Applying Classification algorithms on sensor data for Fault Detection

Introduction

One of the numerous advantages of Smart Factories is the real time visibility into operating parameters of your equipments on the floor, transmitted by sensors. Leveraging this data optimally is key to harvesting true benefits from your Smart manufacturing investment. In this post, we will discuss how we can apply Machine Learning methodologies on data generated by equipment sensors to detect faulty equipment condition or diagnose the cause of the fault.

The Process

The Analytics process for fault detection is a four step process, as shown in the diagram below. We will skip the first one, Step 0, which pertains to collecting and processing the data and focus on the analysis portion.

Capture.JPG

Monitoring the health of your equipments

There are two aspects of fault detection:

  • Determining whether the equipment is faulty or not based on certain operating characteristics
  • If it is already determined that the equipment is faulty,  determine the source of the fault

Now, for both of the above, the very first step obviously is to determine the condition indicators- the parameters or characteristics that will be extracted from the system data.

Step 1: Identifying conditional indicators

condition indicator is essentially a feature of  the equipment that you can find in the data generated by the equipment/system. The value of this indicator changes in a predictable way in different operational modes. In case of fault detection, the value of condition indicator can help distinguishing normal from faulty operation or can help predict remaining useful life commonly known as RUL in predictive maintenance).

An experienced modeler would explore and experiment with different condition indicators to find the one that works best. In my experience, a combination of condition indicators can provide better demarcation between faulty and non-faulty conditions.

Some of the approaches that you can use to determine indicators are (not a comprehensive list):

  • Simple metrics or values, such as the mean value of the data over time. Ex: Heat generated by a servo motor rotor
  • Signal analysis, where you analyze the signal data variations over time. Ex: Frequency of the peak magnitude in a signal spectrum
  • Leverage Feature selection approaches which help you reduce large data sets by eliminating features that are irrelevant to the analysis you are trying to perform.

 

One of the widely used feature selection approaches is Principal component analysis (PCA), which finds the linear combination of independent data variables that account for the greatest variation in observed values. In our scenario, principal component analysis can help us determine which features or combination of features are most effective for separating the different healthy and faulty conditions represented in our data.

Step 2 : Apply the Machine Learning algorithm and train model

Many types of Machine Learning algorithms can be applied to the fault detection problem in manufacturing but we will focus here only on classification algorithms.

Classification is a type of supervised machine learning in which an algorithm “learns” to classify new observations from examples of labeled data.

The definition above is just to set the context. Plenty of online resources are available if you want to do a deep dive into classification algorithms. In the context of fault detection and diagnosis, you can pass condition indicators derived from sensors data and their corresponding fault labels to an algorithm-fitting function that trains the classifier.

As an example, assume that you have created a table of condition-indicator values for each member that covers different healthy and faulty conditions. You can pass this data to a function that fits a classifier model. This training data trains the classifier model to take a set of condition-indicator values extracted from a new data set, and guess which healthy or faulty condition applies to the data.

In practice, you use a portion of your data for training, and reserve a disjoint portion for validating the trained classifier.

Some of the classification algorithms that we can use for fault detection are below. These are all widely used algorithms so if you want to develop an understanding of these, there is plenty of material available online.

  • Support Vector Machine (SVM): Leveraging SVM, you can train a binary classification model to distinguish between two states, in this case the presence or absence of a fault condition.
  • Classification Tree : Using a classification tree, you can train a multiclass classification model by reducing the problem to a set of binary decision trees.
  • Linear Regression: Linear Regression can help you train a classifier using high-dimensional training data. This function can be useful when you have a large number of condition indicators that you are not able to reduce using PCA.

Step 3 : Deploy and Integrate

Once you determine that the model is ready to go, you can deploy it in production so that the algorithm can work on actual stream of real time sensor data.

Digital Twins are valuable not only because you can monitor your processes but also because you can apply advanced algorithms on the data to enhance the efficiency of your processes. Once you have significant fault detection data available, you can then use it to train a predictive maintenance algorithm. We will cover that in a seperate post.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s