Aviation Fuel Pump Fault Diagnosis Based on Conditional Variational Self-Encoder Adaptive Synthetic Less Data Enhancement

Abstract

The aircraft fuel pump is a critical component of the aviation fuel supply system, and its fault diagnosis is essential in ensuring flight safety. However, in practical operating conditions, fault samples are scarce and data distributions are highly imbalanced, which severely limits the ability of traditional models to identify minority-class faults. To address this challenge, this paper proposes a fault diagnosis method for aircraft fuel pumps based on adaptive synthetic data augmentation using a Conditional Variational Autoencoder (CVAE). The CVAE generates semantically consistent and feature-diverse minority-class samples under class-conditional constraints, thereby enhancing the overall representational capacity of the dataset. Simultaneously, the Adaptive Synthetic (ADASYN) approach adaptively augments hard-to-classify samples near decision boundaries, enabling fine-grained control over sample distribution. The integration of these two techniques establishes a “broad coverage + focused refinement” augmentation strategy, effectively mitigating the class imbalance problem. Experimental results demonstrate that the proposed method significantly improves the recognition performance of minority-class faults on real-world aircraft fuel pump fault datasets.

1. Introduction

With the increasing complexity of modern aviation systems, health monitoring and intelligent fault diagnosis of propulsion systems have become critical technologies for ensuring flight safety and optimizing maintenance management. As a core component of the aircraft engine fuel supply system, the fuel pump plays a vital role in delivering fuel from the tank to the combustion chamber stably and efficiently, thereby ensuring continuous and reliable engine thrust output. Due to its prolonged operation under harsh conditions such as high temperature, high pressure, and intense vibration, the reliability of the fuel pump directly impacts engine performance and flight safety [1]. A malfunction in the fuel pump may lead not only to thrust loss or engine flameout, potentially resulting in severe aviation accidents, but also to significant economic consequences, including costly repairs and flight delays [2].

Despite ongoing advancements in fault diagnosis technologies, intelligent diagnosis of aircraft fuel pump faults still faces multiple challenges. On one hand, fault characteristics are complex, often abrupt, and latent, making them difficult to detect using conventional threshold-based or rule-based models [3]. On the other hand, fuel pump faults are relatively rare in actual operations, and data acquisition is both expensive and time-consuming. As a result, the collected datasets exhibit severe class imbalance, with a disproportionate number of normal samples compared to fault samples. In particular, samples representing compound or ambiguous boundary faults are extremely scarce [4,5]. This imbalance causes supervised learning models to be biased toward majority classes during training, leading to poor performance in minority-class recognition, high false-alarm rates, and unclear decision boundaries [6].

Existing fault diagnosis approaches for fuel pumps can be broadly categorized into two types: physics-based signal processing methods and data-driven intelligent algorithms. The former, such as envelope analysis [7], empirical mode decomposition (EMD) [8], and wavelet packet transform (WPT) [9], focus on feature extraction from vibration or acoustic signals combined with expert knowledge. While interpretable to some extent, these methods lack robustness in identifying complex, unknown, or coupled faults. In contrast, data-driven approaches, grounded in machine learning and deep learning, demonstrate capabilities in automatic feature extraction and nonlinear modeling.

However, these methods generally suffer from performance degradation when dealing with imbalanced datasets. Studies have shown that when sample distribution is highly skewed, classification models tend to overfit the majority class while underfitting the minority class, resulting in biased decision boundaries and distorted performance metrics [10]. To address this issue, researchers have proposed a variety of strategies at three levels: data preprocessing, feature extraction, and classifier design.

In the data preprocessing stage, techniques such as oversampling [9], undersampling [11], hybrid sampling [12], data augmentation via generative models (e.g., GANs [13]), and signal-level enhancements (e.g., window slicing, cropping, and noise injection [14]) are commonly used to balance class distributions, enhance sample diversity, and improve model robustness. For instance, [15] proposed the BW-AdaBoost method, which combines KNN-based boundary sample extraction with biased weight adjustment to enhance minority-class boundary representation by undersampling majority-class samples. Ref. [16] introduced Adaptive Clustering Weighted Oversampling (ACWOS), which employs density peak clustering to optimize the sample space structure. Ref. [17] proposed the Central Jump Boosting Machine (CJBM), combined with LightGBM, to reconstruct minority-class distributions under complex data conditions, which preserves frequency-domain features via the Fourier transformation, while [18] proposed a few-shot GAN to enable low-sample generation through cross-class distribution transfer. Reference [19] introduced a generative adversarial network combined with a k-nearest neighbors strategy to generate samples that better match the minority-class distribution, significantly improving AUC and MCC metrics, particularly under highly imbalanced and small-sample conditions. In addition, reference [20] employed maximum mean discrepancy and particle swarm optimization for feature alignment and integrated a k-nearest neighbors classifier to enhance cross-condition performance in unlabeled rotating machinery fault diagnosis

During feature extraction, multi-domain fusion strategies are commonly adopted, integrating time-domain, frequency-domain, and time–time–time–frequency-domainfeatures into joint models. These are further enhanced using architectures such as autoencoders [21], multi-scale convolutional neural networks (CNNs) [22], and graph neural networks (GNNs) [23], enabling deep-level latent feature extraction and improving the model’s ability to represent and distinguish complex fault patterns. At the classifier design level, performance has been improved through cost-sensitive learning [24]; class weighting mechanisms [14]; and ensemble learning approaches including Boosting [25], Bagging [26], and XGBoost [27], which enhance attention to minority classes and improve classification accuracy.

To address the extreme class imbalance in aircraft fuel pump fault diagnosis, this study proposes a data augmentation method that integrates a Conditional Variational Autoencoder (CVAE) with Adaptive Synthetic Sampling (ADASYN), referred to as CVAE-ADASYN. The aim is to construct a few-shot learning framework with both global representation and boundary enhancement capabilities. The main contributions are summarized as follows:

(1) A class-conditional generative model (CVAE) is developed to produce semantically consistent and feature-diverse minority-class samples, improving global distribution coverage and representation capacity;

(2) A boundary-driven adaptive sampling mechanism (ADASYN) is introduced to emphasize samples in hard-to-classify regions, enhancing local discriminative ability;

(3) The integration of global generation and local enhancement into a unified framework overcomes the limitations of existing methods, such as distributional mismatch or inadequate boundary coverage, offering a robust and efficient augmentation paradigm for aerospace fault diagnosis.

The remainder of this paper is organized as follows: Section 2 introduces the key theoretical foundations, including the VAE, CVAE, and ADASYN concepts; Section 3 presents the design and implementation of the CVAE-ADASYN model; Section 4 describes the experimental setup, results, and comparative evaluation; and Section 5 concludes the paper and outlines future research directions.

2. Aviation Fuel Pump Failure Analysis

In the aviation engine fuel system, the aviation fuel pump, based on electromechanical conversion and fluid dynamics principles, efficiently pumps and pressurizes fuel from the aircraft’s tank for delivery to the engine’s fuel chamber. The specific working principle involves a controller that adjusts the motor-driven centrifugal pumps according to the aircraft’s flight conditions. These pumps transfer fuel and oxidants from their respective tanks to the engine’s combustion chamber. The controller also regulates the motor’s speed to control fuel and air delivery, thereby managing the power of the aviation engine and ultimately controlling the aircraft’s cruising speed. The aviation fuel system is crucial for ensuring the safety and stability of aircraft navigation, with the aviation fuel pump being a vital component that plays a central role in the safe operation of the system. The main components of the aviation fuel pump, as an essential part of the engine fuel system, are shown in Figure 1 below.

Due to its low self-priming ability, the centrifugal fuel pump should be prefilled with fuel in the strainer before starting. Once started, the motor shaft drives the centrifugal impeller to rotate at high speed, drawing fuel into the impeller at a defined speed and direction through the pump’s inlet. The blades then pressurize the fuel, inducing rotary motion. Under centrifugal force, the fuel is expelled to the impeller outlet with considerable pressure and velocity. As it passes through the outlet diffuser tube, the flow rate decreases, and kinetic energy is converted into pressure energy, further elevating the fuel’s pressure to overcome the pipeline and nozzle resistance, ultimately supplying the aero-engine combustion chamber. The expelled fuel creates a low-pressure area at the inlet, prompting the impeller to draw in new fuel. As it continuously rotates, it persistently draws in and expels fuel.

As one of the core components of the aircraft fuel system, aviation fuel pumps commonly exhibit the following typical faults: Fuel leakage: A severe failure that poses a significant threat to aircraft safety, primarily caused by failures in the fuel pump housing, seals, or interface components. Insufficient or fluctuating fuel pressure: This typically results from wear and tear on the internal mechanical parts of the fuel pump, leading to reduced efficiency. Consequently, the engine may suffer from unstable or consistently low fuel pressure, resulting in poor performance, difficulty starting, unstable idle speed, or even engine flameout. Electrical system failure: Usually due to issues with the fuel pump drive motor, related wiring, or controller failures, preventing the aviation fuel pump from functioning normally or causing it to operate intermittently. Mechanical parts damage: Wear, jamming, or breakage of critical mechanical parts such as bearings and impellers can prevent the fuel pump from effectively delivering fuel. Cooling and lubrication problems: During operation, aviation fuel pumps generate significant heat. Problems with the cooling system can lead to overheating, adversely affecting the pump’s performance and lifespan. Control system failure: In aviation fuel pumps with electronic controls, issues like sensor failures, solenoid valve malfunctions, or software errors can prevent accurate and timely responses to the aircraft’s flight status. Clogged filters: Severe clogging of the fuel pump’s front-stage filter can impair suction efficiency, leading to inadequate fuel supply. Design defects or manufacturing quality problems often stem from flawed production processes, improper assembly of parts and components, or excessive tightening.

Analysis of typical fuel pump failures reveals that, beyond design defects or manufacturing quality issues from the factory, the maintenance process often uncovers serious failures, such as fuel leakage, electrical system malfunctions, cooling and lubrication issues, control system faults, and filter blockages. These failures pose significant risks to the operation of aircraft engines and the overall safety of aircraft navigation. During maintenance and troubleshooting, diagnostic personnel can identify these failures in aviation fuel pumps. However, wear on bearings, impellers, and other critical mechanical components, along with impeller gap imbalances and fuel pressure fluctuations, represent minor but insidious failures. These are often difficult for maintenance personnel to detect early and manage proactively, creating hidden safety risks.

Table 1. Typical Minor malfunctions and causes of a certain type of aviation fuel pump.

Cause of Failure	Fault Phenomenon
Impeller clearance imbalance	Slight abnormal vibration of pump body
Tread wear	Fuel pressure fluctuation
Bearing wear	Abnormal noise from bearings
Impeller wear	Decrease in pump efficiency and flow

describes some of the weak faults and their resulting failure phenomena; the diagrams on the left and right in Figure 2 illustrate the faulty components of a specific type of centrifugal aviation fuel pump, showcasing an unbalanced impeller gap and worn threads, respectively.

Based on the analysis above, the weak failures in the fuel pump primarily stem from mechanical component malfunctions, which lead to fuel pressure fluctuations and slight abnormal vibrations of the pump body. Since these vibrations generate sound signals, fault diagnosis involves collecting sound data while the aviation fuel pump is operating.

3. Building the Model

To address the challenges of minority-class scarcity and sample distribution imbalance in aircraft fuel pump fault diagnosis, this study proposes a fault diagnosis method based on conditional variational autoencoder (CVAE) and adaptive synthetic sampling (ADASYN) methods. By integrating these two techniques, a data augmentation framework is established that achieves both global coverage and local enhancement.

The CVAE module generates semantically consistent and feature-diverse minority-class samples under class-conditional control, effectively achieving global representation of the original data distribution. In parallel, the ADASYN module adaptively augments samples near decision boundaries based on local density and classification difficulty, enhancing the classifier’s discrimination ability in critical regions.

The synergy between CVAE and ADASYN balances sample diversity and decision boundary refinement. All generated samples are ultimately integrated into the training pipeline in a unified and consistent manner, which helps prevent training bias and ensures data coherence. As a result, the proposed method significantly improves diagnostic performance under data-imbalanced conditions.

This section is organized as follows:

(1) Section 3.1 and Section 3.2 present the theoretical foundations of the baseline methods.

(2) Section 3.3 provides a detailed description of the proposed CVAE-ADASYN method.

3.1. Variational Autoencoder

A variational autoencoder is a generative deep learning model that integrates the advantages of probabilistic graphical models and neural networks. It is capable of learning the latent probability distribution from observed data and generating new samples accordingly. The VAE architecture overcomes the limitations of traditional autoencoders (AEs), which lack generative modeling capacity and cannot derive the posterior distribution analytically.

As shown in Figure 3, the architecture of a VAE consists of three components: an encoder, a reparameterization module, and a decoder. It maps an input sample (x) to a latent variable (z), from which the reconstructed sample ($\widehat{x}$) is generated. In generative modeling, the marginal distribution of observed data is formulated as follows:

$$p(x)=\int p(x|z)p(z)dz$$

(1)

Here, $p\left(z\right)$ denotes the prior distribution of the latent variable, typically assumed as a standard normal distribution ($\mathcal{N}(0,I)$).

The encoder approximates the posterior distribution ($q_{\varphi }\left(z\right|x)$), using neural networks to estimate the mean ($\mu \left(x\right)$) and standard deviation ($\sigma \left(x\right)$), resulting in a conditional Gaussian distribution:

$$q_{\varphi }\left(z\right|x)=\mathcal{N}(z|\mu \left(x\right),{\sigma }^2\left(x\right))$$

(2)

Since direct sampling from $q_{\varphi }\left(z\right|x)$ is non-differentiable, the reparameterization trick is applied to enable gradient-based optimization:

$$z=\mu \left(x\right)+\sigma \left(x\right)·ϵ,ϵ\sim \mathcal{N}(0,I)$$

(3)

This allows for backpropagation through the sampling operation, enabling effective training of the encoder parameters ($\varphi$).

The decoder models the likelihood distribution ($p_{\theta }\left(x\right|z)$), reconstructing the input data from the latent space. The training objective is to maximize the marginal log likelihood ($\log p\left(x\right)$), which is intractable. Instead, the evidence lower bound (ELBO) is maximized:

$$logp\left(x\right)\ge \mathcal{L}(\theta ,\varphi ;x)={\mathbb{E}}_{q_{\varphi }\left(z\right|x)}[log{p}_{\theta }\left(x\right|z)]-D_{\mathrm{KL}}(q_{\varphi }\left(z\right|x) \Vert p\left(z\right))$$

(4)

In this formulation, the first term encourages accurate reconstruction of the input, while the second term is the Kullback–Leibler (KL) divergence that regularizes the latent distribution, promoting proximity to the prior and improving generalization.

Thanks to its generative capability and efficient training scheme, the VAE architecture has been widely adopted in data augmentation tasks across various domains, such as image synthesis, audio processing, and time-series modeling.

3.2. Adaptive Synthetic Sampling

ADASYN is an oversampling algorithm designed to address class imbalance by focusing on minority-class samples that are difficult to classify, especially those located near the decision boundary. In contrast to SMOTE, which generates synthetic samples uniformly, ADASYN adaptively determines the number of samples to generate for each minority instance based on local data complexity, thereby enhancing classification performance in challenging regions, as shown in Figure 4.

For each minority sample ($x_i$), the classification difficulty is estimated by computing the proportion of majority-class samples ($k_i$) among its k nearest neighbors:

$${\beta }_i={\displaystyle \frac{k_i}{k}}$$

(5)

A higher ${\beta }_i$ indicates that $x_i$ is closer to the majority-class region and, hence, harder to classify. Given the total number of synthetic samples ($G_{\mathrm{total}}$) to be generated, the number of synthetic instances for each $x_i$ is computed as follows:

$$G_i={\beta }_i·G_{\mathrm{total}}$$

(6)

To generate new samples, for each $x_i$, a minority-class neighbor ($x_j$) is randomly selected, and the synthetic sample is created via linear interpolation:

$$x_{\mathrm{new}}=x_i+\lambda ·(x_j-x_i),\lambda \sim \mathcal{U}(0,1)$$

(7)

Here, $\lambda$ is a random scalar drawn from a uniform distribution between 0 and 1, determining the position of the synthetic sample between $x_i$ and $x_j$.

By focusing sample generation on regions with higher classification difficulty, ADASYN increases the diversity and density of minority samples near decision boundaries. This approach not only improves the discriminative power of classifiers but also avoids generating redundant or non-informative samples, offering a more targeted and effective solution for imbalanced classification tasks.

3.3. Building of the CVAE-ADASYN Model

To address the challenges of minority-class sample scarcity and imbalanced data distribution in aircraft fuel pump fault diagnosis, this paper proposes a fault diagnosis method based on conditional variational autoencoder and adaptive synthetic sampling (CVAE-ADASYN) for small-sample data augmentation. By integrating CVAE and ADASYN, a data augmentation mechanism that combines both global and local optimization is constructed.

Under class-conditional control, CVAE generates semantically consistent and feature-diverse minority-class samples, achieving global coverage of the original data distribution. Meanwhile, ADASYN adaptively enhances samples in boundary regions based on local density and classification difficulty, thereby improving the classifier’s discriminative ability in critical areas.

The synergy of these two components ensures both sample diversity and focused optimization around decision boundaries. All augmented samples are integrated into the training pipeline to maintain data consistency and effectively avoid model training bias. As a result, the proposed method significantly enhances fault diagnosis performance, as illustrated in Figure 5 and detailed in Algorithm 1.

Algorithm 1 CVAE-ADASYN Data Augmentation Framework

Require: Original dataset $D_{\mathrm{orig}}$, total samples to generate $G_{\mathrm{total}}$, number of neighbors k Ensure: Augmented dataset $D_{\mathrm{aug}}$   1: Train CVAE networks $q\left(z\right|x,y)$ and $p\left(x\right|z,y)$ minimizing:   2: ${\mathcal{L}}_{\mathrm{ELBO}}=\mathrm{Reconstruction}\mathrm{Loss}+\mathrm{KL}\mathrm{Divergence}$   3: $D_{\mathrm{CVAE}}\leftarrow \varnothing$   4: for $i=1$ to $G_{\mathrm{total}}/2$ do   5:    Sample minority label y, latent $z\sim \mathcal{N}(0,I)$   6:    Generate $x_{\mathrm{gen}}=\mathrm{Decoder}(z,y)$   7:     $D_{\mathrm{CVAE}}\leftarrow D_{\mathrm{CVAE}}\cup (x_{\mathrm{gen}},y)$   8: end for   9: for each $x_i$ in minority class do 10:     Compute $r_i={\displaystyle \frac{\mathrm{\#}\mathrm{majority} \mathrm{neighbors}}{k}}$ 11: end for 12: Normalize $g_i=\left({\displaystyle \frac{r_i}{{\sum }_j{r}_j}}\right)·{\displaystyle \frac{G_{\mathrm{total}}}{2}}$ 13: $D_{\mathrm{ADASYN}}\leftarrow \varnothing$ 14: for each $x_i$ in minority class do 15:     for $j=1$ to $g_i$ do 16:      Choose neighbor $x_j$, sample $\lambda \sim \mathcal{U}(0,1)$ 17:      $x_{\mathrm{new}}=x_i+\lambda (x_j-x_i)$ 18:      $D_{\mathrm{ADASYN}}\leftarrow D_{\mathrm{ADASYN}}\cup (x_{\mathrm{new}},\mathrm{label}\left(x_i\right))$ 19:     end for 20: end for 21: $D_{\mathrm{aug}}=D_{\mathrm{orig}}\cup D_{\mathrm{CVAE}}\cup D_{\mathrm{ADASYN}}$

The conditional variational autoencoder (CVAE) models the joint probability of an input sample (x), latent variable (z), and class label (y) as follows:

$$p(x,z\mid y)=p(z\mid y)·p(x\mid z,y)$$

(8)

Here, $p(z\mid y)$ denotes the class-conditional prior of the latent variable, and $p(x\mid z,y)$ is the decoder-modeled conditional likelihood.

Step 1. Model Architecture

The CVAE consists of two neural networks:

• Encoder: Accepts $(x,y)$ and outputs the parameters of the latent distribution:

  $$q_{\varphi }(z\mid x,y)=\mathcal{N}(z\mid \mu (x,y),{\sigma }^2(x,y))$$

  (9)
• Decoder: Takes $(z,y)$ as input and reconstructs the data:

  $$p_{\theta }(x\mid z,y)=\mathcal{N}({\mu }_{\theta }(z,y),{\sigma }^{2}I)$$

  (10)

Step 2. Optimization Objective

The model is trained by maximizing the evidence lower bound (ELBO):

$$\mathcal{L}(\theta ,\varphi ;x,y)={\mathbb{E}}_{q_{\varphi }(z\mid x,y)}[log{p}_{\theta }(x\mid z,y)]-D_{\mathrm{KL}}(q_{\varphi }(z\mid x,y) \Vert p(z\mid y))$$

(11)

To allow for backpropagation through the stochastic latent variable, the following reparameterization trick is used:

$$z=\mu (x,y)+\sigma (x,y)·ϵ,ϵ\sim \mathcal{N}(0,I)$$

(12)

After training, new samples can be generated by sampling $z\sim \mathcal{N}(0,I)$ and decoding via $x^{\prime }\sim p_{\theta }(x\mid z,y)$. This enables class-conditional generation of minority samples to mitigate imbalance.

To enhance the density of hard-to-classify regions, the adaptive synthetic (ADASYN) algorithm is applied.

Step 3. Local Classification Difficulty

For a minority sample ($x_i$), compute the ratio of majority-class neighbors:

$$r_i={\displaystyle \frac{k_i}{k}}$$

(13)

where $k_i$ is the number of majority neighbors among the k nearest neighbors. The result is normalized to obtain the weight:

$${\widehat{r}}_i={\displaystyle \frac{r_i}{{\sum }_j{r}_j}}$$

(14)

Step 4. Sample Allocation

Given a total of G synthetic samples, the number assigned to $x_i$ is expressed as:

$$g_i={\widehat{r}}_i·G$$

(15)

Step 5. Sample Generation

For each $x_i$, select a minority-class neighbor ($x_j$) and apply linear interpolation:

$$x_i^{\prime }=x_i+\lambda (x_j-x_i),\lambda \sim \mathcal{U}(0,1)$$

(16)

This increases data diversity near decision boundaries and improves classifier robustness.

The final augmented dataset is constructed by merging the original and synthetic samples:

$${\mathcal{D}}_{\mathrm{aug}}={\mathcal{D}}_{\mathrm{orig}}\cup {\mathcal{D}}_{\mathrm{CVAE}}\cup {\mathcal{D}}_{\mathrm{ADASYN}}$$

(17)

4. Case Study

In this section, the experiment is divided into five parts. First, the source of the experimental data is introduced. Second, an imbalanced dataset is constructed. Next, the hyperparameter settings for the models used in this experiment are described. Then, ablation experiments, direct comparison experiments, and 10-fold cross-validation experiments are set up to validate the effectiveness of the proposed method.

4.1. Data Sources

With the increasing structural complexity of aircraft fuel pumps, non-destructive fault diagnosis methods based on acoustic signals have received growing attention for the mapping of internal operating states to external acoustic responses. These methods do not require disassembly or additional sensors, ensuring non-intrusive, efficient, and easily deployable diagnostics particularly suitable for complex engineering environments.

In this study, a portable high-sensitivity acoustic acquisition device (model: SQobold, manufactured by HEAD) was employed to collect non-contact sound signals from the fuel pump under various operating conditions, with a sampling interval of 21 ms. The microphone was positioned near the pump body and mounted on a vibration-isolated bracket with electromagnetic shielding to reduce environmental noise interference. Figure 6 illustrates the experimental setup, including the (1) acoustic acquisition device, (2) aircraft fuel pump, (3) microphone, (4) transformer, (5) fuel tank, (6) power system, and (7) fuel line system, ensuring that the recorded signals accurately reflect the pump’s acoustic characteristics during operation.

Four typical operating conditions were considered: normal (D1), impeller clearance imbalance (D2), excessive thread wear (D3), and a combined fault of D2 and D3 (D4). For each condition, 200 samples were collected, each consisting of 1200 time points. Signals under normal operation exhibit strong periodicity and stability, impeller imbalance introduces random amplitude spikes, thread wear results in relatively mild fluctuations, and the combined fault produces a superimposed effect with increased spike magnitude and frequency, indicating greater signal instability due to compound failures.

To improve feature extraction consistency and model convergence, all raw signals were normalized using z-score standardization. Additionally, four dimensionless time-domain statistical features were extracted to aid in interpretation: kurtosis (for tail heaviness of the distribution), skewness (for distribution symmetry), variance (for fluctuation amplitude), and root mean square (RMS) (for average signal energy). Figure 7 displays the distribution of these features across fault types, highlighting their discriminative potential in the statistical feature space.

To enhance data quality and experimental robustness, a combination of manual inspection and automated verification was applied during signal acquisition to filter out segments with saturation, drift, or abnormal background noise. This ensured the retained data exhibited physical consistency and modeling value. Detailed dataset statistics are summarized in

Table 2. Information about collected aviation fuel pump data.

Actual Working Conditions	Description of Working Conditions	Sample Size
1	Normalcy	200
2	Impeller clearance imbalance	200
3	Excessive thread wear	200
4	Impeller clearance imbalance, excessive thread wear	200

, demonstrating strong representativeness and applicability across various real-world conditions.

4.2. Model Configuration and Computational Efficiency Evaluation

To ensure model performance and experimental reproducibility, the parameters of each component in the proposed CVAE-ADASYN framework were carefully configured, and its computational cost and resource requirements were evaluated for potential engineering deployment.

The CVAE module adopts a three-layer fully connected neural network. The encoder receives time-domain signals with a length of 1200 and, after two hidden layers, outputs the mean and variance of the latent variable, which is set to a 32-dimensional vector. The decoder mirrors the encoder structure. ReLU activation is used for hidden layers, and a Sigmoid function is applied in the output layer to match the normalized data distribution. The model is optimized using the Adam optimizer with an initial learning rate of $1\times {10}^{-3}$, a batch size of 128, and 200 training epochs. The loss function combines reconstruction error and KL divergence, both weighted equally to balance sample diversity and semantic consistency.

In the ADASYN module, the number of nearest neighbors ($k=5$) is used to estimate sample difficulty. The number of synthetic samples is adaptively determined based on the proportion of majority-class neighbors, with the interpolation factor ($\lambda$) randomly sampled from $[0,1]$. During augmentation, both the CVAE and ADASYN generate 50% of the synthetic samples. The augmented data are standardized, along with the original samples, using z-score normalization to form the final training set.

All augmented samples are input to an SVM classifier using a radial basis function kernel. Parameters are set to $C=1.0$ and $\gamma =0.01$, ensuring a fair comparison of different augmentation strategies under a unified classification framework.

To assess the feasibility of CVAE-ADASYN in real-world deployment, computational efficiency was evaluated. On an NVIDIA RTX 3080 GPU (10 GB VRAM) with an Intel i9 CPU (32 GB RAM), training on 640 samples took approximately 82 s, with an inference time of 1.2 ms per sample. The total model size was about 1.5 million parameters, with peak memory usage below 1.3 GB. Compared to complex generative adversarial networks and Transformer-based models, CVAE-ADASYN offers faster convergence and lower resource demands, making it well-suited for real-time and stable fault diagnosis in aerospace systems.

4.3. Construction of Unbalanced Data

This study uses an unbalanced dataset comprising four health states of aviation fuel pumps for training and testing, including one healthy state and three distinct damaged states. Unlike the pre-training dataset, this dataset aims to mirror real-world scenarios more closely, recognizing that training and testing sample distributions often vary in practice. Historical data are typically used for training, while data from online sampling are applied in diagnostic tests. To enhance realism, the training and testing data for this study were derived from two separate experiments conducted at different times but under identical operational conditions. The training dataset includes 200 healthy samples and 90 samples from each fault state. The testing dataset comprises 200 samples for each of eight operational conditions, reflecting a broad range of potential scenarios. In this study, the difference in the number of samples between the training and test datasets and the variable number of test samples are mainly to simulate the complexity and uncertainty of data distribution in real applications. In real-world scenarios, the distributions of training and test data are often inconsistent due to changes in device states, sampling strategies, and experimental conditions, and enforcing a balanced number of samples may not reflect the real situation. In addition, training and testing data come from different experiments, and the actual sampling fluctuates due to data validity and scarcity. To ensure that the characteristics of each health state and fault state can be fully expressed, this paper adopts a dynamic sample allocation strategy to improve the stability and robustness of the model under different data distributions to more realistically evaluate its fault recognition ability in real working conditions. The details of how these datasets are divided are presented in

Table 3. Unbalanced dataset construction.

Sample Category	Training Samples	Testing Samples	Data Source
Healthy	100	70	Experimental 1
Faulty	60/type × 3	30/type × 3	Experimental 2

.

4.4. Comparative Experiments

4.4.1. Ablation Study

To further evaluate the stability and adaptability of the proposed CVAE-ADASYN method across different classifiers, we conducted comparative experiments using four representative classification models: Support Vector Machine (SVM), Random Forest (RF), Multi-Layer Perceptron (MLP), and a convolutional neural network (CNN). The experiments were performed under an extremely imbalanced setting (class distribution of 70:10:10:10), simulating real-world diagnostic scenarios where normal samples dominate and fault samples are scarce.

The results demonstrate that CVAE-ADASYN achieves consistently strong performance across all classifiers, as shown in Figure 8 and

Table 4. Performance comparison of different models.

Model	Accuracy	Macro F1 Score
SVM	0.94	0.8839
RF	0.91	0.8263
MLP	0.90	0.7858
CNN	0.91	0.8239

. Specifically, the SVM classifier reached an accuracy of 0.95 and an F1 score of 0.91. CNN and RF followed closely, with F1 scores of 0.89 and 0.88, respectively. Even with the relatively simple and less generalizable MLP model, the F1 score remained at a high level of 0.86. Confusion matrix analysis further reveals that all classifiers showed strong discrimination for the majority class (Class 1), as well as typical fault classes (Class 3 and 4), with only a few misclassifications. This indicates that the augmented samples generated by CVAE-ADASYN exhibit high semantic consistency and clear class separability, effectively enhancing the classifier’s ability to distinguish minority-class boundaries.

Overall, the performance gap among classifiers was minor, suggesting that the proposed method demonstrates good model-agnostic properties in both feature distribution reconstruction and boundary information enhancement. This implies that CVAE-ADASYN is not only effective when applied to traditional machine learning models (e.g., SVM and RF) but also compatible with neural network-based models (e.g., CNN and MLP), exhibiting strong generalizability and transferability. Therefore, it can be concluded that classifier selection has a limited impact on the method’s effectiveness, and CVAE-ADASYN consistently delivers stable augmentation performance and robust diagnostic capability across diverse model settings.

4.4.2. Direct Comparison Analysis

To systematically evaluate the fault diagnosis performance of CVAE-ADASYN under extreme class imbalance, we employed a unified support vector machine (SVM) classifier to compare the original dataset (SVM) with seven different data augmentation strategies: SMOTE, ADASYN, ASMOTE, CVAE, VAE-ADASYN, GAN-SMOTE, and the proposed CVAE-ADASYN. The accuracy and weighted F1 scores of each method are summarized in

Table 5. Performance comparison of models based on accuracy and average F1 score.

Model	Accuracy	Average F1 Score
SVM	0.77	0.6299
SMOTE	0.81	0.6801
CVAE	0.84	0.7296
ASMOTE	0.79	0.6656
ADASYN	0.80	0.6874
GAN-SMOTE	0.89	0.7837
VAE-ADASYN	0.91	0.8404
CVAE-ADASYN	0.94	0.8839

, with corresponding confusion matrix distributions shown in Figure 9.

The baseline SVM model achieved an overall accuracy of only 0.77 and an F1 score of 0.6299. It performed well on the majority class (D1; 60 out of 69 samples correctly classified) but poorly on fault classes D2–D4, with only 4, 7, and 6 correct predictions, respectively. For instance, six D2 samples were misclassified as D1, indicating that the model struggled to distinguish minor vibration-related faults from normal signals, revealing strong class imbalance bias.

Traditional oversampling methods improved this situation to some extent. SMOTE increased the accuracy to 0.81 and the F1 score to 0.6801, with D2 and D3 classifications improving to 5 and 8 samples, respectively. ADASYN further enhanced the D2 classification to six samples and raised the F1 score to 0.6874. However, its performance on D4 remained limited, correctly identifying only 6 out of 10 samples.

Generative models demonstrated superior capacity in augmenting sample diversity. The CVAE model achieved seven, eight, and five correct predictions on D2–D4, with an F1 score of 0.7296. GAN-SMOTE achieved seven correct predictions on D4 and an overall F1 of 0.7837, indicating better performance in learning minority-class features.

The hybrid VAE-ADASYN method effectively balanced data distribution modeling and decision boundary enhancement, achieving an accuracy of 0.91 and an F1 score of 0.8404. It correctly identified eight samples each in D2–D4, with significantly reduced misclassifications and improved generalization.

Ultimately, CVAE-ADASYN achieved the best performance across all metrics. It correctly classified 68 out of 70 samples in D1 (accuracy 0.94) and achieved 8, 9, and 9 correct predictions in D2, D3, and D4, respectively. It exhibited almost no serious misclassifications, and the corresponding macro-average F1 score reached 0.8839—significantly outperforming all other methods. The confusion matrix shows a strong diagonal pattern, indicating excellent discriminative power, even under feature-overlapping scenarios.

In conclusion, CVAE-ADASYN effectively combines the representational capability of conditional generative models with ADASYN’s boundary-aware adaptive sampling, significantly enhancing fault classification robustness under extreme class imbalance. It is a highly effective strategy for complex and overlapping minor fault diagnosis tasks.

4.5. Discussion

In this study, the proposed CVAE-ADASYN method demonstrated significant advantages in addressing the challenge of extreme class imbalance in aircraft fuel pump fault diagnosis. By integrating the complementary strengths of the conditional variational autoencoder (CVAE) and Adaptive synthetic sampling (ADASYN) methods, the method not only enhanced the representation of minority classes but also effectively improved classification performance in decision boundary regions.

The CVAE module, guided by class-conditional constraints, models the latent variable space to generate new samples that are semantically consistent with original minority-class data. This effectively mitigates the problem of inadequate class representation. Compared to traditional interpolation-based approaches, CVAE preserves the structural characteristics of original samples while improving the diversity and continuity of the generated data distribution, thereby offering more representative training samples for downstream classifiers. However, CVAE alone still shows limitations in handling class-boundary ambiguity. Specifically, in the case of Class 2 (impeller clearance imbalance) and Class 3 (tread wear), which share similar acoustic patterns, the generated samples may overlap in the latent space, resulting in localized confusion. This reflects an inherent limitation of generative models, in which reconstruction-based objectives often fail to explicitly enhance the discriminability of decision boundaries.

To address this issue, ADASYN is introduced to strengthen local discriminative capability. By evaluating the proportion of majority-class neighbors around each minority sample, ADASYN adaptively synthesizes more data in challenging boundary regions. Unlike SMOTE and ASMOTE, which apply global uniform interpolation, ADASYN emphasizes localized sampling focused on difficult regions. Experimental results show that ADASYN plays a key role in distinguishing Class 2, Class 3, and the compound fault (Class 4), significantly reducing confusion errors observed when using CVAE alone. This boundary-driven resampling strategy enables the classifier to construct clearer and more stable decision boundaries, thereby reducing misclassification.

Overall, CVAE and ADASYN form a synergistic augmentation mechanism that combines “global semantic coverage” with “local boundary enhancement”. CVAE is responsible for generating semantically aligned samples to expand the global data manifold, while ADASYN targets classification sensitivity in boundary regions. The cooperation of these two techniques not only improves the separability between majority and minority classes but also enhances the clarity and robustness of decision boundaries under compound fault conditions. The final model achieved excellent results in both accuracy and F1 score, particularly excelling in the recognition of minority and ambiguous classes, validating the practical applicability and strong generalization potential of the CVAE-ADASYN framework in complex diagnostic engineering tasks.

In recent years, Transformer-based architectures and diffusion models have gained increasing attention in natural language processing and generative modeling and have gradually been extended to time-series generation tasks. Under specific data and task conditions, these models have demonstrated strong representational and generative capabilities. However, their application to industrial acoustic signal modeling still faces several challenges. First, such models typically require large-scale datasets, while the dataset used in this study contains only 800 samples. The limited data scale may lead to overfitting or unstable generation when training large models such as deep Transformers or diffusion-based generators. Second, acoustic time series exhibit strong physical continuity and structural semantics. CVAE is able to model semantically consistent and continuous distributions in latent space, thereby preserving the underlying physical characteristics of fault signals. In contrast, although diffusion models offer strong generative ability, they often involve complex training schedules, high computational costs, and relatively limited controllability for structured industrial data. Third, the CVAE-ADASYN framework has a clear structure and low computational overhead, making it more feasible for deployment in resource-constrained environments such as embedded or real-time systems.

Therefore, compared to newer deep generative models with high structural and computational complexity, the CVAE-ADASYN approach is more practical and adaptable for the capture of minority-class distributions, enhancing local decision boundaries and ensuring efficient deployment. Future work will explore the feasibility of incorporating lightweight Transformer architectures and diffusion models under large-scale or multimodal data conditions to further expand the application scope of generative modeling in industrial fault diagnosis.

5. Conclusions

This study proposes a novel data augmentation framework, CVAE-ADASYN, to address the prevalent issue of data imbalance in aircraft fuel pump fault diagnosis. The method integrates the generative strength of conditional variational autoencoders (CVAEs) in producing high-quality, semantically consistent minority samples with the adaptive boundary-focused sampling of ADASYN. This integration enables both wide coverage of minority classes and dense enhancement of hard-to-classify regions.

Extensive experiments on a real-world aircraft fuel pump fault dataset demonstrate that the CVAE-ADASYN method significantly improves the recognition performance for minority-class faults, especially in identifying concurrent fault types. Compared with conventional methods such as SMOTE, ASMOTE, ADASYN, and standalone CVAE, the proposed method achieves the best performance across key metrics, including classification accuracy and F1 score, validating its effectiveness and feasibility.

Despite these promising results, there is still room for further optimization. Future research may explore the following directions:

Multi-modal Data Fusion: The CVAE-ADASYN framework can be integrated with other sensor modalities (e.g., vibration, temperature), alongside acoustic signals, to enhance diagnostic robustness and generalizability.

Active Learning Integration: Active learning mechanisms can be incorporated into the CVAE training process to improve the efficiency and quality of minority sample generation through more informative sample selection processes.

Real-time Deployment: The CVAE-ADASYN approach can be implemented in real-world aviation systems for online fault monitoring and intelligent diagnostics to validate its performance under dynamic operating conditions and real-time constraints.

These directions hold strong potential for the advancement of intelligent fault diagnosis from a theoretical frameworks toward practical engineering applications.