid values deviating upward (~248 Hz) and downward (~212 Hz) more sharply than the healthy case. These fluctuations reveal that twist deformation not only increases spectral energy but also redistributes energy across the frequency spectrum, causing momentary shifts in the center of gravity of the spectrum. This phenomenon is consistent with the broadband spectral scattering and modulation caused by aerodynamic imbalance [47]. Although the centroid difference is less uniform than the energy difference, the twist blade consistently exhibits more abrupt centroid excursions, confirming that the spectral centroid acts as a secondary but meaningful indicator of twist-induced spectral distortion.

3.2.2. CUSUM Detection Results

Figure 9 shows the one-sided CUSUM of the short-time spectral energy, using the healthy blade as the statistical baseline (mean and standard deviation estimated from the healthy series). The CUSUM for the healthy blade remains essentially flat and close to zero for the entire record, confirming that the healthy energy fluctuations are statistically consistent with the baseline. In contrast, the twist-fault blade exhibits a steadily increasing CUSUM curve. Starting from zero, the CUSUM grows monotonically with each subsequent window, reaching a value of approximately 0.06. This behavior indicates that the short-time energy estimates for the twist blade lie persistently above the healthy mean, causing the deviations to accumulate rather than cancel out. Although no sharp ‘jump’ is observed—as the twist fault acts as a global increase in energy rather than an abrupt event—the diverging CUSUM trend clearly separates the twist condition from the healthy baseline. This confirms that the method effectively detects the consistent excess of spectral energy introduced by the fault [17,48].

Figure 10 presents the CUSUM of the spectral centroid using the healthy blade as the reference. For the healthy condition, the CUSUM remains at zero for the first three windows, then exhibits a modest peak of about 1.9 around 0.26 s before returning to zero in the final windows. This indicates that the healthy centroid experiences only a brief, moderate deviation from its baseline value, consistent with normal turbulence and measurement variability. The twist-faulted blade, however, shows a markedly different pattern. After a small rise in the second window, the centroid CUSUM displays a pronounced spike at 0.26 s, reaching nearly 7.8—more than four times the peak value observed in the healthy case. This sharp excursion reflects a strong, localized redistribution of spectral mass caused by the twisted blade, which temporarily shifts the dominant frequency content away from the baseline centroid [47]. The subsequent rapid drop back to zero indicates that this is a transient but significant event. Compared with spectral energy, the centroid-based CUSUM provides a much more distinct ‘change point’ signature, demonstrating that twist deformation affects not only the overall energy level but also the frequency balance of the vibration spectrum in a way that is highly detectable by CUSUM [17,48].

4. Discussion

4.1. Case 1: Erosion Fault Detection Using DWT and BOCPD

The Case 1 analysis evaluated the capability of DWT features, combined with BOCPD, to discriminate between a healthy blade and a surface-eroded blade operating at 5 m/s. The global DWT (db4) features demonstrated that the eroded condition exhibits systematically higher wavelet energies across all detail levels (D1–D4) compared to the healthy condition [43]. The differences were particularly pronounced at the D2 and D3 levels, where energy increased by approximately 45% and 34%, respectively. These bands correspond to mid-frequency components associated with airflow irregularities and structural surface disturbances, confirming that erosion induces detectable modifications in the vibration spectrum due to aerodynamic degradation [46]. To further examine how these differences manifest over time, short-time windowed DWT analysis was performed for the Energy D2 and Energy D3 features. The resulting temporal plots showed a clear and consistent separation between the healthy and erosion signals. However, the subsequent BOCPD analysis did not identify distinct or localized change-points within either signal. This outcome is not a limitation of the method but rather a reflection of the fault characteristics: erosion manifests as a persistent, global change in the overall vibration energy rather than a time-localized event. Since BOCPD is inherently designed to detect abrupt regime shifts [18], it does not assign high posterior probability to the smooth, constant deviations observed here.

It is also important to note that the performance of BOCPD is influenced by its parameterization, particularly the choice of hazard function and the assumed observation model. In this study, a constant hazard function was adopted due to the absence of prior information regarding the timing of potential regime changes. Sensitivity checks with different hazard rate values revealed that higher hazard probabilities led to more frequent but weak detections, whereas lower values resulted in smoother posterior distributions without pronounced change-point peaks. Across all tested configurations, BOCPD consistently failed to identify statistically significant change points for the erosion case. This confirms that the non-detection is primarily attributable to the steady-state and non-transient nature of early-stage surface erosion rather than suboptimal parameter tuning.

Consequently, this study concludes that feature-based classification using DWT energies is highly suitable for erosion detection, whereas online change-point analysis does not enhance performance for this specific, non-transient fault scenario.

4.2. Case 2: Twist Fault Detection Using FFT and CUSUM

The twist-fault case study demonstrates that frequency-domain analysis, particularly using short-time FFT features, is highly effective for characterizing the aerodynamic and structural disturbances introduced by a twisted blade. The short-time spectral features extracted from six analysis windows reveal clear and consistent differences between the healthy and faulted signals. Spectral energy across all windows is systematically higher for the twist fault, reflecting the additional broadband vibration generated by asymmetric aerodynamic loading [47]. This behavior is physically consistent with blade pitch-angle asymmetry, where uneven lift and drag forces among the blades induce aerodynamic imbalance and inject excess vibration energy into the system.

However, this increase in spectral energy manifests primarily as a global signal offset rather than a sudden deviation. Consequently, when CUSUM is applied using the healthy blade as the baseline, the twist-fault energy produces a steadily increasing cumulative sum. This reveals that energy-based detection is sensitive to persistent bias—confirming the presence of fault-induced spectral amplification—but offers limited temporal localization [17,48].

In contrast, the spectral centroid provides a much more dynamic signature of the twist fault. The CUSUM of the spectral centroid exhibits a sharp, localized spike reaching nearly four times the magnitude observed for the healthy blade. This response highlights the centroid’s ability to capture transient spectral redistribution, where uneven blade angles lead to momentary shifts in the energy-weighted frequency content of the vibration signal. Such sharp centroid excursions are consistent with aerodynamic perturbations induced by pitch-angle asymmetry, in which unbalanced lift generation causes transient redistribution of vibration energy across the frequency spectrum [47].

Therefore, the combination of short-time spectral analysis and CUSUM demonstrates that centroid-based CUSUM is particularly well suited for detecting twist faults, as it provides a clear and interpretable change-point signature even within short vibration records.

4.3. Rationale for TM–CPD Selection by Fault Type

The selection of different trend monitoring (TM) and change point detection (CPD) combinations in this study is guided by the underlying physical characteristics of the investigated blade faults rather than by algorithmic preference. This fault-aware perspective is consistent with recent systematic analyses of TM–CPD integration for wind turbine condition monitoring, including our recent review study [24].

Early-stage surface erosion represents a gradual degradation mechanism that primarily increases surface roughness along the blade’s leading edge. Such faults introduce low-amplitude, broadband, and multi-scale variations in vibration energy without producing abrupt statistical transitions. As summarized in [24], wavelet-based trend monitoring is particularly suitable for capturing these subtle energy redistributions across multiple frequency bands. Bayesian Online Change Point Detection (BOCPD) is therefore employed to probabilistically assess whether such gradual variations correspond to statistically significant regime changes. The absence of distinct change points observed in this case is a meaningful diagnostic outcome, confirming that BOCPD does not provide additional benefit for very mild, steady-state erosion under low wind-speed conditions, rather than indicating an inappropriate methodological choice.

In contrast, twist faults introduce blade pitch-angle asymmetry, leading to persistent aerodynamic imbalance and frequency redistribution in the vibration response. These faults produce sustained deviations in frequency-domain features, particularly spectral energy and spectral centroid. As discussed in [24], short-time FFT is effective for characterizing such spectral behavior, while the CUSUM algorithm is well suited for detecting cumulative deviations relative to a healthy baseline. Unlike BOCPD, which targets abrupt regime shifts, CUSUM reliably captures the persistent bias and localized spectral fluctuations induced by twist-related aerodynamic instability.

Overall, this fault-dependent pairing of TM and CPD methods reflects the complementary strengths of the selected techniques and aligns with best-practice TM–CPD combinations reported in the recent literature [24]. The results demonstrate that CPD should not be applied uniformly across all fault types; instead, its effectiveness depends strongly on whether the fault induces abrupt, time-localized changes or smooth, steady-state deviations in the monitored features.

5. Conclusions

This study examined the integration of trend monitoring (TM) and change point detection (CPD) for vibration-based wind turbine blade fault diagnostics, emphasizing how fault characteristics govern the effectiveness of different analytical strategies. Rather than treating TM–CPD integration as a universally optimal solution, the results demonstrate that its diagnostic value is fundamentally dependent on the physical nature and temporal evolution of the fault.

The findings indicate that faults characterized by gradual and persistent degradation primarily manifest as global shifts in vibration features. In such scenarios, feature-based trend monitoring—particularly time–frequency energy indicators—provides reliable fault discrimination, whereas advanced change point detection techniques contribute limited additional benefit due to the absence of abrupt statistical transitions. In contrast, faults that induce aerodynamic asymmetry and dynamic instability generate localized, time-varying spectral redistributions, for which the combination of frequency-domain trend monitoring and CPD enables both accurate detection and temporal localization of fault-related deviations.

These relationships between fault type, signal characteristics, and suitable diagnostic methods are systematically summarized in

Table 5. Comparative summary of Case 1 (Erosion Fault) and Case 2 (Twist Fault).

Aspect	Case 1—Erosion Fault (5 m/s)	Case 2—Twist Fault (5.3 m/s)
Operating condition	Constant wind speed 5 m/s; early-stage leading-edge surface erosion (uniform sandpaper-induced abrasion)	Wind speed 5.3 m/s; blade angle misalignment (one blade twisted)
Fault nature	Surface roughness increases at the leading edge, altering local aerodynamic smoothness	Twist deformation, causing asymmetric aerodynamic loading
Signal characteristics	Subtle but persistent global energy increase; nearly overlapping time-domain amplitudes	Clear increase in vibration amplitude and spectral variability compared to healthy
Analytical method	DWT (db4) + BOCPD	Short-time FFT + CUSUM
Key features used	Wavelet Energy (D2, D3)	Spectral Energy and Spectral Centroid
Main plot observations	Stable separation in DWT energy trends; BOCPD shows no distinct internal change points	Energy consistently higher; centroid exhibits sharp local spikes; CUSUM shows clear divergence
Detection type	Weak global trend; no distinct faults detected	Strong global elevation (energy) + clear localized anomaly (centroid)
Sensitivity focus	Broad-band, low-frequency structural changes	Frequency redistribution and spectral imbalance
Strengths	Good for analyzing multi-scale transients; useful for faults with sharp shocks	Excellent at capturing both persistent bias and transient spectral shifts
Limitations	Erosion fault too weak; BOCPD not effective with small feature differences	Requires short-time windowing and feature extraction
Best suited fault type	Gradual surface degradation and early aerodynamic wear	Twist, misalignment, and aerodynamic imbalance faults
Overall finding	DWT + BOCPD fails to reveal erosion signatures at low wind speeds due to minimal feature separation	FFT + CUSUM robustly identifies twist faults via both amplitude increase and centroid-based change-point signatures

, which serves as a practical guideline for selecting appropriate TM–CPD combinations based on expected fault behavior. The table highlights that CPD should be applied selectively—primarily in cases where faults introduce time-localized or nonstationary changes—rather than indiscriminately across all degradation modes.

Overall, this work demonstrates that fault-aware integration of TM and CPD improves diagnostic interpretability, reduces unnecessary algorithmic complexity, and supports more informed maintenance decision-making. While the present investigation is based on controlled experimental data and limited-duration signals, the insights gained establish a robust foundation for future studies involving extended monitoring periods, additional fault classes, and real-world operating conditions.

Despite the clarity of the observed diagnostic trends, it should be noted that the experimental case studies employed relatively short vibration records and a limited number of samples, reflecting the constraints of laboratory-scale datasets. Consequently, the reported findings should be interpreted as proof-of-concept demonstrations rather than statistically exhaustive validation across all operating conditions. While the selected datasets are sufficient to illustrate the contrasting behaviors of gradual and transient blade faults, larger-scale datasets with extended monitoring durations and broader variability in wind speed, loading conditions, and fault severity are required to fully assess robustness, generalizability, and detection sensitivity. Addressing these limitations constitutes an important direction for future work toward deployment in real-world wind farm environments.

6. Prospects and Future Applications

The fault-aware integration of TM and CPD proposed in this study offers several promising prospects for practical wind turbine condition monitoring. By demonstrating that CPD provides substantial diagnostic benefits primarily for faults that induce dynamic spectral instability, the framework enables selective and computationally efficient deployment of CPD algorithms in online monitoring systems. This targeted use is particularly relevant for offshore wind turbines, where minimizing false alarms and reducing unnecessary maintenance interventions are critical.

The proposed methodology is well suited for integration into vibration-based structural health monitoring platforms and can support physics-informed digital twin frameworks by providing interpretable indicators for validating model predictions and detecting abnormal operating states. In addition, the fault-dependent TM–CPD selection strategy may be extended to other rotating machinery components and renewable energy systems exhibiting mixed steady-state and dynamic fault behaviors.

Future research will focus on applying the proposed framework to longer monitoring periods, additional blade fault types, and real-world operational datasets with varying wind and loading conditions. In particular, the integration of CFD-based aerodynamic simulations with vibration-based TM–CPD diagnostics is identified as a promising future direction to establish a more direct link between flow-induced phenomena and the observed spectral changes. The integration of aerodynamic simulations and higher-density sensor configurations is also identified as an important direction to further strengthen the physical interpretation of vibration-based diagnostics.