{"id":{"repo_id":"exeter","oai_identifier":"oai:figshare.com:article/31832809"},"canonical_url":"https://search.dev.ndltd.org/etd/exeter/oai:figshare.com:article/31832809","repository":{"repo_id":"exeter","name":"University of Exeter","base_url":"https://api.figshare.com/v2/oai"},"display":{"title":"Realising data-centric UAV autonomy through learning-based prediction and feedback integration","abstract":"Unmanned aerial vehicles (UAVs) operate under non-linear, time-varying dynamics and distribution shift, where classical controllers depend on accurate models and extensive gain tuning. This dissertation develops an end-to-end pipeline — from raw flight logs to a deployable closed loop — that combines data-driven prediction, evolutionary optimisation, and classical feedback to deliver reliable and interpretable autonomy. A data-driven virtual UAV is learned from real flight data using a nonlinear autoregressive model with exogenous inputs (NARX). Evaluation is horizon-sensitive and employs MSE, RMSE, R squared, and dynamic time warping (DTW) to separate transient and steady-tate behaviour and expose error structure. Under identical dataset splits and preprocessing, a companion LSTM baseline is trained for comparison; the NARX surrogate achieves higher accuracy on representative lateral and vertical manoeuvres and exhibits regular error growth with horizon length. The surrogate is first embedded in a genetic algorithm (GA) sequence designer that explores short actuator sequences and trades a small amount of fidelity for smoothness and robustness; signal conditioning — including Hodrick-Prescott filtering — suppresses high-frequency irregularities while preserving temporal structure. While effective over short horizons, this open-loop synthesis drifts as the horizon grows and is sensitive to unmodelled disturbances. Motivated by these limitations, the core contribution, NeuroFlyPID-Fusion, places the NARX predictor inside a high-rate PID feedback loop. A multi-output reconstruction maps corrections to roll, pitch, yaw, and thrust, reducing inter-axis coupling. The fused design yields smoother actuation, sharper transients, and stabilised residuals under distribution shift. Behaviour awareness is introduced via lightweight segmentation (take-off, landing, forward/backward/sideways motion) with soft transitions and short hysteresis windows, keeping identification local to each primitive. Overall, the pipeline provides a practical route from data to flight-worthy autonomy with predictable behaviour and a low tuning burden.<p></p>","abstract_html":"Unmanned aerial vehicles (UAVs) operate under non-linear, time-varying dynamics and distribution shift, where classical controllers depend on accurate models and extensive gain tuning. This dissertation develops an end-to-end pipeline — from raw flight logs to a deployable closed loop — that combines data-driven prediction, evolutionary optimisation, and classical feedback to deliver reliable and interpretable autonomy. A data-driven virtual UAV is learned from real flight data using a nonlinear autoregressive model with exogenous inputs (NARX). Evaluation is horizon-sensitive and employs MSE, RMSE, R squared, and dynamic time warping (DTW) to separate transient and steady-tate behaviour and expose error structure. Under identical dataset splits and preprocessing, a companion LSTM baseline is trained for comparison; the NARX surrogate achieves higher accuracy on representative lateral and vertical manoeuvres and exhibits regular error growth with horizon length. The surrogate is first embedded in a genetic algorithm (GA) sequence designer that explores short actuator sequences and trades a small amount of fidelity for smoothness and robustness; signal conditioning — including Hodrick-Prescott filtering — suppresses high-frequency irregularities while preserving temporal structure. While effective over short horizons, this open-loop synthesis drifts as the horizon grows and is sensitive to unmodelled disturbances. Motivated by these limitations, the core contribution, NeuroFlyPID-Fusion, places the NARX predictor inside a high-rate PID feedback loop. A multi-output reconstruction maps corrections to roll, pitch, yaw, and thrust, reducing inter-axis coupling. The fused design yields smoother actuation, sharper transients, and stabilised residuals under distribution shift. Behaviour awareness is introduced via lightweight segmentation (take-off, landing, forward/backward/sideways motion) with soft transitions and short hysteresis windows, keeping identification local to each primitive. Overall, the pipeline provides a practical route from data to flight-worthy autonomy with predictable behaviour and a low tuning burden.&lt;p&gt;&lt;/p&gt;","abstract_has_math":false,"creators":["Shuyan Dong (21052235)"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026-03-17T00:00:00Z","date_published":"2026-03-17T00:00:00Z","updated_at":"2026-07-27T19:33:50Z","subjects":["Hybrid data-driven control","Closed-loop control","Data-driven UAV model","Tiime series analysis","Autonomy control","PID control","Motion prediction"],"languages":[],"rights":["CC BY","Open Access after 2027-09-23"],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["10779/exe.31832809.v1"],"render_values":[{"text":"10779/exe.31832809.v1","href":null,"code":true}]}]},"links":{"outbound_url":null,"outbound_label":null,"outbound_source":null},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Shuyan Dong (21052235)"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2026-03-17T00:00:00Z"]},{"key":"dc:relation","label":"Dc Relation","values":["https://figshare.com/articles/thesis/Realising_data-centric_UAV_autonomy_through_learning-based_prediction_and_feedback_integration/31832809"]},{"key":"dc:type","label":"Dc Type","values":["Text","Thesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Hybrid data-driven control","Closed-loop control","Data-driven UAV model","Tiime series analysis","Autonomy control","PID control","Motion prediction"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:rights","label":"Dc Rights","values":["CC BY","Open Access after 2027-09-23"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["10779/exe.31832809.v1"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Unmanned aerial vehicles (UAVs) operate under non-linear, time-varying dynamics and distribution shift, where classical controllers depend on accurate models and extensive gain tuning. This dissertation develops an end-to-end pipeline — from raw flight logs to a deployable closed loop — that combines data-driven prediction, evolutionary optimisation, and classical feedback to deliver reliable and interpretable autonomy. A data-driven virtual UAV is learned from real flight data using a nonlinear autoregressive model with exogenous inputs (NARX). Evaluation is horizon-sensitive and employs MSE, RMSE, R squared, and dynamic time warping (DTW) to separate transient and steady-tate behaviour and expose error structure. Under identical dataset splits and preprocessing, a companion LSTM baseline is trained for comparison; the NARX surrogate achieves higher accuracy on representative lateral and vertical manoeuvres and exhibits regular error growth with horizon length. The surrogate is first embedded in a genetic algorithm (GA) sequence designer that explores short actuator sequences and trades a small amount of fidelity for smoothness and robustness; signal conditioning — including Hodrick-Prescott filtering — suppresses high-frequency irregularities while preserving temporal structure. While effective over short horizons, this open-loop synthesis drifts as the horizon grows and is sensitive to unmodelled disturbances. Motivated by these limitations, the core contribution, NeuroFlyPID-Fusion, places the NARX predictor inside a high-rate PID feedback loop. A multi-output reconstruction maps corrections to roll, pitch, yaw, and thrust, reducing inter-axis coupling. The fused design yields smoother actuation, sharper transients, and stabilised residuals under distribution shift. Behaviour awareness is introduced via lightweight segmentation (take-off, landing, forward/backward/sideways motion) with soft transitions and short hysteresis windows, keeping identification local to each primitive. Overall, the pipeline provides a practical route from data to flight-worthy autonomy with predictable behaviour and a low tuning burden.<p></p>"]},{"key":"dc:title","label":"Title","values":["Realising data-centric UAV autonomy through learning-based prediction and feedback integration"]}]}],"canonical_facts":{"dc:creator":["Shuyan Dong (21052235)"],"dc:date":["2026-03-17T00:00:00Z"],"dc:description":["Unmanned aerial vehicles (UAVs) operate under non-linear, time-varying dynamics and distribution shift, where classical controllers depend on accurate models and extensive gain tuning. This dissertation develops an end-to-end pipeline — from raw flight logs to a deployable closed loop — that combines data-driven prediction, evolutionary optimisation, and classical feedback to deliver reliable and interpretable autonomy. A data-driven virtual UAV is learned from real flight data using a nonlinear autoregressive model with exogenous inputs (NARX). Evaluation is horizon-sensitive and employs MSE, RMSE, R squared, and dynamic time warping (DTW) to separate transient and steady-tate behaviour and expose error structure. Under identical dataset splits and preprocessing, a companion LSTM baseline is trained for comparison; the NARX surrogate achieves higher accuracy on representative lateral and vertical manoeuvres and exhibits regular error growth with horizon length. The surrogate is first embedded in a genetic algorithm (GA) sequence designer that explores short actuator sequences and trades a small amount of fidelity for smoothness and robustness; signal conditioning — including Hodrick-Prescott filtering — suppresses high-frequency irregularities while preserving temporal structure. While effective over short horizons, this open-loop synthesis drifts as the horizon grows and is sensitive to unmodelled disturbances. Motivated by these limitations, the core contribution, NeuroFlyPID-Fusion, places the NARX predictor inside a high-rate PID feedback loop. A multi-output reconstruction maps corrections to roll, pitch, yaw, and thrust, reducing inter-axis coupling. The fused design yields smoother actuation, sharper transients, and stabilised residuals under distribution shift. Behaviour awareness is introduced via lightweight segmentation (take-off, landing, forward/backward/sideways motion) with soft transitions and short hysteresis windows, keeping identification local to each primitive. Overall, the pipeline provides a practical route from data to flight-worthy autonomy with predictable behaviour and a low tuning burden.<p></p>"],"dc:identifier":["10779/exe.31832809.v1"],"dc:relation":["https://figshare.com/articles/thesis/Realising_data-centric_UAV_autonomy_through_learning-based_prediction_and_feedback_integration/31832809"],"dc:rights":["CC BY","Open Access after 2027-09-23"],"dc:subject":["Hybrid data-driven control","Closed-loop control","Data-driven UAV model","Tiime series analysis","Autonomy control","PID control","Motion prediction"],"dc:title":["Realising data-centric UAV autonomy through learning-based prediction and feedback integration"],"dc:type":["Text","Thesis"]},"updated_at":"2026-07-27T19:33:50Z"}