Nondestructive Evaluation of Aging Failure in Insulation Coatings by Ultrasonic Guided Wave Based on Signal Processing and Machine Learning

Abstract

In the field of nondestructive evaluation (NDE) using ultrasonic guided waves, accurately assessing the aging failure of insulation coatings remains a challenging and prominent research topic. While the application of ultrasonic guided waves in material testing has been extensively explored in the existing literature, there is still a significant gap in quantitatively evaluating the aging failure of insulation coatings. This study innovatively proposes an NDE method for assessing insulation coating aging failure by integrating signal processing and machine learning technologies, thereby effectively addressing both theoretical and practical gaps in this domain. The proposed method not only enhances the accuracy of detecting insulation coating aging failure but also introduces new approaches to non-destructive testing technology in related fields. To achieve this, an accelerated aging experiment was conducted to construct a cable database encompassing various degrees of damage. The effects of aging time, temperature, mechanical stress, and preset defects on coating degradation were systematically investigated. Experimental results indicate that aging time exhibits a three-stage nonlinear evolution pattern, with 50 days marking the critical inflection point for damage accumulation. Temperature significantly influences coating damage, with 130 °C identified as the critical threshold for performance mutation. Aging at 160 °C for 100 days conforms to the time-temperature superposition principle. Additionally, mechanical stress concentration accelerates coating failure when the bending angle is ≥90°. Among preset defects, cut defects were most destructive, increasing crack density by 5.8 times compared to defect-free samples and reducing cable life to 40% of its original value. This study employs Hilbert–Huang Transform (HHT) for noise reduction in ultrasonic guided wave signals. Compared to Fast Fourier Transform (FFT), HHT demonstrates superior performance in feature extraction from ultrasonic guided wave signals. By combining HHT with machine learning techniques, we developed a hybrid prediction model—HHT-LightGBM-PSO-SVM. The model achieved prediction accuracies of 94.05% on the training set and 88.36% on the test set, significantly outperforming models constructed with unclassified data. The LightGBM classification model exhibited the highest classification accuracy and AUC value (0.94), highlighting its effectiveness in predicting coating aging damage. This research not only improves the accuracy of detecting insulation coating aging failure but also provides a novel technical means for aviation cable health monitoring. Furthermore, it offers theoretical support and practical references for nondestructive testing and life prediction of complex systems. Future studies will focus on optimizing model parameters, incorporating additional environmental factors such as humidity and vibration to enhance prediction accuracy, and exploring lightweight algorithms for real-time monitoring.

1. Introduction

With the rapid development of aerospace technology, the safety and reliability of aircraft cables, as the key components connecting various aircraft systems and transmitting electrical energy and signals, are directly related to the overall performance and operational safety of the aircraft. During the long-term use of cables, the insulation layer may be gradually damaged due to a variety of factors such as environmental erosion, mechanical wear, and aging, which may lead to short-circuit, signal interruption, and other faults, and in serious cases, may even lead to flight accidents [1,2,3]. Therefore, regular, efficient, and non-destructive testing of aircraft cables, timely detection, and early warning of potential failures, is of great significance to ensure flight safety, extend the service life of the cables, and reduce maintenance costs.

For the detection needs of aircraft cables, there exist a variety of nondestructive testing methods, each with its characteristics and limitations [4,5]. The visual inspection method is simple and intuitive, but it is difficult to find subtle defects [6,7]. The insulation resistance testing method can effectively determine the insulation status of the cable, but it is limited by the testing conditions [8]. The pulsed spark discharge method [9,10], time-domain reflectometry [11,12] and frequency-domain reflectometry [13,14] are highly accurate in locating cable faults, but they are complicated to operate, susceptible to interference, or have limited scope of application. Although the impedance detection method is simple, the error is large, and the operation is cumbersome [15,16]. The extended-spectrum time-domain reflection method is suitable for online detection, but the detection efficiency of complex cable structures is low [17]. All of the above traditional inspection methods must be carried out in the non-electrified state of the wire (except for the visual inspection method and the extended-spectrum time-domain reflectance method), which is no longer adapted to the fast-changing and rapidly developing aerospace field. In addition to some of the above aircraft cable detection methods, there are some other detection methods such as the traveling wave method [18], breakdown method [19], noise domain reflection method [20], Brillouin optical time domain reflection method [21], and other intelligent detection methods [22], and some other forms of these cable detection methods, which also have their own limitations.

Based on traditional detection methods, ultrasonic-guided wave detection technology, as an emerging nondestructive testing method, has shown significant advantages in recent years in the detection of pipelines, anchors, and other structures [23,24]. Based on the theory of elastic mechanics and wave dynamics, Narendar, Ghorbanpour, and others have utilized ultrasonic-guided wave propagation characteristics in the medium, which can realize efficient and remote defects within the structure detection [25,26,27,28]. For the detection needs of aircraft cable insulation integrity, ultrasonic guided wave technology not only overcomes the shortcomings of traditional methods, improves the detection length and efficiency, but also can be used to provide early warning before the failure of the cable, effectively eliminating potential risks. However, the ultrasonic-guided wave signal is susceptible to cable structure, environmental noise, and other factors, and the mapping relationship between the aging of the insulation coating and the ultrasonic-guided wave characteristics of the parameters is complex, and it is difficult to establish an accurate mathematical model to describe it. The core problem of this study is how to establish the accurate mapping relationship between the ultrasonic conductive wave signature parameters and the aging state of the insulation coating, and to realize the quantitative evaluation of the aging failure of the insulation coating. Therefore, this study assumes that the organic combination of optimization algorithm and neural network can significantly improve the predictive performance of the model and provide a theoretical basis and technical support for the intelligent operation and maintenance of aircraft cables.

With the rapid development of artificial intelligence technology, artificial neural networks (ANN) provide new ideas for the modeling and prediction of complex nonlinear systems through their powerful nonlinear mapping ability and self-learning ability [29,30]. Introducing ANN into the field of ultrasonic guided wave signal processing can effectively extract the signal features and establish the mapping relationship between the signal features and the aging state of insulating coatings, to realize the intelligent assessment of the aging state of insulating coatings. However, a single ANN model often suffers from insufficient generalization ability and tends to fall into local optimization [31,32].

To overcome the limitations of a single neural network model, this study proposes a hybrid ANN-based ultrasonic guided wave nondestructive characterization method for the aging failure of insulating coatings on cable cores for aircraft. The method organically combines optimization algorithms (Genetic Algorithm (GA), Particle Swarm Algorithm (PSO), etc. [33,34]) with the underlying neural network models (BP neural network, Support Vector Machine (SVM), etc. [35,36]), and takes full advantage of the strengths of each model to construct a hybrid ANN model, to improve the prediction accuracy and generalization ability of the model. Akram et al. [37] used Support Vector Machine (SVM) to classify ultrasonic guided wave signals, and realized the preliminary evaluation of the corrosion degree of metal pipes. However, these studies mainly focus on the detection of metal materials, and there is still a lack of effective methods for the quantitative evaluation of aging failure of insulation coatings. Compared with Akram et al.’s research [37], this study not only applies machine learning technology to the processing and classification of ultrasonic guided wave signals but further realizes the quantitative evaluation of aging failure of insulation coatings. At the same time, this study will explore the ultrasonic guided wave propagation characteristics of the insulating coating of cable cores under different aging degrees from multiple perspectives, establish the mapping relationship between the ultrasonic guided wave characteristic parameters and the aging state of the insulating coating, and provide a theoretical basis and technical support for the intelligent operation and maintenance of aircraft cables. The aim of this study is to propose a nondestructive detection method for aging failure of aircraft cable insulation layer based on ultrasonic and artificial intelligence (AI) technology. The method will combine the detection accuracy of ultrasonic technology and the data processing capability of artificial intelligence technology to achieve the accurate evaluation of the aging state of the aircraft cable insulation layer.

In this study, by constructing a database of aircraft cables with different degrees of damage, combined with the ultrasonic guided wave detection technology, we systematically studied the defect evolution law and its detection method during the aging process of cables. The study firstly simulates the aging process of cables in the actual service environment through accelerated aging experiments, presets defects such as abrasion and cuts, and records new defects such as cracking and peeling of coatings generated during the aging process. On this basis, the defects were accurately measured and classified using metallurgical microscopy and scanning electron microscopy. To address the noise problem in the ultrasonic inspection data, this study innovatively used the Fast Fourier Transform (FFT) and Hilbert–Huang Transform (HHT) for signal preprocessing [38,39]. On this basis, this study compares the performance of three classification algorithms, decision tree, XGBoost, and LightGBM, in defect classification [40,41,42]. The modeling effects before and after data classification are compared to provide a reliable database for subsequent damage prediction. In terms of machine learning model construction, this study adopts two algorithms, BP neural network and SVM, and also introduces GA and PSO to optimize the model parameters and analyzes and compares the effects of different algorithms. This study not only provides a new technical means for the health monitoring of aviation cable but also provides a valuable reference for the nondestructive testing of other complex systems. By revealing the mapping relationship between ultrasonic guided wave signature parameters and insulation coating aging state, this study is expected to provide a theoretical basis and technical support for intelligent operation and maintenance of aircraft cables, which has important theoretical significance and engineering application value.

2. Materials and Methods

2.1. Specimen Preparation and Defect Detection

The samples in this study were obtained from cables replaced in a type of aircraft after 10 years of service. The cable is not faulty but has aged to a certain extent. In general, constructing machine learning models with a single-source specimen as an input is not of general significance and generalization value. For this reason, in this study, aging experiments are used to construct a database of cables with different levels of damage. The insulating coating material used in this study is irradiated crosslinked ethylene copolytetrafluoroethylene (ETFE). ETFE is a fluoropolymer with excellent chemical resistance, weather resistance, low dielectric constant and good mechanical properties, which make it an ideal choice for insulating coatings for aviation cables. To accelerate the aging process, in this paper, the outer protective layer and shielding layer are peeled off before aging experiments are conducted. The aging experiment is performed according to the industrial aging standard, and the parameters of the aging experiment are shown in

Table 1. Table of aging experiment parameters [43].

Aging Time/Day	Aging Temperature/°C	Bending Angle/°	Preset Defects
0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100	Room temperature, 100, 130, 160	0, 45, 90, 135, 180	No defects, abrasions, cuts

[43]. In the production process of cable cores. In some specimens, defects such as abrasions and cuts are preset in advance. The aging time does not exceed 100 days, and part of the specimens are removed from inside the aging experiment box every 10 days for defect detection and ultrasonic-guided wave characterization. To more closely match the aging in real-world applications, the aerospace cables were bent to different degrees in this study, and the degrees of bending are listed in Table 1. In addition to the preset defects, two new defects, coating cracking, and coating peeling, are generated during the aging process. When the aging temperature was discussed, low temperature was not paid attention to in this study. This study aims to solve the problem of possible aging failure of insulation coatings in high temperature environments in electric power, aerospace, and other industrial fields. Considering that equipment in these fields often needs to operate at high temperature, studying high temperature aging is crucial to improve equipment reliability and safety. In view of the key problem that the low temperature aging time is too long, the study chooses high temperature conditions for research because it can observe obvious aging phenomenon in a short time so as to improve the research efficiency and reduce the cost. The main goal is to use signal processing and machine learning techniques to achieve non-destructive evaluation of the aging failure of insulation coatings by constructing an ultrasonic guided wave-based detection method, and use high temperature to accelerate the aging process and shorten the experimental period to collect enough data to train and verify the model. The process of preparing specimens of cable cores for aircraft and this study is shown in Figure 1.

After the specimens were prepared, defects were measured, classified, and calibrated using a metallurgical microscope (VHX-7000, Keens (China) Co., Ltd., Shanghai, China) and a scanning electron microscope (SEM5000X, Guoyi Quantum Technology Co., Ltd., Hefei, China), and 660 sets of data were obtained. Among them, 220 groups of specimens each were without defects, with wear defects, and with cut defects. Each group of specimens was inspected using ultrasound, and the principle of ultrasonic inspection flaw detection is shown in Figure 2.

2.2. Data Processing

In this work, ultrasonic waves were utilized to measure the specimens with different damage degrees, but the raw data of the inspection usually contain more clutter and noise signals, and the ultrasonic data need to be pre-processed. Combined with the existing research data, we propose to use FFT and HHT to filter the ultrasound signals and compare the processing effects of the two. The FFT is a highly efficient and fast computational method of the Discrete Fourier Transform (DFT), which is used to reduce the number of multiplications required by the computer to compute the DFT. Especially when dealing with large datasets, the advantages of FFT are more obvious. FFT can provide information about the frequency components of the signal, which is very effective for the analysis of smooth signals, and for non-smooth signals (e.g., signals whose frequency varies over time), FFT cannot provide enough information. For this reason, HHT is chosen to compare with FFT, which is particularly suitable for analyzing nonlinear and nonsmooth signals because it can capture the local time-varying characteristics of signals, and the empirical modal decomposition (EMD) step of HHT can automatically generate “bases” according to the actual characteristics of signals, which makes HHT highly adaptive. Thereafter, the gap between the two noise reduction algorithms is comparatively analyzed, and the more effective model is used for subsequent prediction.

2.3. Machine Learning Model Building

Existing research data show that machine learning algorithms are highly advantageous in large-scale data processing. For this reason, excellent machine learning is introduced to process ultrasound data in this work. Combining the characteristics of ultrasound data, the BP neural network and SVM are chosen to construct the model [35,36]. BP neural network is a multi-layer feed-forward neural network, which trains the network by forward propagation of signals and backward propagation of errors. The principle lies in the use of gradient descent to adjust the weights and biases in the network to minimize the error between the actual output and the desired output. The advantage of the BP neural network lies in the powerful nonlinear mapping ability and self-learning ability, which is suitable for complex nonlinear problems. However, it also suffers from the disadvantages of long training time, difficult parameter selection, and poor model interpretability. SVM, on the other hand, is a classification method based on statistical learning theory, which aims to find an optimal hyperplane to separate samples of different classes and maximize the spacing between the hyperplane and the two classes of samples. SVM maps the input data to a high-dimensional space by means of the kernel function to provide better linear segregation. Its advantages include high efficiency, strong generalization ability, and adaptability to small sample datasets. However, SVM also suffers from the disadvantages of long training time (for large-scale datasets), sensitive parameter selection, sensitivity to noise, and does not directly provide probability estimates.

To solve the problem of difficult parameter selection of the two models, in this work, we choose to use GA and PSO to optimize the initial weights and thresholds of both [33,34]. Genetic algorithms to simulate the mechanism of natural selection and genetics by encoding the solution space as a chromosome and gradually searching for the optimal solution using selection, crossover, and mutation operations, which are characterized by strong global search capability, good parallelism, robustness, and wide applicability, but the computation time may be long, and the solution has randomness. Particle swarm algorithm, on the other hand, is based on group intelligence, simulates the foraging behavior of bird flocks, and updates the speed and position of the particles through the information sharing and cooperation between particles, using the individual optimum and global optimum to search for the optimal solution, which is characterized by the fast convergence speed, simple and easy to implement, flexible parameter adjustment, and strong local search ability, but it may be trapped in the local optimum due to the overly strong interactions. The genetic algorithm focuses more on global search and parallelism, while the particle swarm algorithm has more advantages in convergence speed and local search ability.

To improve the prediction accuracy of the model, a classification algorithm is used in this study to classify the defects before making predictions. Meanwhile, to ensure accuracy when classifying, this study compares and analyzes the effects of three algorithms: decision tree classification, XGBoost classification, and LightGBM classification [39,40,41]. The decision tree is based on tree structure and categorizes data through node judgment. Its advantages are intuitive and easy to understand, robust to noise, and can handle multiple classification and feature selection. However, it is prone to overfitting and requires pruning; it is sensitive to noise, and the model complexity increases with data. XGBoost is based on a gradient boosting tree, which iteratively trains weak classifiers (e.g., decision trees) to optimize the model and improve prediction accuracy. It handles large-scale data efficiently, can process in parallel, and provides feature importance assessment. However, parameter tuning is complex, memory consumption is large, and complex models or noisy data are easy to overfit. LightGBM uses a histogram algorithm and depth-first construction of trees to accelerate training and improve efficiency. It has fast training speed, better results than XGBoost, etc., and supports large-scale data processing. However, it is sensitive to parameters and may overfit when there are few features. The best-performing model is selected to classify the ultrasound detection data to ensure more accurate subsequent prediction.

In this work, a total of 660 sets of data were obtained, 80% of which were randomly selected as the training set and the remaining 20% as the test set. The choice of this scale is based on common machine learning practices designed to ensure that the model can perform well on previously unseen data. The training process uses the mean square error (MSE) as the function iteration target, all iterating in the direction of error reduction. To ensure the reliability of the model while controlling reasonable training practices, the number of iterations was set to 500. To verify the accuracy of the model, the root mean square error (RMSE) and the correction coefficient of determination (R²_adj) were used for evaluation. In this case, R²_adj needs to be solved using the coefficient of determination (R²). MSE, RMSE, R²_adj, and R² are defined as follows:

$$MSE={\displaystyle \sum _{i=1}^n{\left(Y_i-{\widehat{Y}}_i\right)}^2}/n$$

(1)

$$RMSE=\sqrt{{\displaystyle \sum _{i=1}^n{\left(Y_i-{\widehat{Y}}_i\right)}^2}/n}$$

(2)

$$R_{adj}^2=1-{\displaystyle \frac{(1-R^2)(n-1)}{n-p-1}}$$

(3)

$$R^2={\left[{\displaystyle \frac{{\displaystyle \sum _{i=1}^n\left({\widehat{Y}}_i-\overline{\widehat{Y}}\right)}\left(Y_i-\overline{Y}\right)}{\sqrt{{\displaystyle \sum _{i=1}^n{\left({\widehat{Y}}_i-\overline{\widehat{Y}}\right)}^2}}\sqrt{{\displaystyle \sum _{i=1}^n{\left(Y_i-\overline{Y}\right)}^2}}}}\right]}^2$$

(4)

where $Y_i$ was the actual value, ${\widehat{Y}}_i$ was the predicted value, n was the number of test sets, $\overline{\widehat{Y}}$ was the mean of the predicted value, $\overline{Y}$ was the mean of the actual value and p was the number of features.

3. Results and Discussion

3.1. Effect of Different Aging Parameters on the Insulation Coating of Cable Cores

Aging experiments were conducted on selected aircraft cable cores to investigate the effects of different parameters on their life spans through the control variable method. During the aging process of the cable core coating, its failure mainly includes cracking and peeling, and the SEM microstructure of the specimen before and after the failure is shown in Figure 3. The macroscopic morphology of the specimen before and after the aging test is shown in Figure 4, which will make the surface of the specimen undergo different degrees of oxidation before and after the test (the color changes from light yellow to dark brown).

When exploring the effect of aging time on the coating, the aging temperature was set at 160 °C, the bending angle was 180°, and no defects were preset. All the examined specimens were counted, and the effects of different aging times on defect damage are shown in Figure 5a. From Figure 5a, it can be seen that the degree of damage to the insulating coating increases nonlinearly with aging time. In the specimen without preset defects, the aging time has a significant cumulative effect on the insulating coating damage, and its effect shows a nonlinear acceleration characteristic. The experiments showed that the aging time showed a three-stage nonlinear evolution on the damage of insulating coatings. At the initial stage, from 0 to 50 days, the crack density and peeling area increase slowly, mainly due to surface oxidation, and then it enters into the accelerated failure stage after 50 days, with the crack density and peeling area increasing to 1.8 strips/mm² and 3.2% (at 80 days), and then it reaches the severe failure stage at 100 days, with the crack density surging to 4.2 strips/mm² (an 11-fold increase compared with that of 50 days), the length of cracks reaching 2.5 mm, and the proportion of peeling area soaring to 2.5 mm, and the percentage of peeled area soared to 18.4%. These results suggest that 50 days is the critical inflection point for damage accumulation, and the exponential increase in the rate of failure at the latter stage may be due to polymer chain breakage and volatilization of the plasticizer, which in turn leads to embrittlement failure.

Increasing temperature accelerates coating failure, and the effect of different aging temperatures on the aging of aircraft cable core coatings is shown in Figure 5b. In exploring the effect of aging temperature on the coating, the aging time was set to 100 days, the bending angle was 180°, and no defects were preset. The aging temperature significantly accelerated the coating deterioration through thermal stresses, with 130 °C as the critical threshold for sudden performance changes. Aging at room temperature (25 °C) triggered only minor damage (crack density 0.8 cracks/mm², peel area 0.8%). When the temperature was raised to 130 °C, the crack density increased abruptly to 3.5 cracks/mm² (338% increase), and the peel area reached 11.2%. Coating damage at 160 °C was equivalent to 100 days of long-term aging (peel area 18.4% in both cases). Crack density and peel area increased approximately twofold for every 30 °C increase in temperature, indicating a sharp increase in the rate of crosslink structure destruction and a sudden drop in thermal stability at high temperatures. The prolongation of both aging time and aging temperature accelerates the degradation of the coating, which is in line with the principle of time-temperature superposition, indicating that the high-temperature acceleration experiment can effectively predict the degradation of the coating’s long-term performance.

Unreasonable placement of cables during use also accelerates the aging of the cores, thus posing a potential risk to the safety of aircraft in service. In the study of the effect of bending angle on the aging life of the coating, the aging time was set to 100 days, the aging temperature was 160 °C, and no defects were preset. The effects of different bending angles on coating aging are shown in Figure 5c. From Figure 5c, it is easy to see that the bending angle dominates the coating failure through mechanical stress concentration, and 90° is the critical angle for mechanical damage. Bending at 0°~45° only triggers micro-cracks (density ≤ 0.5 strips/mm², stripping area ≤ 0.2%), and 90° bending increases the crack density to 2.1 strips/mm² (an increase of 320% compared with 45°), and the stripping area reaches 3.5%. At 180° bending, the damage peeling area was 18.4%. The stress concentration caused by 180° bending can also lead to the aging failure of the coating, indicating that the stress concentration caused by large-angle bending can be equivalent to high-temperature thermal stress, and they jointly accelerate the aging failure of the coating by destroying the cross-linked polymer structure.

Cable cores may be manufactured with defects such as abrasions or cuts in the coating, which may accelerate the aging of the coating and lead to failure of the core within its expected lifetime. In exploring the effect of such defects on coating aging, the aging time was set to 100 days, the aging temperature to 160 °C, and the bending angle to 180°. The effect of different preset defects on coating aging is shown in Figure 5d. From Figure 5d, it is easy to see that the preset defect type significantly changes the damage extension path, and the cut defect is the most destructive. The defect-free specimens were dominated by uniform oxidation (crack density of 4.2 cracks/mm², peeling area of 18.4%). Wear defects increased the crack density to 12.0 cracks/mm² due to surface roughening (186% increase compared to the non-defective specimens), with a stripped area of 24.0%. The cut defects resulted in a crack density of 24.5 cracks/mm² (5.8 times higher than that of the non-defective group) due to stress concentration, with 30.5% of the stripped area. Cracks of the cut defects expanded rapidly during the aging experiment, which shortened the lifetime to 40% of the non-defective group. The above experimental results show that the cut defects increase the crack density to 5.8 times of the non-defective group through stress concentration, indicating that the initial defects are the key risk factor for coating life prediction and need to be avoided as a priority in the cable design stage. Therefore, improving the quality control of cable manufacturing, especially in terms of reducing defects such as cutting, is crucial. This not only extends the service life of the cable but also reduces maintenance costs caused by frequent cable replacement.

Based on the above experimental results, in actual service, it is necessary to shorten the inspection period to less than 50 days for the high-temperature region above 130 °C and ≥90° bending section to avoid safety accidents due to the accelerated failure of the cable coating. The real service parameters, such as humidity and vibration, cannot be equivalently replaced in the experimental process, and the real aging rate of the coating may be higher. Subsequent aging tests will be re-designed to more closely match the real service environment. For practical application, the HHT time-frequency analysis technique can be accelerated and combined with portable guided wave equipment to realize in situ quantitative assessment of coating damage.

In this study, significant effects of aging time and temperature on coating failure were indeed observed, and these effects can largely be attributed to aging mechanisms such as polymer chain breakage and plasticizer evaporation. Polymer chain fracture refers to the fracture of polymer molecular chain under the action of high temperature, mechanical stress, or chemical corrosion, resulting in the reduction in polymer molecular weight and material properties. This change will make the coating material hard and brittle, which is more prone to damage such as cracks and peeling. Plasticizer evaporation refers to the gradual volatilization of the plasticizer in the coating under high temperature or long-term exposure, resulting in the loss of flexibility of the coating material and becoming more prone to cracking. In the experiment of this paper, the damage of insulation coating by aging time presents a three-stage nonlinear evolution, in which 50 days is the critical inflection point of damage accumulation. During the accelerated failure period after 50 days, the crack density and peel area increased significantly, which may be related to the polymer chain fracture and the intensification of plasticizer volatilization. At the same time, it is also observed that the failure rate of the crosslinked structure increases sharply, and the thermal stability drops sharply under a high temperature environment, which further confirms the existence of aging mechanisms such as polymer chain fracture and plasticizer evaporation and their effects on the coating properties. In order to model the aging process of the coating more accurately, this paper needs to have a deeper understanding of the specific mode of action and the relationship between these aging mechanisms. In future studies, this paper will aim to quantitatively evaluate the degree of polymer chain breakage and plasticizer evaporation through chemical analysis, physical testing, and other means, as well as their specific effects on coating properties. At the same time, this paper will also try to incorporate these aging mechanisms into the coating aging model to improve the accuracy and reliability of the model.

3.2. Ultrasonic Guided Wave Detection and Signal Processing Results

During the aging experiment, the specimens were taken out for ultrasonic-guided wave detection at intervals of 10 days. The raw signals of ultrasonic-guided wave detection are shown in Figure 6a,b. It can be seen from Figure 6a,b that the original ultrasonic signals have more clutter regardless of whether the specimens have defects or not, which leads to the difficulty of extracting the feature signals therein and also makes the subsequent modeling process more difficult. For this reason, both FFT and HHT are used for signal processing in this study. The preprocessing results of the original signals are shown in Figure 6c–f. As can be seen from Figure 6c,e, for the specimen without damage, the effect of noise reduction using HHT is much better than that of FFT. Similarly, for the specimen with damage, the noise reduction effect of HHT is also better than that of FFT, and it can be seen from the noise reduction results of HHT that the ultrasonic guided wave signals can be effectively characterized after noise reduction when there are bends or damages in the specimen, and different types of defects will have different characteristic signal performances. From Figure 6f, it is not difficult to find that the degree of damage is different, the amplitude of the performance will also change accordingly. Therefore, the ultrasonic guided wave signal after HHT noise reduction can be used as a characterization means for detecting cable coatings. Ultrasonic guided wave combined with HHT noise reduction has a high sensitivity for bending and damage, and it can be used for early detection of micro-defects in cable coatings.

In order to more fully evaluate the application effect of FFT and HHT in the aging detection of cable coatings, a supplementary experiment was conducted to compare the performance of the two algorithms in terms of computation time, memory usage, and real-time applicability. The experimental results show that the FFT algorithm has an obvious advantage in the calculation time, the average calculation time is 0.0015 s, and the average calculation time of HHT algorithm is 0.012 s. In terms of memory usage, the memory required by the FFT algorithm is 1.8 MB, which is lower than the 3.5 MB of the HHT algorithm. In terms of real-time applicability, FFT algorithm is more suitable for real-time signal processing applications because of its lower computational complexity and faster processing speed. However, it should be emphasized that in the specific application of cable coating aging detection, HHT algorithm, although it has higher computational complexity and memory requirements, has better noise reduction effect and can extract characteristic signals more effectively, so it has higher sensitivity for detecting microdefects of cable coating. Therefore, when selecting a signal processing algorithm, it is necessary to comprehensively consider the computational efficiency, memory requirements, noise reduction effect, and real-time applicability of the algorithm to find the most suitable algorithm for the application scenario.

3.3. Machine Learning Modeling Results

After obtaining the noise-reduced data, the ultrasonic guided wave prediction cable defect regression prediction model is constructed using the noise-reduced data as input and the statistically obtained damage parameters as output. When the prediction model was constructed in the previous stage, the ultrasonic-guided wave signals of certain defects were closer together, resulting in frequent misjudgment of damage. This can lead to the occurrence of cable life before the replacement of the situation. Therefore, to further improve the accuracy of prediction, this study first classifies the data obtained from inspection. The categorized data can then be used to predict the degree of damage to improve the accuracy and robustness of the model. In this study, three classification algorithms, decision tree, XGBoost, and LightGBM, are selected for modeling, respectively, and the training results of the three classification models are shown in Figure 7. From Figure 7, we can see that the training model constructed by LightGBM has an accuracy and precision close to 1, the training model constructed by XGBoost has an accuracy above 0.9, and the training model constructed by decision tree does not exceed 0.9 in terms of accuracy and precision, etc. Among the three algorithms of decision tree, XGBoost, and LightGBM, the model constructed by LightGBM had the best training results, followed by the XGBoost model, and the decision tree model had the worst training results.

The effect of the training model cannot illustrate the prediction effect, and it is necessary to use new data to test the prediction effect of the model, and the prediction effect of the three models is shown in Figure 8. To distinguish the evaluation method from the training model, Receiver Operating Characteristic Curve (ROC) curve and Area Under the Curve (AUC) values are used in this study for the comparison of the prediction effect of the models. The results of the comparison of the predictive effectiveness of the different classification models are shown in Figure 8. From Figure 8, we can see that the decision tree model has an AUC value of 0.72, the worst performance, the AOC curve is close to the diagonal, and the classification ability is limited, the XGBoost model has an AUC value of 0.85, the performance is medium, the AOC curve is significantly higher than the decision tree, and the classification ability is stronger, and the LightGBM model has an AUC value of 0.94, the optimal performance, and the AOC curve is the closest to the upper-left corner, which is significantly better than the former two. Classification ability is significantly better than the first two. The results show that the LightGBM algorithm still maintains an excellent classification effect, and its prediction accuracy reaches 0.94 for unknown data.

The classified data of the LightGBM model is used as the input of the damage prediction model, and the results of the aging experiment are used as the output to construct the aging damage prediction model for the coating of the wire core of aviation cables. (1) To prove whether the classification model is effective or not, the unclassified data are used as the input, and the results of the aging experiment are used as the output to construct the aging damage prediction model for the coating of the wire cores of aviation cables in the study. (2) The training results of the above models are shown in

Table 2. Training effects of aging injury prediction Model 1.

Model	Optimization Algorithms	RMSE	R2adj	Training Time/s
BP	None	0.25	0.82	310
	GA	0.06	0.98	520
SVM	None	0.20	0.89	290
	PSO	0.03	0.99	480

and

Table 3. Training effects of aging injury prediction Model 2.

Model	Optimization Algorithms	RMSE	R2adj	Training Time/s
BP	None	0.42	0.65	420
	GA	0.24	0.83	670
SVM	None	0.32	0.73	390
	PSO	0.21	0.87	580

. As can be seen from Table 2 and Table 3, the optimization algorithm can substantially improve the accuracy of the model while reducing the model error, regardless of whether the data are classified or not. However, the introduction of optimized models also requires longer model training time. In Model 1, the GA improves the accuracy of the BP algorithm by 19.5% and increases the training time by 67.7%. The PSO algorithm improves the accuracy of the SVM algorithm by 11.2% and increases the training time by 65.5%. In Model 2, the role of the GA and PSO algorithm remains essentially the same as in Model 1. This shows the high stability of the GA and PSO algorithm. After multiplying the training model accuracies in Table 2 by 0.95 and comparing them with the values in Table 3, it is not difficult to find that the models constructed from unclassified data are less accurate in prediction than those constructed from classified data, and at the same time require more training time. This may be because, in the unclassified data, there are very close ultrasound-guided waveform data but with different types of outputs, which makes the prediction accuracy lower and is prone to over-learning, thus increasing the training time. Therefore, there is a strong need to pre-classify the data to enhance the reliability and robustness of the model. The subsequent prediction analysis will focus on the effect of Model 1.

After the construction of the training model was completed, the remaining 20% of the data were used for the validation of the model’s prediction of the unknown data. The model validation results are shown in

Table 4. Prediction performance of models on unseen data.

Model	Optimization Algorithms	RMSE	R2adj
BP	None	0.28	0.81
	GA	0.19	0.90
SVM	None	0.22	0.87
	PSO	0.15	0.94

. From Table 4, it can be seen that the PSO-SVM model still maintains the best prediction performance, followed by GA-BP, while the prediction performance of the unoptimized model is a bit worse. In comparing Table 2 and Table 4, it can be seen that for unknown data, the prediction accuracy of all models is lower than that of the trained models. This indicates that even for the most excellent models, there will still be a gap between their prediction accuracy and the actual value when facing unknown data. Therefore, optimizing and adjusting the model parameters to improve the accuracy and robustness of the models is one of the focuses of the subsequent work.

Combining the results of the classification and prediction models, the HHT-LightGBM-PSO-SVM model constructed in this study has the best prediction performance, with an accuracy of 94.05% for the training model and 88.36% for the prediction model, possessing a much lower error (close to 0). The prediction of this model exceeds the model with unclassified data in Table 3, indicating that the modeling approach combining the classification model with the prediction model has better performance in the detection of aircraft cable coatings. To reduce this difference, this study will focus on the solution of this problem in the subsequent research.

The aging experiments show that the aging of aircraft cable insulation coatings is a result of a combination of aging time, temperature, mechanical stress, and pre-existing defects. The aging time exhibits a nonlinear evolution in three stages, with a critical threshold of 50 days, marking the transition from surface oxidation to structural damage. This phenomenon is attributed to polymer chain breakage and volatilization of plasticizers accelerating the embrittlement of the coating. Similarly, temperature is a catalyst for aging degradation, with a critical threshold of 130 °C, above which the crosslinked structure is rapidly degraded. In addition, 160 °C and 100-day aging are consistent with the time-temperature superposition principle, verifying the feasibility of accelerated aging tests for long-term performance prediction. Mechanical stresses at bending angles ≥ 90° exacerbated coating damage. Pre-existing defects, especially cuts, act as stress concentrations, accelerating crack extension and reducing coating life by 60%. These findings emphasize the synergistic effects of thermal, mechanical, and defect-induced damage and the need for multiparameter optimization in cable design and maintenance. For example, in high-temperature areas or areas with significant bends, inspection intervals should be reduced to 50 days to prevent accelerated failure. In addition, the presence of defects such as cuts or abrasions requires stricter quality control during manufacturing to minimize the risk of premature failure.

HHT-based signal processing combined with machine learning significantly improves the accuracy and reliability of coating health monitoring. HHT outperforms FFT in terms of noise reduction and feature extraction and can accurately identify ultrasonic distortions caused by bends and defects. This capability is critical for the early detection of micro defects, as demonstrated by the clear differentiation of damage levels in the signals processed by HHT. In addition, the hybrid machine learning framework (LightGBM-PSO-SVM) achieved better prediction performance with 94.05% training accuracy and 88.36% validation accuracy. Using LightGBM for defect classification improves the robustness of the model by reducing misclassification errors, while particle swarm optimization enhances the generalization ability of SVM for small sample defects. However, the performance gap between the training dataset and the validation dataset emphasizes the need to further improve the model to increase adaptability to real-world variability. Future work should focus on incorporating additional environmental factors (e.g., humidity, vibration) and exploring interpretable AI techniques to bridge the prediction accuracy gap. Additionally, the development of lightweight algorithms for use in embedded systems could facilitate real-time monitoring of aircraft cables, further improving the safety and reliability of operations. With the rapid development of the Internet of Things and embedded technologies, real-time monitoring of cable status has become possible. In order to reduce the computational burden of the algorithm and improve real-time performance, this study will explore more lightweight machine learning algorithms, such as simplified models based on feature selection, online learning algorithms, etc. These algorithms can not only reduce the consumption of computing resources but also improve the efficiency and accuracy of real-time monitoring. By deploying these algorithms on embedded devices, real-time monitoring and early warning of cable status can be realized, thus improving the safety and reliability of cable operation.

4. Conclusions

In this study, through systematic aging experiments on aircraft cable cores, the effects of key parameters such as aging time, temperature, bending angle, and preset defects on coating life are deeply investigated, and an efficient aging damage prediction model is constructed by using advanced ultrasonic guided-wave technology and signal processing methods combined with machine learning algorithms. The main conclusions of this study are as follows:

• Significant effects of aging parameters on coating life: Experiments show that aging time and temperature have nonlinear accelerating effects on coating damage, with 50 days as the critical inflection point for damage accumulation and 130 °C as the critical threshold for sudden change in performance. Mechanical stress concentration significantly accelerates coating failure at bending angles greater than 90°. Preset defects, especially cut defects, significantly enhance the crack density and peeling area through stress concentration, shortening the cable life to 40% of the defect-free group.
• Effectiveness of ultrasonic guided wave and HHT noise reduction: The study confirms that HHT noise reduction is superior to FFT in processing ultrasonic guided wave signals and can effectively extract the characteristic signals to realize in situ quantitative assessment of coating damage. The noise-reduced ultrasonic guided wave signal has higher sensitivity to bending and damage and is suitable for the early detection of micro-defects in cable coatings.
• Excellent prediction performance of the machine learning model: the LightGBM classification model exhibits the highest classification accuracy and AUC value (0.94), and the prediction model (LightGBM-PSO-SVM) constructed by combining the PSO-SVM regression algorithm reaches an accuracy of 94.05% on the training set and 88.36% on the test set, with an error close to 0, which is significantly better than that of the unclassified data constructed. Although the accuracy on the test set is still high, there is a certain gap compared to the training set, which may indicate that the model has an overfitting problem. In future studies, we will focus on further optimizing the model structure, adjusting parameters, and increasing data diversity to reduce the risk of overfitting and improve the generalization ability of the model.

Although this study has achieved significant results in the prediction of coating aging damage of aircraft cable cores, there is still room for improvement. Future research will be devoted to further optimizing the model parameters to improve the accuracy and robustness of the model, so that it can give more accurate prediction results in the face of unknown data. At the same time, considering the influence of humidity, vibration, and other complex factors in the actual service environment, the aging test will be redesigned to more closely match the real service conditions, to more accurately assess the life of the cable coating. In addition, accelerating the promotion of HHT time-frequency analysis technology combined with the application of portable guided wave equipment to realize in situ quantitative assessment of coating damage will provide a strong guarantee for the safe service of aircraft cables. In addition, for the real-time applicability of the model, lightweight algorithms need to be developed to reduce computational complexity and resource consumption. This will help to implement these algorithms in resource-constrained embedded systems so as to realize real-time monitoring and early warning of aging damage of aircraft cable core coatings. Future research should delve into how to implement these lightweight algorithms in embedded systems and evaluate their performance and reliability in practical applications, and accelerate the application of HHT time-frequency analysis technology combined with portable guided wave equipment to achieve in situ quantitative evaluation of coating damage, which will provide a strong guarantee for the safe service of aircraft cables.