This paper addresses the problem of anomaly detection in multi-source heterogeneous data within the ETL (Extract-Transform-Load) process and proposes an intelligent detection framework that integrates temporal modeling and attention mechanisms. The method achieves effective dynamic aggregation of multidimensional features and temporal dependency modeling in ETL logs through the coordinated design of feature encoding, gated recurrent modeling, and multi-head attention allocation. At the feature level, the model uses a unified encoding structure to map raw logs, monitoring metrics, and task status information into a high-dimensional latent space, ensuring consistency of feature scales and completeness of information. At the temporal level, a GRU-based time modeling structure is introduced to capture long-term dependencies, enhancing the model's ability to perceive the evolution of anomaly patterns. At the attention level, a multi-head mechanism is applied to weight different time segments and feature dimensions, enabling adaptive focus on key moments and important features. Finally, the model combines anomaly scoring with distribution consistency constraints to achieve accurate identification and discrimination of potential anomalies. Experimental results show that the proposed framework significantly outperforms traditional rule-based detection, statistical methods, and basic deep models across various ETL task scenarios, demonstrating higher detection accuracy, stability, and generalization capability. The findings verify the effectiveness of integrating temporal modeling and attention mechanisms for anomaly detection in complex data streams and provide a feasible solution for building reliable and scalable intelligent ETL monitoring systems.

This study addresses the challenges of market non-stationarity, risk uncertainty, and dynamic inter-asset relationships in dynamic portfolio optimization by proposing an adaptive investment decision model based on multi-agent reinforcement learning. The research constructs a multi-agent architecture that combines centralized training with decentralized execution, allowing each agent to perform local strategy optimization while sharing global information to achieve a dynamic balance between return and risk. The model embeds multi-dimensional features such as market states, price fluctuations, trading volume, and risk indicators into the reinforcement learning framework, where a reward function drives policy iteration and enhances learning efficiency and decision stability in complex environments. A dynamic risk penalty mechanism and return adjustment term are introduced to effectively suppress excessive risk-taking behavior under high volatility and improve system robustness. Experimental results show that the proposed model outperforms traditional reinforcement learning methods in annualized return, maximum drawdown, Sharpe ratio, and Hit Ratio, maintaining strong return stability and risk control in multi-asset and multi-timescale trading environments. The findings confirm the effectiveness of multi-agent collaborative learning in dynamic asset allocation and risk-constrained optimization, providing methodological support for building intelligent and data-driven investment decision systems.

This study addresses the problems of low scheduling efficiency, unbalanced resource utilization, and delayed system response in ETL task execution under heterogeneous computing environments. It proposes a dynamic resource allocation and execution optimization model based on hierarchical reinforcement learning. The model establishes a collaborative decision-making mechanism by constructing high-level and low-level policy networks to achieve global task planning and local node control. The high-level policy is responsible for modeling task priorities and global resource constraints, while the low-level policy focuses on resource scheduling and performance optimization at specific execution nodes. This enables optimal allocation under conditions of multi-task concurrency and dynamic resource variation. The model takes multi-dimensional system states as input and optimizes key indicators such as average task completion time, maximum completion time, task waiting time, and load balancing index through a joint reward mechanism, forming a self-learning and adaptive scheduling strategy. Experimental results show that the proposed method demonstrates high stability and generalization ability across different hyperparameter and environment settings. It significantly outperforms traditional heuristic and single-layer reinforcement learning algorithms, effectively reducing task latency and improving overall system throughput. Furthermore, sensitivity analyses on learning rate, optimizer, exploration rate decay factor, and input noise confirm the robustness and controllability of the model in complex dynamic scenarios. This research provides an efficient solution for intelligent data processing and adaptive resource scheduling, offering both theoretical and practical value for building sustainable and high-performance data computing infrastructures.