Frame the gap around a system and a decision
Engineering doctorates in artificial intelligence, cyber-physical systems (CPS) and automation often begin with a technology rather than a problem: "apply deep learning to predictive maintenance" or "build a digital twin of a production line". A defensible methodology starts instead from a physical system, a quantity that matters about it, and a decision that a better estimate of that quantity would improve. For example: which pump bearings in a given plant will fail within the next maintenance interval, and how much unplanned downtime would a condition-based schedule avoid?
Reference architectures help you locate the contribution. Lee et al. (2015) proposed a five-level CPS architecture for manufacturing, running from smart connection and data-to-information conversion through a cyber level to cognition and configuration. Monostori et al. (2016) reviewed CPS in manufacturing more broadly, and Tao et al. (2018) set out a framework in which a digital twin links physical and virtual product models across design, manufacturing and service. Use such frameworks to say which level or link your thesis improves, not as evidence that a technology works.
What Indian Industry 4.0 studies have done
The verified Indian studies reviewed here are mostly studies of adoption: what stops firms from using Industry 4.0 technologies and what they need first. They are useful for framing the industrial context of a thesis, and they show a characteristic methodological pattern, but they are not evidence about how well particular AI or CPS techniques perform. Technical method guidance in this article therefore draws on international literature.
At NIT Rourkela, Jena and Patel (2022) used a questionnaire survey of Indian industry to identify industrial system requirements and barriers to Industry 4.0 implementation, then ranked both with fuzzy multi-criteria decision-making and examined the relationships between them. At Poornima University and Banasthali Vidyapith in Rajasthan, Rawat et al. (2024) studied barriers in Indian small and medium enterprises using interpretive structural modelling (ISM) and MICMAC analysis within the technology-organisation-environment framework. They identified a lack of government regulations and a lack of awareness of the benefits of digital infrastructure as root barriers, and found that concern about the future workforce had high dependence but low driving power, contrary to earlier studies. Kamble et al. (2018) similarly analysed the driving and dependence power of barriers to Industry 4.0 adoption in Indian manufacturing.
Three methodological lessons follow. First, these designs rest on expert judgement, so the selection, number and independence of experts determine how far the results generalise, and the published abstracts do not always report them. Second, ISM and fuzzy rankings are sensitive to how pairwise judgements are aggregated, so a thesis using them should test whether the structure changes under reasonable alternatives. Third, adoption studies are best combined with technical work: a barrier analysis can justify why a low-cost, explainable monitoring system matters for Indian SMEs, and the technical chapters can then test such a system.
| Study (institution) | Data and methods | Main contribution | Methodological caution |
|---|---|---|---|
| Jena & Patel, 2022 (NIT Rourkela) | Questionnaire survey; fuzzy multi-criteria decision-making | Ranked system requirements and barriers and related them | Sample and sectors to be reported; sensitivity of fuzzy weights |
| Rawat et al., 2024 (Poornima University; Banasthali Vidyapith) | ISM and MICMAC within the TOE framework | Root barriers for Indian SMEs; workforce concern as dependent | Expert panel composition; aggregation of judgements |
| Kamble et al., 2018 | Driving and dependence power analysis of barriers | Barrier structure in Indian manufacturing | Expert judgement, not measured performance |
Choosing and combining designs
Most AI and CPS theses combine several of the designs below. State which one answers each research question, and why.
| Design | Typical data | What it can show | Main threat |
|---|---|---|---|
| Laboratory experiment (test rig) | Sensor signals under controlled faults or loads | Whether a method detects known conditions | Rig conditions may not resemble the field |
| Simulation or digital twin | Model outputs under designed scenarios | Behaviour across conditions too costly to test physically | Model error and unvalidated assumptions |
| Field study in a plant | Historian, PLC and maintenance records | Performance on real machines and decisions | Missing labels, drift, confounding by operations |
| Survey or expert method | Questionnaires, pairwise expert judgements | Priorities, barriers and requirements | Panel selection and aggregation |
Whatever the design, describe the measurement chain before the model. For sensor data, state the sensor type and mounting, sampling rate, filtering, synchronisation between channels and how faulty or missing readings were detected. For labels taken from maintenance logs, state who recorded them, how failure dates were assigned and how uncertain they are, since a model cannot be more accurate than the labels it learns from. For survey and expert methods, state how respondents were selected, their roles and experience, the response rate and how disagreements between experts were aggregated. These details are often missing from published work, and supplying them is an easy way to make a thesis stronger than the papers it cites.
Field data from Indian plants are often the most valuable and the hardest to obtain. Agree access, data ownership, confidentiality and publication rights in writing with the industrial partner before data collection, and record them in the methodology chapter.
A walkthrough: designing a leakage-free evaluation
The following project is illustrative. A candidate wants to predict bearing faults in centrifugal pumps using vibration data from twelve pumps at one plant over eighteen months, with maintenance records as labels.
- Define the prediction task precisely. For example, whether a pump will need a bearing replacement within 30 days, predicted from the preceding 7 days of vibration features.
- Define the unit of independence. Windows from the same pump are strongly correlated, so the pump, not the window, is the unit to hold out.
- Split by pump and by time. Train on some pumps and earlier months; test on held-out pumps and later months. A random split of windows would put near-identical data in both sets.
- Fit preprocessing inside the training data. Scaling, feature selection and threshold tuning are learned on the training folds only (Varma & Simon, 2006).
- Compare with simple baselines. A threshold on overall vibration level or a fixed replacement interval shows whether the model adds value.
- Report decision metrics. Precision and recall at the operating threshold, false alarms per month and the lead time gained, not only accuracy.
| Split strategy | Question it answers | Risk |
|---|---|---|
| Random split of windows | Can the model interpolate within machines it has seen? | Severe leakage; optimistic results |
| Hold out whole pumps | Will it work on new machines of the same type? | Few pumps means wide uncertainty |
| Hold out later months | Will it work in the future at this plant? | Drift and changed operating regimes |
| Hold out another plant | Will it transfer to a new site? | Hardest to obtain; most convincing |
Kapoor and Narayanan (2023) documented data leakage as a widespread cause of irreproducible results in machine-learning-based science and proposed model info sheets that require authors to justify the separation between training and test data. Writing such a justification into your methodology chapter is one of the most effective ways to pre-empt an examiner's questions.
Sensitivity, uncertainty and explanation
Simulation models and digital twins need two kinds of checking. Validation compares model outputs with held-out measurements. Sensitivity analysis asks which uncertain inputs drive the outputs. Global methods, which vary all inputs together across their plausible ranges, are preferable to changing one factor at a time, and Monte Carlo sampling is the usual way of propagating input uncertainty to the outputs (Saltelli et al., 2008). State the input ranges and distributions and report outputs with intervals.
Explainability methods such as Shapley-value attributions show which features drove a prediction. They help engineers check that a model relies on physically plausible signals, but they explain the model, not the machine, and they do not replace validation on unseen equipment.
Computational comparisons should also control implementation. When you compare your algorithm with published ones, run all of them in the same environment and at a similar level of optimisation, and report the hardware and software versions (Kriegel et al., 2017).
Working within Indian requirements
The UGC 2022 doctoral regulations set coursework of at least 12 credits, including research methodology and research and publication ethics, semester reviews before a Research Advisory Committee and an open pre-submission presentation (University Grants Commission, 2022). Use the advisory committee reviews to fix your evaluation design early, since changing the split strategy after results are known undermines the analysis.
Funding for engineering research now flows in part through the Anusandhan National Research Foundation, set up under the ANRF Act, 2023 (Anusandhan National Research Foundation, n.d.). If your data include information about identifiable people, such as operators, drivers or building occupants, check your institution's guidance under India's Digital Personal Data Protection Act, 2023, and seek ethics review where people are involved. Industrial data that are commercially sensitive need a written agreement on what can be published or shared.
What to report and common mistakes
| Report | Why it matters |
|---|---|
| The physical system, operating conditions and decision supported | Defines where results apply |
| Data sources, sampling rates, labelling and missing data | Lets readers judge data quality |
| Unit of independence and split strategy | Shows the evaluation is free of leakage |
| Baselines and decision-relevant metrics | Shows whether the method adds practical value |
| Simulation or twin validation, sensitivity and uncertainty | Shows how far model outputs can be trusted |
| Expert panel selection and aggregation, for survey methods | Determines generalisability |
| Hardware, software versions, code and data availability | Allows reproduction |
- Describing tools and buzzwords instead of the system, question and evaluation.
- Randomly splitting correlated sensor windows into training and test sets.
- Tuning thresholds or selecting features on the test data.
- Reporting accuracy on heavily imbalanced fault data without precision, recall or false-alarm rates.
- Presenting a digital twin without validation against held-out measurements.
- Citing a framework paper as evidence that a technology improves performance.
Key takeaways
- Start from a physical system, a measurable quantity and a decision, and use architectures only to locate your contribution (Lee et al., 2015; Tao et al., 2018).
- Indian adoption studies give useful context but rest on expert judgement, not measured technical performance (Jena & Patel, 2022; Rawat et al., 2024).
- Hold out whole machines, later periods or other sites, and keep all preprocessing inside the training data (Kapoor & Narayanan, 2023; Varma & Simon, 2006).
- Validate simulations and digital twins against held-out measurements and report global sensitivity and uncertainty (Saltelli et al., 2008).
- Compare algorithms in the same environment and report hardware and software versions (Kriegel et al., 2017).
- Use UGC advisory committee reviews to settle the evaluation design early (University Grants Commission, 2022).
