80-Day Thesis Plan: Virtual Sensing for Position Prediction in Linear Motors

Structured roadmap, deliverables, dataset plan, benchmarks, and an 80‑day schedule.

Scope and success criteria

Topic framing

This topic sits at the intersection of virtual sensing / soft sensing (software-based estimation of an unmeasured variable), and sensorless drives (estimating position/speed from electrical measurements, typically voltages/currents). Virtual sensors are widely motivated by lower cost and maintainability compared with purely physical sensing.

For linear motors, the motivation is often even stronger because a “long-enough” linear position measurement system (e.g., long scale/encoder) can be mechanically exposed and maintenance-intensive; many works note vulnerability/maintenance concerns for linear encoders in harsh environments and position sensorless control as a meaningful alternative direction.

Note about the start date you gave
You wrote “Today is Sunday 22 February 2026.” In Prague local time, the current date is Monday 23 February 2026 (one-day difference). This plan is anchored to Sunday 22 February 2026 as requested; if you start on 23 February 2026, shift everything by +1 day.

What “position prediction using only the power entries” should mean

To keep the thesis technically precise and aligned with established practice, define estimator inputs as only electrical/drive variables available at the converter–motor interface, for example:

Phase currents (measured)
DC bus voltage (measured)
PWM duty cycles or commanded phase voltages (known from control)
Optional: switching state, sampling time, temperature (if already measured as part of the drive)

This matches the common sensorless paradigm that rotor/mover position information can be extracted by analyzing electrical variables (voltage and current) at the motor port.

Labels vs inputs (important for defensibility)

Explicitly separate what you may use for training labels vs what is allowed as estimator input.

Ground truth (labels only):
- Encoder / linear scale position \(x\) for the mover
- Optional derived label: electrical position \(\theta_e\) if you know pole pitch and motor geometry
Estimator inputs (allowed):
- Phase currents
- DC bus voltage
- Duty cycles / commanded voltages

This mirrors the input structure used in practical “virtual position sensor” workflows where a neural model is trained to map \(\alpha\beta\) voltages/currents to electrical position. A common pattern is: \[ (V_\alpha, V_\beta, I_\alpha, I_\beta) \;\to\; \theta_e \]

Clear deliverables that “finish” the thesis in 80 days

By Day 80, you should have all of the following completed:

Literature-backed state of the art + gap statement
Include: virtual sensing overview, sensorless position estimation methods, what is known for rotary machines vs linear motors, and the “viability” gap (risk + integration cost in real deployment).
A reproducible dataset pipeline
A dataset (measured or simulated+measured) that includes:
- Inputs: electrical variables only
- Ground truth: mover/rotor position (encoder/scale) used only for labels and evaluation
- Splits: train/val/test + at least one “unseen condition” test
Implementation of a comparison benchmark
“Most relevant state-of-the-art approaches based on AI” plus at least one classical baseline. Classical baselines are essential to make “viability” defensible (you compare to what industry already trusts).
Viability analysis (your thesis keyword)
Structured argument with measurements and a risk/cost model:
- Accuracy and robustness
- Failure modes and safety hooks
- Integration cost (engineering time, compute, commissioning effort, retraining/maintenance)
If feasible: a short hardware demonstration
Even a minimal experiment is valuable: log real signals and show offline inference; if time allows, show near-real-time inference.

Evidence-based technical roadmap

Conceptual map of methods you should cover

A strong thesis in this topic typically organizes methods as:

Model-based sensorless estimation (classical)

Back‑EMF observers + PLL: good at mid/high speed; weak at low speed because back‑EMF becomes small and noise-sensitive.
Flux-linkage observers / integrators: integration can reduce some noise but introduces drift/DC bias; initialization matters.
Sliding Mode Observers (SMO) and variants: robust to disturbances; practical issues include chattering and filtering design.
Kalman filtering (EKF/UKF): high-quality state observers; compute cost must be discussed.

AI / data-driven virtual sensing (core comparison set)

Feedforward regressors and compact MLPs
Sequence models (LSTM/GRU/RNN) for dynamic mapping from electrical signals \(\to\) position
Hybrid “grey-box” methods: observer + ML correction (often stronger in robustness when physics is partially known)

How to connect “virtual sensing” to “sensorless motor drives” in writing

You can justify that sensorless estimation in drives is essentially a domain-specific virtual sensor: it removes the mechanical position sensor by reconstructing position from existing electrical measurements; literature even uses phrasing like “virtual position sensor.”

Linear motor specificity (your differentiation)

Do not write the thesis as “PMSM but linear.” Instead emphasize linear-motor complications that influence viability:

Long-scale position measurement requirement and maintenance exposure
End effects, cogging/force ripple, rail length segmentation (if applicable)
Typical operating profile: frequent start/stop, low-speed tracking, reversals (stresses low-speed sensorless performance)

Dataset and experimentation strategy

Minimum viable experiment design

To finish in 80 days, design a dataset that is small enough to collect but rich enough to test viability:

Trajectories: at least 3 families (trapezoidal move, S-curve, oscillatory tracking)
Operating regimes: at least 3 (low speed, mid-speed, near lab safe maximum)
Variation factor (pick one you can implement safely):
- Added payload mass
- Different friction condition (e.g., lubrication state, if allowed)
- Supply voltage variation (if the bench supports it safely)
Logging rate: high enough to preserve electrical dynamics (kHz-class is typical). If you cannot log that fast, document it as a “viability constraint.”

“When feasible” hardware demo without overcommitting

Tier A (almost always feasible): offline inference on recorded data + plots of error vs time/speed
Tier B (stretch): near-real-time inference (PC/embedded) with latency measurement

Comparison and evaluation framework

Recommended minimum set to compare

To meet “implement and compare most relevant state-of-the-art approaches based on AI,” while still supporting a viability argument:

Classical baselines (at least one, ideally two):
- SMO + PLL baseline (common in sensorless literature; linear motor variants exist)
- EKF or UKF observer baseline (high quality but heavier compute)
AI methods (at least three):
- Compact MLP regressor to \(\\sin(\theta)\) and \(\\cos(\theta)\) (handles angle wrap naturally)
- LSTM/GRU sequence model (dynamic mapping; useful when ripple/friction introduces time-dependence)
- Hybrid: classical observer output + ML correction (grey-box style)

Metrics that match the “viability” theme

Do not stop at RMSE. Include:

Position error: MAE, RMSE, max absolute error, 95th percentile
Dynamic error: error during acceleration, reversals, and low-speed creep
Latency: inference time and effective delay (important for control viability)
Robustness: performance under an “unseen” operating condition
Failure detection: when the estimator becomes unreliable, can it self-diagnose?

Viability structure you can defend in writing

Replacement viability: Can a virtual sensor safely replace the encoder across required regimes?
Redundancy viability: Can it reliably detect encoder faults/deviations?
Integration cost: commissioning time + compute requirements + maintenance (retraining) overhead

A useful general statement for your Literature Review is that sensorless/virtual-sensing approaches can reduce parts and cost, but may have drawbacks in low-speed stability/accuracy—often the hardest region.

The 80-day schedule

This schedule assumes steady work most days (even if not full-time). Each block includes both research and writing, because waiting to write at the end is a common failure mode in short thesis timelines.

Days / Dates	Theme	Key tasks
Days 1–7 Feb 22 – Feb 28, 2026	Setup, scope lock, thesis skeleton	Create a thesis repository (text + code) and a reproducible environment (requirements + scripts). Write a 2–3 page “scope lock” document: Define “virtual sensing” vs “sensorless control.” Define “position prediction” (current estimation vs future prediction horizon). Declare constraints: “inputs are electrical variables only.” Build the thesis skeleton (all headings, figure/table placeholders). Start a living literature matrix (method class, inputs, motor type, speed regime, validation type, pros/cons).
Days 8–14 Mar 1 – Mar 7	Literature deep dive + gap statement	Read and summarize seed references plus added ones. Draft the Literature Review in method families: Virtual sensing / soft sensing foundations Sensorless control families by operating region (zero/low vs mid/high speed) Linear motor constraints and what changes vs rotary Produce a one-page “research gap and thesis contribution” statement tied to viability + linear motors.
Days 15–21 Mar 8 – Mar 14	Measurement + dataset design	Choose dataset source path: Path A: real bench data (preferred for real-scenario viability) Path B: simulation-first, then limited bench confirmation Write a measurement plan: signals, sampling, synchronization; trajectories/regimes; safety limits/abort conditions. Implement the data logging pipeline (even if initially CSV). Draft the Dataset chapter outline.
Days 22–28 Mar 15 – Mar 21	Pilot data + preprocessing	Collect a small pilot dataset (enough to train a toy model). Implement preprocessing: Clarke transform to \(\alpha\beta\) Normalization strategy Windowing for sequence models Train/val/test split strategy (include an “unseen condition” split) Write “Data acquisition and preprocessing” while it’s fresh.
Days 29–35 Mar 22 – Mar 28	Classical baseline	Implement at least one classical estimator baseline suitable for your available signals: SMO + PLL style baseline (or flux observer baseline) Validate on the pilot dataset and produce first plots. Write the “baseline methods” subsection (equations, assumptions, tuning, compute cost).
Days 36–42 Mar 29 – Apr 4	AI model 1: fast baseline	Train a compact MLP to estimate \(\\sin(\theta)\), \(\\cos(\theta)\), or position directly. Establish experiment harness: fixed seeds, hyperparameter logging, consistent metrics/plots. Write the “AI method 1” subsection (architecture + training procedure).
Days 43–49 Apr 5 – Apr 11	Sequence model for dynamics	Train an LSTM/GRU model for position estimation. Add ablations: window length sensitivity; input set sensitivity (currents only vs currents+voltages). Start “compute feasibility” note: inference time on target platform (PC now; embedded later if feasible).
Days 50–56 Apr 12 – Apr 18	Hybrid model + robustness	Build a hybrid approach: observer output + ML correction, or ML predicts residual between observer and truth (grey-box soft sensor concept) Run robustness tests: unseen trajectory; unseen load condition; noise injection (documented). Draft the Results chapter structure (figures first, text second).
Days 57–63 Apr 19 – Apr 25	Main dataset expansion	Collect the full dataset (or remaining portion). Retrain all models on the final dataset with frozen evaluation protocol. Generate final plots/tables: error vs speed; error during reversals/acceleration; worst-case error distribution. Write the core results narrative (what wins, where, why).
Days 64–70 Apr 26 – May 2	Viability analysis (risk + cost)	Create a “viability checklist” and fill it with evidence: When does the method fail? How does it behave at low speed? What fallback is required? Write a cost model (even if approximate): encoder + installation + maintenance vs compute + engineering + retraining. Write a risk matrix: severity \(\times\) likelihood \(\times\) detectability + mitigation strategies.
Days 71–77 May 3 – May 9	Hardware demonstration window	Tier A demo (target): offline inference on bench data with clear plots and reproducible scripts. Tier B demo (stretch): near-real-time inference and measured latency. Write the Experiment chapter (setup, instrumentation, safety constraints, limitations).
Days 78–80 May 10 – May 12	Finalization	Full thesis pass: symbols/units/definitions consistency; figure readability; citation completeness. Rewrite Abstract + conclusion based on final results (not early expectations). Prepare defense/presentation outline (10–15 slide story: problem → approach → results → viability decision).

Expanded literature starter list

A curated expansion aligned with (a) virtual sensing, (b) sensorless position estimation, and (c) linear motors. It mixes reviews (for chapter structure) and implementable papers (for your benchmark).

Your provided seeds (keep)

Krishnakumar et al., review on sensorless/PWM/control possibilities for permanent magnet motors (EV context).
Zbib et al., back‑EMF position sensing techniques for brushless PM motors (review chapter, Springer Nature).
Kosierb et al., virtual sensors review (robotics angle), hosted on SSRN.

Virtual sensing and soft-sensing foundations

Martin et al., “Virtual Sensors” (cost and lifecycle framing; widely cited).
Sen et al., virtual sensors stepping in for faulty sensors (industrial validation angle).
Barabás et al., modeling/developing virtual sensors (design patterns and fusion framing).

Sensorless position estimation (organize by operating region)

Li et al., review categorizing rotor position estimation approaches and discussing sensor drawbacks/cost/maintenance.
Janiszewski, sensorless MPC with UKF (observer-class baseline; pros/cons).
Urbanski et al., ANN-assisted position estimation in zero/very-low speed region (low-speed challenges).

AI-based sensorless estimation (implementable comparators)

Putra et al., modified Elman NN as a machine-learning observer; includes experimental validation.
Gao & Zong, RNN-based rotor position estimation addressing harmonics and wide operating range.
Gamazo‑Real et al., ANN-based sensorless estimation using terminal voltages (methodology ideas even if BLDC).

Linear motor–specific sensorless / observer work (must-have)

Yu et al., PMSLM position estimation; discusses encoder vulnerability and back‑EMF limitations at low speed; proposes improved observer.
Guo et al., PMLSM sensorless control improving SMO filtering/chattering; experimental results.
Li et al., PMLSM sensorless MRAS flux observation (linear-motor-specific; useful classical baseline section).
Zhang et al., PMSLM sensorless control with SMO + improved PLL (linear motor specific).

Practical “how to implement” references

Microchip Technology application note on PMSM sensorless FOC using sliding mode (practical constraints, startup, back‑EMF limitations).
MathWorks documentation on NN virtual position sensing workflow: \( (V_\alpha,V_\beta,I_\alpha,I_\beta) \to \theta_e \).

Scope and success criteria

Topic framing

What “position prediction using only the power entries” should mean

Labels vs inputs (important for defensibility)

Clear deliverables that “finish” the thesis in 80 days

Evidence-based technical roadmap

Conceptual map of methods you should cover

Model-based sensorless estimation (classical)

AI / data-driven virtual sensing (core comparison set)

How to connect “virtual sensing” to “sensorless motor drives” in writing

Linear motor specificity (your differentiation)

Dataset and experimentation strategy

Minimum viable experiment design

“When feasible” hardware demo without overcommitting

Comparison and evaluation framework

Recommended minimum set to compare

Metrics that match the “viability” theme

Viability structure you can defend in writing

The 80-day schedule

Expanded literature starter list

Your provided seeds (keep)

Virtual sensing and soft-sensing foundations

Sensorless position estimation (organize by operating region)

AI-based sensorless estimation (implementable comparators)

Linear motor–specific sensorless / observer work (must-have)

Practical “how to implement” references

Links captured in the source document