What is AI-driven root cause analysis?
Root cause analysis is the process of figuring out the reasons behind occurring issues in production systems. For example:
What control parameters are causing defects?
Why is one production line performing better than another production line?
What are the top three factors affecting pump failure?
Traditionally, root cause analysis is done by a data analyst using data collected from the production system. However, today’s production systems are very complex and can include hundreds of parameters, so manual analysis can be daunting, if not impossible.
Modern production systems can reap the benefits from AI-driven root cause analysis. AI is very good at figuring out patterns in large, complex data sets. An AI-driven root cause analysis involves the following steps:
Data collection. Data is collected from the production system over hundreds of runs. Control parameters, sensor values, and any other relevant data are collected in real-time in each run.
Data labeling. The outcome of each run is labeled. A simple label is “good” vs. “defective”, but this can be further refined to include multiple categories of defects.
Model training. Now we have the data for each run and its labeled outcome, we can train an AI model. There are many different types of AI models, so depending on the type of problem, the optimal AI model is chosen.
Model testing. We withhold some data sets from training; they are used for testing instead. For example, if we have data for 100 runs, 80 could be used for training, and 20 could be withheld for testing. The test data is put into the AI model, and we compare the AI model’s output to what they are supposed to be. The percentage of correct predictions is the confidence score of the AI model.
Once the AI model achieves a satisfactory confidence score, it can be used for root cause analysis to figure out which parameters in the model cause “defective” outcomes. The power of AI is that even if there are hundreds of parameters, it can pick out the desired parameters.
When applying AI-driven root cause analysis, the below considerations need to be taken into account:
AI relies on analyzing substantial amounts of data. If we could only collect data from a few tens of runs, then the AI model will likely be inaccurate. Generally, the more data we can collect, the better the AI-driven root cause analysis will be.
Data quality is very important. The most time-consuming part of building an AI solution is not building the AI model; it is collecting, validating, cleaning, and engineering the data. This process needs to be automated to prepare data at large scale.
AI models are not physics-driven, but it is important to understand and validate their outcomes from the physics perspective. Otherwise, the AI models could produce results that don’t make sense.