A Hybrid Metaheuristic and Deep Learning Framework for Robust Predictive Maintenance in Wind Energy Systems

The world is moving toward more sustainable energy systems, and wind power is now a big part of the energy mix of the future. Wind energy has more than 900 GW of capacity around the world as of 2023. This shows how important it is to make energy [1]. Wind energy is becoming more popular because it has a low levelized cost of energy (LCOE). The LCOE is made up of a lot of Operations and Maintenance (O&M) costs, which make up 20–35% of the total lifecycle cost of a wind turbine [2]. Wind turbines work in tough and often hard-to-reach places, which makes the costs go up because the parts wear out faster and there are more unexpected outages. The gearbox, generator bearings, and blades are some of the most important subsystems that can fail in a big way. The repair or replacement that comes next can cause downtime that lasts for weeks and costs hundreds of thousands of dollars per incident [3].

The transition from reactive and preventive maintenance strategies to condition-based and predictive maintenance (PdM) signifies a significant change in asset management practices. PdM uses data analytics and machine learning to figure out how equipment will break down over time, which makes it easier to schedule maintenance at the best times—before the equipment breaks down, but not too often [4]. Modern wind turbines have full Supervisory Control and Data Acquisition (SCADA) systems that record a lot of information at regular intervals, like temperature, vibration, rotational speed, and power output. The data-rich environment is the best place to put data-driven PdM models into action [5].

Despite the availability of data, significant challenges remain. The SCADA data is characterized by high dimensionality, noise, and strong multicollinearity, which complicates the identification of the most prognostically significant features [6]. Second, the time patterns that show the start of a fault are often subtle and hidden by normal operational changes, so we need models that can accurately capture long-range dependencies [7]. The class imbalance problem, which is when there is a lot more normal operation data than fault events, makes it very hard for learning algorithms to work [8].

Recent research has examined various AI methodologies to address these challenges. Standard machine learning models, including Support Vector Machines (SVMs) and Random Forests, have shown moderate efficacy in practical applications [9]. Recent enhancements in deep learning models, such as Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs), have augmented their capability to identify patterns in time-series data occurring across temporal and spatial dimensions [10, 11]. Numerous industrial applications, such as detecting plant diseases [12], show that CNNs have successfully moved from image recognition to analyzing data from multiple sensors.

How well deep learning models work depends a lot on how you tune the hyperparameters and choose the features. It is possible to solve this problem with metaheuristic optimization. Swarm intelligence algorithms, like the Artificial Bee Colony (ABC), are very good at finding near-optimal solutions to NP-hard problems, like scheduling jobs in a job shop [13-16]. These problems have complex, high-dimensional search spaces. The main advantage of these methods is that they can find a good balance between diversification, which means looking into new parts of the search space, and intensification, which means making the most of known good areas [16]. This principle is relevant to the difficulties of feature selection and hyperparameter optimization in PdM.

The architectural framework for data handling is essential. The integration of IoT, edge, and cloud computing, as illustrated in precision livestock farming [17] and smart agriculture [18, 19], offers a comprehensive framework for managing the data pipeline in wind turbine PdM. Edge devices facilitate preliminary data filtering and real-time anomaly detection, whereas the cloud offers the computational resources necessary for training complex, metaheuristic-optimized prognostic models. This study presents several significant contributions:

1. We propose a novel end-to-end PdM framework that integrates a multi-hybridized ABC algorithm for feature selection and model optimization with a deep hybrid convolutional neural network-LSTM network for fault prognostics.
2. We use the ideas of "multi-hybridization" from scheduling problems [16] to make a better and stronger optimizer for the PdM domain.
3. We suggest a scalable IoT-edge-cloud data architecture that is specifically made for wind farm operations. This will make it easier to process data and deploy models.
4. We show a thorough empirical evaluation using a real-world SCADA dataset that shows how our proposed framework works better than a number of well-known benchmarks.

This paper is structured as follows: Section 2 examines the pertinent literature, providing a review of existing studies and establishing the research context. Section 3 outlines the proposed methodology and framework in detail, explaining the core approach adopted. Section 4 outlines the experimental setup and the dataset utilized, describing the implementation conditions and data sources. Section 5 delineates and analyzes the results, presenting and discussing the key findings. Section 6 concludes the paper by summarizing the primary contributions and proposes directions for future research.

Figure 1 shows the main structure of the proposed PdM framework. The system comprises four main layers: the Data Acquisition Layer, the Edge Processing Layer, the Cloud Analytics Layer, and the Decision Support Layer.

Figure 1. Proposed IoT-edge-cloud framework for wind turbine PdM

3.1 Data acquisition and preprocessing

The raw SCADA dataset comprises continuous sensor readings collected at 10-minute intervals from multiple wind turbines. To ensure data quality and prepare it for subsequent analysis, a comprehensive preprocessing pipeline was implemented. Each step was carefully designed to preserve fault signatures while addressing common data quality issues in industrial time-series.

1) Missing Value Imputation: Initial analysis revealed approximately 2.3% missing values distributed non-randomly across the dataset. Samples with excessive missing data (> 50% missing features) were deemed unreliable and discarded (affecting < 0.5% of total samples). For the remaining gaps, a k-Nearest Neighbors (k-NN) imputation with k = 5 and Euclidean distance was employed.

2) This method was selected over simpler approaches (mean/median imputation) for its ability to capture local correlations between features in multivariate time-series data [31]. The parameter k = 5 was determined through cross-validation on a validation subset, balancing computational efficiency with imputation accuracy.

3) Signal Denoising: SCADA measurements often contain high-frequency noise from electrical interference and sensor limitations. To enhance signal-to-noise ratio while preserving important fault indicators, a Savitzky-Golay (SG) filter with window length = 15 samples (2.5 hours) and polynomial order = 3 was applied independently to each sensor channel. This configuration was optimized to smooth random fluctuations while maintaining the integrity of rapid changes indicative of incipient faults [32]. The SG filter's advantage over simple moving averages lies in its ability to preserve higher-order moments of peak shapes critical for fault detection.

4) Feature Normalization: Given the diverse measurement units and ranges across SCADA parameters (temperatures in ℃, pressures in kPa, vibrations in mm/s², etc.), Min-Max normalization was applied to transform all features to a [0, 1] range: x_norm = (x−min(x)) / (max(x)−min(x)), where x represents a feature vector. This scaling ensures equal contribution of all features during model training and accelerates convergence of gradient-based optimization algorithms. The normalization parameters (min and max values) were computed only from the training set and subsequently applied to the test set to prevent data leakage.

The preprocessed dataset, denoted as Dpreprocessed, served as input for the subsequent feature selection and modeling stages.

3.2 Multi-hybridized ABC for feature selection (MH-ABC-FS)

This study introduces a new feature selection method grounded in the principles of multi-hybridization as referenced in the study [16]. The standard ABC algorithm is composed of employed bees, onlooker bees, and scout bees. We enhance it by hybridizing it with two mechanisms: (1) a Simulated Annealing (SA) inspired acceptance criterion for onlooker bees to improve intensification, and (2) a Differential Evolution (DE) mutation strategy for scout bees to enhance diversification.

The term "multi-hybridization" in our work refers to the strategic integration of two distinct metaheuristic mechanisms—Simulated Annealing (SA) and Differential Evolution (DE)—into the core ABC framework. This design addresses fundamental limitations of standard ABC in complex optimization landscapes like feature selection for PdM. Standard ABC can suffer from slow convergence (insufficient intensification) or premature convergence to local optima (insufficient diversification) [28]. Simple hybrid variants like ABC-PSO typically integrate only one additional mechanism.

Our MH-ABC uniquely combines:

1) SA-inspired acceptance in the onlooker bee phase to enhance intensification: By occasionally accepting worse solutions based on a probabilistic criterion, the algorithm can escape local optima while maintaining directed search capability.

2) DE mutation in the scout bee phase to enhance diversification: Instead of random re-initialization, DE's vector-based mutation provides guided exploration of new regions in the search space.

This balanced approach draws directly from successful applications in job shop scheduling [16], where similar search space complexity exists. Table 2 compares our MH-ABC against standard ABC and ABC-PSO hybrids, highlighting the theoretical advantages of our multi-hybrid strategy.

Table 2. Comparison of ABC variants for feature selection

Having established the theoretical rationale for multi-hybridization, we now formalize the specific adaptation of the MH-ABC framework to the feature selection (FS) task. The MH-ABC-FS algorithm transforms the combinatorial search for an optimal feature subset into an optimization problem solvable by the bee colony metaphor. Each component of the standard ABC—employed bees, onlooker bees, and scout bees—is redefined with hybrid enhancements to navigate the binary search space efficiently. The algorithm proceeds through the following five core components, which collectively balance the exploration of new feature combinations with the exploitation of high-performing subsets.

1) Solution Representation: A solution (food source) is represented as a binary vector F=[f₁,f₂,...,f_n], where f_i=1 indicates that the i-th feature is selected and 0 otherwise.

2) Fitness Function: The fitness of a solution F is evaluated using a simplified, fast-training proxy model (a small Random Forest) on the training data using only the selected features. The fitness is defined as: Fitness(F)=α*Accuracy_proxy+(1−α)*1−|F|/n), where Accuracy_proxy is the classification accuracy of the proxy model, |F| is the number of selected features, and α is a weighting parameter set to 0.8 to prioritize accuracy.

The choice of a small Random Forest (RF) as the proxy model for fitness evaluation is motivated by several key considerations pertinent to wrapper-based feature selection. First, RF's inherent robustness to varying feature scales and non-linear relationships makes it well-suited for heterogeneous SCADA data without necessitating extensive normalization or linearity assumptions, unlike simpler models such as Logistic Regression [9]. Second, to maintain computational efficiency within the iterative metaheuristic search, the proxy RF is configured as a "small" ensemble—typically comprising 10 to 50 trees with constrained depth. This configuration offers a favorable trade-off: it provides a sufficiently reliable and stable performance estimate to guide the search, while being computationally frugal and less prone to overfitting on the small, dynamically evolving feature subsets evaluated at each iteration—a risk associated with using the final, more complex CNN-LSTM model as the evaluator. This practice aligns with established methodologies in wrapper-based feature selection, where tree-based models are frequently employed as evaluators due to their robustness and built-in mechanisms for assessing feature importance [26, 27]. It is acknowledged that the proxy model influences the search landscape; however, the final prognostic performance is determined by the comprehensive CNN-LSTM model trained on the selected feature subset. The effectiveness of the overall approach, and by extension the suitability of the proxy-guided search, is validated by the ablation study (see Section 5.3), which demonstrates a significant performance drop when the MH-ABC-FS component is removed.

3) Hybrid Employed Bee Phase: Employed bees modify their current solution F_i to produce a new candidate solution V_i by flipping bits based on a neighborhood search. If Fitness (V_i) > Fitness (F_i), it replaces F_i.

4) Hybrid Onlooker Bee Phase: Onlooker bees select a solution F_i with a probability proportional to its fitness. They then perform the same neighborhood search. However, we introduce an SA-like criterion: even if V_i is worse, it is accepted with a probability p=exp(−ΔFitness/T), where T is a temperature parameter that decreases over iterations. This helps escape local optima.

5) Hybrid Scout Bee Phase: If a solution cannot be improved after a predetermined limit, it is abandoned. Instead of generating a completely random solution, the scout bee applies a DE/rand/1 mutation: F_new = F_r1 + β*(F_r2−F_r3), where r₁, r₂, r₃ are distinct random indices and β is a scaling factor. The result is then binarized. This introduces a more guided form of diversification.

The output of the MH-ABC-FS process is an optimal feature subset F_opt, which is used for all subsequent modeling.

The key parameters for the proposed MH-ABC-FS (Modified Hybrid ABC - Feature Selection) algorithm, along with their descriptions and the values used in our experiments, are defined in Table 3.

Table 3. MH-ABC-FS algorithm parameters

3.3 Hybrid CNN-LSTM prognostic model

The selected features F_opt are fed into a hybrid CNN-LSTM network for fault classification. The model is designed to automatically learn both spatial correlations among different sensors and the temporal evolution leading to a fault.

1) Input Layer: Accepts a time-series window of length L (e.g., 100 time steps = 16.7 hours) comprising the selected features.

2) The CNN Feature Extraction Block comprises two 1D convolutional layers utilizing ReLU activation, aimed at extracting local temporal patterns and inter-sensor relationships. The initial layer comprises 64 filters with a kernel size of 3, while the subsequent layer contains 128 filters, also with a kernel size of 3. A 1D Max-Pooling layer of size 2 follows each layer to reduce dimensionality and enhance translational invariance.

3) The feature maps from the CNN block are flattened and reshaped into a sequence, subsequently input into a two-layer stacked LSTM network. The initial LSTM layer comprises 100 units and is configured to return sequences, whereas the subsequent layer consists of 50 units. This structure effectively captures long-term temporal dependencies in the extracted features.

4) Output Layer: The LSTM layer's final output is processed through a Dense layer utilizing a softmax activation function, generating a probability distribution across the target classes (e.g., 'Normal', 'Warning', 'Critical').

3.4 MH-ABC for hyperparameter tuning

The hyperparameters of the CNN-LSTM model have a big effect on how well it works. We employ a separate instance of the MH-ABC algorithm (as described in Section 3.2, but with a continuous solution space) to optimize the following hyperparameters: Learning rate, CNN kernel size, number of LSTM units, and dropout rate. The fitness for this optimization is the final validation accuracy of the fully trained CNN-LSTM model.

3.5 Implementation within the IoT-Edge-Cloud architecture

To balance real-time responsiveness with deep computational analysis, the system is deployed across a three-tier IoT-Edge-Cloud architecture. This structure ensures immediate local action while leveraging the cloud's power for complex, long-term prognostics.

1) Edge Layer: This is on the controller of the turbine or a local gateway. It preprocesses data in real time and runs a lightweight version of the model (like a simple anomaly detector) to send alerts right away.

2) Cloud Layer: This is where the historical database and the full MH-ABC-FS and optimized Hybrid CNN-LSTM model are stored. Here, feature selection, model (re)training, and long-term prognostics are all done, which takes a lot of resources.

3) Decision Support Layer: A web-based dashboard shows the wind farm operators the prognostic results, RUL estimates, and maintenance suggestions.

5.1 Feature selection results

The proposed MH-ABC-FS algorithm selected 28 out of the original 127 features as most relevant for predicting bearing and gearbox failures. The convergence curve of the MH-ABC-FS, compared to standard ABC and a PSO-based FS, is shown in Figure 2. MH-ABC-FS achieved a higher fitness value and converged more smoothly, demonstrating the benefit of its hybrid mechanisms.

The MH-ABC-FS algorithm identified a compact set of 10 most predictive features from the original 127. The top 10 selected features, ranked by the algorithm's importance metric, are presented in Table 5.

The features selected by MH-ABC-FS (Table 4) align closely with domain knowledge of wind turbine failure mechanisms. The prominence of temperature features (Gbox_Bear_Temp_Avg, Gen_Bear_Temp_Avg, Gbox_Oil_Temp_Avg) reflects the well-documented thermal stress preceding bearing and gearbox failures [3]. Vibration measurements (Nac_Vib_X/Y_Avg) are critical for detecting mechanical imbalances and component wear. Operational parameters (Rtr_RPM_Avg, Wind_Speed_Avg, Prod_Latest_Avg) provide necessary context for distinguishing fault signatures from normal operational variations. This alignment with physical failure models enhances the interpretability and practical credibility of our feature selection approach.

Figure 2. Convergence curves of feature selection algorithms

Table 5. Top 10 selected features by MH-ABC-FS

5.2 Prognostic performance comparison

The overall performance of all models on the test set is summarized in Table 6. The proposed MH-ABC-CNN-LSTM model consistently outperformed all benchmarks across all metrics.

Table 6. Performance comparison of different models on the test set

The confusion matrices for the standard CNN-LSTM and our proposed model are shown in Figure 3. The proposed model significantly reduced the number of false negatives (missed faults), which is the most critical type of error in PdM.

Figure 3. Confusion matrices for standard CNN-LSTM and proposed MH-ABC-CNN-LSTM

5.3 Ablation study

To dissect the contribution of each component in our framework, we conducted an ablation study. Table 7 shows that both the MH-ABC feature selection and the MH-ABC hyperparameter tuning have a big effect on the final performance. Removing either component leads to a noticeable drop in performance, confirming the synergy of the integrated approach.

Table 7. Ablation study results

5.4 Computational cost analysis

Training and optimizing the proposed framework is computationally intensive. All experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 4090 GPU (24 GB VRAM) and an Intel Core i9-13900K CPU, with 64 GB of system memory. The average training time for the different deep learning-based models is shown in Table 8. The proposed MH-ABC-CNN-LSTM requires the longest training duration, which constitutes a one-time or periodic cost acceptable for cloud-based model development and retraining.

Table 8. Computational cost comparison

Crucially, for operational deployment, the inference latency is the key metric. Table 8 also reports the average inference time per sample. The results demonstrate that the inference time of our optimized model is virtually identical to that of the standard CNN-LSTM (differing by less than 2 ms), confirming that the metaheuristic optimization incurs no significant overhead during prediction. This low and stable inference latency makes the model suitable for near-real-time fault prognosis at the edge, where prompt alerts are essential.

5.5 Discussion

The enhanced efficacy of our proposed framework is due to three primary factors:

1. The MH-ABC-FS algorithm, guided by balanced search principles from the study [16], effectively identifies a compact and highly discriminative feature set. This removed noise and redundancy, enabling the subsequent deep learning model to concentrate on the most significant signals.
2. Optimized Model Architecture: The hyperparameter tuning based on MH-ABC identified an optimal configuration for the CNN-LSTM network, specifically designed for the chosen features and the temporal characteristics of the fault data. This is consistent with the findings in study [16] that multi-hybridization results in more precise and robust solutions.
3. Synergistic Hybridization: The integration of CNN for spatial feature extraction and LSTM for temporal modeling, when optimized effectively, demonstrated significant efficacy in representing the intricate degradation processes in wind turbines.

The findings provide robust evidence for the hypothesis that the transfer of advanced optimization methodologies between complex domains, such as scheduling and PdM, can result in significant performance improvements. The architecture of the framework, influenced by IoT systems in agriculture [17, 18], offers a scalable and feasible approach for real-world implementation.

It is important to position the proposed MH-ABC framework within the broader landscape of optimization techniques for deep learning in PdM. While recent studies have explored advanced methods such as gradient-based neural architecture search (NAS) for automated model design and Bayesian optimization for hyperparameter tuning, these approaches often require substantial computational budgets and may struggle with the mixed discrete-continuous, high-dimensional search spaces inherent in joint feature selection and model optimization. In contrast, metaheuristic approaches like MH-ABC offer a flexible and robust alternative, capable of handling complex, non-differentiable objectives without relying on gradient information. The selection of MH-ABC, specifically, was guided by its demonstrated efficacy in achieving a superior exploration-exploitation balance in similarly constrained optimization problems, such as job-shop scheduling [16]. Future research will involve a direct, comprehensive comparison with these state-of-the-art optimizers, including contemporary metaheuristics like the Whale Optimization Algorithm and Transformer-based optimization methods, to further delineate the specific scenarios where each approach excels.