Smart factory - module 4

Advantages of smart PPC

• more profitability

• better production and quality

• more production flexibility

• expansion into new markets

• Developping new services for existing productions

Challenges of Smart PPC

• skilled labor shortage

• uncertainty about return of inverstement (ROI)

• lack of data availability

• legal system and IT infrastructure

How does the workforce change with industry 4.0

more highly skilled intellectual workers(in tech and other cognitive areas), less manual skills

Change in the role of AI with industry 4.0

instead of being a differentiator between who does well and who doesnt, it’ll be a basic requirement to even enter the magnet

used in predictive maintenance, knowledge based systems (instad of experts of subjects), adaptive robotics (learns tasks in real time instead of being programmed)

Vertical depth vs horizontal scale, which will drive the future of production

horizontal scale: broad genaral produts that can be used in a lot of different industries

vertical scale: Specialised products with one specific use

vertical scale will grow more

Types of production environment

• discrete - built unit by unit on assembly line, can be undone

• hybrid - has both a process stage and a discrete manufacturing stages

• process - continuous flow of production, production cannot be undone ( like recipes )

AI

Umbrella term for systems that exhibit intelligent behavior.

Why is AI (machine learning specifically) necessary in production planning and control

it enables data-drive smart and adaptable planning and control of systems

helps with:

• increasing complexity of production

• adaptability

• faster changing circumstances

Enables

▪ Optimized planning & scheduling ▪ Autonomous decision making ▪ Resource efficiency & cost reduction ▪ Real-time adaptability

3 different machine learning methods

• unsupervised learning - finds patterns, clusters, or hidden structures in unlabeled data ( clustering, dimension reduction)

• supervised learning - learns from labeled data to predict outcomes (regression, classification into discrete classes)

• reinforcement learning - learning via trial and error by maximising cumulative rewards (independant learning with problem solving strategy)

machine learning

Subset of AI focused on learning patterns from data without explicit programming.

deep learning

A Subset of ML using deep, multilayered neural networks to process data and mimic human thought , represent the world as nested hierarchy of concepts

LASSO regression and at what stage of PPC can it be used

falls in supervised learning

Least absolute shrinkage and selection operator, models linear relationship btwn target variable y and 1+ predictor variables x, and adds a penalty to model to eliminate irrelevant variables. Parametric model

used typically in sales and operations planning to know which variables are actually relevant to better efficiency ((S&OP))

Advantages and disadvantages of Lasso model

• Pros: Reduces overfitting, eliminates irrelevant features, improves interpretability.

• Cons: Unstable with correlated variables (selects randomly), introduces bias (shrinks valid coefficients), parameter selection is critical.

Correlation Analysis and at what stage of PPC can it be used

falls in unsupervised learning

examines the relationship between two or more variables without presupposing causality, determines strength and direction of relationship. Statistical data analysis model

eg Pearson, Spearman

would be used in quality control (production optimisation or market reasearch as well) (S&OP)

pros and cons of correlation analysis

Pros:

▪ Simplicity

▪ Versatility

▪ Visualization

cons:

▪ Spurius correlation (not actual correlation, there is hidden '3rd factor for example or its pure chance)

▪ Linear constraint

▪ Susceptibility to erros

Multilayer Perceptron (Neural Networks) MLP

Falls in supervised learning, deep learning

neural networks to solve complex tasks like preidctions, classification or regression. all neurons are connected via weighted connection. One way flow (from input data to output)

pattern recognition (S&OP)

pros and cons of Multilayer Perceptron

Pros:

• Universality

• Flexibility

• Scalability

Cons:

• Computional Effort

• Dependency on data

• Overfitting

Clustering

unsupervised learning

Data points are grouped so that points in same cluster are similar and points in different clusters are dissimilar

can be used in quality anaysis, customer segmentation, process grouping

Dimension reduction

unsupervised learning

Data transforms complex data with multiple variables into simpler form with less variables but still keeping all significant features

can be used to evaluated component health/wear with more simple graph

Type of clustering and how it works

K - means clustering

1. choose number of clusers you want : k

2. chose k random center points

3. assign points to closest center point

4. update centroids

5. Repeat 3 and 4 until cluster centers no longer change(convergence)

Scheduling in terms of traditional vs smart approaches

traditional: based on empirical data, limited flexibility for complex processes , high dependance of individual experts

smart approaches: use reinforcement learning to predict best schedule, MCTS (Monte Carlo Tree Search)

Scheduling algorithm

1. Sort jobs by Shortest Processing time (SPT)

2. Machine assignment - jobs assigned in order of SPT to available machines

3. Creation of Gantt chart

Gantt chart

visual project management tool that plots tasks across horizontal timeline, showing start and end dates and task dependencies

Monte Carlo Tree Search (MTCS)

tree search algorithm for decision making processes with lots of possible options, finds optimal decisions in incomplete or uncertain environments

pros and cons of MCTS

Pros

• different parts of tree can be worked on at the same time (speeds up simulation)

• dynamic condition assessment (updates strategy continuously as it gets feedback from current simulation)

Cons

• high computationa effort for deep trees

• Depends on number of simulations ran, so it can account for randomness

Reinforment learning

agent learns by interacting with its environment.The agent performs actions, receives rewards or Description punishments, and adjusts its strategy (policy) to maximize cumulative rewards in the long run

Pros and Cons of reinforcement learning

Pros

• flexibitlity

• learning

• long term planning (cummulative reward over time)

Cons

• computational effort

• dependance of rewad signals (depends entirely on how well humans designed reward function)

• exploration vs exploitation (agent must balance exploiting what it already knows and exploring new unknown actions - excessive exploitation makes agent settle on suboptimal good enough option, excessive exploration wastes time)

Matrix production system

stations arranged in grid pattern and product can flow multiple different sequences. flexible machine utilisation( if press 1 is busy, easily go to press 2)

VDI Guidelines

The VDI-Guidelines are technical guidelines produced by the Association of German Engineers (VDI) to create standards an best practices for a variety of technical and engineering topics.