Learning Paradigms for Data-driven Fault Diagnosis in Parallel Delta Robots: A Review

In this paper, we present a comprehensive analysis of intelligent fault diagnosis techniques for parallel Delta machines, with a particular focus on challenges related to data scarcity and distribution shifts.The primary objective of this review is to provide a systematic analysis of data-driven learning paradigms for sensor-based fault diagnosis in parallel Delta robots under limited labeled data and changing operating conditions.Unlike previous reviews that mainly summarize machine learning algorithms, in this work, we emphasize how heterogeneous sensor signals, multi-sensor information fusion, and intelligent sensing concepts improve diagnostic reliability and predictive maintenance in smart manufacturing environments.The overall framework of intelligent fault diagnosis is first outlined.Different learning paradigms-namely, supervised learning, semi-supervised learning, few-shot learning, and transfer learning-are then evaluated, with emphasis on their applicability under limited annotation conditions, diverse fault detection requirements, and varying operational environments.Hybrid deep learning models and ensemble learning strategies are further examined, demonstrating their effectiveness in capturing complex spatiotemporal dependencies.The role of gradient-based optimization methods and automated hyperparameter tuning is also analyzed in enhancing model robustness and performance.In addition, performance metrics and evaluation methodologies for both classification and regression tasks are reviewed, providing a consistent basis for benchmarking diagnostic systems.Overall, in this review, we provide a sensor-oriented perspective on intelligent fault diagnosis by integrating heterogeneous sensor information, advanced learning paradigms, and intelligent sensing concepts, while identifying current limitations and future research directions toward reliable predictive maintenance in smart manufacturing.intelligent production.(1) These systems are widely deployed in high-speed pick-and-place operations, precision assembly, and additive manufacturing processes, where both accuracy and throughput are critical.Nevertheless, their inherently complex kinematic structures and tightly coupled mechanical components make them susceptible to faults such as kinematic deviations, bearing stress accumulation, and belt wear, all of which can directly degrade product quality and process stability.Effective fault diagnosis is therefore crucial for maintaining system reliability and operational safety, while simultaneously reducing downtime, maintenance costs, and unexpected failures associated with progressive equipment degradation.(2) However, the extensive data requirements of deep learning techniques often conflict with practical constraints in real industrial environments.In particular, the annotation of large-scale industrial signal datasets typically requires significant manual effort and domain expertise, leading to substantial time and cost overheads.(3) Moreover, deliberately operating critical equipment until failure is generally infeasible owing to safety risks and economic consequences, resulting in a limited availability of labeled fault data.(4) This data scarcity not only restricts the training of datadriven models but also intensifies challenges such as overfitting and class imbalance, where normal operating data dominate over rare fault conditions.Another critical challenge is domain shift, which arises when discrepancies exist between training and testing data distributions.Such discrepancies are commonly induced by variations in operational conditions, including changes in speed, load, environmental factors, and task configurations.As a result, models trained under controlled or laboratory conditions often exhibit significant performance degradation when deployed in dynamic and nonstationary industrial environments.(5) In Delta robot applications, for example, variations in motion profiles or payload conditions can substantially reduce diagnostic accuracy when models are transferred from static experimental settings to real-world production scenarios.To address these challenges, recent research has increasingly shifted from conventional supervised learning approaches toward data-efficient learning paradigms.These paradigms aim to enhance generalization under limited data conditions by exploiting additional information sources, such as unlabeled data, domain knowledge, or cross-domain similarities, and by enabling knowledge transfer across related tasks.As a result, they provide effective solutions for mitigating data scarcity, domain shift, and class imbalance in industrial fault diagnosis.(6) This work focuses on data-efficient learning strategies for fault diagnosis in Delta robots, with three primary contributions.The novelty of this review lies in its sensor-oriented perspective.Rather than simply comparing artificial intelligence algorithms, in this work, we systematically analyze how heterogeneous sensing technologies, including vibration, attitude, motor current, and encoder sensors, interact with different learning paradigms to improve diagnostic robustness under practical industrial constraints.We further identify current limitations in sensor data availability, sensing reliability, and cross-domain generalization, and discuss how recent advances in fewshot learning, transfer learning, self-supervised learning, and digital twin technologies can alleviate these challenges.From the perspective of sensor technology, we also highlight how multi-sensor information, including vibration sensors, attitude sensors, motor current sensors, and encoder measurements, serves as the fundamental data source for intelligent fault diagnosis in parallel Delta robots.Rather than focusing solely on learning algorithms, the reviewed