In general, the analysis of specific complex sport techniques is qualitatively assessed by expert trainers, occasionally supported by biomechanical analysis obtained from ad-hoc acquisition systems. However, this standard methodology often relies on subjective evaluation and requires the analysis of numerous biomechanical parameters, which hinders the identification of technical errors. This work aims to develop a methodology for quantitatively evaluating and diagnosing complex techniques, in particular, Taekwondo kicking techniques, using an autoencoder implemented through a Gated Recurrent Unit (GRU) neural network. To this end, hip, knee, and ankle joint angles and angular velocities of both lower limbs were captured from multiple kicking executions performed by Taekwondo experts (black belt holders). Subsequently, the GRU-based autoencoder was employed to categorize kicks based on the technical accuracy of their execution as standard or non-standard. Finally, an ad-hoc error analysis was conducted on kicks classified as non-standard to identify the biomechanical parameters contributing most to the differences between standard and non-standard kicks. Based on the F1-score, a metric that combines precision and recall, the results showed that the implemented autoencoder accurately detected non-standard kicks with a score exceeding 90%. Furthermore, error analysis successfully identified the parameters most influential in non-standards classification, enabling the improvement of kicking technique in non-expert individuals.