How To Play With Predictive Maintenance（PdM）

Predictive maintenance can be regarded as a relatively hot application scenario in intelligent manufacturing, involving many advanced technologies centered on machine learning.I hope to help you understand the overall process and principles of predictive maintenance of equipment.

What is predictive maintenance?

The predictive maintenance we are going to discuss refers to the use of various data such as state information and environmental information of equipment operation, mainly based on Mathematical Statistics Models to predict faults.The significance of predictive maintenance is to avoid over-maintenance and save equipment maintenance costs.

Implementation steps

Prerequisites: There must be a certain amount of data accumulation.

The data accumulation mentioned here includes two aspects: the type and quantity of data.

In terms of categories, there must be at least two types of data to be able to establish and train predictive models, namely fault data and status data of equipment operations. In the process of building and training a model, the former is the output of the model and the latter is the input of the model.

In terms of quantity, how much data is enough?

The answer may be disappointing – there is no way to determine how much data is ultimately needed before any models are built and tested, and there are few simple rules of thumb at this stage.

Therefore, for the amount of data, it can only be more beneficial. And, we can be certain that some data requires as much data as possible, such as device failure data, device status data (voltage, current, vibration, etc.) within a certain time window before the device fails.

Different prediction results and modeling methods

The predicted results can be divided into the following two types:

 Discrete predictions：the result is one of a series of finite values, such as "yes" or "no." For example, whether a device will fail during a certain period of time in the future. This will be modeled using a classification model.

The continuous prediction result：the result is a numerical value. For example, at what point in the future, the device will fail, or what is the remaining useful life of the device. This will be modeled using a regression model.

Compared with the regression model, the classification model gives relatively simple prediction results, only "yes" or "no", but the amount of data required is also less.

Model evaluation index

For the classification model, accuracy, recall, etc. can be used as evaluation indicators.

*The accuracy rate describes how many times the total prediction is made if the model has made a total of N predictions. The recall rate describes how many times the model predicts the number of N failures that actually occurred. Of course, these two indicators are finally displayed by the ratio.

Modeling process

The modeling process for predictive maintenance is the same as for general predictive analysis.

Suppose we have collected a certain amount of historical data, including fault data and data on the operational status of the equipment.The regression model is taken as an example to briefly introduce the modeling steps of the prediction model. Of course, this part of the core work requires our data scientists to appear.

Select model

Data scientists first experiment with various algorithm models to fit our existing data. At this stage, it is common to start with a simple model, such as a linear model, and use it as a baseline for comparison with other models.

It should be noted here that the four steps of “select model”, “data preprocessing”, “feature engineering” and “hyperparametric optimization” are not carried out by waterfall, but are iteratively performed.The model originally selected may change later.

Data preprocessing

The purpose of this part of the work is to convert the raw data into the data format required for model input, including the unification of the units of measurement for various types of data, or the exclusion of some apparently problematic data. Of course, some pre-processing work can also be done between model selections, such as the exclusion of error data.

Feature engineering

If we simply understand the model as a function expression y = f (x1, x2, x3.....), then the feature is the argument x1, x2....

Feature engineering can be understood as the process of processing model input variables. There are two processes in this process:

One is to add features, that is, based on the original original variables, and then apply various methods to generate new independent variables, such as the average of x1 and x2. However, it should be noted that the new variables generated must be meaningful in the context of predictive maintenance.

The other is to reduce features. For example, knowledge of the application device selects an independent variable that is closely related to the prediction result among a large number of input variables.

Hyperparametric optimization

The optimization of hyperparameters is to optimize some parameters of the model we selected, so that the predictive performance indicators of the model are more accurate.

If the feature engineering is the work done on the argument x, the hyperparameter optimization is the work done on the function f - adjusting the various parameters of f.

Model evaluation

After building a model based on historical data, we need to evaluate the model, that is, use other data that has not been used in the modeling process to test the model and see how the prediction works. The data of this part of the test model is a part of the data that is generally divided from the historical data before the model is built and trained, usually called test set data.

Model deployment

After the model is established, it needs to be deployed to the actual production system to run, continuously receive data collected from the equipment layer, and perform predictive analysis.

The deployment of the model involves technologies such as data acquisition at the industrial site and enterprise service bus.

Predictive maintenance and condition-based maintenance

You can combine the two to work together.

“Predictive maintenance” is based on mathematical models and is based entirely on data. The "state-based maintenance" is based on the device mechanism model and relies on the understanding of relevant domain knowledge. Therefore, the combination of the two can often achieve better results.For example, in feature engineering, the mechanism model, that is, the domain knowledge of the equipment, is often used to generate new features with practical significance, and accelerate the process of model establishment and training.

Difficulties in making predictive maintenance

The biggest difficulty is the lack of data on the industrial site.

Models that build and train predictive maintenance require at least two types of historical data—fault data (that is, data for unplanned downtime) and equipment operational status data (such as voltage, current, and the like).And the amount of these two types of data should be large enough so that the trained model is more accurate.

However, the probability of downtime for many devices is very low, and may not be several times a year. Therefore, the collection of such data takes a long time.

In addition, many industrial sites do not have a complete equipment data acquisition system, or even if there are some systems such as SCADA, more real-time status monitoring is completed, and the status data of the equipment operation is only saved for a short time ( 3 Months) which is need to be saved by long-term.

Reference material

《Mastering Predictive Analytics With R》

《Machine Learning Techniques for Predictive Maintenance》

Benjamin YANG

Tel:+86 18669037219

Email:[email protected]