1 Introduction

Wind energy is rising and favouring renewable energy to reduce greenhouse gas emissions. Figure 1 shows the wind energy capacity until 2018 and the projection to 2023 according to the Data from Council, G.W.E. Global Wind Report (Ohlenforst and Council 2018). The global wind energy production was approximately 51 GW in 2018, being 47 GW onshore and the rest offshore (Murdock et al. 2019).

Fig. 1
figure 1

Wind energy capacity and projection (Ohlenforst and Council 2018)

Full size image

Wind energy is generated by Wind Turbines (WTs), which are becoming bigger and more complex to maximize energy production (Mao et al. 2020; Delorm et al. 2016). Offshore wind turbines require also complex maintenance procedures to avoid failures that lead to production losses, increasing costs, penalties and downtimes (Gómez 2017; Wang et al. 2020a; An and Jiang 2014). The maintenance procedures depend on data analysis based on the monitoring of the main parameters of the principal components to set their conditions (Zhang and Lang 2020; Ucheniya 2024). Supervisory Control and Data Acquisition (SCADA) is mainly employed for any WT (Santos et al. 2015a), and a Condition Monitoring System (CMS) is used for specific devices, where its use has been demonstrated from the reliability and economic point of view (He et al. 2020).

By capturing the uncertainty and variability inherent in wind turbine operations, fuzzy logic systems enable a more adaptive and resilient approach to maintenance. For offshore turbines, fuzzy logic is used to plan proactive maintenance in calmer weather. Offshore turbines are subject to more unpredictable environmental factors such as wave height, salt corrosion, and extreme winds (Chacón et al. 2023). These factors complicate the traditional methods of reliability analysis (Marquez 2006; Garcia Marquez 2019). Onshore wind turbines are more susceptible to variable terrain effects, leading to localized wind patterns. Fuzzy logic can effectively manage these uncertainties by creating flexible rules that capture subtle differences in wind speed, direction, and turbulence due to topography (Hermosa Gonzalez et al. 2014). Onshore wind turbines experience more diverse fault types due to environmental variations. The harsher offshore environment causes frequent issues like blade erosion, pitch system failures, and corrosion (Pedregal et al. 2009). Fuzzy logic is crucial in both environments but is adjusted for land-based irregularities onshore and harsher conditions offshore. Fault and false alarm detection plays a role in predictive maintenance, with onshore systems focusing on varied fault types and offshore systems on weather-induced wear (Márquez et al. 2003; Muñoz et al. 2017). Offshore systems benefit significantly from fuzzy logic for predictive maintenance, while onshore systems use it for efficiency in cost and scheduling.

Catelani et al. (Catelani et al. 2020) suggested integrating various methods into a unified system to enhance the detection and identification of abnormalities and malfunctions in wind turbines using SCADA data analysis. The outcomes enable timely detection of failures, allowing for appropriate maintenance tasks to be carried out. A similar study was done in references (Jain and Yamé 2020; Razavi-Far and Kinnaert 2013). Yang et al. developed an effective method for SCADA data analysis based on the search for correlations between SCADA data, doing a quantitative evaluation of WT condition monitoring under varying operational conditions, and validating the results by a set of tests. Bangalore and Tjernberg (Bangalore and Tjernberg 2015) introduced a self-evolving maintenance schedule framework for WT maintenance management and severe fault detection. They used an Artificial Neural Network (ANN) approach for condition monitoring of gearbox bearings employing SCADA data of onshore WTs. References (An and Jiang 2014; Mazidi et al. 2017a) show a similar study. Igba et al. (Igba et al. 2016) proposed some techniques based on three models: signal correlation; Extreme vibration, and; Root Mean Square (RMS) intensity. The approach was validated by a time domain approach applied to CMS data. Uzunoğlu (Uzunoğlu 2019) employed the SCADA data for preventive maintenance tasks by using the Bayesian update strategy. Baraldi et al. (Baraldi et al. 2008) genetic algorithms for signal grouping in sensor validation. Authors compare filter and wrapper approaches. A similar study was done in Baraldi et al. (2011) for grouping condition monitoring signals of nuclear power plant components. Roverso et al. (Roverso et al. 2007) presented solutions for plant-wide on-line calibration monitoring.

The gearbox is the most studied component in WTs: Kim et al. (Kim et al. 2011) studied the failure detection of the gearbox. They developed algorithms based on clustering algorithms and analysis of the main parameters; Feng et al. (Feng et al. 2013) analysed two different cases, 2 MW WT variable speed and 1.3 MW two-speed. The SCADA and CMS data of the Gearbox with faults were studied; Wang and Infield (Wang and Infield 2013) presented a Non-linear Condition Estimation Technique (NSET) model to analyse the condition of the WT gearbox by using historical data. They used Welch’s t-test normal behaviour models and a time series filter to identify failures in the gearbox. Other principal components of the WT are also studied: Kusiak and Verma (Kusiak and Verma 2012) employed models for predicting faults in the bearings of the WT generator 1.5 h before their appearances. These models are based on Neural Networks (NN) algorithms applied to SCADA data of 24 WTs over four months. Park et al. studied the planetary gear by transmission error (Park et al. 2016), and Mazidi et al. employed a hybrid approach of NN and a proportional hazards model (Mazidi et al. 2017b) to a similar dataset. Wang et al. (Wang et al. 2020b) employ the wavelet packet pre-processing NN applied to sound intensity analysis for an engine-fault-diagnosis system. NN was also used by Xiang et al. (Xiang et al. 2020) based on long short-term memory with weight amplification to predict the gear remaining useful life. The bearing was also considered by Wei et al. (Wei et al. 2020), utilizing n Cluster-MWMOTE and MFO-optimized LS-SVM. Finally, the planetary gearbox was analysed by He et al. (He et al. 2019) employing a deep learning-based approach.

The aforementioned research shows that the main advances are focused on multivariable analysis to increase the accuracy of fault detection. It is being applied approach for analytics based on artificial intelligence, mainly in expert systems (Sánchez et al. 2024). This paper proposed a multivariable analysis employing Fuzzy Logic for alarm detection, where the literature is focused on fault detection. A survey of Fuzzy Logic and its applications was presented by Mittal et al. (Mittal et al. 2020) and Lam (Lam 2018). Mao et al. (Mao et al. 2019) used Genetic Algorithms together with Fuzzy logic for decision-making optimization considering maintenance costs, energy interaction costs and the cost of pollutant treatment in microgrid problems. A similar approach was presented by Khaniyev et al. (2019) taking into account type-2 Fuzzy parameters, You et al. (You et al. 2019) used a Fuzzy Random Uncertainty, and Chebouba et al. (2021) considered a multi-objective system reliability optimization problem. Gao et al. (2019) employed the Fuzzy and interval variables based on entropy theory for reliability-based design optimization. Zhong et al (Zhong et al. 2019) studied offshore wind farms and their preventive maintenance scheduling by a reliability-and-cost-based Fuzzy approach. Shayeghi and Y. Hashe (Shayeghi and Hashemi 2015) employed Fuzzy Logic to minimize the annualized cost of the photovoltaic-wind hybrid system. They considered a power supply probability index of loss. Schlechtingen and Santos (Schlechtingen and Santos 2014) proposed Adaptive Neuro-Fuzzy Interference Systems (ANFIS) based on normal behaviour models applied to SCADA data. Schlechtingen et al. (Schlechtingen et al. 2013) presented also case studies where the previous research was employed. Alves et al. (Alves et al. 2017) studied fault prediction based on wireless sensor and actuator networks by Fuzzy Logic and Fuzzy NN. Han et al. (Han et al. 2019) employed interval number grey Fuzzy to decision-making in maintenance. Simani and Castaldi (Simani and Castaldi 2019) considered noise and disturbance effects and partial knowledge of system dynamics in WTs. They studied the fault detection and isolation by NN and Fuzzy Logic prototypes. Liu et al. (Liu et al. 2019) employed a bi-level Fuzzy stochastic expectation modelling and optimization for energy storage systems planning. It was applied to virtual power plants by Monte Carlo simulation, being optimized by Quadratic Programming and Genetic Algorithm. The AHP-Fuzzy decision method was employed by Bai et al. (2016) on a WT blade under the full-scale fatigue test to analyse the stiffness degradation, where it was considered the environmental temperature impact. Tautz-Weinert and Watson (Tautz-Weinert and Watson 2016) modelled the wind farm reliability by simulating failures by linear combinations, adaptive Neuro-Fuzzy inference systems, artificial NN, Gaussian process regression and support vector machines (see also similar research in references (Herp et al. 2016; Yadav et al. 2024)). Vibration-based structural damage localization was studied by a novel multi-step approach by Fuzzy Logic in reference (Cross and Ma 2015). The severity estimation in WTs was also considered. Table 1 shows the main research studies done in Fuzzy Logic applied to WT Maintenance.

Table 1 Fuzzy logic research in WT maintenance
Full size table

The literature review shows that the SCADA system generates alarms, and each alarm corresponds to a maintenance task (Marugán et al. 2017; Hijazy and Zempléni 2020). False alarms occur when the data analysis is done incorrectly, resulting in unnecessary downtime and, therefore, reducing the energy production, together with maintenance tasks, resources, penalties and, consequently, costs to the company (Hijazy and Zempléni 2020; Yang et al. 2014). Advanced analytics are necessary to reduce the false alarms, and ensure the reliability, safety and availability of the WTs (May and McMillan 2013; Forcina et al. 2020). The complexity of the problem, due to the variety and number of signals, makes it a complex problem that cannot be solved with classic methods, or, in other cases, the computational cost is very high. In these cases, the Fuzzy Logic has been demonstrated for efficiency and accuracy, according to the previous literature, and is recommended to be used. This is the main reason that this paper has employed this technique.

The global average downtime in WTs is 1–7 days (Reder et al. 2016). Other issues include breakdowns and maintenance events, that cost the industry more than €15 billion per year in lost productivity (Cusidó et al. 2021). The use of increasingly efficient control monitoring systems is a subject of intense study in the field of WTs technology research (Yang et al. 2021). Most studies have focused on SCADA signals rather than SCADA alarms. The SCADA database collects alarm information. When signals from essential components surpass threshold limits, alarms are generated and recorded. SCADA alarms indicate to operators possible emergency events and decrease the risk of catastrophic failure. Models use SCADA or simulated data, with classification accounting for 67% of the methods and regression accounting for the remainder 33% (Peco Chacón et al. 2021).

According to the state of the art, false alarms have not been studied enough. Benmessaoud et al. presented a new technique for identifying alarms based on Fuzzy Logic and data collected by the SCADA system. Maintenance strategies were proposed according to the nature of alarms at the Fuzzy system output. The signals analysed were different to the signals considered in this paper, and also the approach to reduce the number of inputs variables to the Fuzzy system and it was obtained the probability of faults. Pliego et al. employed ANN and a reinforcement learning agent (Marugán et al. 2024) for false alarm detection applied to a real dataset from a SCADA system together with a CMS based on vibration.

The literature shows that many existing alarm systems rely on static thresholds for alerting, which may not adapt well to changing environmental conditions or operational states. This rigidity can increase the likelihood of false alarms during normal fluctuations. Developing robust machine learning models for alarm detection often requires substantial amounts of labeled training data. The scarcity of labeled data, particularly for rare fault events, can hinder the model's ability to generalize effectively. Processing large volumes of data in real time poses computational challenges. If the detection algorithm is not efficient, it may introduce delays, affecting timely maintenance responses. Integrating new detection methods with legacy SCADA systems can be complex and may require significant modifications, posing barriers to implementation. Addressing these limitations requires ongoing research, better data management practices, and the development of more adaptive and intelligent algorithms that can improve the reliability of alarm detection in wind turbine operations.

This article examines real-time monitoring of a wind farm using big data gathered from a SCADA system. The SCADA system facilitates informed decision-making regarding maintenance types. It generates alarms based on the collected data; however, false alarms can lead to unnecessary interventions by the maintenance team, resulting in production losses and increased costs. Reducing these false alarms can enhance wind farm maintenance management. In this paper, we introduce a novel approach for alarm identification using Fuzzy Logic based on SCADA data. The alarms can be classified into two categories: orange alarms, indicating faults that necessitate preventive maintenance, and red alarms, signaling critical conditions that could lead to system failures. This paper presents as main novelty the false alarm detection employing SCADA data. An approach based on Fuzzy Logic is employed by using as the input data the correlations of the variables of the SCADA to reduce the number of the variables. Another contribution is the false alarm detection in WTs. It is employed and validated by a real case study. The approach works online and with a large amount of data, or Big Data.

The rest of the paper is presented as follows: Sect. 2 presents the real case study analysed in this paper; Sect. 3 describes the approach based on Fuzzy Logic. The main results and the validation are shown and studied in Sect. 4; Finally, the main findings of the study are listed in Sect. 5.

2 Real case study

The SCADA data used in this study is derived from the European project OPTIMUS. It consists of data gathered at 10-min intervals over a period of 24 months. Pearson’s correlation was employed to reduce the use of 37 variables system to 12 variables. With this purpose, the SCADA data will be preprocessed before being inputted into the fuzzy system. This preprocessing phase begins with calculating the simple moving average (SMA) of the SCADA signals. Next, the differences between the signals and the SMA are computed, which will serve as input vectors for the fuzzy system. After deriving these new variables, the subsequent step is to establish control laws for them. This classification facilitates the determination of membership functions for the fuzzy system's inputs and the generation of the fuzzy system itself. Each physical parameter collected by SCADA can be represented by the difference between peak values and the mean. The fuzzy inference system uses various rules to generate the probabilities of alarm occurrences at the output. Finally, the new variables will be statistically analyzed to establish two thresholds that divide the graph into three distinct regions. The type of correlation is set in this paper according to the criteria given in reference (Pliego Marugán and García Márquez 2019).

The primary and basic research on this issue was presented in references (May and McMillan 2013), where it is presented the Matrix and details of these results. It is now detailed in this new version. Therefore, this paper considers only the parameters related to the kinematic and to the pitch and brake systems, being 8, that are classified into two groups:

  • The first group includes the variables related to the kinematic condition of the WT, being: speed of the main shaft (Vr); Vibration of the main shaft (Vibr); Bearing temperature of the gearbox (Tbg); Oil temperature of the gearbox (Tg), and; Hydraulic unit oil temperature (Thu). Figure 2 shows the input data over time.

  • The second group includes the variables related to the condition of safety systems. It is divided into two subgroups: the first sub-group includes the pitch system, considering the general accumulator pressure (Pga); The second sub-group comprises the braking safety system, which is responsible for the protection of the WT in case of risk and corresponds to the braking pressure (Pbr).

Fig. 2
figure 2

Kinematic variables over the time signals: a Dynamic Variables. b Temperature Variables

Full size image

In addition to these variables, the wind speed (Vel) is used to set rules related to the mentioned variables ([Vel, Thu] for the control of the yaw angle, and [Vel, Pga] for the control of the pitch angle). Figure 3 shows the Pbr and Pga input data over time.

Fig. 3
figure 3

Pbr and Pga over time

Full size image

In this paper, the signal is studied by considering the difference between the signal and the SMA: it shows if the signal is close or not to the normal value that it should have in conditions free of faults. Figure 4 shows an example (in this case Pga) of the modified variable before being considered by the Fuzzy system. The signal plotted on the top with noise is Pga. Pga is filtered by SMA with a 2 h period. The bottom signal shows the absolute difference between Pga and its SMA, called the difference of Pga (DPga).

Fig. 4
figure 4

Preprocessing of SCADA signals

Full size image

Equations (1–8) show the new signals considered as inputs of the Fuzzy system, following the case for Pga.

$$ DVel = \left| {Vel - SMA\left( {Vel} \right)} \right| $$
(1)
$$ DVr = \left| {Vr - SMA\left( {Vr} \right)} \right| $$
(2)
$$ DVibr = \left| {Vibr - SMA\left( {Vibr} \right)} \right| $$
(3)
$$ DTbg = \left| {Tbg - SMA\left( {Tbg} \right)} \right| $$
(4)
$$ DTg = \left| {Tg - SMA\left( {Tg} \right)} \right| $$
(5)
$$ DThu = \left| {Thu - SMA\left( {Thu} \right)} \right| $$
(6)
$$ DPga = \left| {Pga - SMA\left( {Pga} \right)} \right| $$
(7)
$$ DPbr = \left| {Pbr - SMA\left( {Pbr} \right)} \right| $$
(8)

Figure 5 and 6 show the new signals from Eqs. (1–8).

Fig. 5
figure 5

a Dynamic Input data in SMA b Temperature Input data in SMA

Full size image
Fig. 6
figure 6

Pressure Input data in SMA

Full size image

The correct variable ranges have been set according to references (Schlechtingen et al. 2013; Windpower 1500), and they are shown in Table 2.

Table 2 Normal operational ranges of the SCADA data (Schlechtingen et al. 2013; Windpower 1500)
Full size table

3 Approach based on fuzzy logic

The volume, variety and structure of the data from the SCADA need robust algorithms to be analyzed (Zadeh 1999). Fuzzy logic is beneficial in multivariable analysis due to its ability to manage uncertainty and ambiguity, enabling nuanced decision-making among complex, interrelated variables. It uses natural language terms, making it easier to model human reasoning, and effectively represents non-linear relationships, enhancing accuracy. By integrating expert knowledge and accommodating qualitative insights, fuzzy logic also allows for more tailored analyses through flexible membership functions. Its robustness to noise ensures reliable performance despite data quality variations, while the capacity to consider multiple inputs simultaneously makes it ideal for complex systems. Additionally, fuzzy models tend to be simpler to design and implement compared to traditional statistical approaches, especially in scenarios lacking clear data patterns.

Fuzzy Logic is employed in this paper due to the properties to analyse this type of SCADA data according to the state of the art. The method does not present a scientific novelty, but the approach and problem presented in this paper have not been yet covered in the literature. Figure 7 shows a scheme of the Fuzzy Logic system considered in this paper based on Abreu and Ribeiro (Abreu and Ribeiro 2003).

Fig. 7
figure 7

The basic configuration of a Fuzzy Logic system

Full size image

The Fuzzy system is based on the fuzzification, Fuzzy inference and defuzzification phases (Fig. 7), which are detailed as follows.

3.1 Fuzzification

The SCADA data is transformed to a linguistic value between 0 and 1, i.e., a Fuzzy subset, according to reference (Jamshidi 2003). This transformation has been done by Sigmoid, Hyperbolic tangent and Exponential functions. The best results in this paper were found by the Sigmoid method. This paper considers mainly three states for each input (Good, Acceptable, and Unacceptable), taking into account n inputs.

3.2 Fuzzy inference

Fuzzy inference is employed as the input data of the Fuzzy Logic operators (and/or), membership functions, and If–Then rules, known as Fuzzy rules, at an output according to reference (Dinu and Ilie 2015). The paper explores diverse approaches for deriving the fuzzy rules discussed (such as gradual rules, certainty rules, possibility rules, and more), each exhibiting distinct inference characteristics and catering to different intended purposes and practical scenarios. In this paper, the rules have been obtained following the suggestions given in references (Dubois and Prade 1996) and (Yager and Filev 1994) for the case presented in this paper. The structure of the Fuzzy rules employed are:

$$ \begin{gathered} {\text{IF}}\,Var_{1} \,{\text{is}}\,{\text{a}}_{11} \,{\text{and/or}}\,Var_{2} \,{\text{is}}\,a_{21} \, \ldots \,{\text{THEN}}\,y\,{\text{is}}\,b_{1} \hfill \\ else \hfill \\ {\text{IF}}\,Var_{1} \,{\text{is}}\,{\text{a}}_{12} \,{\text{and/or}}\,Var_{2} \,{\text{is}}\,a_{22} \, \ldots \,{\text{THEN}}\,y\,{\text{is}}\,b_{2} \hfill \\ else \hfill \\ {\text{IF}}\,Var_{1} \,{\text{is}}\,{\text{a}}_{1n} \,{\text{and/or}}\,Var_{2} \,{\text{is}}\,a_{2n} \, \ldots \,{\text{THEN}}\,y\,{\text{is}}\,b_{n} \hfill \\ \end{gathered} $$

being Var1, …, Varn the variables used as Fuzzy inputs (antecedent), where Vari is detailed in Sect. 3; y is a single output (consequent), and; a11, a21, a12, a22... a1n, a2n are the Fuzzy sets (Khanna and Cheema 2013). All feasible combinations of the Vari provide the number of rules utilized in the Fuzzy inference, and they will be also related to the quantity of linguistic factors defining the membership functions of the input data, such as if there are n input variables and 3 fuzzy linguistic variables, then there will be 3n rules. This paper employed the Mamdani and Sugeno methods in the Fuzzy inference, where the Mamdani-Type provided better results.

3.3 Defuzzification

The Fuzzy sets and the membership degrees are employed in the defuzzification process to provide a quantitative result. In this paper, the maximum membership and centroid techniques are employed in this phase according to reference (Zhang 2010).

Figure 8 shows the alarm identification flowchart.

Fig. 8
figure 8

Alarm identification flowchart

Full size image

The main problem is the use of a large number of variables and volume of data from the SCADA to be analysed by the Fuzzy Logic system, which generates a large number of Fuzzy rules and, therefore, increases the computational cost of solving the problem. This paper employs a novelty with regards to the state-of-the-art Pearson’s correlation as a statistical method to reduce the input data by using a linear correlation between two input variables to reduce the inputs in the Fuzzy logic system. Other similar methods in the literature can be also applied, for example, Mander's overlap coefficient (MOC) and Pearson's coefficient share a similar mathematical basis, but they differ in their approach: MOC considers absolute intensities, while Pearson's coefficient focuses on variations from the average. According to Adler et al. (Adler and Parmryd 2010), for the problem presented in this paper, is recommended the Pearson correlation coefficient due to the Model of Concern (MOC) lacking responsiveness to significant shifts in the data and presenting challenges in interpretation, the correlation coefficient (denoted as 'r') between the discrete variables x and y can be calculated using Eq. (9).

$$ r = \frac{{N\left( {\mathop \sum \nolimits_{n = 1}^{N} xy} \right) - \left( {\mathop \sum \nolimits_{n = 1}^{N} x} \right)\left( {\mathop \sum \nolimits_{n = 1}^{N} y} \right)}}{{\sqrt {\left( {N\mathop \sum \nolimits_{n = 1}^{N} x^{2} - \left( {\mathop \sum \nolimits_{n = 1}^{N} x} \right)^{2} } \right)\left( {N\mathop \sum \nolimits_{n = 1}^{N} y^{2} - \left( {\mathop \sum \nolimits_{n = 1}^{N} y} \right)^{2} } \right)} }} $$
(9)

The type of correlation is set in this paper according to the criteria given in reference (Pliego Marugán and García Márquez 2019), being:

  • Weak correlation 0.3 ≤|r|< 0.5

  • Moderate correlation: 0.5 ≤|r|< 0.7

  • Strong correlation: |r|≥ 0.7

  • Perfect correlation |r|= 1

Several research studies delve into examining specific associations between pairs of variables. For instance, Yang and Jiang (Yang and Jiang 2013) evaluate the relationships between windspeed and power, windspeed and temperature, power and temperature, and generator speed and temperature. Similarly, Gill et al. (Gill et al. 2012) focus on the link between active energy and wind speed using power curve copula modelling approaches. Igba et al. (Igba et al. 2016) focus on the correlation between vibration levels and power output.

The variables with perfect and strong correlations have been selected as inputs of the Fuzzy system. The new input vectors of the Fuzzy system are obtained by the differences between the signal values and the corresponding Simple Moving Average (SMA). The reason is that a fault presents data in the signal that are farther from the average signal, i.e., the average signal is considered as the commissioned condition of the component.

3.4 Thresholds and clusters

To set the membership functions range in the inference system requires thresholds for the new input variables. The thresholds are used to cluster the data into three groups by a static analysis of the new variables. Figure 9 shows the thresholds and the clusters for temperature: good, acceptable, and unacceptable.

Fig. 9
figure 9

Clustering of the data for temperature: good, acceptable, unacceptable

Full size image

3.5 Alarms

The Fuzzy inference system uses various rules to determine the likelihood of alarms. The results from the Fuzzy Logic system are divided into three distinct scenarios, where the following criteria are taken from reference (Benmessaoud et al. 2017):

  • Green condition (No alarm): The evaluated parameters are within acceptable ranges, indicating that the system has no faults. This is classified when the Fuzzy Logic output has a probability of less than 0.5.

  • Orange condition (Warning): This indicates issues that are not severe enough to disrupt maintenance schedules and can be addressed through preventive maintenance tasks. This condition applies when the Fuzzy Logic output probability is between 0.5 and 0.75.

  • Red condition (Alarms): This represents critical states where failures might occur, necessitating immediate attention and urgent repairs to bring the system back to safe operating levels. This condition is triggered when the Fuzzy Logic output probability exceeds 0.75.

4 Results

Rules involve multiple input variables combined using logical operators such as AND, OR, and NOT.

  • AND: Represents intersection (e.g., both conditions must be true).

  • OR: Represents union (e.g., at least one condition must be true).

  • NOT: Represents negation (e.g., the condition must not be true).

While a larger number of rules can provide a more comprehensive model, it may also complicate the system and increase computational requirements. The variables considered correspond to the distance of the value measured by the SCADA to the SMA. The greater distance to the average, the more probability of being an abnormal measure, and therefore, the more probability of generating an alarm. Table 3 shows the rules considered in the case study presented in this study.

Table 3 Rules for the fuzzy system
Full size table

Rules shown in Table 3 are the same as the following “If-AND….-THEN…” rules:

  • IF DVel is Good and DVr is Good and DVibr is Good and DTbg is Good and DTg is Good and DThu is Good THEN the output is green.

  • IF DVel is Good and DVr is Good and DVibr is Acceptable and DTbg is Acceptable and DTg is Good and DThu is Good THEN the output is orange.

  • IF DVel is Unacceptable and DVr is Good and DVibr is Good and DTbg is Unacceptable and DTg is Unacceptable and DThu is Acceptable THEN the output is red.

  • If DVel or DVr or DVibr or DTbg or DTg or DThu is Unacceptable THEN the output is red.

Figure 10 shows an example of the probability of the alarm with regards to DThu and DVr. The probability is low for intermediate values of DThu and DVr, and it rises when the values are high. For the values of DThu less than certain values, the alarm probability is also very high, in this case, the high probably is due to the weather conditions.

Fig. 10
figure 10

Alarm probability with regards to DThu and DVr

Full size image

There are 63 alarms given by the SCADA system: 11 given by the wind velocity; 6 by the main shaft speed; and 6 by the main shaft vibration. The validation of the approach has been done by analysing together the red alarms obtained by the Fuzzy Logic system and the alarms given by the SCADA system. The approach can detect 63 alarms, and it also detects other 7 alarms, i.e., 70 alarms are detected by the approach. This study cannot conclude if these 63 alarms are false or not, but the approach can reinforce or not if an alarm given by the SCADA system can be true or false. The approach also generates 7 new alarms that the SCADA does not detect. The detailed results are shown as follows:

  • Case 1 (kinematic and temperature variables as inputs): A total of 81,460 data points have been analyzed. The outcomes are:

  • 39408 green alarms, i.e., not alarm, 48.38%.

  • 41993 orange alarms, 51.54%.

  • 59 red alarms, 7.10%.

  • Case 2 (pressure variables as input): A total of 103,184 inputs have been studied. The outcomes are:

  • 103169 green alarms, 99.99%.

  • 04 orange alarms, 0.0038%.

  • 11 red alarms, 0.0062%.

The results of the experiments are shown in Fig. 11: Fig. 11.a and 11. b show the probability of the alarm over time for Case 1 and Case 2, respectively.

Fig. 11
figure 11

Probability of the alarm: a. Kinematic and temperatures variables; b Pressures

Full size image

4.1 Support vector machine (SVM)

The validation of the results will be done by employing SVM, which has been employed for false alarm detection in references (Chacón et al. 2023; Peco Chacon and García Márquez 2023). Support Vector Machine (SVM) is a type of supervised learning algorithm often applied in fault detection and the analysis of complex data sets (Géron 2019). SVM is used for both classification and regression tasks, aiming to identify boundaries that maximize the margin of separation between distinct data classes.

In the linear scenario, the hyperplane is identified by maximizing the margin between classes. This approach divides the data into two separate classes using two parallel hyperplanes (Santos et al. 2015b). Here, the classes correspond to either alarm activation or no alarm activation. To handle non-linear transformations, the “kernel trick” is applied (Santos et al. 2015b). SVM operates in its dual form to sidestep the infinite dimensions that can arise from inner products. Solving the dual problem requires only the kernel’s functional form, not the specific transformation basis functions. In this study, a polynomial kernel was selected, as it has shown better performance in similar studies (Chacón et al. 2023; Peco Chacon and García Márquez 2023).

4.2 Validation results

The detailed results for red, orange and green alarms previously presented are shown in Table 4 for Case 1 and Table 5 for Case 2.

Table 4 Case 1 (kinematic and temperature variables as inputs): A total of 81,460 data points have been analyzed
Full size table
Table 5 Case 2 (pressure variables as input): a total of 103,184 inputs have been studied
Full size table

Moreover, converting the data gathered from SCADA into alarm probabilities can aid in decision-making. This approach helps to minimize false alarms by allowing the Fuzzy system's response to either confirm or reject an alarm triggered by the SCADA system. The multivariable analysis leads to the detection also other alarms that the SCADA system has not detected.

The findings have significant implications for SCADA system performance and false alarm detection. By enhancing the accuracy of alarm identification, they can reduce unnecessary interventions, leading to improved operational efficiency and cost savings. This, in turn, boosts the reliability of the SCADA system, as fewer false alarms foster greater trust among operators. Additionally, optimized alarm detection can facilitate timely maintenance actions, minimizing downtime and ensuring smoother wind farm operations. Ultimately, these advancements contribute to better decision-making and resource management in monitoring and controlling complex systems.

This approach provides an option to the existing alarm systems that often rely on static thresholds, making them inflexible to changing environmental conditions, which can lead to increased false alarms during normal fluctuations. Real-time data processing presents computational challenges, as inefficient algorithms can delay maintenance responses, which is addressed by this approach. Additionally, integrating new detection methods with legacy SCADA systems can be complex, hindering implementation. To overcome these limitations, ongoing research, improved data management, and the creation of adaptive algorithms are essential for enhancing alarm detection reliability in wind turbine operations.

5 Conclusions

Fuzzy Logic is proposed to analyse the data from the SCADA system of a wind turbine. The main purpose is to determine the probability of the alarm for a wind turbine by using the Fuzzy system as a function of multivariable input data. Pearson’s correlation technique has been employed to reduce the 37 variables collected by the SCADA system to only 12 variables, minimizing the number of rules in the expert system, i.e., the computational cost. The 12 variables are grouped into 3 families according to: kinematic; the pitch and brake system; and the energy generation. The probabilities calculated at the output of the Fuzzy Logic are classified as green, orange and red alarms. Green and orange alarms are dispatched, i.e., no action for green alarms. Red alarms are validated with regards to the real alarms given by the SCADA system, where all the alarms are identified, but there are also a few more alarms detected.

The results have been validated employing SVM, where the MAPE is less than 1.67% for Case 1, and around 0% in Case 2 for green and orange alarms detection. For Case 2, 11 read alarms are detected by Fuzzy Logic and 10 by SVM.

The findings significantly impact SCADA system performance and false alarm detection by improving alarm identification accuracy, which reduces unnecessary interventions and enhances operational efficiency and cost savings. This increased reliability fosters greater trust among operators and enables timely maintenance, minimizing downtime and ensuring smoother wind farm operations. Overall, these advancements lead to better decision-making and resource management in managing complex systems.

It is suggested for future search works to apply this approach to other types of wind turbines, and other industries, and to consider other variables. The approach could be improved using new artificial intelligence approaches to increase the accuracy of the results. New techniques based on Artificial Intelligence could be applied to this complex problem.