An Integrated Pedagogical–Technical Framework for Deep Learning in Vocational Education and Training: A Systematic Review

Deep learning concept

Deep Learning (DL) is a subfield of machine learning that utilizes layered artificial neural networks to learn data representations hierarchically, from simple to complex features automatically without manual feature engineering (Feng, 2022; Jiang, 2022; Nguyen, 2022). The DL architecture consists of input layers, hidden layers, and output layers that perform nonlinear transformations through activation functions to generate predictions or classifications based on data patterns. The model training process takes place through forward propagation and backpropagation mechanisms with gradient-based optimization such as SGD or Adam, and is supported by stabilization techniques such as regularization and batch normalization to improve model accuracy and generalization.

DL surpasses traditional algorithms in its ability to automatically and adaptively extract data representations (Kang, 2020; Zhou, 2021). Various architectures have been developed according to data characteristics, such as multilayer perceptrons for tabular data, convolutional neural networks for visual data, and recurrent neural networks and transformers for sequential data. Advances in big data and GPU/TPU computing have expanded the application of DL across domains, including education. In the vocational context, DL is understood as a representation learning framework that enables learning pattern analysis, performance prediction, and data-driven instructional decision-making.

Deep learning applications in vocational learning

In vocational education, DL serves a strategic role in building personalized, adaptive, competency-based learning systems. One of its main implementations is Intelligent Tutoring Systems (ITS), which can analyze student learning traces, identify misconceptions, and dynamically adjust learning paths (Parker & Roumell, 2020; Wang & Ran, 2022). This approach supports the principle of mastery learning, where students must achieve mastery of a skill before proceeding to the next stage, aligning with the characteristics of practice-based education and competency completion.

In addition, DL is utilized in psychomotor training through the integration of Virtual Reality (VR) and Augmented Reality (AR). Models such as convolutional neural networks are capable of analyzing movements, work procedures, and practical performance in real-time to provide precise corrective feedback (Chan et al., 2022; Polat & Ekren, 2023). DL is also used in educational predictive analytics and career recommendations, such as predicting dropout risk and mapping student competency suitability to industry needs (Nor et al., 2023). This utilization strengthens the function of vocational education in facilitating an effective transition from learning to the world of work.

Authentic and immersive learning in the digital age

Authentic Learning is the main pedagogical foundation in contemporary vocational education because it places students in the context of solving real problems relevant to the world of work (Martínez-Argüelles et al., 2023; Wong et al., 2022). This approach emphasizes industry-based projects, case investigations, team collaboration, and the application of professional standards. In authentic learning designs, students not only understand technical procedures but also the rationale and context for applying skills, thereby strengthening the connection between school and workplace practices.

To enhance this experience, immersive learning using VR and AR provides realistic yet safe simulations of work environments. The integration of this technological approach with DL enables granular analysis of practical performance, adaptive difficulty level adjustments, and dynamic simulation of complex work scenarios (Chan et al., 2022; Polat & Ekren, 2023). The integration of DL, authentic learning, and immersive technology forms the conceptual foundation for developing a pedagogical-technical framework for AI-based vocational learning.

Study design

The literature search was conducted through major academic databases relevant to the field of educational technology and vocational education, namely Scopus, ScienceDirect, IEEE Xplore, SpringerLink, and Web of Science. To broaden the scope of the search and capture cross-indexed publications, the search process was supported by Publish or Perish software as a citation-based discovery tool that searches Google Scholar-indexed sources. The search was conducted from June 2025 to January 2026 (including early access articles) to ensure the recency of the reviewed literature. The search strategy used a combination of keywords that were terminologically consistent with the parameters in the Deep Learning model, including: “deep learning”,: vocational education”, “vocational training”, “skill acquisition”, “authentic learning”, “immersive learning”, “mastery learning”, and “contextual learning”. Keywords were combined using Boolean operators (AND, OR) and truncation techniques to increase search sensitivity and coverage, for example: “deep learning” AND “vocational education” OR “vocational training” AND “immersive learning”. This strategy is designed to capture studies that discuss the integration of Deep Learning technical capabilities such as data-driven adaptation, intelligent feedback, and simulation fidelity with competency-based vocational pedagogical approaches.

The thematic classification of selected literature in the Table 1. shows that Deep Learning parameters in vocational learning are not evenly distributed across the dimension of immersive technology alone, but also include adaptive assessment mechanisms, mastery learning, and contextual and collaborative integration in the VET ecosystem. This distribution confirms that the DL parameter model developed is pedagogical-systemic, not technology-centric.

Inclusion-exclusion criteria

The inclusion criteria for this study are peer-reviewed journal articles and conference proceedings that are available in full text, discuss Deep Learning as a system or approach in the design, adaptation, or evaluation of vocational learning, and relate it to pedagogical approaches consistent with the model, namely authentic learning, immersive learning, mastery learning, and/or contextual learning. Articles that discuss AI or DL in general without explicit relevance to vocational learning or work competency mastery, focus on general education or non-vocational higher education, are technical in nature without adequate pedagogical or conceptual analysis, or are not available in full text, were excluded from the synthesis. A summary of the complete selection criteria is presented in Table 2.

Data collection

Literature was searched in six major academic databases: Scopus, ScienceDirect, IEEE Xplore, SpringerLink, Web of Science, and Google Scholar through Publish or Perish as described in the Study Design section. The search time frame also followed the same criteria, namely from 2016 to January 2026. The study selection process followed the PRISMA 2020 protocol through the stages of identification, screening, eligibility assessment, and inclusion. The inclusion criteria included: articles discussing the implementation, parameters, or architecture of Deep Learning in the context of VET; full text available; published in English; journal articles or conference proceedings of reputable publications. The screening was carried out in stages by two independent reviewers through a review of the title, abstract, and full text. Disagreements were resolved through deliberative discussion until consensus was reached, and if necessary, a third reviewer was involved to make the final decision. The PRISMA flowchart and details of the number of articles are presented in Figure 1.

Figure 1. PRISMA framework flowchart.

Data analysis

Following the eligibility and quality assessment stages, 16 articles were identified as meeting the inclusion criteria and were used in the qualitative synthesis. Since each article represented one empirical study or review study, the number of studies included was equivalent to the number of study reports (n = 16). The synthesis process focused on mapping the pedagogical and technical parameters of Deep Learning that form adaptive vocational learning systems. The main dimensions identified include data-driven learning personalization (adaptive mastery learning), high-fidelity simulation environments (immersive and authentic learning), and competency-based assessment (authentic assessment). This systematic review provides an empirical and conceptual foundation for formulating a Deep Learning parameter model in the context of vocational education and training.

Basic parameters: learning and training

The fundamental parameters that determine the effectiveness of Deep Learning (DL) applications in vocational learning are the model training processes that are inherently related to the quality and quantity of the datasets used. Literature analysis confirms that model performance does not only depend on its architectural sophistication, but is critically determined by the relevance and authenticity of training data specific to the vocational domain (Meyer, 2023; Veber et al., 2022). In this context, data quality is defined by the alignment between the data and the predetermined learning outcomes, where the data must represent the tasks and performance standards applicable in the workplace. A DL model for welding competency assessment requires a dataset consisting of thousands of visual or sensory samples that have been carefully annotated by industry experts (Chan et al., 2022; Yunus, 2025). This process transfers tacit knowledge and professional standards into the model, thereby ensuring the validity of the assessment. The training dataset transforms into a digital representation of the curriculum and industry competency standards, whose position as key parameters must be comprehensively validated before initiating the technical training process.

Once a high-quality dataset has been constructed, technical parameters such as learning rate, batch size, and epoch require precise calibration (fine-tuning). However, this optimization process has a dual purpose that is both technical and pedagogical. In pedagogical terms, an equally essential goal is to achieve computational efficiency to minimize latency in delivering feedback to students. In this framework, a rapid practice-feedback-correction cycle is vital to prevent the internalization of errors and accelerate the achievement of competency mastery. Therefore, fine-tuning these parameters is not merely technical optimization, but rather a pedagogical calibration that balances depth of analysis and speed of interaction, in order to realize a responsive and effective practical learning experience.

Pedagogical parameter 1: mastery learning in vocational education and training

The optimization of Deep Learning (DL) models in VET significantly depends on their ability to support mastery learning (Mukhtar et al., 2026; Parker & Roumell, 2020). This approach is relevant to VET as it ensures mastery of essential prior competencies before learners advance to more complex. A key parameter of DL in this framework is its capacity to function as a formative assessment engine that operates dynamically and individually. By analyzing student response data in real-time, including error patterns, response times, and interactions with the material, DL models are able to precisely identify competency lack in each student. This automatic diagnostic capability overcomes a major implementation challenge of mastery learning in conventional classrooms, namely the limitation of teachers in providing simultaneous individual attention and feedback.

Once competency lack are identified, the DL function shifts to adaptive learning pathways. Based on specific diagnoses, the system does not simply repeat the same material, but automatically provides relevant remedial interventions. These interventions can take the form of micro-learning modules, alternative demonstration videos, additional exercises, or even practice scenarios in a Virtual Reality environment that focus on aspects that have not yet been mastered (Mukhtar et al., 2026). Only after validation of mastery does the system open access to the next learning module. Thus, DL ensures the formation of a solid foundation of knowledge and skills, which are essential elements for long-term success in the vocational domain. In this way, DL supports mastery learning to ensure work readiness. The system will not graduate students before they meet industry standards, thereby ensuring that students have valid entry-level skills when entering the workforce.

Pedagogical parameter 2: contextual learning in vocational education and training

The following fundamental parameter is the ability of DL to support contextual learning. Contextual learning emphasizes the acquisition of knowledge and skills through their relevance to real-world situations, which has been shown to significantly increase student motivation and learning retention (Haryani et al., 2021; Parker & Roumell, 2020; Sephokgole et al., 2021). In the VET context, the effectiveness of a DL model critically depends on its ability to be trained using datasets that are not only large but also contextually grounded in industry practice. These datasets can be technical case studies from industry, data from teaching factory environments, interaction logs from work-based learning scenarios, or even labor market trend analyses for entrepreneurship learning.

By training models on this authentic data, DL architectures inherently internalize the patterns, anomalies, and complexities of industrial problems, making them a digital proxy for industrial reality (Ruhanen et al., 2021). Through continuously updated datasets, DL enables dynamic curricula that train students to update their occupational skills, which is a vital aspect of long-term employability. Once successfully trained with contextually enriched data, DL models can function as generators of relevant dynamic learning scenarios. Unlike static teaching materials, DL-based systems can generate problem-based questions, simulate project challenges, or even run interactive dialogues through chatbots that simulate real professional interactions.

Pedagogical parameter 3: authentic learning (practice with deep learning)

The next fundamental pedagogical parameter is the capacity of Deep Learning (DL) to realize authentic learning (Paeßens et al., 2023; Wong et al., 2022). Authentic learning requires students to engage in solving complex and ill-defined vocational problems, often through collaborative processes that reflect the dynamics of the modern workplace (Weijzen et al., 2024). Within this framework, a key parameter of DL lies in its capacity to underpin hands-on, simulation-based learning environments and digital learning playgrounds distinguished by their authenticity. When trained on real-world datasets, DL models generate virtual practice spaces that are not only safe for experimentation but also possess sufficient verisimilitude to cultivate self-efficacy and facilitate skill transfer to authentic workplace contexts (Lowell & Tagare, 2023). Finally, being involved in authentic learning environments powered by DL correlates positively with an increase in perceived employability.

Authentic learning powered by DL facilitates the development of transferble competencies, which are skills that can be transferred between professional roles (Martínez-Argüelles et al., 2023; Meyers, 2009). In addition, AI-based collaborative simulations can train interpersonal skills and work habits, which are key indicators of employability. The main advantage of DL-based practices lies in the seamless integration of learning activities with authentic assessment mechanisms. Analysis shows that the implementation of valid and reliable authentic assessments remains a significant challenge in conventional VET systems (Paeßens et al., 2023; Wong et al., 2022). DL offers a solution with its ability to conduct technology-based multidimensional assessments that can measure not only the final results, but also the problem-solving process carried out by students, including aspects of collaboration (Mayer et al., 2023; Paeßens et al., 2023). By analyzing a variety of performance data from student interactions in simulations, such as sequence of actions, efficiency, and decision-making patterns, DL models can provide detailed and objective formative feedback.

Pedagogical parameter 4: immersive learning in vocational education and training

A further pedagogical parameter significantly enhanced by Deep Learning (DL) is immersive learning, which utilizes technologies such as VR, AR, and MR to create interactive and high representationally accurate learning environments (Mulders et al., 2024; Veber et al., 2022). The use of immersive technology in VET has a significantly positive impact on students’ behavioral, cognitive, and affective outcomes (Boel, 2022; Lowell & Tagare, 2023; Thomann, 2024). The key parameter for the application of DL in this context is its ability to function as an intelligent simulation engine that can create and manage a safe, repetitive, and efficient virtual practice environment. Thus, DL not only supports visualization but actively builds a digital learning playground where students can engage in hands-on practice to master complex procedures without real-world consequences.

Furthermore, the most significant transformative role of DL in immersive learning is its ability to turn static simulations into adaptive learning experiences. The main strength of DL in immersive environments is its ability to generate unexpected scenarios. This directly trains students’ adaptability so they can be more responsible in finding and keeping a job. According to Wu (2024), instead of following a programd linear arrangement, the DL model can analyze students’ performance in real-time and dynamically adjust the difficulty level or even introduce unexpected new challenges. This adaptability has been essential for maintaining student engagement and optimizing the development of complex and sustainable problem-solving skills. DL leverages granular simultaneous interaction data from the simulation environment, including movement patterns, action choices, and responses to feedback, to generate multidimensional and adaptive technology-based authentic assessments tailored to the competency profiles of learners (Mayer et al., 2023; Paeßens et al., 2023).

Conceptual model of deep learning integration in vocational education and training institution learning

The conceptual model of deep learning in Figure 2 shows integration in VET institutions is built on the integration of pedagogical parameters and technical capabilities to create an adaptive and authentic learning ecosystem. Based on contextual datasets that represent industry competencies and scenarios, this model trains DL architecture as an intelligent machine that simultaneously drives four pedagogical pillars: (1) mastery learning through real-time performance analysis, personalized learning paths, and instant formative feedback; (2) authentic learning, in which the DL model presents contextual work scenarios that validly reflect the complexities of the industrial world; and (3) immersive learning through VR/AR-based high-precision simulations that enable the practice of complex psychomotor skills in a safe and controlled environment, and (4) authentic assessments that evaluate the problem-solving process in a multidimensional way, not just the end result. These four pillars work in a continuous cycle, with assessment data enriching the adaptive pathways, thereby forming a dynamic, student-centered model focused on enhancing employability, work readiness, and career adaptability of graduates, while also bridging the competency gap between vocational education and the demands of the modern industry.

Figure 2. Conceptual model of integrated pedagogical–technical framework for deep learning in vocational education and training.

Implication and suggestions

These findings shift the focus of DL evaluation from sole accuracy to a socio-technical-pedagogical framework that positions complete, authentic, contextual, and immersive learning as the main criteria. Practically: (1) VET needs to develop a curriculum based on experiential learning that produces authentic data and prepares teachers as AI-based facilitators; (2) policymakers should direct investments not only toward hardware but also toward curating relevant national vocational datasets; (3) the industry acts as a core partner in validating, providing data, and contributing domain expertise throughout the DL development cycle.

This SLR is limited by 16 articles and potential publication bias, while the rapid pace of DL innovation makes some technical parameters quickly obsolete. Moving forward, it is possible to: conduct empirical/quasi-experimental studies in the context of VET institutions to measure the impact of pedagogically parameterized DL on competence and employability; develop low-cost and scalable DL models (e.g., mobile AR); research ethical aspects of algorithmic bias and student data privacy, and to build sustainable industry-academia partnerships for dataset curation. The long-term implication of applying these DL parameters is the enhancement of individuals’ capacity to secure and maintain decent employment. Technology is not just a teaching aid, but an instrument for building career resilience in the era of Industry 5.0.