Manufacturing System Algorithm

Development, Modification & Behavioral Modeling

User Model · Machine Model · Transition Matrices · Time Parameters (Tc, Tp)
· Probability Analysis

Prepared by: Advanced Manufacturing Systems Lab

Date: April 2026 | Version 2.1

OEE Target Machines Algorithm Ver. Model States

≥ 85 % 4 Types v2.1 Modified 13 Total

Abstract: This document presents a comprehensive algorithmic framework for manufacturing system
analysis and optimization. The framework integrates User Behavior Models, Machine Behavior Models,
stochastic Transition Matrices, and time-parameter analysis (Tc, Tp) with probabilistic methods. The
modified algorithm addresses real-world issues including machine failures, operator variability, scheduling
conflicts, and dynamic re-routing, producing measurable improvements in OEE, throughput, and
work-in-progress (WIP) inventory.
 1. System Architecture Overview
The manufacturing system algorithm is structured as a multi-layer integrated model. Each layer captures a
distinct aspect of the production environment — human operators, machines, and environmental
constraints — and communicates through shared state variables and probability distributions.

Figure 1: Integrated Manufacturing System Architecture showing User Model, Machine Model, Algorithm Core, and
Output modules.

The core algorithm operates in discrete time steps. At each step it samples user and machine states,
computes effective cycle times, updates transition probabilities, and aggregates performance KPIs. A
fault-handler module intercepts failure events and triggers dynamic re-routing.
 2. User Behavior Model
The User Behavior Model (UBM) represents operator cognitive and physical states during a production
shift. Six discrete states model the full range of operator activity.

2.1 User States

State Description Avg Duration Effect on Tc

IDLE Awaiting job assignment 5–15 min +0%

SETUP Configuring machine/tools 8–20 min +25% setup overhead

OPERATING Actively running machine 45–120 min Baseline

MONITORING Supervising auto-mode 30–90 min -5% (automation gain)

ERROR Responding to fault/alarm 2–30 min +60% (downtime)

BREAK Scheduled/unscheduled rest 10–30 min Machine halted

2.2 User State Diagram

Figure 2: User Behavior State Machine with transition labels and representative probabilities.

2.3 Mathematical Formulation

The operator state at time t+1 is determined by a Markov chain:

U(t+1) = argmax_j { P_U(U(t), j) + ε_j }

where P_U is the user transition probability matrix (6×6) and ε_j is a Gumbel-distributed exploration noise
term accounting for unpredictable human behaviour. The effective cycle time modifier due to user state is:

∆Tc_user(U) = Tc_base × α_U where α_IDLE=0, α_SETUP=0.25, α_ERROR=0.60,

α_BREAK=1.0
 3. User Transition Model
Operator transitions are modelled as a discrete-time Markov chain (DTMC). The 6×6 transition matrix
P_U is estimated from historical shift logs and validated against holdout data using a chi-square
goodness-of-fit test (p > 0.05 acceptance threshold).

Figure 3: User Transition Probability Matrix (heat map). Rows=From State, Columns=To State. Values represent
single-step transition probabilities; each row sums to 1.0.

Key observations:

• IDLE → SETUP has highest off-diagonal probability (0.80), reflecting prompt job assignment under
normal conditions.
• OPERATING → MONITORING (0.20) models automation handover.
• ERROR → IDLE (0.95) captures rapid fault clearance capability.
• BREAK → IDLE (1.00) is deterministic: all breaks return to IDLE.

Steady-State Distribution π_U (computed as π = π·P_U):

IDLE SETUP OPERATING MONITORING ERROR BREAK

12.4% 9.1% 52.3% 18.7% 4.2% 3.3%

 4. Machine Behavior Model
The Machine Behavior Model (MBM) captures the lifecycle of manufacturing equipment through seven
states. Each machine type (CNC, Robot, Assembly, Conveyor) has its own parameterized instance.

Figure 4: Machine Behavior State Transition Diagram with labeled transitions and parameter summary.

Machine State Definitions

State Trigger Output Rate Tc Multiplier

OFF Powered down / End of shift 0% N/A

SETUP New job/tool change required 0% N/A (overhead)

RUNNING Normal production 100% 1.00

IDLE Starved (no parts) / Blocked 0% Wasted time

DEGRADED Wear detected; reduced capacity 65–80% 1.20–1.45

FAILED Critical fault; production stopped 0% Downtime cost

REPAIR Maintenance in progress 0% MTTR dependent

 5. Machine Transition Model
Machine state transitions follow a continuous-time Markov chain (CTMC) embedded at discrete
simulation steps. The 7×7 matrix P_M is derived from:

• Historical MTBF / MTTR records from SCADA logs

• Weibull failure analysis (shape β, scale η)
• Expert elicitation for degradation probabilities

Figure 5: Machine Transition Probability Matrix. High RUNNING → RUNNING (0.85) reflects stable operation; FAILED
→ REPAIR (0.90) confirms efficient maintenance response.

Reliability Metrics:
Machine MTBF (min) MTTR (min) Availability λ (1/min) µ (1/min)

CNC Lathe 480 60 0.889 0.00208 0.0167

Robot Arm 720 45 0.941 0.00139 0.0222

Assembly Jig 360 30 0.923 0.00278 0.0333

Conveyor 1440 20 0.986 0.00069 0.0500

 6. Model Modification: Tc, Tp, Time & Probability
The key innovation of the modified algorithm is dynamic parameter adjustment. Instead of fixed Tc and
Tp values, the model adapts them in real time based on observed system states:

6.1 Effective Cycle Time (Tc_eff)

Tc_eff(t) = Tc_nominal + ∆Tc_user(U(t)) + ∆Tc_machine(M(t)) + ∆Tc_env(t)

• ∆Tc_user: operator-state overhead (+0 to +60%)

• ∆Tc_machine: machine-state overhead (+0 for RUNNING, +20-45% for DEGRADED)
• ∆Tc_env: environmental factors (shift fatigue, ambient temp, material variation)

6.2 Stochastic Process Time (Tp)

Tp(t) ~ N(µ_tp · γ(t), σ_tp) where γ(t) = state-dependent scale factor

The stochastic Tp captures natural process variability. The mean µ_tp is scaled by γ(t) which increases
during degraded or post-maintenance states.

6.3 Dynamic Probability Update

P_U(t+1) = (1-α)·P_U(t) + α·P_U_empirical(t) [α = 0.05 learning rate]

P_M(t+1) = P_M_base · exp(-λ_wear · t_running)

Transition probabilities are updated online using exponential smoothing (user model) and wear-adjusted
Markov kernels (machine model). This allows the algorithm to adapt to changing conditions without full
re-estimation.

Figure 6: (a) Tc vs Tp comparison per machine type with efficiency %; (b) Weibull hazard and reliability curves; (c)
Normal distribution of Tp with ±1σ bounds.
 7. Modified Algorithm – Flowchart & Pseudocode

Figure 7: Complete flowchart of the modified manufacturing system algorithm including fault-handler branch and online
update loop.

Pseudocode Summary
INITIALIZE: Tc, Tp, P_U[6x6], P_M[7x7], λ, µ, OEE=0, t=0 WHILE t < T_simulation:
U(t) ← sample_user_state(P_U, U(t-1)) M(t) ← sample_machine_state(P_M, M(t-1)) IF
U(t)=ERROR OR M(t)=FAILED: INVOKE fault_handler() → update routing, Tc, Tp ELSE:
Tc_eff ← Tc_nominal + ∆Tc_user(U) + ∆Tc_machine(M) Tp(t) ← sample_N(µ_tp × γ(M),
σ_tp) utilisation ← Tp(t) / Tc_eff UPDATE: P_U, P_M using online learning COMPUTE:
OEE(t), WIP(t), Throughput(t) LOG: all metrics t ← t + ∆t OUTPUT: OEE, WIP,
Throughput, state_histograms
 8. Performance Metrics & Simulation Results

Figure 8: (a) OEE decomposition; (b) Throughput comparison baseline vs modified; (c) Machine state occupancy; (d)
OEE sensitivity to Tc/Tp ratio.

Algorithm Improvement Summary:

KPI Baseline Alg. Modified Alg. Improvement

OEE 72.4% 76.1% +3.7 pp

Throughput 0.81 p/min 0.89 p/min +9.9%

Avg WIP 14.2 units 11.6 units -18.3%

MTTR (effective) 68 min 54 min -20.6%

Fault Detection Lag 4.2 min 1.8 min -57.1%

Schedule Adherence 81% 91% +10 pp

9. Manufacturing Issues Addressed

Machine Variability: Stochastic Tp sampling captures natural process variation; Tc_eff adapts to machine
health state dynamically.

Operator Unpredictability: UBM Markov chain models shift-to-shift behavioural variation; online P_U
update tracks learning curves and fatigue.
Unplanned Downtime: Weibull hazard model predicts failure probability; fault-handler triggers re-routing
before full failure.

Schedule Disruption: Dynamic probability updates re-prioritize jobs when a machine or operator
becomes unavailable.

WIP Accumulation: OEE-aware throughput balancing reduces bottleneck-induced WIP by 18.3% versus
the baseline algorithm.

Quality Defects: Degraded-state quality multiplier penalises production in M=DEGRADED, flagging parts
for inspection.

Energy Waste: Idle-state detection triggers machine sleep mode, cutting idle energy consumption by up
to 30%.

10. Conclusion & Future Work

The modified manufacturing system algorithm successfully integrates User Behavior Models (6-state
DTMC), Machine Behavior Models (7-state CTMC), dynamic Tc/Tp parameterization, and online
probability updates into a unified framework. Simulation results demonstrate statistically significant
improvements across all KPIs.

Future Directions:

• Deep Reinforcement Learning (DRL) for real-time scheduling decisions

• Digital-twin integration for live SCADA data ingestion
• Multi-machine interaction modelling (blocking / starving effects)
• Bayesian online learning for faster P_U / P_M adaptation
• Energy-aware OEE metric incorporating power consumption