Federated Learning for Cross-Campus Engineering Learning Analytics without Centralizing Learner Records

Abstract

Engineering education increasingly depends on learning traces generated across learning-management systems, programming laboratories, CAD/BIM environments, simulators, and shared equipment. Combining these records centrally can support programme-level analysis, but creates new exposure of identifiable learner histories and can reinforce uneven data practices across campuses. This article specifies a cross-campus federated learning architecture for engineering learning analytics in which source records, learner identifiers, and evidence artifacts remain within each institution. A governed task registry defines purpose, schema version, model, permitted features, retention, and output constraints. Local adapters map approved events to a common competency vocabulary; training uses learner-level contribution clipping and differential-privacy accounting; secure aggregation reveals only a thresholded sum of institution updates; and release requires utility, subgroup, calibration, and governance checks. The design distinguishes data locality from privacy: gradients and model updates can leak information, secure aggregation does not protect against institutional insiders, and campus-level differential privacy does not by itself protect individual learners. The proposed workflow therefore combines local access controls, student-level privacy accounting, authenticated task configuration, semantic versioning, and human review. A synthetic engineering-course scenario illustrates the protocol, and a conformance and evaluation plan defines tests for privacy, robustness, fairness, interoperability, and educational usefulness. No deployment results or learner outcomes are claimed. The architecture positions analytics as decision support for educators and students, not as an autonomous basis for adverse academic action.