AI-Powered Wavelet System Spots Hidden Cable Faults Before They Trigger Blackouts

Buried beneath city streets and inside some of the most safety-critical industrial sites on Earth, underground transmission cables carry enormous amounts of electrical power with almost no physical visibility. When one of these cables fails, the fault is invisible to the eye, hidden under meters of soil or concrete, and the consequences can range from prolonged blackouts to dangerous conditions in facilities such as nuclear power plants. Detecting and classifying such faults quickly and reliably is one of the most stubborn challenges in power system protection, because the electrical signatures of faults are tangled up with normal switching transients, noise, and the sheer complexity of real-time operating conditions. A new study published in Cluster Computing proposes a computational intelligence framework that its authors say can detect and classify underground cable faults with high accuracy while remaining light enough to run inside a real protective relay.

The research, led by Kabiru Dahiru Ibrahim of the College of Nuclear and Automation Engineering at Chengdu University of Technology, together with colleagues from Chengdu University of Technology, Khalifa University, the University of Dundee, and the University of Electronic Science and Technology of China, introduces a two-stage framework abbreviated WT-PSO-RBFNN. The name encodes its three building blocks: a wavelet transform for extracting fault signatures from current waveforms, particle swarm optimization for tuning a neural network, and a radial basis function neural network for making the final classification decision. The work was supported by the National Natural Science Foundation of China under grant U23A20651, and it targets a setting where the stakes are unusually high: power networks in critical installations, including nuclear facilities, where fast and dependable fault isolation is not optional but essential to safety.

The first stage of the framework addresses the hardest part of the problem, which is simply deciding whether a fault has occurred at all. Underground cable faults announce themselves through brief, high-frequency transients superimposed on the three-phase current waveforms, but these transients are fleeting and easily masked by routine switching events. The researchers pass the acquired current signals through a discrete wavelet transform using the Daubechies db4 mother wavelet, decomposed to level three, with adaptive thresholding applied to the resulting coefficients. Wavelet decomposition acts like a mathematical microscope: unlike a Fourier transform, which assumes a signal repeats indefinitely and therefore blurs the exact moment a transient begins, the wavelet transform localizes information in both time and frequency. Multi-resolution analysis splits the signal into approximation coefficients that capture slow trends and detail coefficients that isolate the sharp, sudden changes most sensitive to fault events. The db4 wavelet is a popular choice in power system protection because its compact support and near-symmetry make it well suited to detecting abrupt discontinuities in current signals without excessive distortion.

Adaptive thresholding is the second ingredient of the detection stage. Rather than applying a fixed cutoff that might miss weak faults or falsely flag benign disturbances, the method adjusts the threshold applied to the wavelet detail coefficients according to the statistical properties of the signal itself. When a genuine fault strikes, the detail coefficients at the finer decomposition levels surge well above the adaptive threshold, providing a clear and computationally cheap detection flag. This design choice matters for practical deployment: a protection relay has milliseconds to decide whether to trip a circuit breaker, so the detection logic must be simple enough to execute in real time on modest hardware. The authors emphasize that the online computational structure of their framework is consistent with real-time relay implementation requirements, a claim that distinguishes it from heavier deep learning approaches that may achieve strong accuracy in offline studies but struggle to meet the latency and memory budgets of field-deployed protective devices.

Once a fault is detected, the second stage must answer two questions: which phases are involved, and is the fault a ground fault? The researchers feed normalized wavelet-domain features into a radial basis function neural network, a class of neural network whose architecture differs fundamentally from the deep convolutional networks that dominate headlines. An RBF network consists of an input layer, a hidden layer of neurons whose activation functions are radial basis functions, typically Gaussian bumps centered at learned points in feature space, and a linear output layer. Each hidden neuron responds strongly when the input vector is close to its center, so the network effectively measures how similar a new fault signature is to the fault signatures it learned during training. This locality of response gives RBF networks fast training, interpretable structure, and, crucially, a very fast inference step, since classification reduces to evaluating a modest number of Gaussian functions and a weighted sum.

The performance of an RBF network, however, depends heavily on its internal parameters: the positions of the Gaussian centers, the widths of those Gaussians, and the output weights. Traditionally these are chosen by heuristics or by grid search, an exhaustive but computationally expensive procedure that samples parameter combinations on a fixed lattice. The Chengdu-led team instead turned to particle swarm optimization, a nature-inspired algorithm in which a population of candidate solutions, called particles, flies through the parameter space, each one remembering the best position it has personally found and being attracted toward the best position discovered by the whole swarm. This social information-sharing dynamic allows the swarm to converge on high-quality parameter settings far more efficiently than exhaustive search, and it is particularly effective for continuous optimization problems like neural network tuning. In the proposed framework, PSO searches for the RBF network parameters that minimize classification error on the training data, producing a network that is better matched to the fault classification task than one tuned by conventional means.

To evaluate the framework rigorously, the authors built a MATLAB/Simulink model of an underground transmission cable using pi-section modeling, in which the cable is represented as a cascade of series impedance and shunt admittance segments that reproduce its distributed electrical behavior. Three-phase current waveforms were generated under a diverse set of fault and no-fault operating conditions, then processed through the wavelet stage to produce the feature vectors used for classification. The dataset was split using a stratified 70/15/15 protocol into training, validation, and test sets, a design that preserves the class proportions in each subset and guards against the misleading results that can arise when rare fault types cluster in one split. Under this protocol, the PSO-optimized network achieved 98.12 percent test accuracy with a mean squared error of 0.0834.

Those numbers become meaningful in comparison with the baseline. Under identical parameter bounds, a grid search-tuned RBFNN reached 96.56 percent accuracy with an MSE of 0.087. The difference of roughly one and a half percentage points may sound modest, but in protection engineering the margin matters: every misclassified fault represents either a missed trip that allows damage to propagate or a false trip that unnecessarily de-energizes a healthy circuit. In a nuclear facility power network, either outcome carries costs that far exceed the inconvenience of an ordinary distribution outage. The comparison also demonstrates a methodological point: the improvement came purely from smarter optimization within the same search space, not from a larger model or more data, which suggests that the PSO tuning found parameter regions that grid search simply could not reach at practical computational cost.

The broader context of this work is a field in transition. The study’s bibliography surveys a decade of approaches to power system fault detection, from time-domain reflectometry and travelling wave methods, which locate faults by timing reflected signals along the cable, to singular value decomposition techniques, support vector machines, and more recently deep architectures including convolutional neural networks, LSTM networks, graph convolutional networks, and hybrid CNN-LSTM models. Deep learning methods have shown impressive results on large datasets, but they bring heavy computational demands, opaque decision-making, and a hunger for training data that underground cable fault scenarios, which are by nature rare and varied, cannot always satisfy. Wavelet-plus-RBF schemes occupy a middle ground: physically motivated feature extraction, compact classifiers, and inference speeds compatible with hardware relays. The new study positions itself squarely in that tradition while adding the swarm-optimized tuning that closes part of the accuracy gap with heavier models.

The authors are careful about the scope of their claims. The evaluation was performed on a simulation model rather than on live grid hardware, and simulated faults, however diverse, cannot capture every degradation mechanism, noise source, and measurement imperfection found in real buried cables. Validation on physical test beds and field data remains the necessary next step before such a framework could be certified for deployment in safety-critical environments. Nevertheless, the study makes a persuasive case that the path to reliable underground cable protection does not necessarily run through ever-larger neural networks. By combining a mathematically principled signal decomposition, an adaptive detection threshold, a compact classifier, and a biologically inspired optimizer, the WT-PSO-RBFNN framework delivers near-state-of-the-art accuracy with an architecture simple enough to live inside the protective relays that guard the hidden arteries of the power grid. For the operators of nuclear facilities and other critical infrastructure, where a fault that goes undetected for even a few extra cycles can cascade into serious consequences, that combination of accuracy, speed, and deployability may prove to be exactly what the field has been looking for.

Subject of Research: Computational intelligence for fault detection and classification in underground power transmission cable networks

Article Title: A computational intelligence approach for reliable fault protection of underground transmission networks

Article References: Ibrahim, K. D., Ukwuoma, C. C., Abubakar, A., Bamisile, O., Qi, H., & Cai, D. (2026). A computational intelligence approach for reliable fault protection of underground transmission networks. Cluster Computing, 29(12), Article 702. https://doi.org/10.1007/s10586-026-06357-6

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06357-6

Keywords: underground cable fault detection, wavelet transform, radial basis function neural network, particle swarm optimization, adaptive thresholding, power system protection, transmission networks, nuclear facility power systems, MATLAB Simulink, fault classification, computational intelligence, real-time relays