A Novel Deep Learning Framework for Predictive Fault Diagnosis in Dry-Type Transformers Using Vibration Signal Data Analysis Technique

The prevailing transformer monitoring infrastructure in industrial environments remains largely fragmented and functionally constrained. Most systems rely on standalone monitoring and alarm units that track only a limited set of fundamental electrical parameters. This decentralized architecture typically lacks the capacity for holistic data acquisition and advanced analytics, rendering continuous online monitoring and periodic performance assessment infrequent. Consequently, the implementation of automated predictive maintenance scheduling is rare and often technologically prohibitive.

This restricted visibility into transformer health introduces substantial operational vulnerabilities. Without timely fault detection, latent issues may escalate undetected, resulting in costly repairs, prolonged downtime, or catastrophic equipment failure necessitating full unit replacement.

Recent studies have underscored both the opportunities and challenges inherent in applying machine learning (ML) to predictive maintenance, particularly for critical rotating machinery such as bearings, motors, gearboxes, and pumps [1-4]. Significant gaps persist in this field, presenting opportunities to enhance model accuracy and improve the flexibility of future predictive maintenance solutions. For instance, Rahman et al. developed an intelligent anomaly detection method for electric motors using a combination of vibration signals and artificial intelligence algorithms [5]. They trained an unsupervised learning model on two distinct motors, a new test motor and an aged industrial motor. Their comprehensive analysis demonstrated that the model achieved high anomaly detection capability for standardized operating conditions using mapped features. However, a key limitation of their work was the exclusive use of data from normal motor conditions due to a scarcity of data on actual fault conditions.

Another approach, presented by Tahir et al. [6] leveraged statistical features from vibration signals, such as root mean square (RMS), mean, variance, skewness, and kurtosis, for model training, a common practice in many studies. A notable contribution of their work was the introduction of a Mean-based Outlier Detection (MOD) method during the data preprocessing phase. This technique was designed to identify and remove outliers—data samples influenced by external factors rather than actual faults—thereby improving model performance. However, their study did not address the classification of similar fault types with varying degrees of severity.

The rapid advancement of deep learning (DL) has spurred a growing interest among researchers in applying DL methods to address the aforementioned challenges [7-9]. Many DL techniques have already been successfully deployed in fault diagnosis. Convolutional Neural Networks (CNNs), in particular, have emerged as a classic and highly effective architecture for image classification tasks and they are considered a good solution to the problem of bearing fault detection [10, 11]. Nevertheless, a key challenge with very deep DL models is the vanishing gradient problem, which can lead to a degradation in performance. To circumvent these issues, Support Vector Machines (SVMs) have been proposed [12]. SVMs, which operate on the principle of linear combination through hyperplanes, are particularly effective in handling complex, high-dimensional data, leading to the development of a series of robust SVM-based models [13-15].

A particularly promising approach for enhancing fault diagnosis involves encoding complex time-series data into visual representations for subsequent DL analysis. The Gramian Angular Field (GAF) methodology, encompassing the Gramian Angular Summation Field (GASF) and Gramian Angular Difference Field (GADF) techniques, has achieved considerable prominence as an approach for transforming univariate time-series data into a two-dimensional image representation. This technique is valued for its ability to preserve the temporal relationships within the data, thereby enhancing feature extraction by DL models. For example, Shen et al. applied GAF to vibration signals for bearing fault diagnosis, demonstrating marked improvements in classification accuracy when combined with CNNs [16]. Similarly, Bai et al. utilized GAF encoding for fault detection in rotating machinery, highlighting its robustness in catching subtle patterns for data in time-series [17]. The GAF with its effectiveness in various tasks of fault diagnosis has been further validated by other studies, confirming its utility in transforming complex vibration signals into a format suitable for DL models [18-20].

This study introduces a novel deep learning-based framework for the predictive fault diagnosis of dry-type transformers. The proposed methodology leverages GAF encoding to convert normalized time-series vibration data into GASF and GADF images. These images are subsequently processed by a CNN for automated feature extraction. The extracted features are then concatenated and fed into a series of machine learning models for fault classification, focusing on three common transformer faults: loose windings, loose mounting bolts, and overload conditions. By integrating Internet of Things (IoT) technologies for real-time data acquisition with advanced artificial intelligence (AI) for predictive analytics, this approach seeks to improve transformer reliability, reduce maintenance costs, and mitigate operational risks. This scalable methodology offers a promising solution for industrial applications and holds the potential to significantly outperform conventional fault diagnosis techniques.

The rest of this paper is organized as follows. Section 2 will present the theoretical framework for the diagnosis of transformer’s faults. Then, Section 3 proposes a novel method to solve such a problem. Section 4 shows experimental results as applied the proposed methodology. Finally, conclusions and future work will be provided in Section 5.

3.1 Data processing and GAF transformation

The Gramian angle field is a mapping technique which converts 1D-waveforms into 2D-photos in Cartesian coordinates. This transformation retains the dynamic features of the original signal following the discrete time domain while also preserving its temporal dependencies. The surface vibration signal of the transformer, once collected, is represented as a 1D time series , where n is the total number of samples acquired during the measurement process at a specified sampling frequency. To standardize the data's magnitude and nullify variations related to the measurement scale, the complete signal undergoes a normalization procedure to fit within the predefined range [−1,1] using Eq. (1):

$x_i^{\mathrm{norm}}=2 \times \frac{x_i-\min (x)}{\max (x)-\min (x)}-1$     (1)

where:

• $x_i$ is the signal value at time index $i$;
• $\min (x)$ and $\max (x)$ are defined as the minimum and maximum amplitudes, respectively, observed across the entire signal range.

Signal normalization is first employed as a critical preprocessing stage to attenuate amplitude-scale artifacts and thereby enhance the convergence stability and rate of the deep learning architecture. Subsequent to normalization, the continuous time-series signal is partitioned into a series of fixed-length, overlapping segments via a sliding window protocol, as depicted in Figure 2. This segmentation strategy serves a dual purpose: it facilitates the extraction of localized temporal dynamics from the signal and simultaneously acts as a data augmentation technique to substantially expand the volume of training instances.

Figure 2. Re-sampling process of the signal

Figure 3. The transformation pipeline from the normalized time-series to GASF/GADF representations

To convert one-dimensional vibration time-series into a format suitable for CNNs, each segmented and normalized sample $x=\left\{x_1, x_2, \ldots, x_n\right\}$, where $x_i \in[-1,1]$, is transformed into a 2D image applying the GAF method as presented in Eq. (2).

$\left\{\begin{array}{c}\phi=\arccos \left(\tilde{x}_i\right),-1 \leq \tilde{x}_i \leq 1, \tilde{x}_i \in \tilde{X} \\ r=\frac{t_i}{N}, t_i \in \mathbb{N}\end{array}\right.$     (2)

where:

• $x_i$ is the signal value at time $t_i$;
• $t_i$ is the timestamp corresponding to each sample;
• N is a constant used to normalize the radius $r_i$ into a reasonable range within the polar coordinate system.

Based on the polar coordinates, two types of GAFs are computed to encode temporal correlations between signal points:

The Gramian Angular Summation Field (GASF)

$G_S=\left(\begin{array}{ccc}\cos \left(\phi_1+\phi_1\right) & \cdots & \cos \left(\phi_1+\phi_N\right) \\ \cos \left(\phi_2+\phi_1\right) & \cdots & \cos \left(\phi_2+\phi_N\right) \\ \vdots & \ddots & \vdots \\ \cos \left(\phi_N+\phi_1\right) & \cdots & \cos \left(\phi_N+\phi_N\right)\end{array}\right)=\tilde{X} \tilde{X}^T-\left(\sqrt{I_N-\tilde{X}^2}\right)^T \sqrt{I_N-\tilde{X}^2}$      (3)

The Gramian Angular Difference Field (GADF)

The resulting GAF images as shown in Figure 3 retain both temporal and amplitude information, making them highly suitable for training CNN models in fault classification tasks.

$G_D=\left(\begin{array}{ccc}\sin \left(\phi_1-\phi_1\right) & \cdots & \sin \left(\phi_1-\phi_N\right) \\ \sin \left(\phi_2-\phi_1\right) & \cdots & \sin \left(\phi_2-\phi_N\right) \\ \vdots & \ddots & \vdots \\ \sin \left(\phi_N-\phi_1\right) & \cdots & \sin \left(\phi_N-\phi_N\right)\end{array}\right)=\left(\sqrt{I_N-\tilde{X}^2}\right)^T X-X^T \sqrt{I_N-\tilde{X}^2}$       (4)

As a result, the GAF images of the transformer corresponding to four different states are also presented in Table 1.

Table 1. GAF images of transformer according to the four different states

3.2 A CNN and classification model

The CNN architecture, detailed in Figure 4, was specifically engineered for feature extraction from the image-based representation of the converted vibration signals. This architecture is structurally organized into a sequence of three distinct convolutional blocks, each interspersed with a pooling layer. The final feature maps are then channeled through a flattening layer, which transitions the data into a vector format suitable for processing by the subsequent fully connected layers.

The raw time signal is converted into a 2D image with 1 channel (grayscale) using the Gramian Angular Field (GAF) method, resulting in an image with a size of 2000 × 2000. Then, this 1-channel image is converted into a 3-channel (RGB) image by duplicating the single channel into three identical channels, resulting in an image with a size of 2000 × 2000 × 3. Finally, the image is resized from 2000 × 2000 × 3 to 64 × 64 × 3 to serve as the input for the CNN model as detailed in Table 2.

Figure 4. An illustration of the CNN architecture

The input layer receives RGB images of size 64 × 64 × 3, representing the image-transformed vibration signals.

The first convolutional layer (Conv2D + ReLU) applies 32 filters of size 3 × 3, using the ReLU non-linear function to detect local features such as edges and corners.

This is followed by a MaxPooling2D layer with a 2 × 2 kernel, reducing the spatial dimensions and computational complexity.

The second convolutional layer (Conv2D + ReLU) increases feature depth with 64 filters of size 3 × 3, capturing more abstract patterns.

Another MaxPooling2D (2 × 2) layer further reduces the dimensionality.

The third convolutional layer (Conv2D + ReLU) uses 128 filters of size 3 × 3, enabling the network to learn complex and high-level representations.

A final MaxPooling2D (2 × 2) layer down samples the spatial information before flattening.

The Flatten thickness transforms the 2D characteristic maps into a 1D feature vector.

Finally, a Dense layer with 256 units and ReLU activation is used to project the learned features into a high-level representation suitable for classification or further analysis.

Table 2. The CNN architecture for feature extraction

Figure 5. A representation of the overall structure

Two independent CNN branches are designed to process GASF and GADF inputs respectively. Each CNN consists of multiple convolutional and pooling layers, followed by a flattening layer to convert the output feature maps into 1D vectors (see Figure 5):

Branch 1: CNN trained on GASF images

Branch 2: CNN trained on GADF images

Let $\mathrm{f}_{\text {GASF}} \in \mathbb{R}^{d_1}$ and $\mathrm{f}_{\text {GADF}} \in \mathbb{R}^{d_2}$ be the flattened feature vectors extracted from the two branches.The feature vectors are concatenated to form a unified representation:

$f_{\text {concat }}=\left[f_{G A S F} ; f_{G A D F}\right] \in R^{d_1+d_2}$     (5)

To evaluate the discriminative capability of the fused features, five classical machine learning classifiers are employed in this study, covering both linear and non-linear modeling paradigms with parameters provided in Table 3:

• Support Vector Machine (SVM): It is highly suitable for classification tasks involving high-dimensional feature vectors and sparse datasets, demonstrating strong performance in scenarios where the number of features significantly exceeds the number of samples.
• Random Forest (RF): An ensemble-based classifier that combines multiple Decision Trees to improve generalization and robustness against overfitting. RF is well-suited for handling noisy and complex feature distributions.
• K-Nearest Neighbors (KNN): This algorithm is an instance-based, non-parametric classifier that determines the class assignment of a novel data point by performing a majority vote among the labels of the K data points that are closest in the defined feature space. The KNN is simple yet effective, especially when the feature distribution preserves local structures.
• Logistic Regression (LR): A linear model that estimates the probability of class membership using the logistic function. It serves as a strong baseline for evaluating feature separability, particularly in linearly separable problems.
• Multi-Layer Perceptron (MLP): A feedforward artificial neural network that consists of an input layer, one or more hidden layers, and an output layer. The MLP is capable of modeling intricate, non-linear dependencies embedded within the dataset.

To select the optimal hyperparameters for the classifiers, the study uses grid search combined with 5-fold cross-validation to optimize the hyperparameters for each model.

Table 3. Optimal hyper parameters for each classifier

These classifiers were selected for their diverse algorithmic foundations (linear, non-linear, and ensemble), allowing for a comprehensive evaluation of the fused feature set.