Improving Vibration-Based Classification in a Wind Turbine Jacket Structure and Rotor Through Multitask Learning and Spectral Representations

This work proposes and evaluates a supervised classification framework based on multichannel vibration signals to analyze two systems associated with wind turbines: rotor imbalance in a laboratory-adapted Enair E30Pro wind turbine and structural damage in a jacket-type structure. The methodology compares single-task and multitask learning models using spectral representations obtained from the Fast Fourier Transform (FFT) in log-magnitude format, the Short-Time Fourier Transform (STFT) in log-power format, and the Welch power spectral density (Welch PSD) in log-power format. During the single-task stage, TabNet (Attentive Interpretable Tabular Learning) models, a deep learning architecture designed for tabular data, were independently trained for each system and representation to establish baseline performance levels. Subsequently, a multitask scheme composed of a shared encoder and two task-specific classification heads was implemented: one for the jacket-type structure and another for the rotor. In addition to the original TabNet baselines, a capacity-matched single-task MLP control was evaluated using the same encoder architecture, preprocessing, optimization settings, epoch budget, and cross-validation folds as the multitask model. Relative to the TabNet baselines, the multitask model achieved higher performance in several configurations, particularly for the more challenging rotor task; however, the capacity-matched control showed that these gains cannot be attributed uniquely to multitask parameter sharing because the multitask model did not consistently outperform an equivalently structured single-task network. For the jacket-type structure, all three spectral representations achieved near-perfect classification performance, with Macro-F1 values above 0.99 in both the single-task and multitask settings. For the rotor imbalance problem, the multitask model using Welch PSD log-power achieved a global accuracy of 0.9125 and a Macro-F1 of 0.8909. Negative-control experiments further showed that replacing the auxiliary-task labels with random permutations reduced target-task performance in several configurations, indicating that informative auxiliary supervision influences the shared optimization process, although this effect did not translate into a consistent predictive advantage over the capacity-matched single-task control. The multitask architecture nevertheless provides a parameter-efficient joint formulation: one shared multitask network contains 4,481,354 trainable parameters, compared with 8,958,218 parameters required by two independent capacity-matched single-task networks, corresponding to an approximately 49.97% reduction in total model parameters. Overall, the results support multitask learning as a compact joint modeling strategy for physically different vibration-based classification problems, while indicating that its principal benefit in the present study is architectural efficiency rather than a consistently superior predictive effect attributable to positive cross-task transfer.