Unsupervised acoustic detection of fatigue-induced damage modes from wind turbine blades

Abstract

This paper proposes a new in-situ damage detection approach for wind turbine blades, which leverages blade-internal non-stationary acoustic pressure fluctuations caused by the mechanical loading as the main source of excitation. This acoustic excitation was leveraged for the detection of fatigue-related damage modes on a full-scale wind turbine blade undergoing edgewise fatigue testing. An unsupervised, data-driven structural health monitoring strategy was developed to learn the normal cavity-internal acoustic sequences generated by the blade’s load cycles and to detect damage-related anomalies in the context of those sequences. A linear cepstral-coefficient based feature set was used to characterize the cavity-internal acoustics and LSTM-autoencoders were trained to accurately reconstruct healthy-case sequences. The reconstruction error was then used to characterize anomalous acoustic patterns within the blade cavity. The technique was able to detect a damage event earlier than a strain-based system by 120,000 load cycles.

Keywords

Structural health monitoring wind turbine blades anomaly detection passive damage detection acoustics

Introduction

Of all wind turbine components, the wind turbine blades (WTBs) are among the most susceptible to damage due to their immense size, operational loads, and environmental exposure (Griffin, 2019; Jiang et al., 2018; RISO/DNV, 2002). Consequently, WTB maintenance can drive up the levelized cost of energy (LCOE) of the wind turbine system and inhibit financial viability of wind energy in comparison to other renewable and non-renewable resources (IRENA, 2012). Through the implementation of reliable structural health monitoring (SHM) techniques, WTB maintenance expenses can be greatly reduced through the early damage detection and the development of preventative maintenance strategies (IRENA, 2019). Although a variety of SHM methodologies have been developed for WTBs, few of these methodologies have been commercially adopted due to their high cost, insensitivity to common damage modes, or susceptibility to environmental noise (Li et al., 2015; Martinez-Luengo et al., 2016; McGugan et al., 2008; Mishnaevsky et al., 2017; Sorensen et al., 2002).

Recently, airborne acoustic excitation-based SHM methods, which detect damage via the analysis of air-propagating acoustic signals emanating from or propagating within the WTB cavity, have emerged in the research community to address the growing need for a reliable WTB SHM system. These acoustic techniques are typically active or passive in nature; in active acoustic SHM techniques, a controlled audible acoustic source signal is emitted by a loudspeaker and used to excite the WTB cavity. Microphones located on the opposite side of the cavity boundary are then used to detect reductions in transmission loss due to damage on the WTB’s outer shell (Arora et al., 2014; Beale et al., 2020a; Regan et al., 2017). Despite the high signal-to-noise ratio afforded by the active-based techniques, they are limited to the detection of cavity-breaching damage modes, such as holes or cracks, which impact the transmission loss over the cavity boundary. In contrast to active approaches, passive acoustic SHM techniques can detect a wider range of damage modes by leveraging the naturally-occurring acoustic excitation of the WTB. This naturally-occurring sound is typically produced by the airflow interaction with the WTB or can be generated by the WTB structure itself due to operational load fluctuations. For techniques relying on airflow-induced excitation, damage is detected by analyzing changes in the WTB cavity-internal acoustics due to sound produced by air leakage into the cavity (Beale et al., 2020b; Traylor et al., 2020), or through the analysis of WTB-external sound produced by the interaction of surface imperfections with airflow (Bouzid et al., 2015; Fazenda, 2011; García Márquez et al., 2022; Sanchez et al., 2021). Techniques relying on structurally-induced passive acoustic excitation leverage the cavity-internal acoustic excitation produced by WTB load fluctuations for damage detection. The WTB load fluctuations are caused by aerodynamic forces or occur repeatedly with every turn of the rotor, as in the case of gravitational loads (Jiang et al., 2018; RISO/DNV, 2002). These load fluctuations lead to the displacement and subsequent “breathing” of the WTB panels (as demonstrated in Figure 1), which results in the emission of audible acoustic energy within the blade cavity.

Figure 1.

Blade displacement due to gravitational load fluctuations (Windmills Tech, 2019).

In addition to generating a WTB cavity-internal acoustic signature, these load fluctuations can also lead to the development of fatigue-related damage modes on the blade. Fatigue-related damage modes include the failure of various adhesive layers, laminate delamination, disbonding at skin and core interfaces, splitting along fibers, in-plane compressive failure, and gelcoat disbonding and cracks, and are estimated to be responsible for up to 97% of WTB failures (Thomsen, 2009). Previous studies have demonstrated that these fatigue-related damage modes are capable of altering the structurally-induced cavity-internal acoustic signature of the WTB compared to its healthy state. In a study by Chen et al. tapping and rubbing noises were observed in a WTB with adhesive joint disbond damage that was subjected to flapwise and edgewise fatigue loading (Chen et al., 2021). Other studies have demonstrated the occurrence of audible sound signatures related to damage propagation in composite materials subjected to loading (Pearson et al., 2018). Only a handful of researchers have attempted to leverage structurally-induced sound for WTB damage detection. Krause and Ostermann utilized blade cavity internal microphones to detect transient cracking noises emitted from a WTB during fatigue testing. Using custom acoustic features, a decision tree was tasked with identifying the occurrence of impulsive crack events by comparing the extracted features with known threshold values (Krause and Ostermann, 2020; Krause et al., 2015). The technique was further augmented by Tsiapoki et al. by combining it with vibration-based SHM, which enabled slowly propagating damage to be detected instead of just the acoustic impulses associated with the cracks (Tsiapoki et al., 2016). In a study previously conducted by Solimine et al. shear web disbonding was detected using a cavity internal microphone by identifying anomalies employing a k-means clustering algorithm and a linear predictor cepstral coefficient based feature set (Solimine et al., 2020).

Though acoustic blade monitoring has recently attracted attention of the research community, the need for further development and in-depth understanding of a robust blade damage monitoring approach still exists. To this point, this paper outlines a new approach that utilizes experimentally collected acoustic sequences in conjunction with an unsupervised data-driven algorithm, enabling the detection of blade structural damage under fatigue loading. Preliminary observations of WTB cavity-internal acoustic measurements suggest that (i) the cavity-internal acoustic signature follows a distinct pattern associated with the load cycles on the blade, and (ii) the onset and progression of damage on the WTB leads to intermittent, temporally-dependent interruptions to this pattern (i.e. damage modes are only acoustically “activated” for a portion of the blade’s load cycle). Thus, this research focuses on the development of a novel acoustics-based approach, which seeks to detect WTB damage through the analysis of cavity-internal acoustic sequences. This contrasts with the previously discussed approaches, which focused on event detection or static excitation conditions, such as continuous airflow (with no alterations in airflow conditions). The analysis of acoustic sequences enables fatigue-related damage modes and load-cycle activated damage signals to be detected.

This research presents a novel extension of the acoustics based methodology that will leverage temporal patterns in the structurally induced acoustic excitation of the WTB cavity for damage detection. In addition, the technique will utilize an unsupervised, data-driven approach since labeled or damage-case data is typically scarce in SHM applications.

The technique proposed utilizes an LSTM autoencoder-based algorithm for detecting anomalies in the cavity-internal acoustics of the blade. Specifically, the normal acoustic patterns within the blade cavity are “learned” by the autoencoder during the encoding/decoding process, and any deviations from this normal pattern will result in an increase in the reconstruction error associated with the autoencoder’s output. Thus, various defects and cracks should theoretically be detectable using this method if they alter the cavity internal acoustics of the blade—the autoencoder will capture the shift in the blade-internal acoustic sequence since it has been tuned to the blade in its healthy state. This makes the technique particularly robust for capturing previously unobserved damage modes (no need to know how the damage modes will alter the acoustics beforehand, just that they will in some way).

The proposed approach is a significant step toward real-time field implementation of the passive acoustic monitoring approach and addresses the following challenging factors: (i) the passive acoustic excitation of the structure will be uncontrolled and therefore unknown, (ii) the acoustic activation of certain damage modes in response to the uncontrolled structurally induced activation will be unknown, and (iii) the acoustic activation of damage will be time-dependent due to the fluctuation of the loads on the blade. The data-driven technique will be used to learn the cavity-internal acoustic signature associated with the normal condition of the blade and detect anomalies in the context of this normal acoustic signature. Due to the uncontrolled nature of the system (which is similar to the conditions that will be expected in an operational environment), all methods will be largely unsupervised and make limited assumptions about the acoustic properties of the system.

To summarize, the approach proposed in this paper (i) leverages, for the first time, structurally induced passive acoustic excitation for the detection of fatigue-related damage modes in the blade, (ii) implements a novel sequence-based anomaly detection and characterization algorithm for the detection of fatigue-related damage modes via the analysis of non-stationary passive acoustic excitation, (iii) quantitatively demonstrates the acoustic sensing approach can detect damage earlier than a strain measurement system with strain gages in the vicinity of progressive blade skin damage for the first time. The non-stationary acoustic excitation refers to sound sources within the blade cavity that fluctuate in magnitude or frequency content over the course of a single measurement.

The rest of the paper is organized as follows. In Section 2, details of the experiments facilitating the collection of acoustic and strain data are presented. In Section 3, the generalized data processing and machine learning approaches along with the features extracted and used are reported. Section 4 includes demonstrated results and their interpretation along with detailed discussions on the significance of the findings. The last section summarizes the impactful outcomes of this investigation and proposes an outlook for future studies.

Experimental investigation

In this study, the passive acoustics-based SHM technique was implemented on a 62 m long WTB undergoing edgewise fatigue testing at the Wind Technology Testing Center (WTTC) in Charleston, MA. During fatigue testing, the WTB was anchored to a concrete test stand and forced into side-to-side oscillations (load cycles) via a hydraulic actuator operating at a frequency of approximately 0.575 Hz. This test was designed to simulate the gravitational loads experienced by the WTB during operation. The periodic displacement of the WTB led to the generation of a distinct acoustic sequence within its cavity and served to provide the passive acoustic excitation needed for damage detection.

During the fatigue tests, acoustic data was collected using custom microphone nodes placed in five different locations throughout the three WTB cavities (leading edge cavity, center cavity, and trailing edge cavity), as outlined in Table 1. These nodes were programed to collect 30 seconds of acoustic data every 10 minutes at a sampling rate of 44.1 kHz. The technique developed predominantly operates within the theoretical frequency range of 0 Hz–22 kHz due to the hardware selected. This frequency range was observed to interact well with the internal geometry of the specific blade geometries dealt with, increasing the success of the technique. However, with a simple hardware modification, our approach can easily capture even higher, ultrasonic ranges. Very few, low-cost, robust acoustic microphones capture audio waves propagating through air, capturing any acoustic aberrations successfully. This intermittent data collection schedule aimed to reduce system power consumption and data volume in order to make the technique feasible for a wireless operational application.

Table 1.

The microphone node placement details, including name, cavity location, and distance from the blade root.

Name	Location	Distance from Root [m]
TE	Trailing Edge Cavity	9.1
CC₁	Center Cavity	4.6
CC₂	Center Cavity	13.7
LE₁	Leading Edge Cavity	4.6
LE₂	Leading Edge Cavity	13.7

In addition to the acoustic data, strain data was also collected on the outer surface of the WTB. Strain measurements are commonly used in fatigue testing environments to monitor WTBs for structural deformations that may be indicative of damage. In total, 124 strain gages (Vishay Precision Group Micro-Measurements WK-06-250BG-350) were bonded to the outer shell of the blade, as depicted in Figure 2 (only a limited number of strain gage positions nearby the damage was depicted for clarity purposes). In this study, this strain data served to provide a ground truth estimation of the onset and progression of structural damage so that the performance of the passive-acoustics based SHM technique could be evaluated. Only the data from the strain gages in the vicinity of reported damage events was included in the analysis, as detailed in Table 2. All strain data was collected at a sampling rate of 100 Hz.

Figure 2.

The microphone and strain gage locations are highlighted in addition to the crack-type damage propagation on the WTB shell.

Table 2.

Strain gage details, including name, cavity location/side, and distance from the blade root.

Name	Location	Side	Distance from Root [m]
8_TE_HP	Trailing Edge	High Pressure	8
9_TE_HP	Trailing Edge	High Pressure	9
10_TE_HP	Trailing Edge	High Pressure	10
11_TE_HP	Trailing Edge	High Pressure	11
13_TE_HP	Trailing Edge	High Pressure	13
15_TE_HP	Trailing Edge	High Pressure	15

The test campaign was initiated on the WTB approximately 579 k load cycles into fatigue testing while the blade was still in its healthy state. Both strain and acoustic data were collected until 1278 k load cycles, at which point the WTB failed due to the onset and progression of fatigue-related structural damage. The damage was a crack in the WTB’s outer shell, which developed on the high pressure side of the trailing edge, approximately 8 m down the length of the WTB as measured from the root. The WTB cavity internal and external damage progression is outlined in Figure 2, along with the microphone node and strain gage locations. Different sections of Figure 2 have been labeled to reflect the damage progression according to the number of load cycles endured by the WTB.

Ultimately, this configuration enabled the cavity-internal acoustics to be measured from the three WTB cavities (delimited by the blade shear webs) while the strain data was collected from the WTB surface. The damage event (cracking) on the WTB was a naturally-occurring, fatigue-related structural failure resulting from load fluctuations on the WTB.

Methodology

The following section details the methods employed for the passive acoustics-based SHM of a WTB. First, the overall technique is described in detail. Then, the signal processing steps are presented along with the feature extraction and selection procedure. Next, the use of self-supervised learning for establishing the normal condition of the cavity-internal acoustics is outlined. Finally, the approach for anomaly detection and characterization is described.

Passive acoustics-based SHM via sequence analysis

As discussed, this research details the development of a novel acoustics-based approach which seeks to detect damage on WTBs through the analysis of cavity-internal acoustic sequences. The basis of this methodology lies in that the acoustic presentation of some damage modes is tied to the load cycles experienced by the WTB. Namely, the audible damage signal may only be “activated” during certain phases of the WTB’s loading cycle and can be detected by analyzing changes in cavity-internal acoustic patterns. Previous studies have confirmed the activation of damage-related noise due to the motion of the WTB during fatigue loading. These studies have found that damage modes like adhesive joint failures and delaminations can result in structurally-induced sound such as snapping, rubbing, or creaking noises (Chen et al., 2021; Solimine et al., 2020). The spectrograms provided in Figure 3 demonstrate how the cavity-internal acoustics changed in the trailing edge cavity due to the onset and progression of damage during the test campaign. It can be seen that the damage signal is “activated” with each load cycle via structurally-induced passive acoustic excitation, and that the anomalous frequency content becomes more prevalent as the damage becomes progressively more severe.

Figure 3.

Spectrograms depicting 5 seconds of acoustic data collected from the trailing edge cavity between the beginning (January 20, 579 k load cycles) and end (March 23, 1278 k load cycles) of testing: (a) healthy condition, (b) continued healthy condition, (c) initial states of damage, (d) damage progression, (e) more damage progression, and (f) near-failure.

The spectrograms verify the distinct pattern associated with each cavity-internal acoustic profile class in the time-frequency domain. In a normal oscillation cycle (which lasts approximately 2 seconds), two increases in the power of the 1–5 kHz band can be observed corresponding to the times at which the blade reaches its maximum displacement as it moves side to side. With the introduction of the “crack breathing” noise in Figure 3(c), larger spikes in power across the audible frequency range (0–20 kHz) can be seen corresponding to the maximum displacement of the blade. This indicates that the acoustic activation of the suspected damage mode is due to either the maximum tension or maximum compression of the blade.

The distinct noise content related to the progression of damage on the structure demonstrates the need for a robust, data-driven damage detection algorithm. Different damage modes, or even the same damage mode throughout its lifetime, will alter the acoustic signature of the blade cavity in different ways. The technique can be considered robust against noise on two fronts: first, the autoencoder is trained using data from a noisy environment. This ensures that any consistent sources of noise are accounted for in the model and should not affect the overall reconstruction error that is used for damage detection. Second, the autocorrelation-based secondary feature set is designed to focus on anomalous acoustic patterns that are repeated with each rotation of the blade. Random or inconsistent noise would not exhibit a repetitive pattern and would thus not affect the extracted features. In addition to detecting the onset and progression of damage, it will also be useful to differentiate between acoustic patterns in the blade, which can aid in the characterization of damage type or progression in the future.

Signal processing and feature extraction

Framing and windowing

The structurally-induced passive acoustic excitation of the WTB produces non-stationary acoustic signals that possess time-evolving statistical properties. In order to properly characterize the dynamic components of sound within the blade cavity, framing was used to divide the acoustic signal in segments of 50 ms in length with an overlap of 50%. The frame length was selected following an analysis of the acoustic signals produced within the WTB. Each frame was windowed using a Hamming window to limit the spectral leakage caused by the abrupt, cut-off signal ends created during the framing process. Once the acoustic signal was framed and windowed, it could be considered quasi-stationary within each frame, thus allowing the dynamic properties of the signal to be properly characterized without being diluted.

Feature extraction: Linear frequency cepstral coefficients

The aim of feature extraction is to summarize the salient signal characteristics in each frame to optimize the input for machine learning. In this study, linear frequency cepstral coefficients (LFCC) were extracted largely due to their previous success in characterizing WTB damage (Solimine and Inalpolat, 2022). LFCC’s are derived from cepstral coefficients, which are calculated from the inverse Fourier transform of the logarithm of a spectrum (equation (1) (Randall, 2011)),

C (τ) = F^{- 1} [\log (X (f)]

(1)

where $X (f) = F [x (t)]$ , $F$ denotes the Discrete Fourier Transform (DFT), and x(t) is the windowed signal. Cepstral coefficients are often referred to as the “spectrum of a spectrum” and contain information on the rate of change of a signal’s spectral bands (Randall, 2017). LFCC’s are a class of cepstral coefficients that are calculated using linearly spaced triangular band-pass filters. Here, the magnitude spectrum of the X(f), | $F$ [x(t)] |, is calculated and passed through a filter bank, which varies depending on the type of cepstral coefficient being calculated. In the case of LFCCs, this filter bank was composed of linearly spaced triangular band-pass filters. The natural log of the filter bank output is then taken, and the discrete cosine transform or inverse Fourier Transform is taken to obtain the de-correlated cepstral coefficients. LFCCs were selected based on their performance compared to common spectral-based features, Mel-frequency cepstral coefficients, and gammatone cepstral coefficients in a previous related WTB study and various studies involving high-frequency acoustic sources of interest (Lei and Lopez-Gonzalo, ; Zhou et al., 2011; Jung et al., 2017). In this research, the cepstral coefficients are expected to be effective at characterizing the acoustics within the highly reverberant cavity-internal environment. Overall, 80 filters were used in the filter bank and a total of 39 cepstral coefficients were extracted based on the results of a previous investigation. These were extracted from the signals associated with each individual microphone.

Ultimately, the feature calculation resulted in each 30-second acoustic measurement being summarized by a feature matrix with m rows and n columns, where m is equivalent to the number of signal frames and n is equivalent to the number of LFCCs extracted for each frame. In other words, each acoustic measurement was summarized by m length sequences of n LFCCs, as seen in equation (2).

F_{matrix} = [\begin{matrix} F_{11}, F_{12}, F_{13}, \dots, F_{1 n} \\ F_{21}, F_{22}, F_{23}, \dots, F_{2 n} \\ \begin{matrix} F_{31}, F_{32}, F_{33}, \dots, F_{3 n} \\ ⋮ \\ F_{m 1}, F_{m 2}, F_{m 3}, \dots, F_{mn} \end{matrix} \end{matrix}]

(2)

This was accomplished using the measurements associated with each individual microphone. These feature matrices were used as input into the machine learning model. They were divided into Training, Validation, and Test sets for the model development, tuning, and evaluation, respectively. The time frames for which these samples were selected are outlined in Table 3. It should be noted that all training and validation samples were extracted from the WTB in its presumably healthy state (priori), which is a feasible assumption in an SHM application. Thus, the machine learning model is only trained and tuned using healthy case data. This is significant since damage is expected to take many forms and the sound produced by damage is impacted by many factors. By using a completely unsupervised, data-driven approach, damage can be detected without ever having to be observed during training.

Table 3.

Training, validation, and test set details.

Data Set Name	Start (cycles)	End (cycles)	# Samples
Train	579,097	700,116	650
Validation	Random Healthy	Random Healthy	100
Test	700,116	1,227,996	950

Self-supervised learning for damage signal enhancement

In this study, self-supervised learning is used as a way to detect acoustic anomalies with respect to the established normal WTB cavity-internal acoustic signature. These types of machine learning models are ideal since they allow anomalies to be identified without the use of labeled training data, which is notoriously scarce in SHM applications. Self-supervised learning is a variety of unsupervised learning where the model outputs are identical to the model inputs. Although these models are trained to map the inputs to the outputs in a supervised fashion, they are considered “unsupervised” because no data labels are required. A common machine learning model that employs this style of learning is a type of neural network known as an autoencoder. Autoencoders are tasked with first mapping, or encoding, the input data into a reduced dimensionality representation, and then attempt to accurately reconstruct, or decode, the input data from the compressed format (Goodfelllow et al., 2016). In doing so, the autoencoder learns the optimal reduced dimensionality representation of the input data (the most salient information) such that it can be reconstructed with minimal error. The typical architecture of an autoencoder is demonstrated in Figure 4.

Figure 4.

This figure represents the typical architecture for an autoencoder. The input layer composed of i input features, x_i, is mapped to a hidden layer, z_i, of reduced dimensionality. The data is then reconstructed, x_i’, in the output layer.

In damage detection tasks, the autoencoder can be trained using only data collected from the structure in its healthy state. Therefore, it will only be conditioned to encode and decode healthy-case data. When damage case or anomalous data is passed through the autoencoder, the input data will not be accurately reconstructed since the data possesses characteristics that the autoencoder has not been trained to recognize or store within its latent space. Damage can then be identified by the high reconstruction error associated with the data. This technique has been leveraged for both vibration and acoustic anomaly detection models (Anaissi and Zandavi, 2010; Bayram et al., 2021; Marchi et al., 2015; Oh and Yun, 2018).

In this research, an autoencoder variety known as a long-short-term-memory (LSTM)-autoencoder, was used in lieu of a conventional autoencoder due to its ability to reconstruct sequences of data instead of just discrete data points. The architecture of these models is similar to that of regular autoencoders except the hidden layers are composed of LSTM units, as shown in Figure 5.

Figure 5.

The LSTM-autoencoder architecture with the encoding and decoding layers composed of LSTM units.

The use of LSTM networks, which are a type of recurrent neural network (RNN), in the hidden layers allows the autoencoder to learn patterns in the data over long sequences. LSTM networks are able to learn over long sequences of data through the use of “gates,” which serve to regulate the flow of information between nodes and along temporal sequences. Each LSTM unit, as pictured in Figure 6, possesses three such gates: an input gate, an output gate, and a forget gate. These gates are designed to differentiate between the data sequence’s significant and insignificant information in the context of the desired output, and update the unit’s “cell state” accordingly at each time step. The cell state acts as the memory for the unit and is the mechanism for passing information across long sequences.

Figure 6.

LSTM unit schematic.

The “input gate,”i, controls what information from the input at a particular time step is used to update the cell state, the “forget gate,”f, decides what information to remove from the cell state, and the “output gate,”o, controls what information is passed on to the next time step. Each of these gates uses sigmoid activation functions, σ, which output a value between 0 and 1. Equations (3a)–(3c) depict the gate activation equations for the LSTM input, forget, and output gates.

i_{t} = σ (w_{i} [h_{t - 1}, x_{t}] + b_{i})

(3a)

f_{t} = σ (w_{f} [h_{t - 1}, x_{t}] + b_{f})

(3b)

o_{t} = σ (w_{o} [h_{t - 1}, x_{t}] + b_{o})

(3c)

In these equations, i, f, and o represent the input, forget and output gates, respectively. The sigmoid activation function is given by σ, and w represents the weights for the respective gate neurons. The output from the previous LSTM unit is given by h_t-1, the input from the current time step is given by x_t, and the biases for the respective gates are given by b. The cell state, c, is updated according to equation (4).

c_{t} = {f_{t}}^{*} c_{t - 1} + {i_{t}}^{*} \tilde{c_{t}}

(4)

Where c_t-1 is the previous cell state, f is the forget gate, i is the input gate, and $\tilde{c}$ represents the interim cell state at time step t. From the equation, it can be seen that the cell state “forgets” information based on the previous cell state, and adds information according to the present input. The interim cell state, $\tilde{c}$ , can be found using equation (5).

\tilde{c_{t}} = \tanh (w_{c} [h_{t - 1}, x_{t}] + b_{c})

(5)

The output, h, is then found from equation (6).

h_{t} = {o_{t}}^{*} \tanh (c^{t})

(6)

Using these nodes in the hidden layers allows the autoencoder to encode and reconstruct a sequence input. If the LSTM-autoencoder is trained using only sequences of data available from the WTB in its healthy condition, it will yield high reconstruction errors when anomalous sequences are input into the model. A similar strategy has been successfully used in the detection of solar energy system anomalies (Pereira and Silveira, 2018), electric motor faults (Principi et al., 2019), and acoustic anomalies (Marchi et al., 2015). This result is demonstrated in Figure 7. When a healthy case feature sequence (Figure 7(a)) is passed through the LSTM-autoencoder, the resulting reconstruction error is low since the model has been trained to accurately reconstruct healthy case sequences. Alternatively, the damage case sequence in Figure 7(b) (which is similar in magnitude to the healthy case sequence before input to the autoencoder) yields a reconstruction error that is high in magnitude and displays a cyclic component. This is because the model cannot accurately reconstruct the input signal due to the presence of anomalous damage information which occurs with each load cycle.

Figure 7.

The pre-autoencoder input and reconstruction error associated with an LFCC feature is plotted for the: (a) healthy and (b) damaged states of the WTB.

Autocorrelation analysis

In this study, the analysis of damage modes that are activated by fatigue loading presents a unique challenge to the task of unsupervised anomaly detection. In these cases, only acoustic anomalies that occur repetitively with each load cycle or rotation of the blade are of interest. These anomalies represent those damage modes that are acoustically activated during a specific phase of the structure’s load cycle and are related to the fatigue failure of the blade. In general, distance metrics alone (such as RMSE) are not effective at characterizing these structurally activated, sequence-based anomalies since their timing, magnitude, and duration are all important factors in differentiating a damage-case acoustic sequence from a healthy-case one. Moreover, while the LSTM-autoencoder is effective at enhancing these anomalies, another technique is required to extract information on the patterns they exhibit.

For the detection of anomalies in the context of dynamic acoustic excitation, the autocorrelation was taken of the reconstruction error sequences for each sample. It was then used in the extraction of secondary features aimed at the characterization of the timing, duration, and magnitude of structurally activated acoustic anomalies. The autocorrelation, by definition, is a mathematical representation of the degree of similarity (or correlation) between a given time series and a lagged version of itself over successive time intervals, as calculated in equation (7).

r_{k} = \frac{\sum_{t = k + 1}^{T} (y_{t} - \bar{y}) (y_{t - k} - \bar{y})}{\sum_{t = 1}^{T} {(y_{t} - \bar{y})}^{2})}

(7)

where r_k is the autocorrelation r at lag k, y_t is the original signal, y_t-k is the time lagged version of the signal, and $\bar{y}$ is the mean of the original signal. The portion of the equation in the numerator is essentially the covariance between the original signal and the lagged version of itself. The denominator is the sum of the squared deviations (or difference between the signal and its mean) of the original signal.

The progression of structural damage impacts the timing, magnitude, and duration of the acoustic damage signal within the blade cavity in addition to its frequency content. By monitoring the autocorrelation of the reconstruction error of each feature, these damage-related changes can be parameterized via the calculation of secondary features and leveraged for anomaly detection and characterization.

Here, peaks are defined as any local maxima of the autocorrelation that exceeds a minimum height of two standard deviations over the average of the autocorrelation. Each peak is ordered according to the lag at which it was calculated. The width of each peak is defined as the distance between the points to the left and right of an identified peak or local maxima. The decision to extract three peak heights/widths was based on the observation of the data—while some acoustic anomalies only occur once per oscillation cycle (typically when the blade has reached its peak displacement on either the left or right side), others can occur twice (during the peak displacements on both the left and right sides) or are distributed throughout the cycle. By extracting information from three peaks, it is possible to account for more than one damage related pattern. Based on a Pearson Correlation Coefficient analysis, the extraction of any information beyond three peaks did not yield any additional information. The average peak height was also extracted to summarize the strength of the pattern throughout the signal duration. In essence, the autocorrelation-based features helped to parameterize patterns in the abnormal information of the periodic components of the cavity-internal acoustic signal. The “abnormal information” refers to the use of the LSTM-autoencoder to emphasize the information in the signal that was different from the expected healthy condition. The “periodic components” refer to the extraction of cepstral coefficients from the raw signal. In addition, the calculation of the autocorrelation-based features helped to further reduce the dimensionality of the dataset since they summarized the information captured by the sequences of frames corresponding to each audio signal. Each 1185 X 39 reconstruction error feature matrix was transformed into a 1 × 273 feature vector, since seven secondary features (see Table 4) were calculated from each of the 39 LFCC sequences ( $39 * 7 = 273$ ). For the remainder of this study, this secondary feature set is denoted as LFCC80*.

Table 4.

Secondary autocorrelation-based features.

Feature	Description
Peak₁	Magnitude of the first autocorrelation peak
Peak₂	Magnitude of the second autocorrelation peak
Peak₃	Magnitude of the third autocorrelation peak
PeakAvg	Average magnitude of all autocorrelation peaks
PeakWidth₁	Width of first autocorrelation peak
PeakWidth₂	Width of second autocorrelation peak
PeakWidth₃	Width of third autocorrelation peak

Unsupervised damage detection

Once the secondary autocorrelation-based features were extracted, the task of anomaly detection was reduced to a relatively simple outlier analysis. In this application, the Mahalanobis Squared Distance (MSD) metric was used. The MSD, D², calculates the distance between an N-dimensional sample vector $x = (x_{1}, x_{2}, x_{3}, \dots x_{N})$ and a set of observations with means $μ = (μ_{1}, μ_{2}, μ_{3}, \dots μ_{N})$ , as given in equation (8).

D^{2} = {(x - μ)}^{T} \cdot Σ^{- 1} (x - μ)

(8)

In this equation, $Σ$ ^-1 is the inverse of the distribution’s covariance matrix. The use of the inverse covariance matrix negates the need for data’s dimensions to be equally weighted or independent of each other, making the MSD an attractive choice for multivariate applications. The secondary autocorrelation-based features calculated from the validation set (composed of presumably healthy samples that were randomly selected, without replacement, from the training set) were used to establish a healthy distribution. The distances of all test set data from this healthy distribution were then calculated using the MSD. The MSD summarized the distances of entire multivariate feature vector from the multivariate healthy distribution with a single distance value, thus reducing the 1 × 273 feature vector to a single value. An anomaly threshold was then set at three standard deviations over the mean of the MSD of the validation data. For the threshold, a limit of three standard deviations was chosen since over 99% of all healthy data would be expected to fall within three standard deviations of the mean. The validation data was used to calculate the threshold since it was unknown to the LSTM-autoencoder model but was still composed of audio samples collected from the blade cavity in its healthy/normal condition.

Results and discussion

Strain data

The data collected from the strain gage located closest to the damage location on the blade (8 m from the root on the high-pressure side of the trailing edge) was analyzed for signs of damage onset and progression. The peak positive strain associated with each load cycle was first extracted using a peak-finding algorithm and the MSD of each sample was subsequently calculated in relationship to the healthy distribution. The healthy distribution was established by computing the peak positive strain values from the strain measurements collected between 579,097 and 700,116 load cycles. In Figure 8, the MSD results are plotted against the load cycles of the blade.

Figure 8.

The MSD values of the peak positive strain for each load cycle are plotted. The first indication of strain abnormality, the first official report of strain abnormality, and the failure of the blade due to damage are highlighted (numbered as 1, 2, and 3).

The MSD values indicate that no abnormalities existed in the strain data collected from the 8 m-HP-TE strain gage until approximately 958,000 load cycles. The first sign of abnormality occurred at 958,400 load cycles (at which time the MSD exceeding the threshold established from a healthy validation set) after which the MSD of each sample began to rise steadily. This indicated that the peak positive strain associated with each load cycle on the blade was steadily deviating from the established normal/healthy condition. A sharp spike in the MSD occurred at approximately 1,195,000 load cycles before dropping off at 1,227,000 load cycles, when the blade was stopped due to the extent of the crack-type damage 8 m from the root. Official records from WTTC personnel first report the onset of damage at 964,000 load cycles and confirm the failure of the blade at 1,227,000 load cycles.

Acoustic data

Following the signal processing, feature calculation, and machine learning steps, the MSD values were calculated for each audio sample collected by each microphone. The MSD results calculated from the LFCC80* feature set for each cavity-internal microphone are presented in Figure 9(a) to (e). It should be noted that a lapse exists between 890,000 and 990,000 load cycles due to maintenance on the microphone sensing system.

Figure 9.

LFCC80* MSD results plotted against number of load cycles for the: (a) TE microphone, (b) CC1 microphone, (c) CC2 microphone, (d) LE1 microphone, and (e) LE2 microphone.

The greatest deviation from the healthy condition occurred in the TE microphone (Figure 9(a)), which was positioned in same cavity where the damage occurred. The CC₂ microphone (Figure 9(c)) also demonstrated some sensitivity to the damage event, although the observed deviation was not as great in magnitude as the TE microphone. Limited deviations from the normal condition were observed in the remaining microphones. Overall, this indicates that the TE and CC₂ microphones observed a shift from the patterns associated with the normal acoustic signature of their respective cavities. Although the damage was located on the trailing edge cavity, it is likely that the CC₂ microphone picked up some residual anomalous acoustic energy transmitted through the structure. Figure 10 presents the TE microphone MSD results in greater detail. Here, all samples that exceeded the threshold (established using the healthy-case validation data) were flagged as anomalies.

Figure 10.

MSD results calculated from the LFCC80* feature set for the TE microphone plotted against the number of load cycles. The first indication of acoustic abnormality is highlighted in addition to the first indication of strain abnormality from Figure 4–24 and the failure of the blade due to damage.

The first indication of abnormality in the blade-cavity internal acoustic sequences occurred about 838,300 load cycles into testing. This was approximately 120,000 load cycles before any abnormality was observed in the strain data and over 130,000 loads cycles before the onset of damage was officially reported. The magnitude of the MSD grew as damage progressed on the blade and only a handful of samples were classified as “normal” after the onset of damage. Overall, the results demonstrate the ability of passive acoustics-based technique to successfully detect damage onset and progression. The greatest abnormalities were observed in the TE microphone, which was consistent with the location of the damage event. In addition, the growth of the MSD was correlated with the progression of damage on the blade, indicating that the structurally induced passive acoustic excitation of the blade shifted in response to fatigue-related structural damage. The data-driven, passive acoustics based technique displayed sensitivity to the damage event over 120,000 load cycles before the strain-based system. While the strain gages in the experiment were strategically placed in areas where damage is most likely to occur, it should be noted that the strain based system may have detected damage onset sooner if a gage had been placed in closer proximity to the damage event. However, it is not feasible to install a dense network of strain gages on an operational wind turbine blade to capture every possible damage event. Thus, the acoustics-based technique is advantageous since a minimal number of acoustic sensors are required to effectively survey the blade for damage.

Conclusion

This paper outlines the development and application of a new non-stationary damage detection approach for the data-driven, passive-acoustic excitation based SHM technique for wind turbine blades. Currently no other in-situ damage detection method exists for wind turbine blades that can detect multiple damage modes within the composite material of wind turbine blades. The technique demonstrated in the paper is intended to detect damage early enough in its progression to prevent catastrophic failure. The acoustics-based technique was compared to the strain-based techniques currently employed by wind turbine blade testing facilities (which are considered the state-of-the-art for in-situ monitoring) and was shown to pick up on damage onset/progression sooner. The benefit of such a technique is to detect damage early enough to prevent catastrophic failure of the blade and afford maintenance crews more time to plan and execute blade repairs than the current methods allow. The non-stationary acoustic excitation refers to mechanically induced sound sources within the blade cavity that fluctuate in magnitude or frequency content over the course of a single measurement. A sequence-based approach was implemented on a full-scale (63 m) blade undergoing edgewise fatigue testing. LFCC features were used to characterize the blade cavity-internal acoustics and LSTM-autoencoders were trained to emphasize differences between the normal (healthy-case) and abnormal (damage-case) acoustic sequences within the blade cavity. The autocorrelation of the reconstruction error associated with the LSTM-autoencoder output was taken and used in the calculation of secondary features for the characterization of anomalous acoustic patterns within the blade cavity. Upon its application on field data, the technique was successful at detecting the onset and progression of damage. Furthermore, it detected damage earlier than the strain-based system (the standard condition monitoring system for fatigue testing) by 120,000 load cycles.

Ultimately, this research represents the first SHM technique for WTBs to leverage structurally-induced acoustic sequences for fatigue-related damage detection. The investigation was conducted utilizing data collected from a fatigue testing campaign meant to emulate operational environments where both the excitation and damage signal are unknown/poorly defined. It was proven that certain damage modes are acoustically activated by the shifting loads of the blade and can be detected by analyzing deviations from the cavity-internal acoustic sequences produced by the blade in its healthy condition. Furthermore, anomalous acoustic sequences within the blade could be detected using an entirely unsupervised, data-driven approach and without the need for damage-mode specific features/algorithms or knowledge of the excitation condition.

Footnotes

Acknowledgements

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Department of Energy. The authors are also indebted to Massachusetts Clean Energy Center’s Wind Technology and Testing Center (WTTC) for utilizing the utility-scale blade throughout the testing effort.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This material is based upon work supported by the Department of Energy’s Wind Energy Technologies Office under grant number EE0008968.

ORCID iD

Murat Inalpolat

References

Anaissi

Zandavi

(2010) Multi-object variational autoencoder: An application for smart infrastructure maintenance. ACM Transactions on Intelligent Systems and Technology 9(4).

Arora

Wijnant

de Boer

(2014) Acoustic-based damage detection method. Applied Acoustics 80: 23–27.

Bayram

Duman

Ince

(2021) Real time detection of acoustic anomalies in industrial processes using sequential autoencoders. Expert Systems 38(1).

Beale

Inalpolat

Niezrecki

(2020a) Active acoustic damage detection of structural cavities using internal acoustic excitations. Structural Health Monitoring 19: 48–65.

Beale

Willis

Niezrecki

, et al. (2020b) Passive acoustic damage detection of structural cavities using flow-induced acoustic excitations. Structural Health Monitoring 19: 751–764.

Bouzid

Tian

Cumanan

, et al. (2015) Structural health monitoring of wind turbine blades: Acoustic source localization using wireless sensor networks. Journal of Sensors 2015: 1–11.

Chen

Semenov

McGugan

, et al. (2021) Fatigue testing of a 14.3m composite blade embedded with artificial defects–Damage growth and structural health monitoring. .

Fazenda

(2011) Acoustic based condition monitoring of turbine blades. In: 18th International Congress on Sound and VibrationRio de Janeiro, Brazil.

García Márquez

Bernalte Sánchez

Segovia Ramírez

(2022) Acoustic inspection system with unmanned aerial vehicles for wind turbines structure health monitoring. Structural Health Monitoring 21: 485–500.

10.

Goodfelllow

Bengio

Courville

(2016) Deep Learning. Cambridge: MIT Press.

11.

Griffin

(2019) R&D Pathways for Supersized Wind Turbine Blades. Berkeley, CA: Lawrence Berkeley National Laboratory.

12.

IRENA (2012) Renewable Energy Technologies: Cost Analysis Series. Abu Dhabi: International Renewable Energy Agency.

13.

IRENA (2019) Renewable Power Generation Costs in 2018. Abu Dhabi: International Renewable Energy Agency.

14.

Jiang

Dong

, et al. (2018) Structural reliability analysis of wind turbines: A review. Energies 10: 2099.

15.

Jung

Kim

Lim

, et al. (2017) Linear scale filterbank for deep neural network based voice activity detection. In: 20th conference of the oriental chapter of the international coordinating committee on speech databases and speech, pp.1–5. Seoul, South Korea: IEEE.

16.

Krause

Ostermann

(2020) Damage detection for wind turbine rotor blades using airborne sound. International Association for Structural Control and Monitoring 27: e2520.

17.

Krause

Preihs

Ostermann

(2015) Acoustic emission damage detection for wind turbine rotor blades using airborne sound. Structural Health Monitoring.

18.

Lei

Lopez-Gonzalo

Mel, linear, and anti-mel frequency cepstral coefficients in broad phonetic regions for telephone speaker recognition. In: 10th international conference of the international speech communication association. Brighton, UK, 2009.

19.

SCM

Song

, et al. (2015) A review of damage detection methods for wind turbine blades. Smart Materials and Structures24: 1–24.

20.

Marchi

Vesperini

Eyben

, et al. (2015) A novel approach for automatic acoustic novelty detection using a denoising autoencoder with bidirectional LSTM neural networks. In: IEEE international conference on acoustics, speech, and signal processing (ICASSP), South Brisbane, QLD, Australia, pp.1996–2000.

21.

Martinez-Luengo

Kolios

Wang

(2016) Structural health monitoring of offshore wind turbines: A review through the statistical pattern recognition paradigm. Renewable and Sustainable Energy Reviews 64: 91–105.

22.

McGugan

Larsen

Sorensen

, et al. (2008) Fundamentals for Remote Condition Monitoring of Offshore Wind Turbines. Roskilde: Risoe National Laboratory for Sustainable Energy.

23.

Mishnaevsky

Branner

Petersen

, et al. (2017) Materials for wind turbine blades: An overview. Materials10(11): 1285.

24.

Yun

(2018) Residual error based anomaly detection using auto-encoder in SMD machine sound. Sensors18: 1–14.

25.

Pearson

Pullin

Szigeti

(2018) Audible acoustics for detecting and locating damage in composite structures. In: 9th European workshop on structural health monitoring. Manchester, UK.

26.

Pereira

Silveira

(2018) Unsupervised anomaly detection in energy time series data using variational recurrent autoencoders with attention. In: 17th IEEE international conference on machine learning and applications, Orlando, United States, 2018, pp.1275–1282.

27.

Principi

Rossetti

Squartini

, et al. (2019) Unsupervised electric motor fault detection by using deep autoencoders. IEEE/CAA Journal of Automatica Sinica 6: 441–451.

28.

Randall

(2011) Vibration Based Condition Monitoring. Hoboken: Wiley.

29.

Randall

(2017) A history of cepstrum analysis and its application to mechanical problems. Mechanical Systems and Signal Processing 97: 3–19.

30.

Regan

Beale

Inalpolat

(2017) Wind turbine blade damage detection using supervised machine learning algorithms. Journal of Vibration and Acoustics139: 1–14.

31.

RISO/DNV (2002) Guidelines for Design of Wind Turbines, 2nd edn. Roskilde: Wind Energy Department, RISO National Laboratory.

32.

Sanchez

PJB

Ramirez

Marquez

FPG

(2021) Wind Turbines Acoustic Inspections performed with UAV and sound frequency domain analysis. In: 7th international conference on control, instrumentation, and automation.

33.

Solimine

Inalpolat

(2022) An unsupervised data-driven approach for wind turbine blade damage detection under passive acoustics-based excitation. Wind Engineering 46: 1311–1330.

34.

Solimine

Niezrecki

Inalpolat

(2020) An experimental investigation into passive acoustic damage detection for structural health monitoring of wind turbine blades. Structural Health Monitoring 19: 1711–1725.

35.

Sorensen

Lading

Sendrup

, et al. (2002) Fundamentals for Remote Structural Health Monitoring of Wind Turbine Blades-a Preproject. Roskilde: Risoe.

36.

Thomsen

(2009) Sandwich materials for wind turbine blades — Present and future. Journal of Sandwich Structures and Materials 11: 7–26.

37.

Traylor

DiPaola

Willis

, et al. (2020) A computational investigation of airfoil aeroacoustics for structural health monitoring of wind turbine blades. Wind Energy 23: 795–809.

38.

Tsiapoki

Krause

Hackell

, et al. (2016) Combining a vibration-based SHM scheme and an airborne sound approach for damage detection on wind turbine rotor blades. In: 8th European workshop on structural health monitoring.

39.

Windmills Tech (2019) Wind Turbine Blades Handbook. Kirt Thomsen and Bladena.

40.

Zhou

Garcia-Romero

Duraiswami

, et al. (2011) Linear versus mel frequency cepstral coefficients for speaker recognition. In: IEEE workshop on automatic speech recognition and understanding, Waikoloa, HI2011, pp.559–564.