Transmission line fault classification is a deceptively simple problem. Given a short window of three-phase voltage and current measurements after a disturbance, decide whether a fault occurred, which phases are involved, and roughly where along the line it happened. Protection engineers have solved versions of this with hand-designed relays for a century. The question for a learning-based approach is whether it can be both more accurate and more general, and the answer depends almost entirely on the representation you hand the network.
Why raw waveforms are a poor input
A three-phase fault window is a small multivariate time series: six channels, a few hundred samples. The obvious approach is a one-dimensional convolutional network over this sequence. It works, in the sense that it trains and reaches reasonable accuracy on the data it was trained on. It also generalises poorly, and the reason is structural.
The discriminative information in a fault is relational. It lies in how the phases move relative to each other, how the magnitude of one channel compares with another at the same instant, and how these relationships evolve over a few cycles. A 1D convolution has a receptive field along time within a channel, and it mixes channels only through the depth of its filters. It has to learn the cross-channel relationships that the physics makes primary, and it learns them in a way that is fragile to changes in fault inception angle, fault resistance, and load.
Time-series imaging
Time-series imaging is a family of transforms that convert a sequence into a two-dimensional array where the relational structure is explicit. The one that proved most useful here encodes, for every pair of time steps, a measure of the relationship between the signal values at those steps. The result is a square image whose texture reflects the signal’s temporal dynamics: periodic behaviour shows up as regular patterns, a fault’s onset shows up as a sharp change in texture, and the type of fault shapes the pattern that follows.
Applying this per channel and stacking the results gives an image with one plane per phase. Now a two-dimensional convolutional network sees exactly what it is good at: local textures and their spatial arrangement. The relational information that a 1D network had to reconstruct is present in the pixel values from the start.
The practical benefits are twofold. Accuracy on held-out fault conditions improves because the representation is less sensitive to the nuisance variables that dominate the raw waveform. And the model becomes easier to inspect, because the regions of the image the network attends to correspond to identifiable intervals of the fault.
Where self-attention belongs
Convolutions are local. A fault’s signature is often not: the relationship between the pre-fault steady state at the start of the window and the post-fault behaviour at the end carries information a small kernel cannot see. Stacking more layers widens the receptive field but dilutes it.
Self-attention over the feature map addresses this directly. Each spatial position computes weights over every other position and aggregates accordingly, so a pixel representing the fault onset can attend to pixels representing the steady state several cycles earlier. In the network I ended up with, a single self-attention block placed after the convolutional stem and before the classifier was enough. Adding more attention increased cost without improving accuracy. The lesson generalises: attention is a tool for the long-range part of the problem, and most of the problem is local.
Localisation as a second head
Classifying the fault type is only half of what a protection system needs. The other half is distance to fault. The two tasks share almost all of their features, so a single network with two heads, one for the categorical fault type and one for a continuous distance estimate, is both more efficient and more accurate than two separate models. The multi-task loss needs balancing, and the distance head benefits from a robust regression loss because a handful of badly estimated cases otherwise dominate the gradient.
Generative models for the unsupervised case
There is a second line of attack that avoids the need for labelled faults altogether: learn a probabilistic model of healthy operation and treat faults as low-likelihood events. A wavelet decomposition as the front end, giving a multi-resolution view of the signal, followed by a probabilistic generative network trained on healthy windows only, gives a detector that flags any departure from normal without having seen an example of it. It cannot say which fault occurred, but it catches faults the supervised model was never trained on, which in a safety-critical setting is not a small advantage. The two approaches belong together, with the generative detector as a first stage and the imaging-plus-attention classifier as the second.
What I would do differently now
Three things. First, I would spend more effort on simulating a wider distribution of operating conditions, because every generalisation failure I encountered traced back to a nuisance variable that was under-represented in training. Second, I would evaluate on measured data from a physical system earlier, since simulated waveforms are cleaner than reality in ways that flatter the model. Third, I would treat the transmission network as a graph from the start. A single line is a special case; the interesting problems, and the ones I am now working on, involve faults propagating through interconnected networks where topology is the primary structure and the per-line signal processing described here is one node’s worth of the picture.