Applications of Machine Learning in Manufacturing

D.T. Pham, A.A. Afify

Manufacturing Engineering Centre, Cardiff University, UK

Contents

• Introduction
• The Classification Learning Problem
• Machine Learning Approaches to Classification Learning
• Inductive Learning
• Instance-based Learning
• Neural Networks
• Genetic Algorithms
• Applications of Machine Learning Techniques in Manufacturing
• Conclusions

Introduction (1)

• In modern manufacturing environments vast amounts of data are being collected in database management systems and data warehouses from all involved areas.
• Many knowledge-based components have also been added to (semi-) automate certain steps in manufacturing.
• A persistent problem is the gathering of the required expert knowledge to implement these knowledge-based components.
• Machine learning techniques can help in automating the time-consuming process of knowledge acquisition that is essential to the development of a knowledge-based system.

Introduction (2)

• Automation would increase the speed and reduce the cost of development by decreasing the amount of time needed from experts and knowledge engineers.
• Automation could also uncover knowledge that might otherwise be overlooked by those involved in the knowledge acquisition process.
• Thus, machine learning can help to systematise and make the knowledge that is embedded in manufacturing data more explicit.
• A great deal of research in machine learning has focused on classification, the task of developing a model, from a set of previously classified examples, that can correctly categorise new examples from the same population.
• Many manufacturing problems fall under the category of classification.

The Classification Learning Problem

• Given:

• a set of pre-classified examples from the problem domain called the training set.

• Learn:

• a model that can classify new examples from the same population.

Machine Learning Approaches to Classification Learning

• Inductive Learning

• Decision Tree Induction

• Rule Induction

• Instance-based Learning

• Nural Networks

• Genetic Algorithms

Inductive Learning

• Categories:

• Decision Tree Induction
• “divide-and-conquer” inductive systems.
• Knowledge is represented as decision trees.
• Rule Induction
• “separate-and-conquer” inductive systems.
• Knowledge is represented as IF-THEN rules.
• Advantages:
• Fast compared to other techniques,
• Simple and their generated models are easy to understand.
• Obtain similar and sometimes better accuracies compared with other classification techniques.

Instance-based Learning (1)

Main characteristics:

• Simply store the training instances for future use.
• Defer generalising beyond these instances until a new instance must be classified.
• Classify a test instance by finding the nearest stored instance according to some similarity function, and assigning the class of the latter to the former.

Instance-based Learning (2)

Advantages:

• Simplicity.
• The ability to model complex target concepts.
• Information present in the training instances is never lost.

Disadvantages:

• Greater storage requirements.
• Higher computational costs when classifying new instances.
• Higher sensitivity to noise and irrelevant attributes.

Neural Networks (1)

Main charachteristics

• Computational models of the brain.
• Function approximation tools which learn the relationship between independent variables and dependent variables.
• Make no assumption about the statistical distribution or properties of the data, and therefore tend to be more useful in practical situations.
• An inherently non-linear approach, giving them much accuracy when modelling complex data patterns.

Neural Networks (2)

Advantages:

• Wide applicability.
• Higher accuracy.
• Robust to noise in the training data.
• Well-suited to complex problems.

Disadvantages:

• Difficulty in understanding the models they produce.

• Computationally expensive to train.

Genetic Algorithms (1)

Main characteristics:

• Optimisation algorithms based on evolution principles.
• Potential solutions represented as strings (genes).
• Best string is found using genetic operators:
• Selection
• Crossover
• Mutation

Genetic Algorithms (2)

Advantages:

• A potentially greater ability to avoid local minima than is possible with the simple greedy search employed by most learning techniques.
• Especially suited to complex tasks.
• Robust to noise in the training data.
• Easily to be parallelised.

Disadvantages:

• Higher computational cost.

Applications of Machine Learning Techniques in Manufacturing (1)

• Dealing with different types of data (numerical, nominal, text and images).
• Handling noise, outliers and fuzzy data.
• Real-time processing.
• Dealing with large data sets and data of high dimensions.
• Producing results that are easy to understand.
• Being simple to implement.

Applications of Machine Learning Techniques in Manufacturing (2)

 Main stages:

• Formulating the problem.
• Determining the representation.
• Collecting the training data.
• Evaluating the learned knowledge.
• Fielding the knowledge base.

Applications of Machine Learning Techniques in Manufacturing (3)

Example Applications

• Design optimisation.
• Production Planning and scheduling.
• Quality and process control.
• Process planning.
• Equipment fault diagnosis.
• Machine condition monitoring.

Applications of Machine Learning Techniques in Manufacturing (4)

Features of Successful Applications:

• The problem is of a sufficient degree of complexity.
• The problem can be formulated to match existing learning algorithms.
• Ample training data are available in an appropriate format.
• The training data are representative.
• The data are free from noise or can be cleansed cost effectively.
• Methods for learning and evaluating performance are suitable for the application under consideration.

Applications of Machine Learning Techniques in Manufacturing (5)

Limitations:

• Most existing algorithms are prohibitively inefficient when it comes to large data sets.
• Most techniques to date handle only data with continuous and nominal values
• Present machine learning models and manufacturing models are very loosely coupled.

Applications of Machine Learning Techniques in Manufacturing (6)

Challenges:

• Developing learning algorithms that can process large data efficiently, while building from them the best possible models.

• Designing algorithms that can learn regularities in other data types, such as text and images.
•  Closer integration between machine learning models and manufacturing models.

Conclusions

• Machine learning algorithms are demonstrably of significant value in a variety of real-world manufacturing applications.

• In these applications, much of the power derives from proper formulation of the problems and from crafting the representation to make learning tractable.