Machine Learning provides a way of resolving problems that might have previously seemed unsolvable.
Problems able to be solved by Machine Learning are wide-ranging and might be, for example, scheduling difficulties, maintenance issues, process problems, product issues or even worker related.
The machine learning training identifies complex patterns that exist in data, over time, from multiple sensors. These patterns are too complex to be recognised by a human. The patterns are stored as models that later allow actionable alerts and decisions to be created from new data from the same sensors.
There are three main types of analysis:
Detection – something has happened
Classification – between types of thing or behaviour
Prediction – the next value or status over time
Hence, your business needs to be examined, problems identified and mapped onto prospective solutions that perform one or more of the above types of analysis.
The stages of using machine learning are usually:
- Collect data.
- Clean the data.
- Do machine learning training to create a model that detects a condition, classifies a situation or predicts a state.
- For actual use, implement a run time (called ‘inference‘) mechanism on a server, smartphone or edge device to use the model to provide the output in response to new data input.
Cleaning the data can be manually time consuming and needs to be done in such a way that it can be automatically repeated during the final ‘in use’ stage. When novices are tasked with machine learning, projects sometimes break down because while data cleaning could be done in an ‘academic’ or ‘test’ situation it wasn’t considered how it could be done in the production system.
Creating a model needs to be repeated using different machine methods and parameters to achieve an anticipated accuracy. This stage is computationally intensive and in some situations can take weeks or months.
Industry usually generates time series data containing sequences of values that indicate a particular state. For example, a motor might have a vibration sensor that’s used to determine bearing health via sequences of oscillation values. Similarly, the usage of a forklift truck might need analysis of sequences of movement data. Supervised learning is when such sequences are pre-classified, usually by a human, prior to the training stage. Such classification can be time consuming, open to human error, costly or impractical. Unsupervised learning, where the model itself determines the states has recently become achievable and offers the most compelling opportunities.
Machine learning has different risks than conventional programming and in some ways can thought of as a type of R&D. The problem you wish to solve might not be solvable using sensor data or take too long. Alternatively, while the problem might be solvable, the resultant accuracy might not be good enough.
It’s not possible to get 100% accuracy and if it’s a case you need this then machine learning isn’t the solution. Some see this as a turn off but remember that conventional safety critical systems and processes usually have some possibility of failing. Because the models often can’t be understood, it’s not always possible to say, in retrospect, why a particular decision was made. Stakeholders, and maybe your customers might not buy into insights generated in a non-transparent way. Some newer ways of machine learning can provide an indication of what inputs cause particular outputs which provides a level of clarification.