[Lecture Note] From Predictive to Prescriptive : Autonomous Maintenance Operation through Failure Twins

[Background] Evolution of maintenance

  • Indstry 5.0 : human-centirc, resilient, sustainable
  • (What has happened?) Scheduled maintenance : Historic data, assessment (risk and benefits)
  • (What is happened?) Condition-based maintenance (CBM)
  • (What happened, What is happening and what will happen?) Predictive maintenance (PdM) : detection, diagnose, predict
  • (What should be done to make something happen?)

PdM (GEneral Processs of Predictive Maintenance)

Physic robot

data

Online Monitoring

Data processing

Prediction model

Deep learning model have bocome the SpA of PdM

However,

  • Require a large amount of labled training data
  • Lack of physical explanation
  • Real : lack physics-baed failure behaviour model

Prescriptive maintenance : State-of-the-att

Most of PsM is passive:

  • Prescribing the maintence

Existing PsM is based on stochastic correlation, but not true causal relationship

To discuss Today

Introduction : digital failure twin

Modeling : DFT modeling considering multiple dependent failure behaviors

Updating :

  • Parameter updating considering

Decison supports (3) PsM:

  • RUL-aware control
  • PsM based on causal inference
  • AI-driven planning and control of robots

Discussions of a few ideas

What is a digital twin, and Why?

What is a digital twin?

simulation models ? monitoring and visualization ?

  • Digital model (standard simulation model) Not communication to physical world
  • Digital shadow : Putting census
  • Digital twin : always catch

Very few papers focus on the third level

Why moving from digital twin to digital failure twin

  • traditonal digtial twins focus mainly on normal system behaviors
  • Failure and Reliablity
    • One happened, other will often happen

Digital failure twin mdeling

  • reference
  • dependence
  • efficient

Reference Model

is like a road map

5 :

  • digital model
  • Failure behavior model
  • Data and knowledge
  • Connection and updating
  • Decision and control

failure behavior modeling

Evolition:

  • Empirical model : Not concerning the clause, just collect the failure data
  • State/time-based model : logical combination of components
  • Health indicator model : more details, identify, check the degradation, continuous state failure
  • Hybrid system mode

Stochastic hybrid system modeling of dependent failure ehavior

model curve + disctrete : continuous and discrete

Efficient estimation of reliability through moment closure

Semi-analytical analytics

A contrete example :

image-20260812170300445

Updating the digital failure twin

  • unable to guarantee 100% accurate

Sim-to-gaps : Motivating example

Developing the digital change

Diagnose catch things in 3-dimensional space

And to use to train a new model

Many monitor

General approach : Simulation model to train deeper model and then evaluate

Obervations :

  • perform well on the validation set, but drops significantly on real-world

Update I : Alignment by model

Add a error block, to change the parameter

To collect the real-data

Re-run

Accurary: 60% - 73%

Update II : Alignment by algorithm

image-20260812170321627

Magic : Blue as accurate as possible, Pink as inaccurate as possible (try to confuse)

Ideal : Pink one become completely confuse

No difference between sitmulation and reality

Improve : 85%

Prescriptive maintenance based on DFT

decision model, full real-time, can be address better and send back

RUL-aware control : Current challenge of robot motion planning

  • Conventional path planning cannot perceive motor degradation

Satisfy the constraint

Optimization problem : minimaxing the energy consumption

Different motor more, more available

Need to maintain different motor at different time

Idea : Cost function in RRT*

Cost(s1,s2)=∣∣x2−x1∣∣2Cost(s_1,s_2)=||x_2-x_1||_2

This is called Path length

Get a hign penalty to prohibit do that

PsM based on causal inferene : Problem settings

To inspect the transformer

To detect the frequency

Collect the obervation data, the core-relation

Data tells the more frequent, shorter lifetime.

Why ? Confining effect.

How to correct ?

Reliability modeling based on propensity score matching

Most similar, and compare

Prescribe, approach much better

Intelligent AI agent for resilience planning and control

Really think why it fails