1. What is Predictive Maintenance?
According to Wikipedia the definition of predictive maintenance is “techniques that designed to help determine the condition of in-service equipment in order to estimate when maintenance should be performed. This approach promises cost savings over routine or time-based preventive maintenance, because tasks are performed only when warranted.” Predictive maintenance periodically monitors machine based on the analysis of data collected through the monitoring process. One of the objectives of predictive maintenance is to verify the condition of the machine in order to anticipate eventual problems that may lead to higher costs.
2. Benefit of Predictive Maintenance
a. Reduction in maintenance costs
Most machine come with manufacturer recommendations on maintenance. The problem with manufacturer recommendations is that they represent an average. So the need for each machine to perform maintenance varies depending on external conditions and workloads. In this scenario, you can maintain machine too frequently. In other words, the machine gets maintenance when it is not required. You waste money and resources doing unnecessary maintenance, to prevent this from happening we can use predictive maintenance (PM) to predict when machine need maintenance.
b. Reduction in machine failures
Regular monitoring of the actual conditions of the machine and process systems can reduce the number of unexpected and catastrophic machine failures. Failing to maintain machine means that the machine will break while operating. Here, the costs could be substantial. Not only do you have the repair costs, but also costs associated with lost production. If a machine on the assembly line goes down, the line cannot produce anything. So, we need to maintain machines when they need maintenance.
c. Increased service life of parts
By doing regular maintenance we keep every part of the machine running well. Doing this will increase the service life of each part, compared to performing careless maintenance. The ability to pre-determine defective parts that require maintenance will help reduce repair time and costs.
3. Predictive Maintenance Using Machine Learning
Each machine has a significant amount of historical data, and there may be some patterns of each machine failure. With the maturity of machine learning techniques, we are now capable in predicting the next failure. There are at least 3 stages that we must go through to be able to apply machine learning methods to perform predictive maintenance:
a. Data Cleaning & Wrangling
This is the initial stage as well as the crucial stage for determining whether we can perform predictive maintenance using machine learning or not. First of all we have to make sure that we have enough data to do the modeling. The data is taken from sensors in the machine. In this data, there must be at least several anomaly events or have a record for several Failure Machine events. A Predictive Maintenance Model cannot be created using only normal sensor data (without failure/anomaly event). By combining historical data when the machine was running normally and when the machine experiences a failure, the algorithm can learn data patterns from each event on the machine, so that we can build a machine learning model which capable of predicting what would occur in the future accurately. After making sure we have enough data, the next process is to remove noise in the data using several transformations and statistical approaches as shown in the chart below.
Figure 1. Data Cleaning Pipeline
b. Feature Engineering
The Machine learning model is only ever as good as the data we train it on. As such a significant proportion of our effort should be focused on creating a dataset that is optimized to maximize the information density of the data. Feature engineering is the method used for achieving this goal. One form of feature engineering is to decompose raw attributes into features that will be easier to interpret patterns from. For example, decomposing dates or timestamp variables into a variety of constituent parts may allow models to discover and exploit relationships. Other methods like removing outliers to ensure that the processed data contains sensible information. Depending on the type of algorithm used, there are many transformations that can be performed on the process engineering features. For example in the chart below there are several different ways of transforming data because they use different algorithms (ML-based & DL-Based).
Figure 2. Feature Engineering Pipeline
Based on the type of case there are three approaches that we can apply:
1. Regression: To Predict the Remaining Useful Life (RUL), or Time to Failure (TTF). The target / label used in this modelling process must be continuous value.
Figure 3. Regression Model
2. Binary classification: To Predict if an asset will fail within certain time frame. The target / label used in this modelling process must be discrete value with 2 distinct values.
3. Multi-class classification: To Predict if an asset will fail in different time windows: E.g., fails in the window [w0+1,w1] days. The target / label used in this modelling process must be discrete value with more than 2 distinct values.
Figure 4. Classification Model
c. Data Modelling
The last stage is modelling. In this process, we train the data that we cleaned in the previous step using several algorithms. For example in the table below we use 2 kinds of algorithms (ML-Based and DL-Based). At this stage it also requires to do hyper-parameter tuning to find the best model and select Performance metrics to assess the quality of the model we have created. After we get the best model then the model is saved for deployment to production.
Figure 5. Model Training Pipeline
For example we can use DMAA AnalytiX Platform to build a modeling pipeline from loading Dataset, doing some feature engineering & selection, select algorithm and finally doing some Hyper-Parameter tuning.
Figure 6. Data Modeling using DMAA AnalytiX Platform
From the above explanation, we can take several important points:
- We can perform predictive maintenance to reduce maintenance costs and machine failures.
- We can use machine learning methods to predict the state of the machine and when the failure occurs.
- There are three types of machine learning models that we can apply: regression (Estimate Remaining Useful Life), Binary classification (estimate failure probability for the next x-days), Multi-class classification (estimate status of the machine for the next x-days).