This article underscores the importance of integrating machine learning analytics to enhance preventive maintenance methodologies, particularly emphasizing condition-based maintenance (CBM) within the wind energy domain. Through empirical evidence derived from wind turbine data, the paper outlines the efficacy and applicability of Machine Learning in Operations (MLOps) for predicting the residual operational life of wind turbine bearings. While the study's principal domain is renewable energy, especially wind power, it employs a specific wind turbine dataset for exhaustive model testing, leading to the proposition of an innovative ensemble model tailored for high-speed wind turbine bearing prognosis. The introduced model, "The Data Interpretation Algorithm for Forecasting Time Series" (DIAFS), crafted for assessing wind turbine bearing conditions, is predicated on an adaptive polynomial model approximation. It emerges as an indispensable asset for maintenance professionals implementing CBM methodologies.
THE E.U.'s 2050 energy roadmap mandates member states to advance infrastructure for longterm energy system decarbonization. Estimates suggest a global population increase of 2 billion by 2050, requiring 47% more energy for a total of 10 billion people. Given the current energy system's inadequacies and climate goals, there's an urgent need for sustainable energy practices (energy.ec.europa.eu). In this context, aspects of the wind farms, especially the components, such as the wind turbines, are critical to investigate. Wind turbines (W.T.) are complex equipment with a design lifetime of approximately 20 years. However, its life span and availability can be augmented by the implementation of condition monitoring and preventive maintenance techniques related to operation and maintenance activities. As a result, this means that with effective operation and maintenance, these W.T.s are operational to be able to produce energy. In addition, Operating and maintenance (O&M) costs are part of a large amount of a wind farm's Levelized Cost of Electricity (LCOE). Thus, the reduction of O&M costs provides possibilities to control the LCOE in an effective way. Therefore, the optimization of operation and maintenance are critical factors in controlling the LCOE.
Consequently, one of the few measures of reducing the cost of production is to cut the cost
of operation, including maintenance. It is the second-highest or even the highest element in
operating expenses in some industries. Swanson [
Additionally, it's well known that condition-based maintenance (CBM) preventive maintenance strategy is preferred wherever it is technically feasible and financially viable. The heart of CBM is condition monitoring (CM), which in principle, involves data acquisition, processing, analysis, interpretation, and extracting useful information. The information helps to identify whether asset health has deviated from the normal. If so, then fault diagnosis and prognosis usually follow. Finally, a decision is taken regarding when and what maintenance tasks are to be performed.
The CBM is now the most widely employed strategy in the Wind Farm industry [
When it comes to renewable energy sources, such as wind farms, the efficiency
improvement in energy production goes hand in hand with the possibility of providing
predictive maintenance to the different energy production equipment. For instance, to provide
service to various equipment, the level of energy production is stable and can utilize its full
potential without fluctuations due to equipment breakdowns. Although, a disadvantage of using
wind farms is that they do not always produce energy. It requires storing energy when
production diminishes because of changed weather conditions. Therefore, battery storage
systems offer the possibility of having a more stable output of power to match the required
demand. Obtaining energy while minimizing pollution/contamination and costs is worrying
many due to climate change and global warming. In this aspect, renewable energy is an
excellent alternative. Continuously, the fluctuations in energy production as a result of different
weather conditions have led to the option of moving the wind turbines offshore since the
conditions are better suited, i.e., higher and stable wind speeds are experienced [
Furthermore, it is well known that difficulties arise when implementing the entire cycle of
condition-based maintenance (CBM). It is connected with specific issues regarding CBM
characteristics, such as the amount of data produced, data integrations, and the lack of experts.
Consequently, these issues have been tried to be solved over time, namely with the support of
various ICTs, such as expert systems, decision support systems, artificial intelligence, and
distributed intelligence [
Consequently, the main contributions of this paper are connected with the use of machine learning (ML) algorithms supported with MLOps, e.g., machine learning in operations. Its contributions are the following:
•A Machine Learning (ML) pipeline that uses training data from wind turbines. This design identifies patterns and creates a model that can handle new data without breaking.
• A new ML model for predictive maintenance for wind turbine bearings. This model monitors the condition of wind turbine bearings using an adaptive polynomial model to approximate time series data. The data is split into two: level and trend. For forecasting, it considers past data trends. We also employed the Holt-Winters Exponential Smoothing method for more precise forecasting. The high R-Squared values of proposed model suggest it effectively predicts new data.
• In addition, two ML stacking models were developed. One combines regression and knearest neighbors (knn) with a regression tree, while the other combines regression with support vector regression (svr) and a regression tree. These combinations reduce predictor correlation and enhance model generalization. Notably, the second model has a low Root Mean Square Error, indicating minimized errors in predictions.
Overall, the approach described in this article is comprehensive and provides an effective solution for wind turbine-bearing data analysis and forecasting. The research methodology is built as follows: 1. Dataset preprocessing. 2. Models' development. 3. Results evaluation. 4. System architecture development. 5. System development and testing.
The contribution mentioned above is presented further in section 4 of the paper. Accordingly, section 2 highlights some aspects of the domain of interest. Then, later in section 3, the machine learning analytical capabilities are explained, and the term MLOps, i.e., machine learning in operations, is presented; hence, the section discusses what would be needed to achieve analytical capabilities supported by the MLOps in the field. Further on, in section 4, a case example using Wind turbine failure data is provided, namely for a bearing fault considering the essential aspects of implementing MLOps in the domain. Finally, in section 5, the discussions and conclusions are given.
Despite the widespread adoption of wind turbines, the effective prediction of bearing failures remains a significant challenge, leading to unplanned downtimes and reduced turbine availability. Therefore, operation and maintenance, and specially condition-based maintenance, become key aspects to consider in this context.
Thus, one of the most common faults in wind turbines is connected with gearbox failures,
but where the bearing errors are overrepresented [
Consequently, all the former aspects are related to operation and maintenance practices,
with considerable potential for cost reductions and innovations. Therefore, it is essential to
consider the operation and maintenance of the wind industry, especially when it is believed to
drastically cover a significant part of the E.U.'s total electricity consumption. Further, the
domain's state of the art and challenges and its wind turbine bearing condition monitoring
methods can be found in [
Availability = 1 – (Unavailable Time)/(Available Time + Unavailable Time)
Thus, low availability can be poor wind farm reliability performance or a below-standard maintenance action. Therefore, an effective operation and maintenance approach (O&M ) is crucial to be able to make the conditions of optimal availability of the wind turbine, which at the same time can produce the expected amount of energy.
Consequently, the efficiency improvement in energy production goes in hand with the
possibility of providing predictive maintenance to the different energy production equipment.
For instance, to provide service to the various equipment by doing so is the level of energy
production stable and can utilize its full potential without fluctuations because of equipment
breakdowns. It is, therefore, crucial to optimize the O&M. Thus, preventive maintenance based
on the CBM approach provides an understanding of when it is suitable to maintain the
equipment to avoid unplanned stoppages; in this case, the energy production and try to keep the
wind turbines available so they can be operational and receive service when it is needed. It is,
therefore, key to understand the various offshore wind turbine operation and maintenance
approaches and how they might impact O&M [
In the application of machine learning methodologies, it is imperative to source appropriate
datasets and subsequently undergo rigorous data preprocessing. This involves the acquisition of
relevant data, subsequent cleansing procedures, and preparation for analytical diagnosis.
Periodic retraining of models is essential to ensure their robustness and accuracy. Furthermore,
for optimal performance, it's paramount that models autonomously adapt their parameters based
on data-driven insights, especially in response to dynamic changes in the input dataset [
Nevertheless, there are efforts to use machine learning methods for wind farms turbine,
especially for blade fault detection or generator temperature monitoring [
At the same time, the prerequisite surrounding infrastructure is vast and multifaceted. The
different business needs to organize continuous cooperation and interaction between all
participants in the processes of working with machine learning models, from business to
engineers and Big Data developers, including Data Scientists and ML specialists. MLOps is
important for industries with needs of streaming data processing, such as wind turbines
remaining useful life forecasting. Thus, MLOps provides organizational and business
capabilities, i.e., automation, engagement, insight/decision-making, and innovation [
This section outlines the process of implementing MLOps for wind turbines equipped with sensors on their bearings. Initially, the MLOps framework facilitates the integration of resources from diverse origins. Additionally, the selection of an appropriate model may vary based on the source of data. A schematic representation of the MLOps pipeline is provided in Figure 2.
The reason for monitoring the ML model is to understand how it solves the business problem. Concerning wind farms, data quality is crucial. Often this data is presented as a time series. In addition, seasonality may be different for different data sources because wind farms are situated in different places with various conditions. Despite its "flexibility" in finding relationships in large datasets, the ML model has many vulnerabilities.
Therefore, effective monitoring of machine learning models is essential for several reasons. First, the quality and structure of input data play a pivotal role in the accuracy of the model. Second, as models evolve, their performance can degrade over time, necessitating consistent evaluations. Third, there are challenges related to interdependent models and unique pipeline configurations. Fourth, there may be instances of abnormal values or predictions that the model has not previously encountered, emphasizing the need for robust outlier detection mechanisms. Additionally, understanding the model's inner workings and decision-making process is vital, particularly in contexts where interpretability is crucial. In many scenarios, there's ambiguity regarding the true values in queries a priori, leading to uncertainty about the precise class or cluster to which a particular component belongs. Furthermore, the time required for model computations, potential unavailability of deployment endpoints, alterations in the application's business logic, susceptibility to cyberattacks, and potential data losses all underscore the importance of rigorous monitoring. Such considerations are fundamental in ensuring the model's accuracy and reliability throughout its lifecycle.
In the present study, we analyzed a dataset obtained from a 2MW wind turbine's high-speed
shaft driven by a 20-tooth pinion gear [
Initially, classical time series models are applied. The polynomial trend model essentially functions as a multiple regression equation, making the methods and procedures of regression analysis, as discussed in the initial segment of this publication, largely relevant for its delineation. The AutoRegressive Integrated Moving Average (ARIMA) constructed time series is illustrated in Fig. 3. It begins with a visualization of the vibration signals in the time domain. Forecasts are generated for the subsequent five values (depicted by the red line). This visualization offers forecasted data without any smoothing.
For time-series forecasts, Root Mean Squared Error (RMSE) measure is used. It is equal to 17.93.
Next, based on vibration level, RUL is calculated.
To decrease RMSE, smoothing is proposed. The adaptive mono-parameter Braun model is used for stationary time series based on simple exponential smoothing: y ̂_(t+1)=S_t,S_t=αy_t+(1-α) S_(t-1),t=1,2,3,…, (1) where y ̂_(t+1) is the prognostic value of time series level in time (t+1), S_t is the exponential mean, α is the adaptation coefficient, and y_t is the current time series value.
In this context, the model's value is derived from a weighted average of the current actual value and preceding model values. This weight is commonly referred to as the smoothing factor. It dictates the rate at which the most recent observable data point diminishes in influence. A lower weight ensures that prior model values exert a stronger influence, leading to a smoother data series. Taking the adaptation coefficient α and the warning period τ, it is necessary to approximate the series using an adaptive polynomial model.
The novel method for predictive maintenance of wind turbine bearing is developed in the paper.
This method consists of the following steps: 1. Zero-order time series analysis (р = 0); 2. First order time series analysis (р = 1); 3. Assess the accuracy and quality of forecasts; 4. Make a forecast.
Step 1.
The procedure for Step 1 was developed as a sequence of the following steps: Let y ̂_0= y_0.
Append array y ̂ using the following formula:
y ̂_(t )= α*y_t+(1-α)* y ̂_(t-1) , where y ̂_(t ) is an actual value, and y _̂(t-1) is the previous number from the prediction array.
Repeat step 2 for all values in the dataset.
Step 2. 1. Let x=1, y ̂_0= y_0, l_0= y_0, and .b_0= y_1-y_0, where y is our initial dataset. 2. Define new level value using the formula: l_x= αy_x+(1- α)(l_(x-1)+b_(x-1) ). 3. Define new trend value: b_x=β(l_x-l_(x-1) )+(1-β)b_(x-1). 4. Define our prediction y ̂_(x+1)=l_x+b_x. 5. Define x=x+1 and repeat steps 2-5 until x<n. (2)
The performance of the time series models on the presented testing data is given below: 1. Exponential Smoothing: MAPE is appr. 8.5 % 2. Proposed method: MAPE is appr. 2.2 % 3. Holt's Trend Method: MAPE is appr. 6.6 % 4. ARIMA: MAPE is appr. 3.1 % 5. TBATS: MAPE is appr. 3.2 %
The Exponential Smoothing model did well by achieving a lower MAPE of 8.5 percent. All the other models outperformed them by producing lower MAPE. However, the DIAFS model emerged as the winner based on its test data with MAPE performance, which was close to 2.2 %.
First of all, kurtosis was calculated. The kurtosis measure describes the tail of distribution – how similar are the outlying values of the distribution to the standard normal distribution? For example, the standard normal distribution has a kurtosis of 0.
Stochastic variables and numerous uncertainties can substantially complicate the problem. To circumvent these challenges, various ensembles are proposed. An ensemble method in both statistics and machine learning leverages multiple trained algorithms to achieve superior predictive performance than what could be attained by any single algorithm alone. In contrast to a statistical ensemble, a machine learning ensemble encompasses a distinct finite set of alternative models but typically allows for much more flexible structures. The central premise is to employ fundamentally diverse models to enhance the capability of processing unfamiliar data. The results derived from regression, KNN, ANN models, and the proposed ensemble of predictive models were subsequently compared.
Bagging is an ensemble technique where models are trained in parallel on different random subsets of the training data. The final decision is derived from the majority voting of the ensemble classifiers, selecting the class predicted by the majority.
Boosting involves training an ensemble of models sequentially, where each subsequent model focuses on instances that the preceding classifier misclassifies. While boosting typically yields more accurate results than bagging, it can be susceptible to overfitting.
Stacking involves partitioning the training set into N blocks. N-1 blocks are used to train a set of base models, while the Nth block, paired with outputs from the base classifiers (referred to as meta-features), trains another model. One limitation of the traditional stacking approach is the disparity between the meta-features in the training sample and the actual responses from specific regressors. In classical stacking, non-overlapping unique values may exist between training and testing meta-features. Our developed stacking model seeks to address this by deforming meta-features based on pairwise multiplication outcomes. These transformed features, combined with the training dataset, feed into a meta-model, mitigating weak predictor result correlations and enhancing model generalization.
Regression analysis utilized three condition indicators: RMS, Kurtosis, and E.I. Performance of each regression model was assessed using RMSE, R^2, and adjusted-R^2. Among the models tested, SVR, polynomial regression, and a single-hidden layer ANN with 12 neurons emerged as the superior weak predictors. Conversely, the tuned KNN model underperformed, as shown in Table 1.
The results derived from our proposed ANN model showcase its capability to predict the
remaining useful life of a bearing, a feat attributed to the synergy between the regression and
ANN models through the optimal condition indicator. The DIAFS model also exhibits
promising results, boasting the highest R-Squared and Adjusted R-Squared values, suggesting
its potential in forecasting novel observations. Although the ANN model provides accurate
predictions and is corroborated by other research [
In the implemented system, forecasts are pre-calculated using a pre-trained ML model for incoming data and stored in a dedicated database. Subsequent input requests access this database for predictions. Architecturally, this mirrors the Lambda pattern, blending "hot" realtime data processing with "cold" historical data processing, typically residing in a Data Lake on Apache Hadoop (Fig. 4).
The Lambda Architecture encompasses both a conventional batch data pipeline for faststreaming real-time data and a serving layer designated for query responses. Ingressed batch data populates a batch layer, prepping it for indexing. Several pre-trained ML models then analyze this data concurrently. The models listed in Table 1 are incorporated. The serving layer facilitates the pre-result voting process, selecting the best model or a suitable combination for the given data. The selected model subsequently processes incoming stream data.
This research introduces a predictive model rooted in MLOps methodology to determine wind turbine bearings' Remaining Useful Life (RUL). The model efficiently detects degradation patterns in real time, adjusting its parameters in response to new data, and offers a fully automated configuration, enabling its deployment across multiple wind turbines. As such, it emerges as an indispensable tool for condition-based maintenance.
Our ensemble stacking model, incorporating regression, SVR, and random forest techniques, has demonstrated commendable generalization capabilities. The model's effectiveness is underscored by its proficiency in leveraging condition indicators, which were appraised using criteria like monotonicity and trendability.
The incorporation of MLOps in this research has facilitated the following: Enhanced innovation through holistic machine learning lifecycle management; Reproducible and robust model iterations tailored for enterprise settings; Efficient tracking using advanced dataset and model registries; Improved traceability and accountability through detailed logging; Streamlined model workflows ensuring consistent delivery; Generation of unbiased models emphasizing feature importance, assessed using uniform distribution metrics.
Our research further proposes new model for monitoring wind turbine bearing conditions. DIAFS's adaptability, anchored in an adaptive polynomial model, facilitates real-time refinement, promising accurate and timely predictions. This algorithm not only refines the precision of forecasting but also augments maintenance strategies. Through DIAFS, potential issues can be preemptively identified, optimizing resource allocation and reducing operational downtimes. Such a data-driven approach, focusing on empirical evidence, leads to significant cost savings, increasing wind turbines' overall efficiency and lifespan. Furthermore, the scalability inherent to DIAFS ensures its applicability within the expanding world of wind energy.
While the primary application of the ensemble model is in wind turbines, its adaptable architecture ensures relevance across varied industrial maintenance scenarios, such as managing Rolling Element Bearings (REBs) failures. Beyond the technical abilities mentioned above as a digital transformation driver, this study recognizes the pivotal role of workforce adaptation to these evolving digital tools.
Lastly, future research will focus on adapting the developed models to analyze raw data in real time. Additionally, it would be interesting to extend the MLOps pipeline to facilitate realtime monitoring of wind turbines, thereby enabling faster fault detection and the potential implementation of predictive maintenance strategies.