A Generalized Autonomous Power Plant Fault Detection Model Using Deep Feature Extraction and Ensemble Machine Learning

Abstract

Ensuring operational reliability and efficiency in steam power plants requires advanced and generalized fault detection methodologies capable of addressing diverse fault scenarios in boiler and turbine systems. This study presents an autonomous fault detection framework that integrates deep feature extraction through Convolutional Autoencoders (CAEs) with the ensemble machine learning technique, Extreme Gradient Boosting (XGBoost). CAEs autonomously extract meaningful and nonlinear features from raw sensor data, eliminating the need for manual feature engineering. Principal Component Analysis (PCA) is employed for dimensionality reduction, enhancing computational efficiency while retaining critical fault-related information. The refined features are then classified using XGBoost, a robust ensemble learning algorithm, ensuring accurate fault detection. The proposed model is validated through real-world case studies on boiler waterwall tube leakage and motor-driven oil pump failure in steam turbines. Results demonstrate the framework’s ability to generalize across diverse fault types, detect anomalies at an early stage, and minimize operational downtime. This study highlights the transformative potential of combining deep feature extraction and ensemble machine learning for scalable, reliable, and efficient fault detection in power plant operations.

1. Introduction

Steam thermal power plants are essential to global energy production, providing a substantial portion of electricity to meet increasing industrial and domestic demands [1,2,3]. These plants operate by converting thermal energy into mechanical energy to drive steam turbines, which are connected to generators. Despite their critical role, the reliability and efficiency of steam power plants are frequently compromised by operational faults, particularly in boilers and steam turbines [4]. Boiler faults, which account for 52% of all power plant failures, are often caused by waterwall tube leakages due to thermal stress, material degradation, and corrosion [5,6]. Similarly, steam turbines are prone to mechanical faults such as rotor imbalances, gear faults, misalignments, and bearing defects [7,8,9]. A significant fault in turbine systems arises from failures in motor-driven oil pumps, which are critical for providing hydraulic lubrication and maintaining the control oil system. Inadequate oil supply, often caused by the malfunction or failure of the electric motor driving the pump, can lead to catastrophic bearing failures in the turbine’s rotating components [10]. Research indicates that approximately 40% of AC motor failures are caused by rolling bearing defects, emphasizing the importance of diagnosing bearing conditions proactively to prevent such failures [11]. Other prevalent faults in AC motors include winding failures, rotor and stator imbalances, broken rotor bars, and eccentricity-related problems [12].

As shown in Figure 1, the survey presents the fault distribution in steam power plants based on forced outages and severity, highlighting boiler and turbine faults as the major contributors [13]. These faults critically impact operational efficiency, lead to increased maintenance costs, and result in prolonged downtime, emphasizing the need for effective fault detection and diagnostic strategies. The significance of robust fault detection strategies has been highlighted in numerous studies. For instance, Babak et al. [14] reviewed the causes of boiler tube failures in steam power plants and stressed the importance of advanced monitoring systems to mitigate their impact on power generation. Similarly, Huang et al. [15] investigated vibration-based diagnostics for steam turbines, revealing that mechanical faults often go undetected until significant performance degradation occurs. This highlights the need for early detection techniques that can minimize operational disruptions.

Traditional fault detection methods, including manual inspections [16], rule-based monitoring [17], and basic signal analysis [18], have provided a foundation for addressing these challenges. Techniques such as vibration analysis and thermal imaging, while effective, remain reactive, labor-intensive, and limited in their ability to predict faults before significant disruptions occur [19,20]. To overcome these limitations, model-based methods have been introduced, utilizing mathematical models to represent normal system behavior and identify anomalies through deviations [21]. Techniques such as observer-based approaches, parameter estimation, and Kalman filters have demonstrated precision but are hindered by their reliance on accurate system modeling and sensitivity to parameter uncertainties [22,23,24]. Moreover, knowledge-based methods have utilized expert systems and fuzzy logic algorithms to detect anomalies based on predefined knowledge [25,26]. However, these approaches face challenges in adapting to evolving systems and handling unforeseen faults, limiting their scalability.

Recent advances in sensor technologies and data-driven methods have revolutionized fault detection in steam power plants [27]. Modern sensors generate high-fidelity data streams, facilitating the development of sophisticated analytical methods. For example, Min et al. [28] demonstrated the utility of piezoelectric sensors for real-time acoustic emission monitoring in steam boilers, achieving precise fault localization. Similarly, Ukil et al. [29] utilized distributed optical fiber sensors to monitor temperature variations in real time, enabling early detection of thermal anomalies. Data-driven methods have further enhanced fault detection by incorporating advanced statistical and machine learning techniques. Statistical techniques like PCA play a pivotal role in managing high-dimensional data. PCA facilitates dimensionality reduction while retaining essential features critical for anomaly detection. For instance, Jungwon et al. [30] applied PCA to detect plugged tubes in superheater banks during power plant startups, enabling timely decision-making and mitigating critical failures such as tube leakage. Miroslaw et al. [31] extended this approach with multiway PCA (MPCA) to model healthy system behavior in steam boilers, creating a confidence ellipsoid that allowed for early leak detection and improved maintenance efficiency. Ajami et al. [32] further demonstrated the efficacy of using PCA in turbine systems through a PCA-based inverse neural network control strategy for fault-tolerant control.

Machine learning techniques have significantly advanced fault detection by integrating robust preprocessing and classification algorithms. Khalid et al. [33] developed a sensor optimization framework for boiler waterwall tube leakage detection, combining correlation analysis with supervised learning for accurate fault classification. Jaswanth et al. [34] and Liang et al. [35] highlighted the robustness of XGBoost, a gradient boosting algorithm, in accurately classifying faults in boilers and turbines. Additionally, Zijun et al. [36] combined XGBoost with Dynamic Time Warping (DTW) for turbine health prognostics, achieving high reliability in detecting early signs of degradation. Similarly, Zhanhong et al. [37] explored a hybrid model combining XGBoost and genetic algorithms for fault detection in complex thermal systems, demonstrating superior adaptability to evolving operating conditions. Despite their success, many machine learning methods depend on manually engineered features, which may limit adaptability in dynamic industrial environments. Deep learning addresses the limitations of manual feature engineering by enabling autonomous feature extraction directly from raw data [38,39]. Hyeongmin et al. [40] proposed the Optimal Temporal Convolutional Auto-Encoder (Opt-TCAE) for boiler fault detection, capturing inter-sensor and temporal relationships to improve accuracy and reduce false alarms. Zhang et al. [41] combined Robust Long Short Term Memory (LSTM) Autoencoders with 1D CAEs to detect boiler leaks, effectively managing corrupted data and capturing dependency patterns. For turbines, Jinxing et al. [42] utilized CAEs to detect anomalies based on reconstruction errors, while Jose et al. [43] demonstrated the efficacy of CAEs in modeling normal operating conditions in gas turbines without labeled data. Despite their strengths, CAEs are not inherently designed for fault classification tasks [44]. While they excel in feature extraction, their effectiveness relies on the quality of the reconstructed latent representations, which can lead to overfitting or poor generalization when datasets are imbalanced or noisy.

The proposed framework directly addresses the challenges of overfitting and poor generalization through the following key contributions: CAEs are employed to autonomously extract meaningful and nonlinear features directly from raw sensor data, eliminating the reliance on manual feature engineering. Manual approaches often introduce biases that can lead to overfitting, whereas CAEs ensure a systematic and unbiased extraction of fault-relevant patterns. Additionally, PCA is integrated to refine these features by reducing dimensionality, focusing the model’s attention on critical fault-related information while discarding noise and redundancy. This step further mitigates overfitting risks and enhances the robustness of the feature set. Finally, the ensemble learning capabilities of XGBoost significantly enhance the generalization of the framework. By combining multiple decision trees, XGBoost effectively handles the challenges posed by noisy and imbalanced datasets, ensuring reliable and accurate fault detection across diverse scenarios. The proposed method is validated through real-world case studies, including boiler waterwall tube leakage and motor-driven oil pump failure in steam turbines, demonstrating its effectiveness and practicality for industrial fault detection applications.

2. Proposed Autonomous Fault Detection Methodology and Theoretical Foundations

2.1. Description of the Proposed Methodology

The proposed methodology, illustrated in Figure 2, begins with the collection of multi-sensor data from thermal power plants, including temperature data for boiler waterwall tube leakage and vibration data for turbine motor-driven oil pump faults. This dataset encompasses both healthy and faulty operating states, providing a robust foundation for analysis. During preprocessing, the raw data undergo normalization and segmentation to ensure consistency and prepare them for analysis. Autonomous feature extraction is performed using a CAE, which compresses high-dimensional data into a compact and informative feature space while preserving fault-relevant patterns. The extracted features are further refined through PCA for dimensionality reduction, optimizing computational efficiency while retaining critical information. The refined feature set is then classified using XGBoost, a robust ensemble learning algorithm, to accurately distinguish between healthy and faulty system states.

2.2. Theoretical Background of Applied Algorithms

(a)

PCA

PCA is a widely employed statistical technique for dimensionality reduction, feature extraction, and data visualization [45]. By transforming high-dimensional data into a lower-dimensional space, PCA preserves the most significant variance in the dataset, enabling efficient processing and analysis of complex systems. It achieves this by identifying principal components, which represent the orthogonal directions of maximum variance within the data. The PCA process begins with the calculation of the covariance matrix to capture the relationships between features in the dataset. For a mean-centered dataset $X$, the covariance matrix $\Sigma$ is computed as [46]:

$$\Sigma ={\displaystyle \frac{1}{n-1}}{X}^{T}X$$

(1)

Here, $X$ is assumed to be mean-centered, ensuring that all features have zero mean. The covariance matrix $\Sigma$ is symmetric and forms the basis for identifying directions of maximum variance in the data. PCA then derives the eigenvectors and eigenvalues of $\Sigma$, where the eigenvectors represent the principal components, and the eigenvalues quantify the variance explained by each component. The original data $X$ are projected onto the new principal component space, resulting in a transformed dataset $Y$, as defined by:

$$Y=XV$$

(2)

where $Y\in R^{n\times k}$ represents the transformed data in the reduced $k$-dimensional space ($k<p$), and $V\in R^{n\times k}$ is the matrix of the top $k$ eigenvectors corresponding to the largest eigenvalues of $\Sigma$. In steam power plants, PCA is extensively utilized for reducing the dimensionality of sensor data while retaining critical information [30,32]. This dimensionality reduction not only enhances computational efficiency but also filters out irrelevant variations, thereby improving signal quality. PCA’s robustness and ability to efficiently manage high-dimensional data make it an indispensable technique for fault detection and detection in complex industrial systems.

(b)

XGBoost

XGBoost is an advanced ensemble machine learning algorithm rooted in the gradient boosting framework [47]. It is widely acclaimed for its exceptional efficiency, scalability, and accuracy, particularly in structured data applications. XGBoost constructs an ensemble of decision trees by iteratively adding models that correct the errors of previous iterations, optimizing both predictive accuracy and generalization performance. The core of XGBoost lies in its objective function, which balances model predictive performance and complexity. The objective function is defined as [48]:

$$Objective=\sum _{i=1}^{n}L\left(y_i,{\widehat{y}}_i\right)+\sum _{t=1}^T\Omega \left(f_t\right)$$

(3)

where $L\left(y_i,{\widehat{y}}_i\right)$ is the loss function measuring the difference between the true value $y_i$ and the predicted value ${\widehat{y}}_i$. $\Omega \left(f_t\right)$ is the regularization term that penalizes model complexity. XGBoost employs gradient descent to minimize the objective function. At each iteration $t$, it fits a new decision tree $f_t\left(x\right)$ to the negative gradients (residuals) of the loss function from the previous iteration:

$$g_i={\displaystyle \frac{\partial L\left(y_i,{{\widehat{y}}_i}^{t-1}\right)}{\left({{\widehat{y}}_i}^{t-1}\right)}}$$

(4)

The new predictions are then updated as:

$${{\widehat{y}}_i}^t={{\widehat{y}}_i}^{t-1}+\eta f_t\left(x_i\right)$$

(5)

where $\eta$ is the learning rate, contributing each new tree to the overall prediction. In industrial applications, particularly in steam power plants, XGBoost has proven to be highly effective in fault detection and classification, owing to its ability to manage high-dimensional data and capture complex relationships [35,36]. Its capacity to identify subtle deviations from normal behavior makes it a powerful tool for early fault detection, enabling timely interventions and minimizing operational disruptions. When integrated with dimensionality reduction techniques such as PCA, XGBoost processes reduced feature sets with remarkable computational efficiency, maintaining high accuracy without overburdening resources.

(c)

SVM

SVM is a versatile algorithm for classification and regression tasks, capable of processing both linear and nonlinear data by utilizing kernel functions [49,50]. A linear kernel (LK), appropriate for linearly separable data, calculates the similarity between data points through the dot product of feature vectors:

$$K\left(x,x^{\prime }\right)=\mathrm{x}·x^{\prime }$$

(6)

where $x\mathrm{a}\mathrm{n}\mathrm{d}x\prime$ are feature vectors. For data requiring nonlinear decision boundaries, the polynomial kernel (PK) transforms data into higher-dimensional spaces, enabling the algorithm to capture complex relationships:

$$K\left(x,x^{\prime }\right)={\left(\mathrm{x}·x^{\prime }+c\right)}^d$$

(7)

Here, $c$ is a constant controlling the trade-off between high-order and low-order terms, and $d$ is the degree of the polynomial. The radial basis function Kernel (RK) is well suited for highly nonlinear data. It maps data into infinite-dimensional space, capturing intricate patterns with the equation:

$$K\left(x,x^{\prime }\right)=\mathrm{e}\mathrm{x}\mathrm{p}\left(-\gamma {‖x-x^{\prime }‖}^2\right)$$

(8)

In this equation, $\gamma$ is a hyperparameter controlling the influence of individual data points, and ${‖x-x^{\prime }‖}^2$ represents the squared Euclidean distance between two feature vectors. SVM’s adaptability through kernel functions makes it versatile for various classification challenges. By scaling features and employing dimensionality reduction techniques, SVM achieves high performance even in complex datasets. In fault detection scenarios, such as in boiler and turbine systems, SVM effectively identifies patterns indicative of faults, contributing to enhanced system reliability and minimized downtime [51]. Its robust decision-making capabilities are particularly advantageous in industrial environments where precise fault classification is critical.

(d)

Artificial neural networks (ANNs)

ANNs are computational models inspired by biological neural networks, designed to approximate complex nonlinear relationships in data [52]. They consist of layers of interconnected neurons, where each neuron applies a weighted sum of inputs followed by an activation function. ANNs learn by minimizing a loss function through backpropagation and adjusting weights and biases iteratively to improve accuracy. The output of a single-layer ANN can be expressed as:

$$y=f\left(\sum _{i=1}^n{w}_i{x}_i+b\right)$$

(9)

where $y$ is the output $w_i{x}_i$ are weights, $x_i$ are inputs, $b$ is the bias, and $f$ is the activation function. In industrial applications, such as fault detection in power plants, ANNs are highly effective due to their adaptability and ability to process large-scale, high-dimensional data [53]. They excel in tasks requiring nonlinear decision boundaries, such as anomaly detection and classification. By utilizing their flexibility and scalability, ANNs provide robust solutions for identifying faults in complex systems, ensuring operational reliability and efficiency.

(e)

CAE

CAEs are a class of deep learning models specifically designed for unsupervised feature extraction and dimensionality reduction [54]. CAEs combine the principles of CNNs and traditional autoencoders, utilizing spatial hierarchies in data to extract meaningful features. Their architecture is particularly well suited for processing high-dimensional structured data, such as images, time series, and acoustic signals, where preserving spatial or temporal relationships is crucial. A CAE typically consists of two main components: an encoder and a decoder [55]. The encoder compresses the input data $X$ into a lower-dimensional latent representation $Z$. It employs convolutional layers to capture local patterns and hierarchies within the data. The encoder’s function is mathematically represented as:

$$Z=f_{encoder}\left(X\right)$$

(10)

where $Z$ represents the latent representation capturing essential features of the input data. The decoder reconstructs the input $X$ from the latent representation $Z$, using transposed convolutional layers (deconvolutions) to restore the data to its original shape. The decoder’s function is given by:

$$\widehat{X}=f_{encoder}\left(Z\right)$$

(11)

The objective of the CAE is to minimize the reconstruction error, which quantifies the difference between the original input $X$ and reconstructed output $\widehat{X}$. This reconstruction error is expressed as:

$$\zeta ={‖X-\widehat{X}‖}^2$$

(12)

In industrial fault detection, CAEs are widely applied to analyze sensor data, acoustic emissions, and time-series signals [40,42]. By learning meaningful representations directly from raw data, CAEs are particularly effective for tasks such as anomaly detection, unsupervised learning, and feature extraction for classification. However, CAEs are not inherently optimized for fault classification, as their performance heavily depends on the quality of reconstructed latent representations. This limitation can lead to overfitting or poor generalization, particularly when datasets are noisy or imbalanced. To address the complexities of fault detection in boiler and turbine systems, this study presents an integrated framework combining CAEs, PCA, and XGBoost. The CAE architecture implemented in this study, as shown in Figure 3, comprises three encoding and three decoding layers, all using ReLU activation functions. This comprehensive approach achieves reliable and efficient fault detection, addressing critical challenges in maintaining the operational reliability of steam power plant systems.

3. Implementation of the Proposed Model on Real-World Steam Power Plant Data

This section presents two real-world case studies, steam turbine fault detection and boiler waterwall tube leakage detection, to validate the proposed autonomous fault detection model. These critical fault scenarios in steam power plants demonstrate the model’s ability to generalize across diverse fault types.

3.1. Data Acquisition

Accurate fault detection in power plants relies heavily on the acquisition of high-quality sensor data that reflects real-world operating conditions. For this study, critical data were obtained from steam power plant systems, focusing on two major fault scenarios: steam turbine motor-driven oil pump failure and boiler waterwall tube leakage.

3.1.1. Steam Turbine Motor-Driven Oil Pump Fault

For this case study, data were collected to analyze the performance and failures of the turbine’s motor-driven oil pump, a critical component of the steam turbine lubrication and control system. This oil pump ensures a continuous supply of hydraulic lubrication and control oil to the turbine’s rotating components, and its failure can result in insufficient oil supply, leading to severe bearing damage and potential turbine failure. To capture relevant fault patterns, three critical sensors were selected by power plant experts to monitor bearing vibration in the horizontal direction. The dataset consists of 8 days of healthy operational data and 8 days of faulty data, recorded at a sampling rate of 1 sample per minute to effectively monitor fault progression over time. As shown in Figure 4, the faulty data exhibit higher fluctuations, indicating significant anomalies in the system during fault conditions.

3.1.2. Boiler Waterwall Tube Leakage

For this case study, three critical sensors were selected by power plant experts to capture the most relevant steam temperature data for analysis. These sensors measured temperatures at the outlets of Superheater I (SH-I), Superheater II (SH-II), and Reheater I (RH-I). The dataset included 17.5 days of healthy operational data and 17.5 days of leakage data, collected at a sampling rate of one sample per second, resulting in a comprehensive dataset for fault detection and analysis. As shown in Figure 5, the healthy data are represented by the green curve, while the leakage data are depicted by the red curve. The leakage data exhibit noticeable fluctuations compared to the relatively stable patterns observed in the healthy data.

3.2. Data Preprocessing

To ensure consistency across the datasets and eliminate biases caused by varying scales, all sensor readings were normalized using Min-Max scaling [56]. This technique transforms the raw data into a standardized range of [0, 1], retaining the relative differences between the values and ensuring that no feature dominates the analysis due to its magnitude. Min–Max scaling was chosen for its simplicity and effectiveness in preserving the integrity of the data while standardizing it for efficient processing [57]. Normalization was performed for both healthy and faulty data collected from all sensors, enabling the models to better identify anomalies and patterns related to fault detection. This preprocessing step ensures that the fault classification process remains robust and reliable by enhancing the model’s ability to extract meaningful features. The normalized data for one sensor is depicted in Figure 6, which illustrates the transformed healthy and faulty states, demonstrating the effectiveness of this approach in preparing data for analysis.

3.3. Model Development and Evaluation

The model development and evaluation process, as illustrated in Figure 7, begins with the collection of raw sensor data from power plant systems, including both temperature and vibration signals to represent healthy and faulty states. The data undergo preprocessing, including cleaning and normalization, to ensure consistency and reliability. The pre-processed data are then randomly split into training (70%), validation (15%), and testing (15%) subsets. During the training phase, three distinct approaches were employed: the ANN, CAE-SVM and CAE-XGBoost models. PCA was utilized for dimensionality reduction, enhancing computational efficiency while retaining essential fault-related features. Once the models achieved convergence, the best-performing ones were saved and evaluated on the unseen testing dataset, ensuring a reliable assessment of their fault detection accuracy and generalization capabilities. To evaluate the proposed model’s robustness and accuracy, multiple performance metrics were employed. Classification accuracy served as the primary metric, indicating the proportion of correctly classified instances. To ensure a more comprehensive assessment, precision, recall, and F1-score were also analyzed. The equations for these metrics are provided in Reference [58]. Additionally, the confusion matrix provided detailed insights into the model’s strengths and areas for improvement by visualizing classification results.

4. Case Studies

This section presents the computational results obtained from applying the proposed fault detection framework to two case studies: steam turbine motor-driven oil pump fault and boiler waterwall tube leakage.

4.1. Turbine Fault Detection

The results of the turbine motor-driven oil pump fault detection, based on autonomous features extracted using the CAE, are presented in this section. A comparison is made for the traditional ANN-based deep learning model with the proposed hybrid models with CAE as a deep feature extractor integrated with machine learning classifiers for fault detection. The evaluation metrics in terms of training, validation, and testing accuracy are shown in Figure 8. The ANN model achieved a high training accuracy of 98.88%; however, its performance on the validation and testing datasets dropped to 93.10%. This indicates that while the ANN could learn patterns in the training data effectively, it struggled to generalize to unseen data, potentially due to overfitting or insufficient capacity to capture more complex fault-related features. The CAE-SVM models displayed varied performance depending on the kernel function utilized. The linear kernel (CAE-SVM-LK) demonstrated results similar to the ANN model, achieving 98.88% accuracy during training and maintaining 93.10% accuracy for validation and testing datasets. In comparison, the polynomial kernel (CAE-SVM-PK) exhibited slightly reduced training accuracy at 98.51%, though it achieved a testing accuracy of 95.55%, reflecting its ability to model more complex relationships. This consistency suggests that the polynomial kernel effectively captured the fault patterns in the turbine extracted by the CAE model without overfitting. The radial kernel (CAE-SVM-RK) performed well during training, achieving 99.25% accuracy, but validation and testing accuracy remained at 93.10%. Among the evaluated models, CAE-XGBoost delivered the most robust performance, achieving a perfect training accuracy of 100% and maintaining high validation and testing accuracies of 96.55%. The ensemble nature of XGBoost, combined with the ability of CAE to capture intricate relationships in the data, enabled it to outperform other models in detecting turbine faults.

Table 1. The computational resources required by each model in terms of computational time and model size.

Model	Training Time (ms)	Testing Time (ms)	Model Size (Mbs)
ANN	1973.44	179.04	0.06
CAE-SVM-LK	2807.62	237.05	1.59
CAE-SVM-PK	2807.62	238.05	1.60
CAE-SVM-RK	2806.62	239.05	1.59
CAE-XGBoost	2823.62	239.05	1.67

presents the computational resource requirements for each model, measured in terms of training time, testing time, and model size. The hybrid models exhibit slightly higher computational times compared to the traditional ANN model; however, this increase is minimal and is well-justified by the significant reduction in overfitting and improvement in overall improvement in turbine fault detection performance. Moreover, the testing times for all hybrid models remain in the millisecond range; thus, the data acquisition rate of one sample per minute demonstrates the suitability of hybrid models for real-time monitoring. Although the hybrid models exhibit larger model sizes (approximately 1.6 MB), this remains negligible in the context of the modern big data era. These results underscore the effectiveness of combining CAE for autonomous feature extraction with a highly scalable and precise classifier like XGBoost.

The confusion matrices presented in Figure 9 illustrate the performance of the developed models on the unseen test dataset, emphasizing their capability to classify turbine faults accurately. The ANN model demonstrated highly unreliable fault classification, achieving 86.21% accuracy in identifying a healthy state and 100% accuracy in identifying a faulty case. Therefore, a 13.79% misclassification rate for healthy states limits the generalization ability of the ANN model. The CAE-SVM models, particularly using linear (LK) and polynomial (PK) kernels, improved the accuracy for healthy state to 93.10%, with a minor misclassification rate of 6.90%, showcasing their enhanced ability to model turbine fault features extracted by the CAE. The radial kernel (RK), however, showed identical performance to the ANN model with 86.21% accuracy for a healthy state. This highlights the kernel-specific limitations in capturing fault patterns through the SVM model. Moreover, the CAE-XGBoost achieved good performance and was consistent with CAE-SVM-LK and CAE-SVM-PK. The CAE-XGBoost model showed 93.10% accuracy for identifying a healthy state and maintained 100% accuracy for identifying the fault in the power plant turbine. This underscores the robustness and precision of the hybrid models by utilizing the CAE-extracted features for turbine fault classification.

Table 2. Performance evaluation of ANN, CAE-SVM and CAE-XGBoost models for turbine fault detection.

Model	Health State	Accuracy (%)	Precision (%)	Recall (%)	F-1 Score (%)
ANN	Healthy	86.21	100.00	86.10	92.59
	Faulty	100.00	87.88	100.00	93.55
CAE-SVM-LK	Healthy	93.10	100.00	93.10	96.43
	Faulty	100.00	93.55	100.00	96.67
CAE-SVM-PK	Healthy	93.10	100.00	93.10	96.43
	Faulty	100.00	93.55	100.00	96.67
CAE-SVM-RK	Healthy	86.21	100.00	86.10	92.59
	Faulty	100.00	87.88	100.00	93.55
CAE-XGBoost	Healthy	93.10	100.00	93.10	96.43
	Faulty	100.00	93.55	100.00	96.67

presents the detailed performance metrics of the ANN, CAE-SVM-based hybrid models, and the CAE-XGBoost model for turbine fault detection, evaluated in terms of accuracy, precision, recall, and F1-score. The ANN model achieved moderate accuracy for the healthy state (86.21%) and perfect accuracy for the faulty state (100%), but its precision and recall values indicate limited generalization. The ANN model also showcased large differences of 13~14% between the two health states in terms of all metrics, thus revealing an overall F1-score of 92.59% for the healthy state and 93.55% for the faulty state. The CAE-SVM models with linear and polynomial kernels demonstrated superior performance, achieving 93.10% accuracy for the healthy state and 100% for the faulty state, with consistently high F1-scores of 96.43% for the healthy state and 96.67% for the faulty state. The radial kernel again showed identical performance to ANN, while CAE-XGBoost showed identical performance to the CAE-SVM-LK and CAE-SVM-PK models. Thus, the detailed evaluation revealed that the hybrid models perform better compared to the traditional ANN model, and the kernel function in SVM plays a significant role in classifying the autonomous features extracted by the CAE model.

Table 3. Comparison of the proposed approach with existing popular methods for turbine fault detection.

Model	Accuracy (%)
SVM [59]	88.10
KNN [59]	86.80
Naive Bayes [59]	93.00
ANN	93.10
CAE-SVM-LK	96.55
CAE-SVM-PK	95.55
CAE-SVM-RK	93.10
CAE-XGBoost	96.55

provides a comparison of the proposed approach with existing popular methods for turbine fault detection, highlighting the superiority of the hybrid models. The CAE-SVM-LK and CAE-XGBoost models achieved the highest accuracy (96.55%), outperforming traditional methods such as SVM (88.10%), KNN (86.80%), and Naive Bayes (93.00%), as well as the standalone ANN (93.10%). This demonstrates the effectiveness of integrating autonomous feature extraction with advanced classifiers for enhanced fault detection performance.

4.2. Boiler Fault Detection

The ability of the three best models for turbine fault detection to generalize across different power plant components has been evaluated in this section. For this purpose, the same three hybrid models have been used for boiler waterwall tube leakage detection. Figure 10 illustrates the training, validation, and testing accuracies of these models for boiler leakage detection. The results highlight the generalization capability of the hybrid models when applied to a different fault detection scenario within the same power plant. Among the models, CAE-XGBoost exhibited the highest generalization ability, achieving 96.80% training accuracy, 94.93% validation accuracy, and 93.20% testing accuracy. Meanwhile, the CAE-SVM-LK showed relatively robust performance, maintaining 93.34%, 93.73%, and 92.00% for training, validation, and testing accuracy, respectively. Moreover, the CAE-SVM-PK model exhibited a slight decline in testing accuracy to 90.53%. These results emphasize the adaptability of the hybrid CAE-based framework, particularly the CAE-XGBoost model, which utilizes autonomous feature extraction and robust ensemble classification to handle diverse fault types.

The confusion matrices presented in Figure 11 illustrate the classification performance of the three hybrid models, CAE-SVM-LK, CAE-SVM-PK, and CAE-XGBoost, on the unseen test dataset for boiler leakage detection. Among these models, CAE-XGBoost demonstrated the most balanced performance, achieving 92.53% accuracy in correctly classifying the healthy state and 93.87% accuracy for the leakage state, with minimal misclassification rates of 7.47% and 6.13%, respectively. The CAE-SVM-LK model also performed reasonably well, with 92.00% classification accuracy for both healthy and faulty states, though it exhibited slightly higher misclassification rates of 8.00% compared to CAE-XGBoost. Moreover, the CAE-SVM-PK model showed reduced accuracy for the healthy state, classifying 85.87% of the instances correctly while achieving a higher 95.20% accuracy for the faulty state. However, it exhibited a misclassification rate of 14.13% for healthy conditions, indicating challenges in generalizing boiler leakage detection. Overall, these results validate the robustness of the CAE-XGBoost model in maintaining high classification accuracy across diverse fault types, demonstrating its superiority in utilizing CAE-extracted features for fault detection in critical components of the power plant.

Table 4. Performance evaluation of the best three hybrid models for boiler fault detection.

Model	Health State	Accuracy (%)	Precision (%)	Recall (%)	F-1 Score (%)
CAE-SVM-LK	Healthy	92.00	92.00	92.00	92.00
	Leakage	92.00	92.00	92.00	92.00
CAE-SVM-PK	Healthy	85.87	94.71	85.87	90.07
	Leakage	95.20	87.07	95.20	90.06
CAE-XGBoost	Healthy	92.53	93.78	92.53	93.15
	Leakage	93.87	92.63	93.87	93.25

provides a comprehensive evaluation of the performance metrics for these three models applied to boiler fault detection. The CAE-SVM-LK model demonstrated balanced performance, achieving 92.00% accuracy, precision, recall, and F1-score for both healthy and leakage states, reflecting its consistent ability to classify healthy and leakage conditions with equal reliability. However, the CAE-SVM-PK model displayed a disparity in its performance metrics, achieving 85.87% accuracy for the healthy state, accompanied by a higher precision of 94.71%, but with a lower recall of 85.87%, indicating a tendency to overestimate healthy state identification. For the leakage state, the CAE-SVM-PK model achieved higher accuracy (95.20%) but at the cost of a slightly reduced precision (87.07%). In contrast, the CAE-XGBoost model demonstrated the most robust and balanced performance across all metrics, achieving 92.53% accuracy for the healthy state with precision and recall of 93.78% and 92.53%, respectively, resulting in an F1-score of 93.15%. Similarly, for the leakage state, it achieved 93.87% accuracy, with precision, recall, and F1-scores consistently above 92%. The robust generalization observed in the CAE-XGBoost model can be attributed to two primary factors: the autonomous feature extraction capability of the CAE and the ensemble-based classification approach of XGBoost. The CAE effectively captures complex fault patterns irrespective of the fault type directly from raw data, eliminating the need for manual feature engineering, while XGBoost incorporates its ensemble nature to combine multiple decision trees, ensuring precise classification and resilience to overfitting. Therefore, these results demonstrate the superior classification capability and generalization ability of the CAE-XGBoost model in accurately diagnosing both turbine fault and boiler leakage, making it the most reliable hybrid model for diverse fault detection scenarios in power plants.

Table 5. Comparison of the proposed approach with existing popular methods for boiler leakage detection.

Model	Accuracy (%)
SVM [33]	90.50
KNN [33]	88.10
Naive Bayes [33]	85.70
Discriminant analysis [33]	88.10
CAE-SVM-LK	92.00
CAE-SVM-PK	90.53
CAE-XGBoost	93.20

provides a comparison of the proposed approach with existing popular methods for boiler leakage detection, emphasizing the advantages of the hybrid models. The CAE-XGBoost model achieved the highest accuracy (93.20%), surpassing traditional methods such as SVM (90.50%), KNN (88.10%), Naive Bayes (85.70%), and Discriminant Analysis (88.10%). These results underscore the effectiveness of combining autonomous feature extraction with advanced classifiers for improved detection accuracy in boiler leakage scenarios.

5. Conclusions

This study proposed a novel autonomous fault detection framework that integrates CAEs, PCA, and XGBoost to address critical challenges in power plant fault detection. The framework was developed through a systematic process, starting with the identification of key fault scenarios—boiler waterwall tube leakage and turbine motor-driven oil pump failure—and the acquisition of high-quality sensor data representing both healthy and faulty states. To ensure consistency and eliminate biases in the data, normalization was applied as the sole preprocessing step, transforming raw sensor readings into a standardized range for effective analysis. Autonomous feature extraction was performed using CAEs, which captured complex fault-related patterns directly from the raw sensor data, eliminating the need for manual feature engineering. PCA was employed to reduce dimensionality while preserving critical fault-related information, and XGBoost provided robust classification, leveraging its ensemble learning capabilities to achieve high accuracy, precision, recall, and F1-scores. Comparative analysis demonstrated its superiority over traditional models, including ANNs and CAE-SVM hybrids, in terms of classification performance and generalization ability. This structured and iterative development approach ensured that the framework met the dual objectives of early fault detection and practical applicability. The CAE-XGBoost model was shown to significantly improve operational efficiency, making it a scalable and robust solution for critical industrial applications. Future work will focus on addressing the challenge of model explainability, which remains a significant limitation in deep learning-based approaches. Developing interpretable methods to provide insights into the decision-making process of the proposed framework will enhance its transparency and foster greater trust for deployment in critical industrial applications.