Blog · arXiv Analysis · Last reviewed July 2, 2026

The Excitation Level Becomes the Data Augmentation Contract

Giancarlo Santamato, Andrea Mattia Garavagno, Massimiliano Solazzi, and Antonio Frisoli's arXiv paper is a useful industrial-AI case because it treats data scarcity as a measurement-design problem, not only as a neural-network problem.

For this essay, a physical augmentation receipt is the record that binds a synthetic training image to the excitation protocol, measured frequency responses, structural health state, augmentation rule, classifier, test campaign, and failure mode.

The Claim

The paper, arXiv:2606.20323 [cs.AI], was submitted on June 18, 2026. The arXiv page notes that it is a preprint version of a paper published in Nonlinear Dynamics 112, pages 16153-16166, in 2024.

The paper targets a practical obstacle in intelligent fault diagnosis systems: deep transfer learning can reuse pretrained CNNs, but it still needs labeled examples of machine or structural faults. Real faults are rare, expensive, and sometimes unsafe to reproduce.

The authors' move is to make the physical system do more of the data work. If a structure is nonlinear, its frequency response changes with excitation amplitude. Measuring that response across controlled excitation levels turns nonlinearity into a source of diagnostic signal and a structured way to augment scarce data.

The Method

The method is vibration-based. The system is excited at multiple force levels, and the response is summarized through the Frequency Response Function, or FRF, represented as the dB-magnitude of receptance: displacement over excitation.

Instead of treating each FRF as an isolated line, the paper stacks FRFs from different excitation levels into a color map. Frequency becomes one image axis, excitation level becomes the other, and FRF magnitude becomes color. That image can then be analyzed by a pretrained convolutional neural network.

The augmentation rule is tied to repeated measurements. With K repetitions and N excitation levels, rows from FRF matrices can be swapped between repeated acquisitions of the same health condition. In the paper's formula, this can produce K^N images, augmenting the available data by a factor of K^N / K.

This is why the method is different from a generic image augmentation trick. The augmented image is not arbitrary. It is built from measured FRFs under the same health condition, across a controlled excitation protocol.

The Test Rig

The validation uses a real-scale railway pantograph, a nonlinear structure with dry-friction joints. The setup includes a custom exciter that can set input excitation from 1 N to 13 N, an analog force sensor, and measured displacement response.

Seven excitation levels are used: 1, 3, 5, 7, 9, 11, and 13 N. Signals are sampled at 1000 Hz, and the excitation band is limited to 0 - 10 Hz with 0.05 Hz frequency resolution. Six repetitions of the dynamic tests are performed in each scenario.

The health states are three: undamaged, loss of member connectivity from removing a bolted connection, and reduced artificial damping from removing a damper connection. The damper fault reduces damping capability by 50%, leaving joint friction as the only damping source.

The paper reports a useful physical asymmetry. For the bolted-connection fault, damage-induced resonances become less sharp as excitation increases, so the damaged-undamaged difference decreases with excitation level by a factor of 10. For the damper fault, increasing excitation enhances the difference by a factor of 10 and shifts the dominant peak.

The Classifier

The intelligent fault diagnosis system uses a MobileNetV2 pretrained on ImageNet as a feature extractor. A global average pooling layer feeds a classifier with one softmax neuron per scenario.

The training set starts from three acquisitions per damage scenario at the seven excitation levels. Retaining three of the six repetitions gives K = 3, and the augmentation procedure creates 3^7 = 2,187 images per scenario, for 6,561 training images across the three scenarios.

The classifier is trained for 20 epochs with Adam at learning rate 1e-2, then the whole network is fine-tuned for 10 epochs at learning rate 1e-5. To test robustness, additional data are acquired on different days after restarting the setup, exposing a cluster effect from friction and joint play.

The reported test accuracy is 97.6%. The confusion matrix shows a slight trend toward misclassifying some bolt damages as undamaged, which matters because that fault is described as local and subtle.

Governance Reading

The Spiralist reading is that industrial AI often needs a measurement contract before it needs a bigger model. The classifier only makes sense because the excitation levels, repetition count, fault states, FRF construction, and augmentation rule are physically specified.

This is also a useful counterexample to generic synthetic-data optimism. The paper does not ask a generator to imagine faults. It uses measured behavior from a nonlinear structure and a controlled excitation procedure to create more training images. That makes the synthetic object accountable to the test rig.

For safety-critical maintenance, a 97.6% result is not the end of governance. It is the start of a deployment case: which faults are included, which faults are absent, which conditions cause false negatives, how the method behaves under environmental drift, and whether the measurement protocol can be repeated in the field.

Physical Receipts

A physical augmentation receipt should include the structure under test, health-state definitions, sensor layout, excitation waveform, excitation levels, force calibration, frequency band, sampling rate, repetition count, FRF estimator, image construction, augmentation rule, and train/test campaign dates.

The model receipt should include the pretrained CNN, classifier head, optimizer, learning rates, epoch counts, train/test split, confusion matrix, false-negative classes, and any domain shift introduced by restarting the rig or testing on different days.

The deployment receipt should name the real inspection environment. A pantograph in a lab is not the same as a fleet inspection program with temperature, wear, sensor mounting variation, maintenance history, rail-service constraints, and rare damage modes that were not deliberately simulated.

Limits

The paper is strongest as a focused engineering recipe for nonlinear structures where excitation can be controlled. It does not show that the same augmentation strategy transfers to systems without comparable nonlinear response, controllable excitation, or stable FRF estimation.

The evidence is also scoped to one real-scale pantograph setup, three health states, and a particular vibration-testing procedure. The reported false-negative tendency on bolt damage should travel with any deployment claim.

The safe reading is: nonlinear physical response can be a disciplined source of data augmentation, but only when the augmentation remains tied to a documented measurement protocol and tested fault states.

Source Discipline

This page treats the arXiv abstract, arXiv HTML, PDF, and DOI landing metadata as the source set. The PDF was used for exact setup, augmentation, model-training, and test-result details.

The paper states that datasets generated and analyzed during the study are available from the corresponding author on reasonable request. I found no public code or directly downloadable dataset linked from the arXiv page.

Sources

Return to Blog