An Intelligent Detection Logic for Fan-Blade Damage to Wind Turbines Based on Mounted-Accelerometer Data

Abstract

Many wind turbines operate in harsh marine or shore environments. This study assists industry by establishing a real-time condition-monitoring and fault-detection system, with rules for recognizing a wind turbine’s abnormal operation mainly caused by different types of fan-blade damage. This system can ensure ideal wind turbine operation by monitoring the health status of the blades, detecting sudden anomalies, and performing maintenance almost in real time. This is especially significant for wind farms in areas subject to frequent natural disasters (e.g., earthquakes and typhoons). Turbines might fail to endure these because the manufacturers have built them according to the standards developed for areas less prone to natural disasters. The system’s rules are established by utilising concepts and methods from data analytics, digital signal processing (DSP) and statistics to analyse data from the accelerometer, which measures the vibration signals in three dimensions on the platform of the wind turbine’s base. The patterns for those cases involving fan-blade damage are found to establish the rules. With the anomalies detected and reported effectively, repairs and maintenance can be carried out on the faulty wind turbines. This enables ‘maintenance by prediction’ actions for unplanned maintenance as a supplement to the ‘predictive maintenance’ tasks for regular planned maintenance.

1. Introduction

1.1. Research Background

Improving the operating efficiency of wind turbines and reducing maintenance costs are goals common to all wind farm operators. However, similar to the erosional effects caused by seawater and weather conditions on the base constructs of offshore or onshore wind turbines, machines and equipment are also prone to failure, leading to a loss of efficiency in power generation operations. However, an integrated perspective to examine these issues is seldom discussed [1,2,3].

The industry developing wind turbines has adopted a conservative attitude, i.e., it lacks integration with related upstream and downstream industries. In some countries, for example, blade manufacturers concentrate only on the R&D (research and development) of the blade shape and materials, while the generator manufacturers care only about the design and manufacture of the generator bodies. The power-converter manufacturers alone bear the final responsibility for power generation efficiency. However, converter manufacturers do not usually have the expertise to integrate blades and generators. Moreover, the international market currently lacks protocols or standard operating procedures (SOPs) for these integrative matters. This is critical for turbine maintenance.

Although avoiding the replacement of faulty major parts is the most efficient way to reduce operating costs, an alternative logic for turbine maintenance entails detecting faults efficiently and replacing the necessary part promptly, in addition to scheduled preventive maintenance (e.g., predictive maintenance). Within the entire power generation system utilising wind as a resource, the blade is an important component. To ensure the normal operation of a wind turbine, the key factors are the durability, safety and reliability of the wind turbine blades, and these have become a concern for wind farm operators and manufacturers. However, to the authors’ knowledge, most fan blade health diagnoses today still rely on subjective human efforts to identify whether the blades are damaged or not, i.e., acoustic or vision-based inspections performed by professionals, and currently there is no rapid and objective method for blade diagnosis [1,2,3].

Benbouzid et al. [4] reviewed the current progress in condition monitoring of wind turbines, from traditional condition monitoring and signal-processing tools to machine learning-based condition monitoring and predictive maintenance using big-data mining. Their systematic review of signal-based and data-driven modelling approaches using intelligence and machine learning approaches examined recent developments in the field and their use in diagnosis, prognosis, health assessment, and predictive maintenance of wind turbines and farms [4]. Natilli et al. [5] developed a multi-scale method for wind turbine gearbox fault detection and tested it in real-world test cases. The novelty of this work is that the detection method was developed using industrial datasets provided by standard SCADA (supervisory control and data acquisition) and TCM (turbine condition monitoring) systems [5].

Papi [6] proposed using an uncertainty-quantification method to model the effect of blade damage on the performance of multi-megawatt wind turbines. The proposed method aims to overcome some of the problems associated with evaluating individual test cases. In fact, treating blade damage as a random phenomenon avoids biases due to specific test cases of combinations of blade damage factors and allows for more general conclusions [6]. This point is critical for this study. Santolamazza et al. [7] presented an approach based on machine learning techniques using data from SCADA systems. Since these systems are usually already implemented on most wind turbines, they provide a lot of data without the need for additional sensors. In particular, they used artificial neural networks (ANNs) to develop models to describe the behaviour of some of the main components of wind turbines, such as gearboxes and generators, and predict operational anomalies. The proposed method was tested on real wind turbines in Italy to verify its effectiveness and applicability and proved to be able to provide important assistance in the maintenance of wind farms [7].

Knowing the status of the blades at any time should be critical to the normal operation of turbines at the lowest cost, because wind has become one of the most popular sources of green power generation (solar being another popular energy source), and more and more wind farms are operating or being constructed worldwide. The offshore equipment is usually expensive and produces electricity almost constantly; however, it may fail, incurring significant operational losses if it cannot be repaired quickly during the downtime [1,2,3].

The repair and subsequent maintenance tasks involve considerable effort and costs, thus exacerbating the problem. Moreover, in some places, such as Penghu (or the Pescadores Islands) and Taiwan, most wind turbines are maintained by non-native manufacturers, who regard internal equipment maintenance as a business secret (and do not disclose it easily). However, a discussion about this critical problem is beyond the scope of this paper. In any case, there is a need to detect and diagnose problems related to offshore wind turbines quickly, in order to launch the subsequent unplanned yet necessary repairs [1,2,3].

1.2. This Study: An Overview

In this study, a complete functioning mechanical condition monitoring system (CMS) is established with an initial stable performance, which measures parameters such as the vibrations and noises of a turbine during operation. The recorded data are stored and reported remotely (from the remote-side CMS). They are organised as datasets (on the server-side CMS), and the characteristics and patterns of/in the signal are then analysed using the established rules and classified as having different levels of fan-blade damage severity. This process automatically detects the health status of a turbine by inspecting its vibration states, rather than through human efforts.

Thus, if effective rules can be established to program the CMS (which embeds these rules) to detect the different severity levels of turbine fan-blade damage based on a real-time database that is regularly imported and updated by the (remote) CMS, the performance of the server-side CMS relies heavily on the effectiveness of the rules. Once an anomaly is detected, the CMS serves to guide the subsequent fan maintenance and repair actions. In this study, we call this ‘maintenance by prediction’, in contrast to ‘predictive maintenance’, which is a type of ‘preventive maintenance’ [1,2,3].

The CMS can thus warn of blade damage almost in real time, allowing wind farm operators to launch and arrange repair operations for maintenance by prediction at the earliest moment. This reduces the turbine’s downtime. As such, this is different from predictive maintenance, which aims to avoid wasting time and money by preventing serious damage by performing regular maintenance in advance. Maintenance by prediction operates by giving sudden fan-blade status warnings (in the event of damage) for launching the actions as soon as possible.

An example may help to illustrate the difference. During predictive maintenance, the engineers are recommended to replace some parts of the turbine, but they may be replaced improperly, ultimately resulting in a machine shutdown. By contrast, the proposed maintenance by prediction mechanism works to detect, warn of, and support addressing an incident in almost real time.

Theoretically, the mechanism proposed by this study to detect wind turbine blade damage in wind turbines aims to support unplanned maintenance and repairs, which are carried out for the faulty wind turbine(s) using rule-based predictions (i.e., the maintenance by prediction actions). If the CMS is programmed with this detection mechanism, most of the equipment manager’s efforts can be focused on the machine’s normal operations, e.g., acquiring or preparing the parts required for equipment repairs, i.e., preventive maintenance. In operational practice, such a CMS can improve the safety (effective operation) rate (or operating efficiency) of the power system, reduce maintenance costs and increase the reliability of the wind turbine [1,2,3,8,9,10,11,12].

Section 2 focuses on the design of the CMS for wind turbines, describing how our research led us to the fault-diagnosis functions for fan blades from the traditional functions used for power generators, while reviewing the relevant literature. Section 3 presents the results from the data analysis upon taking real sampling data and introduces the methods used ‘by example’ (on the fly). Although the methods being applied are individually common in data processing, statistics, and digital signal processing (DSP), hybridising them for the purpose of the application described in this study is novel. The rules are established in Section 4 based on the analytical results, followed by a discussion of key findings. Section 5 concludes this study.

2. Condition Monitoring System (CMS) and Methods

This section describes the development of the CMS, which involves remote-side and server-side subsystems. Because the remote-side CMS is mounted on the structure of a wind turbine on the base platform, we start from an overview of the wind turbines.

2.1. Wind Turbines

A wind turbine is mainly composed of control systems, transmission mechanisms, generators, converters, blades, towers, power cables, converters and transformers. The internal transmission mechanisms include gearboxes, hubs and steering systems. The performance of wind turbines varies from one model to another.

As discussed previously, the installation site’s climate conditions and the natural disasters which may occur affect wind turbines’ operation, resulting in various causes of sudden failure, which is to be resolved by maintenance by prediction. In addition to these external factors, some wind turbines may have inherent (internal) drawbacks, such as inconsistent design specifications, low manufacturing quality and inappropriate design specifications for the sites on which they operate. All of these factors could cause them to operate in an unstable manner and accelerate the possible damage to their components, including the blades.

The reasons for fan-blade damage are manifold, but each can be categorised as either internal or external. Internal damage may be caused by kinetic or mechanical reasons. For example, machine wear due to operation over a long period will result in the deformation of parts due to stress. Also, during fan operation on normal days, the tower column will produce different deformations and displacements due to different operating conditions, tower conditions, and conditions of the base. External damage, by contrast, can result from the sudden occurrence of natural disasters, e.g., large typhoons and earthquakes. These are more harmful to the blades, and the consequences are usually unexpected. While the events causing damage are all occasional, they meet the condition that the tool provided by this study can address. See these further in Section 4.3.

2.2. CMS Hardware

The central component of the remote-side CMS hardware uses Adlink’s USB-2405, with which each channel can be connected to an accelerometer or a microphone. The USB-2405 is a 24-bit USB interface for dynamic signal data acquisition. The USB-2405 has a USB-interface power supply and BNC connector; there are four analogue input channels, and each channel provides up to 128 kS/s of simultaneous sampling. The USB-2405 also features selectable AC or DC coupled inputs and a built-in high-precision 2 mA excitation current measurement integrated electronic piezoelectric (IEPE) sensor. Each channel can be connected to an accelerometer or microphone for measuring vibrations or noises. In the measurement system, each channel was connected to a PCB 601A01 accelerometers with a sensitivity of 100 mV/g and a frequency band range of 0.27 to 10,000 Hz.

Relatedly, the CMS can be divided into vibration and noise detection subsystems, and each subsystem has its effective fault detection and diagnostic methods. The integration of technologies inspecting the wind turbine near the accelerometer or microphone enables the operation status monitoring functions and the preliminary remote-side fault-diagnosis functions. However, it is difficult to derive effective rules that can be used for such fault diagnosis on the remote side. This is discussed in the next subsection.

2.3. Fault-Diagnosis Functions for Wind Turbines

This subsection illustrates the fault-diagnosis process and the analysis functions of the CMS for a wind turbine. When the machine experiences abnormal vibration, the diagnostic analysis can be performed according to the following guidelines [13,14,15,16,17,18,19]:

(A) Understand the machine’s construction. For an abnormally functioning machine, it is necessary to understand its primary structure, component transmission methods and even an overview of the system to determine the possible failure modes and the appropriate measurement method and position.

(B) Observe abnormalities. When an exception occurs in a machine, it is necessary to understand the machine’s operational conditions at that moment and to observe the abnormal phenomenon that is causing the device to operate abnormally.

(C) Abnormal phenomenon measurement. It is necessary to select appropriate sensors and instrumentation equipment, measure the vibration signal of the unusual machine phenomenon, and process the measured signal through spectrum analysis.

(D) Abnormal signal analysis. The abnormal signals measured should be analysed to determine the most severe locations and conditions and the components that are causing the problems. The reasons for these problems should be investigated. If necessary, further measurements should be made to confirm the causes of the issues.

(E) Verification of the results. After confirming the cause (or causes) of the abnormality, and after repairing or replacing the machine’s faulty components, it is necessary to observe or measure again to determine whether the abnormality still exists and whether there are any signs of improvement. If the situation has not improved, the above steps must be repeated until the problem is resolved.

The following list summarises the judging criteria for abnormal situations that have traditionally been applied for fault diagnosis on the remote side (or ‘device side’, ‘client side’, or ‘site side’) CMS for the power generator component(s). Usually, a subset of these criteria is sufficient to program the diagnostic function. Appendix A [1,2,3,17,18] presents detailed descriptions.

• Vibrations generated by the generator caused by an uneven magnetic force acting on the rotor or the stator.
• Vibrations generated by the generator or vibrations caused by an abnormal rotor or rotor coils.
• Vibrations generated by the generator or vibrations caused by an abnormal stator or stator coils [1,2,3,20].
• Unbalanced weight of the shaft of the driveshaft [1,2,3].
• Bent shaft of the driveshaft.
• Improper installation of the shaft of the driveshaft [1,2,3].
• Misalignment of the driveshaft with the coupling connecting the machine during installation, causing the machine’s severe vibration [1,2,3,13,14,15].
• Shaft diameter too small during bearing installation, resulting in the bearing’s movement relative to the inner ring of the bearing [1,2,3].
• Poor shoulder angle during bearing installation, resulting in the centre of the inner and outer rings no longer being on the same straight line [1,2,3].
• Ball bearing causing temperature rise and vibration during mounting of the bearing due to insufficient lubrication or improper lubrication viscosity [1,2,3,13,14,15].
• Gearbox [1,2,3,13,14,15,20,21].
• Bearing damage [22,23].

While the faults listed above are mainly mechanical problems that can be judged using vibration signals (with the means of judging these detailed in Appendix A), the acoustic interface is another main interface used to understand whether a wind turbine is faulty, specifically through noise measurements. Appendix B provides further details about this process [1,2,3,17,18]. Theoretically, then, the measurement results from either interface, or both, can be considered [1,2,3,13,14,15,16,17,18,19,21] on the remote side.

However, the following statements are critical for explaining the reasons this research is conducted recording only the vibration data (and not the noise data) on the ‘remote side’ of the CMS and utilising the collected datasets on the ‘server side’.

• Using the datasets recorded from an interface is sufficient. According to the above review summary, vibration and noise are the two interfaces that can be used to understand a wind turbine’s physical deterioration. However, it has been proven that most of the noises generated by a turbine are caused by vibrations.
• The criteria referenced above do not support judging the faulty conditions of the blades as addressed by this study. The experimental results from our extensive laboratory tests performed prior to this study showed that the criteria f above or judging the faulty conditions of power generators were ineffective for those of blades (i.e., they usually resulted in more incorrect judgments than correct ones) on the remote side. We referred to the usual criteria for power generators, and thus they might not be suitable for other components of turbines (even though a turbine always includes at least one generator) (see Section 2.4).

From the above discussion, as using the vibration datasets is sufficient (and using these is better than using the noise datasets due to the causal relationship), and as the traditional criteria for diagnosing a power generator do not apply in diagnosing the component of turbine blades on the remote side, we decided to seek clues to establish the judgement rules from the vibration datasets received and gathered on the server side.

2.4. Developing Wind Turbines: A Briefing

To demonstrate the role of wind turbine blades in designing and maintaining a wind turbine, as a supplemental review, Figure 1 shows the life cycle of wind turbine development.

After a wind turbine is set up, if it becomes inoperable, most wind turbine manufacturers will assert that it was because the turbine was not properly maintained. Then, to argue with the manufacturer, the engineers in a green energy operator company would like to design a simulation program based on a given turbine manufacturer’s model and run it to determine if the design was flawed. However, it is questionable whether or not the operator can successfully request the design drawings and related data for subsequent operations, maintenance, monitoring and analysis from the manufacturer in practice, let alone build up a simulated turbine. Wind turbine manufacturers believe that wind turbine design drawings and related technologies are trade secrets, so technology transfers are impossible for them; a chance of negotiation exists only if the business deal involves a large-scale wind farm. In reality, however, most cases do not contain this possibility, even in a large-scale wind farm construction project (to the authors’ knowledge).

Therefore, based on the operation and maintenance phase in Figure 1, we must rely on the condition monitoring approach using a CMS (both remote side and server side) and perform the repair or replacement tasks (i.e., the unplanned maintenance outside of the regular, periodic maintenance tasks) when this system detects a sudden faulty status of the turbine while it is operating.

Blade design is the most commonly addressed topic in the design phase for wind turbines because it is usually used to distinguish the brand and type of the turbine directly (visually) and because the technologies for the design of all other components (generator, tower, convertor and controller) were mature long before wind turbines appeared. Thus, this study focuses on the detection of blades’ sudden faulty status. Before this, it is necessary to seek clues for establishing the effective rules for the detection process based on the vibration data transmitted to the server side in near real time from the remote side.

An effective CMS providing accurate predictive detections and posing precise warnings for subsequent maintenance by prediction actions cannot proceed without these rules. Strictly speaking, the methods to establish these rules also fall within the scope of the failure and cause-of-failure analyses phase in Figure 1, in addition to the traditional tasks usually defined based upon the testing failures. The next section presents the results obtained from the data analysis, which forms the basis for establishing the rules.

3. Results

In this section, we summarise the process for establishing effective rules and present the key results obtained from the data analysis (omitting the results from the extensive trial analysis). The process follows the methods introduced in Section 4.3 and Section 4.4, while the flows and methods for the experimental conditions, data collection, data preprocessing and data curation processes are detailed in Section 4.1 and Section 4.2.

3.1. Source Datasets and Visualisation

After the collected datasets were preprocessed and curated, the first phase involved visualising the source time-domain datasets, so that observations could be made in order to perform some initial comparisons. For this purpose, the datasets of the vibration data updated on the server-side CMS were sampled and rendered for 3.0-blade (full blades), 2.5-blade, and 2.0-blade wind turbine settings. In Appendix C, a full coverage for these is shown in Figure A1, Figure A2 and Figure A3, respectively. For space reasons, only several subfigures are preserved here in Figure 2, Figure 3 and Figure 4 as example cases, which are plotted for the data being recorded along the X, Y and Z axes subject to the ‘no wind imposed’ setting.

3.2. Using the Local Regression Method (LRM) and the Smooth-Line Approach

Next, based on the first 1024 data entries in each dataset (data sequence), we found the smooth line to represent the trend of the data points using the local regression method (LRM). In Appendix D, we plotted it as a bold blue line in Figure A4, Figure A5 and Figure A6 for different wind-speeds subject to the wind turbine settings of 3.0-blade (full blades), 2.5-blade and 2.0-blade configurations, respectively. For space reason, only several subfigures used for subsequent discussions are preserved here in Figure 5, Figure 6 and Figure 7.

In these visualisations, we used a small grey circle, instead of points, to denote each original data point, so that the intensity of the data points could be observed in terms of their overlaps. Through the presentation of these points, we might easily find that the cycle of a waveform roughly concurs with the interval between the ‘peaks’ or the interval between the ‘valleys’ of the original acceleration data. As seen in most of these figures (except for the acceleration signals recorded along the X-axis and the Y-axis for the 2.5-blade turbine with no wind, i.e., Figure 6a,b), a waveform containing several repeated waves with a highly static time cycle appeared.

However, despite some 2.5-blade cases being distinguished easily, the 3.0-blade cases and the 2.0-blade cases are not dissimilar visually. Therefore, no rule can be established to differentiate these three cases through the eyes till now, let alone this vision-based process lacks a mathematical foundation. Here, only the observed static time cycle is worthwhile, i.e., in each subfigure, each pair of two nearby peaks, or each pair of two nearby valleys, was almost equally spaced, so the interval for nearby peaks or that for nearby valleys was consistent as well.

However, from these graphs we also saw that (1) the number of data points outside the waveform, (2) the degree to which they were outside the waveform, (3) the concentration of these data points and (4) the proportion of data outliers were all different (i.e., they varied from case to case). For example, in general, on the same axis, there were more acceleration data outliers in Figure A5 for the 2.5-blade setting than in Figure A4 for the 3.0-blade setting and in Figure A6 for the 2.0-blade setting.

Take the settings with 0.6 m/s wind-speed arbitrarily as an example. In Figure A5a (upper-right subfigure), there are 14 outliers (out of the first 1024 data entries), according to the normal outlier equations below:

$$\{\begin{array}{c}Outliers(VD)=U(VD)\backslash \left\{vd\in VD|vd=\left[Q_1-1.5\times IQR,Q_3+1.5\times IQR\right]\right\}\\ IQR=Q_3-Q_1\end{array}$$

At the same time, in the upper-right subfigure in Figure A4a (no blade broken), imposing the same wind-speed, there is no (0) outlier. And surprisingly, in the upper-right subfigure in Figure A6a (a full blade broken) with the same wind-speed being imposed, there is no (0) outlier either.

In addition, in Appendix D in general, on the X axis and the Y axis, apart from the acceleration magnitude ranges, it seems that accelerations with the 2.0-blade setting (e.g., Figure 7a,b) were more concentrated than those with the 3.0-blade setting (e.g., Figure 5a,b). Combined with the previous observations, these provided further clues for obtaining the effective information for constructing the judgement rules in the next subsection.

3.3. Further Transformation: Resampling and Resmoothing

In this part of the analysis, we found that the initial 1024 data points in each data sample (i.e., the ‘data digest’) were sufficient to establish the final rules to identify the different (faulty) situations of turbine blade malfunction after further data transformations using resampling and resmoothing. After conducting experiments, we also found that resampling eight consecutive data points as a representative one (i.e., eight original sampling time units as a ‘clock’) was appropriate. Therefore, three processes were run as follows.

First, each shortened dataset was resampled using the new clock. We let the original dataset be a(t), where t = $t_j,j=0,1,...,1023$ was the original sampling time sequence, and then defined the resampled dataset as $\overline{a}(\overline{t})$, where $\overline{t}$ was the redefined clock sequence and ${\overline{t}}_i=\{t_{8i},t_{8i+1},t_{8i+2},\cdots ,t_{8i+7}\},i=0,1,...,128$. For each clock, the information from the original data was preserved as follows:

$$m({\overline{t}}_i)={\displaystyle \sum _{j=0}^7\frac{a(t_{8i+j})}{8}},$$

$$v({\overline{t}}_i)=\frac{1}{8}{\displaystyle \sum _{j=0}^7{(a(t_{8i+j})-m({\overline{t}}_i))}^2},$$

where $m({\overline{t}}_i)$ and $v({\overline{t}}_i)$ are, respectively, the mean of the original data sequence on the redefined clock sequence $\overline{t}$.

Next, the curve-like piecewise line was approached using LRM, and it was in fact a predictor that also produced the ‘theoretical value’ of the acceleration degree at any specified time point t, i.e., P(t). Therefore, when the data was re-considered using a ‘clock’, this clock was also applied to the predictor function. We named this new predictor a function called $\overline{P}(\overline{t})$, where $\overline{t}$ was the redefined clock sequence and ${\overline{t}}_i=\{t_{8i},t_{8i+1},t_{8i+2},\cdots ,t_{8i+7}\}$. Therefore, $\overline{P}(\overline{t})$ could be simply redefined as:

$$\overline{P}(\overline{t})=\overline{P}(m({\overline{t}}_i))$$

However, in the above equation, the function $\overline{P}(·)$ was identical to $P(·)$ because both used local regression as the smoother function.

Third, the computational results are rendered in Figure A7, Figure A8 and Figure A9 for the 3.0-blade, 2.5-blade and 2.0 blade settings, respectively. Once again, for these settings, only the no-wind-imposed cases on the 3 axes are presented in Figure 8, Figure 9 and Figure 10 for simplicity. These simplified plots were qualified to clarify the features among the different faulty blade cases and establish the rules to distinguish them automatically.

As can be seen, these plots provided clear ways to compare and establish the rules to identify the two malfunctioning blade cases against that with full blades working normally.

4. Establishment of Rules and Discussion

Based on these results, the rules to judge whether a blade on the wind turbine was half-broken (i.e., the 2.5-blade case), normal with full-blades running (i.e., the 3.0-blade case), or completely missing a blade (i.e., the 2.0-blade case), could be established.

4.1. First Rule to Judge the 2.5-Blade Case

First, certain features of the 2.5-blade cases emerged as salient. From the graphs in Figure A8, e.g., Figure 9b, it was easily observed that many of the 2.5-blade cases had distorted or abnormal waveshapes after the data points (t, a) were smoothly interpolated using LRM, compared to the 3.0 or 2.0 cases in Figure 8b and Figure 10b. An extreme case of this could be seen in the plot for the 2.5-blade, Y axis, no-wind case, while other trivial cases were the 2.5-blade, X axis, no wind; 2.5-blade, X axis, wind-speed = 12 m/s; 2.5-blade, X axis, wind-speed = 18 m/s; 2.5-blade, Y axis, wind-speed = 12 m/s and 2.5-blade, Y axis, wind-speed = 18 m/s cases, as seen in Figure A8. As seen in the figures, almost all of these cases appeared based on the data series recorded on the X and Y axes, despite the level of the peaks/valleys also being slightly jittered based on the data recorded on the Z axis subject to certain wind-speeds, e.g., ‘2.5-blade, Z axis, no wind’ and ‘2.5-blade, Z axis, wind-speed = 18 m/s’.

However, despite its simplicity, a rule based on the visualisation process was difficult to implement because it was established through human-based pattern recognition. For example, to what extent could the so-called ‘distortion’ and ‘abnormality’ be justified for a wave-like plot? Therefore, a numerical rule needed to be established so that an algorithm could be implemented based on the results for automatic detection in the future. This relied on the true mean and variance values of the acceleration source data that corresponded to a peak or valley in a resampled and smoothed waveform, which were displayed as bold vertical red line segments in the plot. Each such line ranged from (mean − variance) to (mean + variance) of a source data slice associated with some peak or valley.

As can be observed, in general, the red line segments were longer in the figures plotted based on the X-axis and Y-axis vibration data subject to the 2.5-blade setting, no matter how great the wind-speed imposed on the turbine, compared to those plotted based on the X-axis and Y-axis vibration data subject to either the 2.0-blade setting or the 3.0-blade setting. Moreover, no such situation was found for the figures plotted based on the Z-axis vibration data subject to the same 2.5-blade setting.

As such, this feature (Rule 1), i.e., occurrences of the long red line segments around the peaks and valleys of the LRM-smoothed and resampled waveform that appear for the X-axis and Y-axis vibration data under different wind-speed settings, can be used to identify whether half (0.5) a blade on the wind turbine, or a part of a blade, is broken (i.e., the 2.5-blade case).

4.2. Second Rule to Judge the 2.0-Blade Case

Next, since the 2.5-blade case could be excluded using the above rule, the remaining problem involved how to distinguish the 2.0-blade (one blade totally missing) case from the 3.0-blade case. A clue to the reasoning was that a turbine having full blades (the 3.0-blade setting) should be heavier than the same turbine with a blade totally missing (the 2.0-blade setting) in the case of no wind (static without other conditions changing). That is, unlike the vibration data recorded along the X axis or Y axis, the Z axis data corresponded to the vertical power (i.e., the weight factor of the turbine) of the turbine interacting with the foundation structure and the land. Therefore, lower accelerations should be detected along the Z axis for a turbine with a full blade missing than for a normal (full-bladed) turbine, and this effect could be clearly compared and displayed when no wind was imposed.

This was reflected in the experimental results. Comparing the acceleration data recorded along the Z axis for the situation with no wind (0 m/s) (see Figure 8c and Figure 10c), the interval of the predicted acceleration values for the 3.0-blade normal case was [–1.4,–2.0] (m/s) (bounded by the peaks and valleys of the waveform). In contrast, the interval of those predictions for the 2.0-blade case was [–1.2,–1.8] (m/s), a lower window.

Therefore, the results supported our theoretical suppositions. These became the second rule to distinguish a turbine with a blade totally missing (i.e., the 2.0-blade case) from a normal turbine (Rule 2): if the case is not ‘a part of a blade is broken’ (which can be detected based on the first rule), the predictive waveform identified from the Z axis data can be used to see whether or not a blade is missing for any possible reason by checking to see if waveform fluctuation of the resampled data, in terms of the interval delimited by the peaks and valleys of the smoothed function (i.e., $\{\max \overline{P}(\overline{t}),\min \overline{P}(\overline{t})\}$), is narrower than usual.

4.3. Discussion

In short, the means and variances of the time-domain data recorded along the X axis or the Y axis around the peaks and valleys of the locally regressed and resampled waveform can be used to determine whether the wind turbine already has a partially broken blade (i.e., the 2.5-blade case). The bold vertical red line segments are measured based on these means and variances, and the extraordinary appearance of these line segments may indicate the faulty 2.5-blade case or perhaps harm to a blade.

Following this rule and excluding the 2.5-blade case, the fluctuation range of the predictive waveform obtained based on the data interval recorded along the Z axis when there is no wind can be used as a measure to distinguish the 2.0-blade case from the normal 3.0-blade case. When this interval rises further upward than usual, the case in which a blade is totally missing (the 2.0-blade case) can be detected.

Combing these two rules, all three of the cases—the normal case and the two faulty cases of wind turbine fan-blade damage—can be explained systematically. As these rules are simple (with limited computational complexity), they can be used to detect these malfunctions almost in real time and to transmit the necessary warning messages to those in charge just in time, given that the vibration datasets for the wind turbines are synchronised routinely (within a short period of time) on the server side of the CMS. This greatly benefits the unplanned maintenance efforts of wind farm operations.

The Circum-Pacific Belt area is prone to super typhoons and strong typhoons, whose wind-speeds may easily reach 150 km/h or above, and thus a situation in which a wind turbine blade breaks apart or falls off completely should not be news to anyone in the green energy industry. Although the proposed set of rules does not fit the case in which a turbine totally collapses (as it is unclear whether the accelerometer and remote-side CMS would still work in this case), in most cases it can serve as a computerised remedy to detect whether a blade is broken or falling off, if the wireless transmission works.

Since another common cause of turbine damage in the studied area involves earthquakes, which are usually as unpredictable as typhoons, the maintenance by prediction mechanism is suggested as a supplement to regular predictive maintenance, even during periods of predictable weather conditions (i.e., to best control that which is controllable). In short, from the above discussions for the Circum-Pacific Belt area and articulating back to the outset of this study (see Section 1.1), it should be clear that the proposed mechanism, even just putting a puzzle piece for detecting the damages on the fan component of a turbine (with respect to the whole integrative perspective of turbine maintenance), can improve the operating efficiency of turbines. This reduces the maintenance costs and benefits the unplanned maintenance of wind farm operations.

Note that in terms of digital signal processing (DSP), we established the set of detection rules based on the converted (original) time-domain data, rather than the data in any other domain, e.g., frequency-domain data. Doing so not only maintains simplicity for future implementation and makes the entire process faithful to the original data but also avoids possible ambiguity. For example, if the upper and lower limits of the predictive waveform window determined by the peaks and valleys are logged, it becomes harder to examine whether there has been an upward shift in the interval from the 3.0-blade setting to the 2.0-blade setting (i.e., for the second rule) just because the window in the logged domain would become narrower.

4.4. Extensive Materials

This subsection provides extensive descriptions for how the rules are justified using tabularised quantitative information.

Table 1. Resampled, resmoothed mean–var. transformed data on the X axis (WS = 6 m/s). (a) Min/var. peak/valley Stats for the 3.0-blade turbine. (b) Min/var. peak/valley stats for the 2.5-blade turbine. (c) Min/var. peak/valley stats for the 2.0-blade turbine.

(a)
ID	Mean	Var.	Peak	Valley	Peak Top	Peak Bot.	Valley Top	Valley Bot.	Peak Range	Valley Range
4	−0.23097	0.016014	0	1	-	-	−0.21496	−0.24699	-	0.032028
19	0.914043	0.003893	1	0	0.917935	0.91015	-	-	0.007785	-
32	−0.28885	0.008199	0	1	-	-	−0.28065	−0.29705	-	0.016398
45	0.66058	0.011351	1	0	0.671931	0.64923	-	-	0.022701	-
59	−0.45689	0.00657	0	1	-	-	−0.45032	−0.46346	-	0.013139
72	0.83469	0.005848	1	0	0.840538	0.828842	-	-	0.011696	-
85	−0.37567	0.010922	0	1	-	-	−0.36475	−0.3866	-	0.021845
99	0.661981	0.006781	1	0	0.668762	0.6552	-	-	0.013562	-
112	−0.51664	0.018861	0	1	-	-	−0.49778	−0.5355	-	0.037722
(b)
ID	Mean	Var.	Peak	Valley	Peak Top	Peak Bot.	Valley Top	Valley Bot.	Peak Range	Valley Range
11	−0.00738	0.112158	0	1	-	-	0.104774	−0.11954	-	0.224315
24	0.879034	0.05588	1	0	0.934914	0.823154	-	-	0.11176	-
38	0.365574	0.043997	0	1	-	-	0.409572	0.321577	-	0.087994
52	0.878567	0.015044	1	0	0.893612	0.863523	-	-	0.030089	-
64	−0.05406	0.017749	0	1	-	-	−0.03631	−0.07181	-	0.035497
77	0.927113	0.884839	1	0	1.811952	0.042274	-	-	1.769678	-
90	0.233009	0.022373	0	1	-	-	0.255381	0.210636	-	0.044746
104	0.990595	0.006383	1	0	0.996978	0.984212	-	-	0.012765	-
117	0.004286	0.00893	0	1	-	-	0.013216	−0.00464	-	0.017861
(c)
ID	Mean	Var.	Peak	Valley	Peak Top	Peak Bot.	Valley Top	Valley Bot.	Peak Range	Valley Range
15	0.846359	0.002904	1	0	0.849264	0.843455	-	-	0.005808	-
28	−0.23564	0.005518	0	1	-	-	−0.23012	−0.24116	-	0.011036
41	0.851494	0.001186	1	0	0.85268	0.850308	-	-	0.002373	-
55	−0.23797	0.008646	0	1	-	-	−0.22933	−0.24662	-	0.017291
68	0.837491	0.00561	1	0	0.8431	0.831881	-	-	0.011219	-
81	−0.27158	0.003622	0	1	-	-	−0.26796	−0.2752	-	0.007243
95	0.774475	0.002352	1	0	0.776827	0.772123	-	-	0.004705	-
108	−0.30332	0.008789	0	1	-	-	−0.29453	−0.31211	-	0.017578
122	0.760005	0.004668	1	0	0.764673	0.755336	-	-	0.009337	-

digests how the first rule was established in Section 4.1, based on the resampled, resmoothed, and mean–var. transformed data on the X axis (see Section 3.3 and Figure A7a, Figure A8a and Figure A9a), with the setting that wind-speed (WS) = 6 m/s.

Table 2. Resampled, resmoothed mean–var. transformed data on the Y axis (WS = 6 m/s). (a) Min/var. peak/valley stats for the 3.0-blade turbine. (b) Min/var. peak/valley stats for the 2.5-blade turbine. (c) Min/var. peak/valley stats for the 2.0-blade turbine.

(a)
ID	Mean	Var.	Peak	Valley	Peak Top	Peak Bot.	Valley Top	Valley Bot.	Peak Range	Valley Range
7	0.553172	0.008447	0	1	0	0	0.561619	0.544724	0	0.016895
20	1.113486	0.009612	1	0	1.123098	1.103874	0	0	0.019224	0
34	0.597753	0.010292	0	1	0	0	0.608045	0.58746	0	0.020585
48	0.91686	0.008372	1	0	0.925231	0.908488	0	0	0.016743	0
61	0.746513	0.019163	0	1	0	0	0.765676	0.72735	0	0.038326
75	1.28571	0.002974	1	0	1.288683	1.282736	0	0	0.005947	0
88	0.474803	0.007407	0	1	0	0	0.48221	0.467396	0	0.014814
101	1.31762	0.01647	1	0	1.33409	1.30115	0	0	0.032939	0
115	0.595876	0.002217	0	1	0	0	0.598092	0.593659	0	0.004433
(b)
ID	Mean	Var.	Peak	Valley	Peak Top	Peak Bot.	Valley Top	Valley Bot.	Peak Range	Valley Range
12	1.145866	0.154981	0	1	0	0	1.300847	0.990884	0	0.309963
25	1.584168	0.018942	1	0	1.60311	1.565226	0	0	0.037884	0
38	1.204525	0.18757	0	1	0	0	1.392095	1.016955	0	0.37514
52	1.810359	0.021382	1	0	1.831741	1.788976	0	0	0.042765	0
65	1.112547	0.109336	0	1	0	0	1.221883	1.003211	0	0.218672
78	2.019655	0.094675	1	0	2.11433	1.92498	0	0	0.18935	0
92	1.097061	0.010047	0	1	0	0	1.107108	1.087014	0	0.020093
105	1.745129	0.010731	1	0	1.75586	1.734399	0	0	0.021461	0
118	1.113955	0.005631	0	1	0	0	1.119586	1.108324	0	0.011263
(c)
ID	Mean	Var.	Peak	Valley	Peak Top	Peak Bot.	Valley Top	Valley Bot.	Peak Range	Valley Range
7	0.532054	0.010882	0	1	0	0	0.542937	0.521172	0	0.021765
20	1.098469	0.00871	1	0	1.107179	1.089759	0	0	0.01742	0
34	0.570066	0.011347	0	1	0	0	0.581412	0.558719	0	0.022693
48	0.934223	0.007859	1	0	0.942082	0.926363	0	0	0.015719	0
61	0.798602	0.016174	0	1	0	0	0.814776	0.782429	0	0.032347
75	1.308704	0.004071	1	0	1.312775	1.304633	0	0	0.008142	0
88	0.445239	0.002191	0	1	0	0	0.447429	0.443048	0	0.004381
101	1.283832	0.013534	1	0	1.297366	1.270298	0	0	0.027068	0
115	0.610893	0.001298	0	1	0	0	0.612191	0.609594	0	0.002597

is numerically presented on the Y axis (see Section 3.3 and Figure A7b, Figure A8b and Figure A9b), because rule 1 says that using the information either on the X-axis or on the Y-axis is okay.

In these tables, ID is the number of the data entry in the resmoothed and resampled dataset in 3.3, based on the retrieval of the initial 1024 signals of each data sequence in 3.2. For example, in Table 2b, the ‘Mean’ and ‘Var.’ for ‘ID’ = 12 means the data entries with clocks [(12 − 1) × 8, (12) × 8 − 1] in the source accelerometer data sequence (which is 0-started) have a mean value of 1.145866 and a variance of 0.154981. ‘Peak’ = 0 and ‘Valley’ = 1 means that a valley appears here (as can be seen, in this table, entries that are neither a peak nor a valley in the figure are not shown here), and this valley is having a value delimited by [‘Valley Top’, ’Valley Bot.’] = [0.990884, 1.300847], while the ‘Valley Range’ of it is 0.309963. For this identified extreme, ‘Peak Top’, ‘Peak Bot.’ and ‘Peak Range’ do not receive any value because they are not peaks (see also in the corresponding subfigure in Figure A8b).

Rule 1 is obvious from these tables. Along the X or Y axis, either the peak range or the valley range of the 2.5-blade turbine is far greater than the peak range or the valley range of a 3.0-blade turbine or a 2.0-blade turbine. This is more salient when a total summary for all cases and the average is given in

Table 3. Summary for peak/valley information for all cases.

Turbine Case	Accel. Axis	WindSpeed	#Peaks/Valleys	Avg. All Peaks	Avg. Peaks’ Variances	Avg. All Valleys	Avg. Valley Variances
3.0	X	00	9	0.419068	0.019037	0.010037	0.020074
3.0	X	06	9	0.767823	0.013936	0.012113	0.024227
3.0	X	12	9	0.862113	0.016336	0.010408	0.020817
3.0	X	18	9	0.948771	0.014	0.010141	0.020282
3.0	Y	00	9	1.027233	0.018741	0.005394	0.010787
3.0	Y	06	9	1.158419	0.018713	0.009505	0.01901
3.0	Y	12	8	1.212972	0.013009	0.014553	0.029105
3.0	Y	18	9	1.364735	0.025744	0.015587	0.031175
3.0	Z	00	9	2.041538	0.013117	0.006551	0.013102
3.0	Z	06	9	3.094599	0.026482	0.018394	0.036788
3.0	Z	12	9	3.197712	0.029044	0.012246	0.024493
3.0	Z	18	9	3.446215	0.050402	0.021442	0.042884
2.5	X	00	8	0.532799	0.010725	0.004935	0.00987
2.5	X	06	9	0.918827	0.481073	0.041041	0.082083
2.5	X	12	8	0.852311	0.113416	0.036543	0.073086
2.5	X	18	9	0.873083	0.231489	0.032187	0.064373
2.5	Y	00	12	1.429386	0.010936	0.006929	0.013859
2.5	Y	06	9	1.789828	0.072865	0.093513	0.187026
2.5	Y	12	9	1.696677	0.032836	0.043012	0.086024
2.5	Y	18	8	1.819861	0.125277	0.040118	0.080236
2.5	Z	00	9	2.242601	0.005715	0.004855	0.00971
2.5	Z	06	9	3.405228	0.036168	0.017841	0.035681
2.5	Z	12	9	3.288273	0.068447	0.048333	0.096666
2.5	Z	18	9	3.69345	0.042324	0.032733	0.065466
2.0	X	00	9	0.46955	0.014728	0.007318	0.014636
2.0	X	06	9	0.813965	0.006688	0.006644	0.013287
2.0	X	12	9	0.769434	0.025687	0.00883	0.017661
2.0	X	18	9	0.904054	0.012814	0.00654	0.01308
2.0	Y	00	9	1.082138	0.015552	0.005682	0.011363
2.0	Y	06	9	1.156307	0.017087	0.008378	0.016757
2.0	Y	12	8	1.213207	0.014997	0.013445	0.026891
2.0	Y	18	9	3.088382	0.023286	0.013972	0.027944
2.0	Z	00	9	1.876956	0.013995	0.007343	0.014686
2.0	Z	06	9	2.940388	0.020016	0.011664	0.023329
2.0	Z	12	9	3.114471	0.014417	0.003698	0.007396
2.0	Z	18	9	3.088382	0.023286	0.013972	0.027944

(e.g., when wind-speed = 6 m/s, on Y axis, 0.072865 (2.5-blade) >> 0.018713 (3.0-blade) > 0.017087 (2.0-blade); 0.187026 (2.5-blade) >> 0.01901 (3.0-blade) > 0.016757 (2.0-blade)). That is, for a 2.5-blade turbine, the variance at the peak or valley along the X or Y axis is at least triple or more than the 3.0-blade or 2.0-blade cases.

Finally, Rule 2 should also be obvious from these tables. See in Table 3. For the 3.0-blade settings, on the Z axis, from the average of all valley values to the average of all peak values (i.e., the fluctuation), the numbers are 2.034987, 3.076206, 3.185465 and 3.424772, respectively, imposing the four wind-speeds. However, for the 2.0-blade settings, on the same axis, these numbers are 1.869613, 2.928724, 3.110773 and 3.074410, respectively. All of these numbers are below those of the 3.0-blade settings (while other settings are identical) by 7% (8.8%, 5%, 2.4% and 11.4%) in average. This quantified rule can be applied, and the reason for this is related to electromechanical conversion: more electricity is produced by the wind turbine when no blade is missing.

5. Conclusions

Green energy has become a major power source over the past two decades. Wind does not pollute and is currently one of the most promising clean and inexhaustible energy sources for power generation [24]. Recent advances in wind energy production have helped to solve practical problems [25] and improve quality of life [26]. Due to the torsional and flexural coupling of the pre-bent blades, the dynamic characteristics of blades made from orthotropic composite materials with conical pre-torsion are much more complex than those of isotropic blades. On the other hand, this means that—compared to other components of a wind turbine whose designs have had time to mature—the relevant technologies and designs for the fan blades are relatively new.

Condition monitoring and fault diagnosis of wind turbines have gained more and more practical value for reducing maintenance costs and improving wind farms’ operational efficiency [27], because more and more wind farms are being constructed and operated globally. Thus, the market has become competitive, and wind farm operators usually need to reduce their operating and maintenance costs in order to make their operations more profitable, and to maintain the sustainable competitive advantage (SCA) of the company, the maintenance strategies must be effective as operations continue. Due to these drivers, the condition monitoring and early fault diagnosis of wind turbines have become required industrial practices because they help improve the reliability and productivity of wind farms [28].

Due to the high maintenance costs (and efforts) incurred, the failure of wind turbine blades during the operation and maintenance phases has become a major problem for the wind power industry. Therefore, the utilisation of quality real-time data [29] and the development of methods to monitor the turbine blades’ integrity is critical [30] because of the novelty of the blades’ designs (which also determine the feature(s) of some turbine types of certain turbine brands).

In order to detect blade damage, after a review, we found that vibrations and noises are the two interfaces through which to determine mechanical faults, and the signals these carry can be analysed to determine the cause of the event. However, after long laboratory trials, we also found that taking these signals and using the existing justification criteria from research on power generators is infeasible and ineffective for diagnosing fan-blade faults using CMS with the wind turbine on the remote side. The rules (criteria) for detecting faulty situations therefore had to be reconsidered, and the idea of detecting faults using remote-side CMS with limited computation power had to be abandoned.

This led to the idea to establish effective new rules for the server-side CMS to detect faulty situations based on the vibration data transmitted and updated from the remote-side CMS, because using the datasets recorded from an interface were sufficient and using the interface recording vibrations would be better. This resulted in the creation of a new plug-in for the failure and cause-of-failure analysis module on the server-side CMS, which detects sudden faulty events and types of fan-blade damage in almost real time (see Figure 1). It can then send an alert message via SMS or email to the engineers on duty, so that they can make repairs immediately.

In this sense, it may help to establish a maintenance by prediction mechanism for unplanned maintenance when any fan blade is out of order, which supplements the common planned tasks carried out for the preventive maintenance of the fan blades. Despite the suggested mechanism playing a supplementary role to regular preventive maintenance being the main possible contribution, it should be particularly noted that this point is also exactly the boundary of this study, i.e., no work related to traditional preventive maintenance is presented. Another boundary of the study should be that the mechanism suggested by this study is for turbines having blades made of isotropic materials (because of current experimental limitations), so there is still room for exploring whether or not it still holds (or is there any other more effective mechanism) for those turbines having orthotropic blades.

All of this relies on the effective rules established to identify the 3.0-blade (full blade or normal), 2.5-blade (half or part of the blade broken) and 2.0-blade (one blade totally missing) cases while the turbines are operating. Fortunately, with the help of a contemporary data-driven approach and the adoption of suitable data processing/analysis methods in both DSP and statistics, we found that watching a continuous (but short, in terms of sampling time) sequence of the vibration dataset was sufficient to establish these rules. Fortunately, despite the considerable time and effort put into the data experiments to determine the rules, these rules are not difficult to implement on any given CMS, because:

(1) No real domain transformation is required, and only data conversions happen: the rules work with the data in the source domain, with just a few conversions (rather than transformations) of the data required for the computation;

(2) The rules are data-pattern-based, rather than learning-based: fixed judgements are made on the observable patterns or characteristics in the converted data, so no other processes, such as training, verifying, or parametric tuning efforts, are needed; and

(3) The rules are simple: only two rules are included (and required) to make the judgements, and they are simple, so they can be designed as additional functions of the CMS without utilising significant run-time computational resources.

The real use of these rules on the server-side CMS is expected in the future, and their true value will be shown when the unplanned maintenance by prediction tasks is someday carried out for the turbine blades. We sincerely hope that in the future, these rules can benefit not only the company operating the wind farms but also the entire wind turbine industry. They have the potential to change the ways people approach CMS design for wind turbines, as the effectiveness of data-driven server-side fault diagnostics has been made evident in this study. Finally, while the proposed mechanism works for shoreside wind turbines, a similar logic can be generalised to other wind turbines (e.g., offshore or inland) as well.