Fault Diagnosis of Aircraft Hydraulic Pipeline Clamps Based on Improved KPCA and WOA–KELM

Abstract

Due to the complexity and diversity of aviation hydraulic pipeline systems, there has been a lack of qualitative formulas or characteristic indicators to describe clamp failures within these systems. In this paper, based on the data-driven idea, an improved KPCA-based feature extraction method is proposed and combined with the optimized KELM for fault diagnosis and condition monitoring of aviation hydraulic line clamps. Firstly, the kernel parameters of KPCA are combined using polynomial and Gaussian kernels based on their proportional weights. Secondly, a GA–PSO (Genetic Algorithm–Particle Swarm Optimization) hybrid algorithm is employed to optimize the kernel parameters, selecting 13 time-domain and 4 frequency-domain feature indicators to form the initial feature dataset, which is then subjected to dimensionality reduction using the improved KPCA. Finally, diagnosis is conducted using a KELM optimized by the whale optimization algorithm. The results indicate that, across multiple diagnostic trials, the average diagnostic accuracy can reach 99.99%, providing a feasible approach for the precise diagnosis of clamp faults in aviation hydraulic pipeline systems.

1. Introduction

Due to the complex structure of hydraulic pipeline clamp systems in aircraft engines, along with the harsh environment characterized by high power and high pressure, and the effects of the fluid–structure interaction, the clamps are subjected to various types of vibration excitations. This results in highly complex vibration signals from the hydraulic pipeline clamps. Pipeline system-induced engine failures account for approximately 50% of all engine failures, and the pipelines are secured and connected by clamps [1,2,3,4,5,6]. Therefore, accurately diagnosing the fault modes of aircraft hydraulic pipeline clamps is crucial for ensuring the normal operation of the hydraulic pipeline system and guaranteeing its performance reliability [7,8,9,10].

In recent years, scholars both domestically and internationally have conducted research on the signal characteristic parameters and models of clamp systems in aircraft hydraulic pipelines. Li investigated loosening faults in pipeline system joints, selecting multiple time-domain and frequency-domain features for integration. This approach effectively combined the corresponding time–frequency domain features to create a comprehensive set of characteristics, enabling a more detailed description of the signal’s features [11]. McFadden et al. used a time-domain analysis to process vibration signals generated during the engagement of hydraulic pipeline clamps and sun gears, employing a method to calculate the time-domain average of vibrations [12]. Chen et al. analyzed and processed signals using time-domain features, although external factors often introduce noise and other signals during the acquisition of time-domain signals [13,14].

Kwong and Edge conducted experiments on clamp vibration models by controlling the length of the pipeline. Their experiments revealed that adjusting the pipeline length and optimizing the position of the clamps could reduce pipeline faults [15,16,17]. W. Bartelmus studied the operational state of aircraft hydraulic pipeline clamp systems, identifying the most critical factor as disturbances from the rotation of the mechanical arm [18]. The specific vibration signals generated by the mechanical arm’s rotation were represented as time–frequency diagrams using Short-Time Fourier Transform (STFT) and the Wigner–Ville distribution.

Wang et al. employed the Kernel Principal Component Analysis (KPCA) for anomaly detection, distinguishing actual faults from abnormal sensor readings, and subsequently used Support Vector Machines (SVMs) for fault diagnosis [19]. Hao Tong Sun et al. proposed a model combining KPCA and SVM. They used KPCA with different kernel functions to extract features from imbalanced data and then established an SVM model to improve recognition accuracy [20]. Su et al. introduced a hybrid method based on Simulated Annealing–Particle Swarm Optimization (SA–PSO) and an improved Kernel-based Extreme Learning Machine (KELM) [21]. This method extracted multi-domain fault features from the time domain, frequency domain, and time–frequency domain, reconstructed high-dimensional feature sets using Variational Mode Decomposition (VMD), and employed Isometric Mapping to reduce the dimensionality of the high-dimensional feature sets. This approach achieved high classification accuracy in rolling bearing fault diagnosis. The finite element model of the clamp was established, and the stiffness test rig was conducted. The stiffness of the clamp was verified by comparing the natural frequencies of the pipeline clamp system. The results showed that the numerical predictions were in good agreement with the experimental data [22,23].

Although scholars have extensively studied and analyzed aircraft hydraulic pipeline systems, there is no unified detection method for pipeline clamp faults. This paper addresses the problem of difficult fault detection in clamps of pipeline systems. It conducts analysis and research in both the time domain and frequency domain, integrating various features in these domains using MATLAB software. Additionally, experiments are conducted to simulate the actual working conditions of aircraft hydraulic pipeline clamps. The resulting data are processed using optimized Kernel Principal Component Analysis (KPCA) for dimensionality reduction, selecting the most fault-representative kernel principal components as the feature set. These features are then fed into a Kernel Extreme Learning Machine (KELM) optimized by the Whale Optimization Algorithm (WOA) for feature extraction and subsequent diagnosis of the collected signals. This research aims to provide a more accurate method for analyzing faults in aircraft hydraulic pipeline clamps, thereby improving the detection accuracy of pipeline clamp faults.

2. Clamp Failure

Hydraulic System and Clamp Failures in Aviation

The hydraulic clamping system in aviation primarily comprises hydraulic pipelines, fixed clamps, pipe joints, seals, and brackets, as depicted in Figure 1. In this system, the spatial arrangement of the aviation engine is typically compact, leading to a highly intricate spatial structure of the pipelines. Consequently, clamps are essential to secure and constrain the connections between pipelines, between pipelines and adjacent components, and between the pipelines and the engine casing. This ensures the proper transmission of hydraulic oil within the pipelines, thereby guaranteeing the normal operation of the aviation engine.

Due to the various vibration excitations experienced by the clamps securing the hydraulic pipelines, the resulting vibration signals are highly complex, exhibiting strong nonlinearity and instability. Furthermore, when the hydraulic pipeline clamp system encounters one or more faults, such as pipeline cracks, pipeline dents, loose clamp bolts, clamp cracks, or clamp liner wear [24], the vibration signals become exceedingly intricate. Consequently, simple fault diagnosis methods are insufficient for accurately identifying the fault type and location. Several typical clamp failures are illustrated in Figure 2.

3. Fault Diagnosis Model

3.1. Improved Kernel Principal Component Analysis

The Kernel Principal Component Analysis (KPCA) is derived by incorporating a kernel function into traditional PCA [25]. The essence of the kernel method involves three main steps: first, the original data are mapped from the original space R to a feature space F through a nonlinear transformation; second, linear feature extraction and learning are conducted within the feature space F; finally, the linear features obtained in the feature space F are transformed back to the original input space R, manifesting as nonlinear features. By applying principal component analysis in the feature space, the directions of maximum variance of the original samples are identified, enabling the dimensionality reduction of the linear dataset in the high-dimensional space. This method, which achieves nonlinear mapping between data space, high-dimensional feature space, and category space through kernel mapping transformations, enhances the ability to process nonlinear data. Figure 3 shows a simple demonstration of a two-dimensional model decomposition using Kernel Principal Component Analysis (KPCA).

According to the theory of selecting an appropriate kernel function, let x_i and x_j represent all data samples in the original space R. Through nonlinear mapping, the corresponding kernel matrix K = [K_ij] _N×N is computed, where K_ij is the i × j sample point in the space.

$$〈x_i,x_j〉\to k(x_i,x_j)=\varphi (x_i)\varphi (x_j),$$

(1)

The kernel matrix is first centered and then corrected. In the feature space F, the covariance matrix of the data samples X can be represented as follows:

$$C_{\varphi }={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N\varphi \left(x_i\right)\varphi {(x_i)}^T},$$

(2)

By introducing the definition of the kernel function:

$$N\lambda K\beta =K^2\beta ,$$

(3)

If the data in the feature space do not satisfy the standard condition for centering, appropriate adjustments are necessary for the kernel matrix. Specifically, setting $\varphi (x_i)\to \varphi (x_i)-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{k=1}^N\varphi \left(x_k\right)}$ allows us to obtain the following:

$$\begin{array}{l}{K}_{ij}^{C_{\varphi }}=((\varphi (x_i)-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{k=1}^N\varphi \left(x_k\right)})(\varphi (x_j)-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{l=1}^N\varphi \left(x_l\right)}))\\ =K_{ij}-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{k=1}^N{K}_{kj}}-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{l=1}^N{K}_{il}}+{\displaystyle \frac{1}{N^2}}{\displaystyle \sum _{k=1}^N{\displaystyle \sum _{l=1}^N{K}_{kl}}}\end{array},$$

(4)

To calculate the projection on feature space vectors, the kernel matrix obtained after feature extraction is denoted as $K^{C_{\varphi }}$. The resulting matrix after extraction is denoted as $Y=K^{C_{\varphi }}\alpha ,\alpha =({\alpha }_1,\dots ,{\alpha }_t)$, which represents the feature components after dimensionality reduction.

The Gaussian kernel exhibits excellent local and global capabilities, providing a better balance between local and global kernel performance. The polynomial kernel also has relatively strong kernel mapping capabilities, making it suitable for most data analysis tasks. Therefore, this paper employs a linear fusion of the Gaussian and polynomial kernels. The expression for the combined kernel function is shown in Equation (5).

$$\begin{array}{l}k(x_i\cdot x_j)=\beta \cdot [\exp (-{‖x_i-x_j‖}^2/2{\sigma }^2)]\\ +(1-\beta )\cdot {[(x_i\cdot x_j)+C]}^{d_1}\end{array},$$

(5)

In this context, β represents the weighting coefficients for the Gaussian and polynomial kernels, with values $0\le \beta \le 1$. The parameter σ is the kernel parameter for the Gaussian function, and d₁ is the kernel parameter for the polynomial function. There are various ways to combine kernel functions, but each combined kernel must satisfy Mercer’s condition. If K₁ and K₂ are kernels on X × X, then it can be proven that the kernel K also satisfies Mercer’s theorem. In the feature space after KPCA projection, the degree of dispersion among the samples is defined as “dispersion”, thus obtaining dispersion.

$$\begin{array}{l}{\overline{S}}_W={\displaystyle \sum _{j=1}^2{\displaystyle \sum _{p\in N}{({\eta }^T{x}_p-{\eta }^T{\overline{\mu }}_j)}^2}}\\ ={\eta }^T[{\displaystyle \sum _{j=1}^2{\displaystyle \sum _{p\in N}(x_p-{\mu }_j){(x_p-{\mu }_j)}^T}}]\eta \\ ={\eta }^T{S}_W\eta \end{array},$$

(6)

Therefore, it follows that ${\overline{S}}_b$ is obtained:

$$\begin{array}{l}{\overline{S}}_b={({\overline{\mu }}_1-{\overline{\mu }}_2)}^2={({\eta }^T{\mu }_1-{\eta }^T{\mu }_2)}^2\\ ={\eta }^T{S}_b\eta \end{array},$$

(7)

Therefore, the discriminant function FC can be obtained as follows:

$$J(\eta )={\displaystyle \frac{{\overline{S}}_W}{{\overline{S}}_b}}={\displaystyle \frac{{\eta }^T{S}_W\eta }{{\eta }^T{S}_b\eta }},$$

(8)

Through Equation (8), it is possible to derive the final function $J(\eta )$, which represents obtaining the optimal kernel parameter values. Therefore, the objective transforms into solving $J(d_1,\sigma ,\beta )$, and through the GA–PSO intelligent fusion algorithm, the optimal kernel parameter $d_1*,\sigma *,\beta *$ can be found, achieving the optimal value of the $J$ function.

3.2. GA–PSO Hybrid Algorithm

In the composite kernel function, each individual kernel has control over its own kernel parameters and is assigned a weight coefficient that determines its contribution. As a result, the requirements for optimal selection are higher, making the optimization process more challenging. This paper optimizes the Kernel Principal Component Analysis (KPCA) of the designed composite kernel function using a hybrid approach of the Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) to select the optimal solution and establish the corresponding FC discriminant function. The KPCA kernel function is formulated through a linear combination of polynomial and radial basis kernels. The GA–PSO hybrid algorithm is then applied to optimize the three associated parameters, obtaining the optimal kernel parameters for the composite kernel KPCA [26,27]. The GA–PSO fusion mechanism is as follows: In the first step, particles are randomly initialized, and optimization is performed for each particle according to the PSO algorithm to achieve the first step of parameter optimization. In the second step, the particles corresponding to the optimal solutions are combined to form a new population, which serves as the chromosome population in the genetic algorithm. The optimized population then undergoes operations such as selection, mutation, and crossover according to the GA mechanism. Finally, the optimal solution obtained after two layers of optimization—through both PSO and GA—is selected. The first optimization step identifies the optimal particle swarm, and this optimal swarm is further refined to select the best “elite” particles. The two-step optimization process enhances the stability of the intelligent hybrid fusion algorithm and improves its ability to extract the optimal parameters. The flowchart of the GA–PSO fusion algorithm is shown in Figure 4.

3.3. Optimizing Kernel Extreme Learning Machine

The Whale Optimization Algorithm (WOA) is inspired by the predatory behavior of humpback whales, where each solution in the search process represents the position of a whale encircling its prey [28,29]. In the traditional Extreme Learning Machine (ELM) algorithm, parameters are determined through random mapping [30]. However, replacing this mapping with a kernel function resolves the instability issues associated with random parameters. Therefore, selecting an appropriate kernel function is a critical step in the KELM model computation [31]. This paper employs the WOA algorithm to optimize the two parameters of the KELM, constructing a WOA–KELM network diagnostic model to identify and diagnose clamp faults in four different states of aviation hydraulic pipelines. The steps and flowchart of the model are illustrated in Figure 5.

3.4. Establish a Diagnostic Model

Experiments were conducted by simulating the actual working conditions of aviation hydraulic pipeline clamps. The data obtained were subjected to dimensionality reduction using the optimized Kernel Principal Component Analysis (KPCA). The kernel principal components with the most significant fault characteristics were selected as the feature set and input into the Kernel Extreme Learning Machine (KELM) optimized by the Whale Optimization Algorithm (WOA). To efficiently and accurately identify and diagnose different fault modes of the clamps, further processing of the collected clamp vibration signals was necessary. This involved extracting both time-domain and frequency-domain characteristic indicators from the vibration signals. The time-domain features used in this study include the mean value, standard deviation, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, root mean square, amplitude factor, waveform factor, impact factor, margin factor, and energy. The frequency-domain features include the mean frequency, centroid frequency, frequency root mean square, and frequency standard deviation. The research model is illustrated in Figure 6.

4. Experimental Results and Analysis

4.1. Fault Simulation Experiments for Hydraulic Pipeline Clamps

The experimental setup platform is shown in Figure 7. The experimental system consists of a hydraulic power system, a hydraulic power system device, an experimental bench excitation system, and an information acquisition system.

The hydraulic pipeline clamp system is installed on the aviation hydraulic pipeline vibration test bench using fixed clamps. Vibration acceleration sensors are fixed at two measurement points. The hydraulic system parameters are set with the experimental conditions limited to a hydraulic pump motor speed of 1500 rpm, an inlet pressure of 10 MPa at the pipeline fixed clamp, and a flow rate of 30 L/min. To simulate three fault conditions in the pipeline clamp system, namely clamp liner wear, clamp base fracture, and clamp bolt loosening, we manually induce wear on the clamp liner, tear the clamp base, and loosen the bolts, respectively. Experiments are conducted for four different clamp states, with the vibration measurement points of the pipeline clamp system distributed as shown in Figure 8.

Vibration signals for different clamp states are labeled according to the specific clamp conditions, as detailed in

Table 1. Description of different clamp datasets for hydraulic pipeline clamp system.

Fault Type	Label	Measurement Point Location
Healthy state of the fixed clamp	KGJK	1
		2
Clamp padding wear fault	KGMS	1
		2
Clamp padding wear fault	KGDL	1
		2
Clamp bolt loosening fault	KGSD	1
		2

.

Due to the proximity of signal acquisition at measurement point 1 to the clamp, which better reflects the true signal of the clamp, this study selects measurement point 1 in the aviation hydraulic pipeline clamp system. Time-domain and frequency-domain waveforms of four states of fixed clamp conditions—healthy state, loose bolt state, base fracture state, and padding wear state—are measured and depicted in Figure 9.

Figure 9A–D present the time-domain and frequency-domain waveforms for the hydraulic pipeline clamps of an aviation engine in four states: healthy (KGJK), loose clamp bolt (KGSD), clamp root fracture (KGDL), and clamp pad wear (KGMS), respectively. From these Figs, it can be observed that the amplitude of the clamp vibration signal in the time domain ranges between −5 g and 5 g. In the frequency domain, a peak at around 177 Hz corresponds to the base pulsation frequency induced by the hydraulic pump, while other amplitudes reflect the frequencies generated under different clamp conditions.

Through the analysis of these time-domain and frequency-domain characteristics, it is evident that the vibration amplitude significantly increases when the clamp pad is worn, with the amplitude nearly tripling. Additionally, the vibration amplitude increases up to fourfold when the clamp bolt is loose, exhibiting noticeable vibration characteristics. The pipeline clamp’s vibration signal, influenced by the forced vibrations from pump source harmonics and the relatively harsh working environment, results in fault characteristic frequencies being interfered with by strong noise and pulsation harmonic multiples.

Therefore, it is difficult to accurately diagnose faults such as clamp pad wear, clamp root fracture, and clamp bolt loosening in the hydraulic pipeline clamps of aviation engines solely based on simple time–frequency domain waveforms of the clamp vibration signals.

4.2. Feature Extraction

Focusing on the observed and analyzed vibration signals of hydraulic pipeline clamps in various states, 40 training sets and 20 test sets were extracted for both the normal state and three fault states. Thirteen time-domain features and four frequency-domain features were used to construct the original feature dataset. Each signal group consists of 1024 sampling points, with 60 groups for each state. The original feature dataset was standardized and analyzed using the improved Kernel Principal Component Analysis (KPCA) method proposed in this paper, extracting nonlinear features.

For the vibration dataset associated with the health states of hydraulic pipe clamps in aviation, including three fault modes—clamp liner wear, clamp root fracture, and clamp bolt loosening—this paper applies the proposed GA–PSO method to optimize the kernel parameters in the combined kernel function of the Kernel Principal Component Analysis (KPCA). The optimization iteration process is shown in Figure 10. It can be observed that within 500 iterations, the fitness value reaches a minimum between 0.85 and 0.86. The larger the number of iterations, the closer the solution’s target value is to the actual function value. The number of iterations also needs to be set appropriately; too few iterations may prevent the algorithm from converging effectively, while too many iterations could increase the runtime of the algorithm, leading to a waste of resources in terms of time. Therefore, the optimal kernel parameter obtained after 500 iterations is selected as $\beta =0.124,d_1=1.27,\sigma =9.48$.

Based on experimental vibration datasets corresponding to the health states of clamps, including the normal state, clamp liner wear, clamp root fracture, and clamp bolt loosening faults, along with experimental tests of clamp conditions under these four modes, this study employs the GA–PSO optimized KPCA. The analytical outcomes are presented in Figure 11.

The feature analysis plot in Figure 11 clearly demonstrates that the three-dimensional features of the four different clamp states are distinctly separated on the projection plots of KPC1, KPC2, and KPC3. Each state accurately projects within its respective range. This indicates that the proposed Kernel Principal Component Analysis (KPCA) method, optimized with a GA–PSO fusion algorithm for combined kernel functions, effectively extracts samples from each fault mode, thereby ensuring robust data preparation for subsequent classification models. The cumulative contribution rates of clamp states across the four different conditions and under various kernel functions are compared in

Table 2. The accumulative contribution rate of core principal component analysis after the fixed clamp was optimized in different states.

Serial Code	WGJK	WGMS	WGDL	WGSD
1	0.9999	0.9999	0.9998	0.9999
2	1.0000	1.0000	0.9999	1.0000
3	1.0000	1.0000	1.0000	1.0000
4	1.0000	1.0000	1.0000	1.0000

and Figure 12.

From Table 2 and Figure 10, it is evident that the cumulative contribution rates of the top ten principal components from the optimized combined Kernel Principal Component Analysis (KPCA) approach reach 0.9999 for each principal component across different states. For nearly all states, the cumulative contribution rates of the three principal components collectively reach 100%. The mapped results show clear classification of categories with distinct inter-class boundaries. The intelligent fusion algorithm proposed in this study optimizes combined kernel functions for KPCA, achieving effective feature extraction using a single principal component.

4.3. Fault Diagnosis and Comparative Analysis

Based on signals collected from vibration experiments on aviation hydraulic pipe clamp faults induced by pump sources, a fault diagnosis model, GA–PSO–KPCA–WOA–KELM, was constructed. To validate the stability and superiority of this proposed model, the impact of KPCA pre- and post-improvement, as well as the network model, on diagnostic performance was investigated. Extracted feature vectors were used as input to the WOA–KELM network diagnostic model, producing classification plots, as shown in Figure 13. For comparative analysis, the accuracy of the pre-optimized and post-optimized KELM fault diagnosis models was evaluated against SVM and BP network models using the same vibration signal data in 10 repeated tests. The statistical comparison of averaged results is presented in

Table 3. Comparison of accuracy rates of different fault diagnosis models.

Diagnostic Model	Standard Deviation	Precision/%
GA–PSO–KPCA–WOA–KELM	0.00005164	0.9999
GA–PSO–KPCA–KELM	0.0042	0.9835
WOA–KELM	0.0074	0.9334
KELM	0.0043	0.8366
GA–PSO–KPCA–BP	0.0051	0.9155
KPCA–BP	0.0062	0.8053
GA–PSO–KPCA–SVM	0.0045	0.9259
KPCA–SVM	0.0098	0.8209

.

From Figure 13 and Table 3, it can be observed that using the feature set derived from GA–PSO–KPCA as input to the WOA–KELM network model achieves an average classification accuracy of 99.99%. At the same time, the standard deviation of repeated experiments for the fault classification method proposed in this paper is the lowest among all models, which demonstrates the robustness and stability of the proposed fault classification method. The average classification accuracies from repeated tests are as follows: the GA–PSO–KPCA–KELM model is 98.35%, the WOA–KELM model is 93.34%, the KELM model is 83.66%, the GA–PSO–KPCA–BP model is 91.55%, the KPCA–BP model is 80.53%, the GA–PSO–KPCA–SVM model is 92.59%, and the KPCA–SVM model is 82.09%. Compared to the diagnostic model proposed in this paper, the average accuracies decreased by 1.64%, 6.65%, 16.33%, 8.44%, 19.46%, 7.4%, and 17.90%, respectively. This indicates that different network models have varying impacts on the classification accuracy of feature indicators, with the WOA–KELM network model demonstrating significant advantages in stability and classification accuracy. It also underscores that the proposed fault diagnosis model in this study enhances the accuracy of clamp fault diagnosis, achieving an accuracy of 99.99%.

5. Conclusions

This paper attempts to simulate the real-world environment of an aerospace hydraulic pipeline as accurately as possible. The main research conclusions on the fault identification and diagnosis of aerospace hydraulic pipeline clamp failures are presented in three sections: theoretical analysis of vibration signal processing and feature extraction methods, experimental testing and analysis of aerospace hydraulic pipeline clamp vibrations, and fault diagnosis of aerospace hydraulic pipeline clamp failures. These conclusions are summarized as follows:

1

The GA–PSO hybrid algorithm optimizes multiple kernel parameters for improved KPCA, establishing a feature extraction method model based on Kernel Principal Component Analysis (KPCA). Through validation studies, it was demonstrated that a single principal component contributes 99.99% of the original sample data attributes, surpassing single-kernel KPCA and unprocessed PCA methods. This approach exhibits excellent feature extraction capability, thereby validating the feasibility of the proposed feature extraction method.

2

This study designed a vibration testing experimental plan for fixed clamps in aviation engine hydraulic pipelines, focusing on the aviation hydraulic pipe clamp system. Vibration experiments were conducted to evaluate the health status of fixed clamps and simulate conditions such as clamp root fractures, liner wear, and clamp bolt loosening. Test data were collected from different clamp states within the same pipeline, and both time-domain and frequency-domain analyses were performed. The paper proposes using the Kernel Principal Component Analysis (KPCA) with parameters optimized using a GA–PSO fusion algorithm to analyze clamp vibration signals. Comparative analyses were conducted between different kernel functions and unoptimized combined kernel functions for KPCA of clamp data. The results validate that the proposed method efficiently and accurately maps and extracts vibration data from clamps. Data analysis demonstrates that the method effectively identifies different clamp fault states across various pipelines, thereby providing robust data support for subsequent fault diagnosis.

3

Experimental vibration signals were subjected to time-domain and frequency-domain feature extraction, resulting in a dataset comprising 13 time-domain features and 4 frequency-domain features. This combined dataset was input into a feature extraction model, yielding processed feature datasets via GA–PSO–KPCA. Subsequently, these features were incorporated into a Whale Optimization Algorithm (WOA)-enhanced Kernel Extreme Learning Machine (KELM) network model to achieve fault classification, recognition, and diagnosis of clamps. The proposed GA–PSO–KPCA–WOA–KELM fault diagnosis model demonstrated an average accuracy of 99.99%. This validates the method’s stability, effectiveness, and feasibility in extracting and identifying fault characteristics in aviation hydraulic pipe clamps.