Based on a comprehensive analysis of industry failure statistics, international standard provisions, and partial discharge mechanisms, this study ultimately selects three representative real-type defects: outer semiconductive layer tip defects, main insulation scratch defects, and stress cone misalignment defects. According to the CIGRE WG B1.57, 2023 service survey, approximately 80% of cable breakdowns originate from installation defects at joints or terminations. Supporting this, experimental data collected in this study show that these three defect types account for roughly 70% of early failure cases in cable joints.

International standards provide explicit risk descriptions regarding residual sharp tips, insulation scratch depth, and stress cone misalignment, reflecting the strict engineering control requirements associated with these defects. From a discharge mechanism perspective, the three selected defects correspond to surface-type, volume-type, and interface-type partial discharges, respectively, forming a minimally complete set of failure mechanisms. This also enables comprehensive validation of the applicability of the proposed multi-source fusion diagnostic algorithm.

Therefore, focusing on these three defect types not only ensures high representativeness but also significantly enhances the practical value of the proposed method in real-time condition assessment and maintenance decision-making for cable systems.

Based on the defect types selected in the preceding section, three real-type defect configurations were implemented on 10 kV XLPE cable intermediate joints using a real-condition experimental platform. The defect fabrication process simulates the actual procedures and conditions encountered during cable installation in power systems, ensuring that the defects closely resemble those found in industrial applications. The setup is illustrated in Figure 1. A 4-m-long defect-free cable was divided into two 2-m segments, which were connected using a CSS-1733J-8.7/15 kV cold-shrink joint. The test cable model was ZH-YJV with a cross-sectional area of 185 mm² and a rated voltage of 8.7/15 kV.

1) Construction of outer semi-conductive layer-tip defect: During the production of intermediate cable joints, the semi-conductive layer is usually beveled using a wallpaper knife. However, if not handled carefully, sharp burrs can form at the beveled edges of the outer semi-conductive layer. This defect model was designed to simulate the discharges caused by the concentration of electric field forces at the internal composite insulation interface of the cable joint. To make the partial-discharge signal more stable and easier to detect, the length of the semi-conductive tip was designed to be 25 mm with a width of 10 mm. After inserting the stress cone, the tip defect protruded 5 mm from the semi-conductive part of the stress cone, disrupting the balanced electric field near the tip because of the presence of a sharp edge.

To minimize the suppression and interference from the discharge defect signals, none of the three defects were coated with silicone grease during fabrication, and no copper mesh was applied during joint construction. The two cable segments were grounded using a grounding strap connected to the outside of the joint.

A real experimental platform for power cables is shown in Figure 2 and comprises three parts: a power supply system, real cable system, and discharge signal measurement system. The power supply system comprised a signal generator and a co-phase high-voltage amplification coupling power supply. The cable system comprised defect-free cables, artificially defective cable joints, and other accessories. The discharge signal measurement system was designed to use multiple types of sensors for a more comprehensive fusion data analysis of defects. Therefore, two signal measurement systems were designed to measure different types of signals. However, the signal collection of both systems was synchronized to ensure data consistency. Meanwhile, to ensure cost-effectiveness and the feasibility of measurement, all equipment used in the experiments consists of commonly employed devices in practical engineering applications.

1) The discharge signal’s time-domain and frequency-domain waveform data were collected via a high-frequency current transformer (HFCT) using a digital storage oscilloscope (DSO). The frequency range of the HFCT sensor was 10 kHz–100 MHz. The DSO is a Tektronix MDO3014 with four channels, maximum sampling rate of 5 G/s, and record length of 10 M. In this study, sampling channel two was selected, with a sampling rate of 5 G/s and a record length of 10 M. During the measurement, the HFCT sensor was clamped to a cable-shield ground wire.

To accurately reflect the discharge patterns, reduce the sampling data volume, and avoid interference from the randomness of the discharges, the partial discharge starting voltage and stable discharge voltage were first determined using a pulse current method partial discharge detector. The measurement system is shown in Figure 3.

The pulse current method partial discharge instrument used was the xPD-1717 model with a working voltage of DC 8.4 V. When the partial discharge signals first appeared on the PC, the discharge-starting voltage U ₀ was recorded. When the discharge became more evident and the discharge signal stabilized, a stable discharge voltage U _m was recorded, as detailed in Table 1.

To ensure that the three jointed cable samples were subjected to identical voltage stress conditions and to induce sufficiently clear and representative defect discharges, a unified voltage application scheme—comprising synchronous ramp-up, voltage hold, and ramp-down phases—was employed in this study. The voltage was gradually increased from zero to 10 kV, during which discharge signals were monitored. Once stable discharge behavior was observed, both HFCT and UHF sensors were used to perform continuous sampling over equal time durations.

As a result, 217 sets of partial discharge (PD) data were obtained for the outer semi-conductive tip defect, 230 sets for the main insulation scratch defect, and 223 sets for the stress cone misalignment defect. In terms of PRPD patterns, 141 were acquired for the outer semi-conductive tip defect, 132 for the insulation scratch defect, and 157 for the stress cone misalignment defect.

Considering that the dataset exhibits a near-balanced distribution—sufficient to ensure classifier performance is not adversely affected—and to avoid irreversible damage to test samples due to over-pressurization, no effort was made to forcibly equalize the number of samples across defect types.

The fundamental concept of wavelet transform is to decompose a signal using wavelet basis functions, which are combinations of a series of low-pass and high-pass filters, to capture different frequency components through scaling and translation. The scale parameter controls the width of the wavelet, while the translation parameter determines its position, thereby effectively separating the noise components from the signal. Subsequently, noise suppression is performed on the decomposed coefficients, and the signal is reconstructed to achieve the purpose of denoising. The specific implementation steps are as follows:

1) Selection of the wavelet basis function: An appropriate wavelet basis function is chosen based on the characteristics of the signal, such as Daubechies wavelets, Haar wavelets, Symlet wavelets, or Morlet wavelets.

The above equations are used to implement the filtering and denoising of defect discharge signals in both the time and frequency domains.

Figure 4 shows the time-domain waveforms collected using an HFCT sensor under a 10 kV voltage level for the three types of real defects after noise filtering.

From the figure, the following can be observed:

1) The discharge amplitude of the outer semi-conductive layer-tip defect was relatively low, with the maximum positive discharge amplitude of approximately 400 mV and the maximum negative discharge amplitude of approximately 580 mV. The discharge shows asymmetry between the positive and negative half-axes, with a noticeable discharge concentration and clear height of the discharge clusters.

Figure 5 shows the frequency-domain spectra of the three types of real defects collected using an HFCT sensor under a 10 kV voltage level after noise filtering.

Evidently, both the outer semi-conductive layer-tip defect and the stress-cone misalignment defect exhibit a relatively narrow frequency domain bandwidth concentrated between seven and 45 MHz. However, the spectrum of the outer semi-conductive layer-tip defect has a bimodal shape, whereas the spectrum of the stress-cone misalignment defect has more denser peaks. Therefore, the peak density of the stress-cone misalignment defect is higher than that of the outer semi-conductive layer-tip defect; however, the outer semi-conductive layer-tip defect has higher amplitude in the spectrum. The spectrum amplitude of the main-insulation scratch defect was significantly higher than those of the other two defects, with the highest peak at 0.013. Its frequency-domain bandwidth was broader, with a maximum bandwidth of 75 MHz.

From the spectrum, it is also possible to understand that the total harmonic distortion of the different defects varies considerably.

As shown in Figure 6, the PRPD patterns of the three types of actual defects at the 10 kV voltage level were collected using the UHF sensor. In the figure, the colors red, yellow, and green represent different ranges of discharge repetition counts: green indicates a discharge count between 0 and 20, yellow represents a count between 20 and 40, and red corresponds to a count greater than 40.

Evidently, the PRPD spectra of the three types of defects exhibit several differences: the discharge points of the outer semi-conductive layer-tip defect are mainly concentrated in the power frequency phase ranges of 50°–120° and 230°–310°. The amplitude distribution of the positive and negative half-cycles of the power frequency shows significant differences, with the positive half-cycle amplitude range between −50 and −40 dB, and the negative half-cycle amplitude range between −50 and −33 dB. The discharge amplitudes near 90° and 270° are higher, with a higher concentration of discharges at lower amplitudes. The overall shape of the spectrum was triangular, with a general leftward shift in the profile. The discharge signals of the main-insulation scratch defects were present in the power frequency phase ranges of 0°–170° and 180°–350° and were mainly distributed between 0°–90° and 180°–270°. The distribution patterns of the positive and negative half-cycles are generally similar; however, the discharge amplitude in the positive half-cycle is slightly higher (approximately 5 dB) than that in the negative half-cycle. The overall shape of the spectrum was trapezoidal, with a leftward shift indicating more intense discharges, and the discharge points were densely distributed across the entire amplitude range. The discharge range of the stress-cone misalignment defect was between the power frequency phases of 60°–200° and 240°–360°, with most of the discharge points concentrated in the second and fourth quadrants. The differences in the contours and distribution of the discharge patterns in the positive and negative half cycles are small. The discharge amplitude is higher and concentrated between −25 and −15 dB, and the distribution of discharge signals roughly forms a “gate” shape.

From the contour features of the PRPD patterns, it is visually evident that there are significant differences between real defects and typical defects. Under more complex actual operating conditions in practical production, it is necessary to investigate defect identification methods tailored to complex discharge patterns, in order to enhance the practical relevance of the results for actual operating conditions.

Based on the above analysis, it can be concluded that the time-domain waveforms, frequency-domain spectra, and PRPD spectra of the different defects have varying degrees of distinguishable features. For spectra with obvious characteristics, the defect type can be identified through a relatively straightforward visual differentiation. However, for spectra that are more difficult to distinguish, feature extraction and algorithms are necessary.

Based on the analysis of the time-series signal in the previous section, the following features were extracted according to the waveform characteristics: The specific meanings and expressions are as follows.

Signal energy: Describes the total power or energy of the signal, representing the magnitude of the signal’s energy. where y(t) is the time-domain signal amplitude, T is the discharge signal period, and E _c is given by Equation 6. E c = ∑ t = 0 T y t 2   (6)

Crest Factor: The ratio of the peak value to the root mean square (RMS) value of the signal, characterizing the relative strength of the spikes to the average energy and the instantaneous peak characteristics in the time-domain signal. The CrestFactor is given by Equation 7. C r e s t F a c t o r = max y t 1 T ∑ t = 0 T − 1 y t 2 (7)

Kurtosis describes the sharpness or concentration of a signal waveform. A waveform with a higher kurtosis indicates sharper peaks, whereas a waveform with a lower kurtosis indicates a flatter signal. Where: E is the mathematical expectation, x is the sample amplitude of the signal, and μ is the mean value of the signal. The Kurtosis is expressed using Equation 8. K u r t o s i s = E x − μ 4 E x − μ 2 2 (8)

Attenuation coefficient: The ratio of the average amplitude of the signal during the initial discharge period to the average amplitude after signal attenuation. The attenuation coefficient A _c of the signal can be expressed using Equation 9, where M ₁ is a segment of the fault time during the initial discharge, and M ₂ is a segment of time after the signal is attenuated. A c = 1 M 1 ∑ i = 1 M 1 y i t / 1 M 2 ∑ j = 1 M 2 y j t (9)

Based on these feature parameters, a four-dimensional time-domain feature vector set was obtained.

From the frequency spectrum, it can be observed that the frequency spectrum amplitude, bandwidth, and frequency distribution differ for various types of defects in real defective cables. Based on these characteristics, the frequency-domain signal features were extracted using the following calculations:

Frequency amplitude ratio: the ratio between the amplitudes of different frequency bands within a certain frequency range. Where N ₁ and N ₂ are the sample numbers in the first and second frequency bands, respectively. P(k) and P(f) represent the signal amplitudes in the first and second frequency bands, respectively. The frequency amplitude ratio F _ar is given by Equation 10. F a r = 1 N 1 ∑ i = 1 N 1 P i k / 1 N 2 ∑ j = 1 N 2 P j f (10)

The total Harmonic Distortion: Reflects the ratio of harmonic components to the fundamental frequency component. Where: N is the number of harmonics, and k ₁ is the fundamental frequency. The total harmonic distortion T _hd is given by Equation 11. T h d = 1 N ∑ n = 2 N y n k 1 2 / y k 1 (11)

Spectral peak density: The ratio of the number of spectral peaks exceeding the mean value within a certain frequency range to the total number of spectral peaks in that range, which describes the complexity of the spectrum. where M is the number of spectra within the frequency range, P(h) is the amplitude of the signal in the frequency band, P _a (h) is the average amplitude of the frequency band. The spectral peak density S _pd is given by Equation 12. S p d = ∑ i = 1 M P i h > P a h / M (12)

Based on these feature parameters, a three-dimensional frequency-domain feature vector set is obtained.

In studies of real-world (in-service) defects, small sample sizes limit the effectiveness of deep-learning approaches such as CNNs and Transformers; on limited datasets these models cannot fully exploit their capacity, and relying on extensive data augmentation or heavy regularisation often leads to pronounced over-fitting and unstable convergence. Concurrently, the large-scale deployment of GPU resources tailored for deep learning proves challenging in actual industrial settings, constraining their widespread adoption. Alternative approaches, such as defect identification models based on Extreme Learning Machines (ELM), exhibit accelerated training speeds. However, ELM’s random initialization of connection weights and bias thresholds renders its recognition accuracy vulnerable to external interference. For industrial applications, Support Vector Machine (SVM)–based diagnostic models are widely favoured for their robustness, high accuracy, feature interpretability, visualisable decision boundaries, and low deployment cost.

The Genetic-Algorithm-optimised SVM (GA-SVM) employs genetic optimisation to tune the hyperparameters of the Support Vector Classifier (SVC), achieving global optima and delivering greater stability when classifying high-dimensional feature vectors (Guohai et al., 2013; Mingyu and Wang, 2022; Zhu and Shen, 2024). Although GA-SVM is not a novel approach, recent domain-specific modifications targeting distinct defect and fault identification scenarios have demonstrated enhanced recognition performance. This substantiates that optimization research based on GA-SVM algorithms continues to hold significant research value (Zhao et al., 2021; Yunfeng et al., 2025).

In the context of multi-source fused-feature identification for real-type defects, the feature sets originate from different measurement devices and characterise the defects from entirely different perspectives. Simply concatenating these heterogeneous features and training a conventional GA-SVM tends to weaken complementary information, making it difficult to achieve the desired recognition accuracy. Accordingly, this paper improves the GA-SVM framework by refining the feature-selection strategy and the genetic crossover–mutation operations to better accommodate multi-source fusion.

The selection operation in the GA-SVM algorithm is typically driven by a single fitness metric, such as classification accuracy or cross-validation error. However, when selecting features for multi-source feature identification, relying solely on a single distribution-based metric may overlook valuable real-time discriminative information provided by the model during the training process.

In this study, the PJS-M method is employed to dynamically amplify the influence of strong features and suppress the influence of weak features during the genetic algorithm iterations through a group-wise weighting mechanism. Specifically, the method adaptively fuses Probabilistic Jensen–Shannon Divergence (PJS) and Margin Contribution (MC) to achieve dual weighting based on both data-driven and model-driven strategies. This guides the GA’s search and crossover-mutation processes to converge toward the most discriminative feature subsets, thereby reducing the impact of redundant features, accelerating convergence, and ultimately improving classification accuracy.

The specific formulas are as shown in Equations 23, 24: PJS k = JSD P k + ∥ P k − (23) JSD P ∥ Q = 1 2 D KL P ∥ P + Q 2 + 1 2 D KL Q ∥ P + Q 2 ∈ 0 , 1 (24)

In the equation, P k + and P k − ​ represent the probability density estimates of the kth feature group for the positive and negative classes, respectively, and D KL ​ denotes the Kullback–Leibler (K-L) divergence. The PJS metric is used to achieve maximal mutual information, quantifying the distributional discrepancy of the same feature group across different classes. The greater the separation between distributions, the stronger the discriminative power of the feature group, and consequently, the larger its assigned weight. MC k = 1 G k ∑ i ∈ G k w i (25)

In the Equation 25, G k ​ denotes the kth feature group (k = 1, … ,K), and w i represents the ith component of the SVM weight vector. MC k ​ indicates the average contribution of the kth feature group to the current SVM margin. This metric reflects the ability of the feature group to minimize structural risk, as a larger margin implies better generalization performance and lower generalization error. The specific formulas are as shown in Equations 26, 27. ξ k t = PJS k α t MC k 1 − α t (26) α t = 1 − t T max (27)

The exponential fusion term ξ k t can be regarded as a linear combination. The temporal decay coefficient α t decreases linearly with respect to generation t, where t denotes the current generation in the genetic algorithm. When t = 0 and α ₀ ≈ 1, the fusion relies entirely on the pure data distribution, helping the GA quickly move away from random noise. As t→T _max and α _Tmax ≈ 0, only the model feedback is retained, guiding the search to focus on the feature groups that truly support the decision margin.

In the early stages of iteration, due to the high diversity of the population, static distributional differences are first used to perform global search. In the later stages, dynamic margin-based feedback is introduced to perform fine-grained local adjustments. The coefficient α t ​ is employed to balance the contributions of PJS and MC, such that in the early generations, α t ​ emphasizes distributional differences, while gradually shifting toward model feedback in later generations.

By adopting a dual data–model-driven strategy and incorporating cross-validated classification accuracy Acc _CV ​, the algorithm retains high-quality individuals while maintaining population diversity, thereby preventing the recognition process from converging to local optima. The specific formulas are as shown in Equations 28, 29. w k t = ξ k t + ε ∑ j = 1 K ξ j t + ε (28) ∑ k = 1 K w k t = 1 (29)

In the equation, w k t ​ denotes the weight coefficient of the kth feature group obtained in the tth iteration. First, the normalized weights for each feature group are calculated according to Equation 28, and based on these weights, either fitness weighting or feature scaling is applied. The selection, crossover, and mutation operators of the genetic algorithm are then employed to complete one evolutionary cycle.

At the end of each generation, the best-performing offspring in the current population is selected to re-estimate the feature group importance and update the raw weight vector w k r a w ​. A local information entropy–guided mechanism is embedded within the genetic search framework, enabling the feature weights to adaptively converge during the evolution process. This continuously amplifies the influence of highly discriminative features while suppressing redundant or noisy ones, thereby accelerating convergence and improving the overall classification performance of the model.

The PJS-M method introduced in Section 5.2 is incorporated to optimize the GA-SVM algorithm. The optimized algorithm, referred to as GAW-SVM, is then used to perform classification of real defect types. The specific identification procedure is outlined as follows:.

1. Acquire the raw time-domain data, frequency-domain data, and PRPD patterns of real defects in cable intermediate joints. Feature extraction is performed separately for each of the three data types, and the extracted features are quantified into a common interval using standardization and normalization techniques.