Maintenance requires all steps possible to maintain or restore the correct operation of equipment or machines. The aim is to eliminate the possibility of failures that can lead to machine breakdown or unscheduled downtimes, or that could escalate to safety problems.
For example, wear, progressive damage, or material deformation due to force that develops in several mechanical parts, for example, roller bearings, O-rings, or gears, is the common cause of failure. Systematic maintenance processes improve machine availability, minimize costs and encourage appropriate maintenance plans to be scheduled. Traditional, preventive maintenance requires the routine monitoring of devices according to a set timeline or fixed target based on the simplistic presumption that faults occur most of the time. However, this strategy does not work most of the time for the reason being failure takes place before planned maintenance or maintenance is performed even though it was not required.
(Centomo et al. 2020, p. 1782.) 2.1.1 Predictive maintenance
Predictive maintenance aims at solving the strategy of the fixed schedule by implementing a procedure to predict specific possible failures. The purpose is to have maintenance unless is really required, i.e., not too soon, or too late. Predictive maintenance has the benefit of substantially reducing maintenance costs by allowing better use of capacities and preventing operation downtimes. (Centomo et al. 2020, p. 1782.)
Predictive maintenance is focused on the prediction of faults based on the data collected by different sensors, for examples vibration, temperature, humidity, or acoustic sensors. Thus, it is important to preset data that describe the different states of the machine for example location sensors or switches as well as the actuator status. Based on preset data and collected data from the decision for maintenance are called out. Due to a large number of data
collection, manual tracking, and decision-making are impossible. That is why machine learning and particularly deep learning are the best fit for data processing. They are often used for predictive maintenance tasks like Remain Useful Life (RUL), Root Cause Analysis also referred to as Fault Diagnosis (FD), Fault Prediction (FP), and Maintenance Strategy Optimization (MSO). But before the use of the Machine Learning (ML) algorithm, all sets of data labeled with the respective fault and patterns or shape of signals have to be discovered. This data is then used as a source of information for predicting when which failure has taken place. Figure 4 illustrates the predictive maintenance working processes and various technologies use to accomplish an effective process. (Çınar et al. 2020, p. 8211;
Klein & Bergmann 2018, p. 3.)
Figure 4. Predictive maintenance working principle and used technologies (Çınar et al.
2020, p. 8211).
The collection of data and creating algorithms from machine learning in predictive maintenance are the most challenging phase in the process. There are two approaches for data collection described in detail in the next sub chapters, and many methods for application of ML algorithm in predictive maintenance such as Artificial Neural Network (ANN), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), Logistic
Regression (LR), Extreme Gradient Boosted Trees (XGBoost), Gradient Boosting Machines (GBM), Linear Regression, Symbolic Regression (SR) (Çınar et al. 2020, p. 8211).
2.1.2 Prescriptive maintenance
Prescriptive maintenance works in cooperation with preventive maintenance and physics-based simulation to indicate not only what and when a breakdown will occur, but also why it will occur when the behavior of equipment had changed. Prescriptive maintenance will take the study a step further by determining alternative choices and their potential effects in order to reduce any harm to the system. The data and analysis will continue in the period prior to the maintenance activity, with the potential consequences and suggestions being continually adjusted and changed, increasing the credibility of the results. The analytical engine will keep monitoring the machine after the maintenance activity is done to see if the maintenance was effective. (Kovacevic 2017.)
A machine learning model that is developed on sensor and service data is required for prescriptive maintenance to be successful. The artificial intelligence model would be increasingly accurate when more high-quality data becomes available, recognizing more indicators of maintenance requirements and failure signatures while providing fewer false positives. During training a prescriptive maintenance algorithm, higher-level information about an industry may be supplied to the machine learning algorithm. This allows the program to consider important factors for example maintenance costs and product downtime.
The machine learning model is trained using specialized hardware, which might be local, or cloud based. The model is code that may be installed on-premises or in the cloud, therefore a means to reach and operate it is necessary. This can be readily connected with various asset management software packages, easing the process of implementing the prescriptive maintenance model's suggestions. At last, prescriptive maintenance needs a company's willingness and ability to put into practice the machine learning suggestions. Hypothetical outcomes created by a prescriptive maintenance program give options that were previously either left by chance or tried and tested. (Aspentech 2021.)
In many sectors and industries, predictive maintenance has been proven to be accurate. A company or organization's physical operations can benefit from the power of machine
learning by implementing prescriptive maintenance suggestions. The difference between prescriptive maintenance and predictive maintenance is that prescriptive maintenance gives a range of alternatives and outcomes from which to choose. In many cases, prescriptive maintenance can also detect capital expenditure needs months before they could even appear to human operators giving time to the company for economy purchases. (Aspentech 2021.)
2.1.3 Methods of data collection
The primary method for determining the health condition of machines is by observing the machine which can be successfully achieved by the implementation of a sensor. The IoT sensor such as accelerometers, gyroscopes, pressure sensors, etc. is normally used for this process. The data coming from the sensor then can be utilized for the ML algorithms for predictive maintenance. However, the implementation of these sensors is not straightforward and vary many cases not appropriate especially for already existing machines and equipment.
The failure data of machines are generally collected by a method called run-to-failure. This method can be very time-consuming and costly for larger sets of data. Once data has been collected, there come difficulties for data handing and drawing conclusions that can be used in predictive maintenance. (Centomo et al. 2020, p. 1786.)
Although there is not enough or sufficient data available for analysis from the actual system, it is possible to generate them. There are four ways to generate the sensor data: fully synthetical, synthetical based on previous data, synthetical based on a virtual simulation model, and finally based on a simplified physical model (Klein & Bergmann 2018, p. 4).
For the generation of fully synthetic data, sensor data is produced using a parameter-based algorithm. This method may slightly drift from its concept because its results are based on the statical model. (Klein & Bergmann 2018, p. 5.) The procedure made by Hahsler et al.
can be used for generating and analyzing fully synthetic data (Hahsler et al. 2017, p. 1-45).
Generation synthetic data based on previous data can be archived by generating new data based on fundamental properties of existing data distribution. This could be achieved by preparing a generative and discriminative neural model by either directly learning the distribution parameters or indirectly using a generative adversarial time-series network.
(Klein & Bergmann 2018, p. 5.)
Synthetical data generation based on a virtual simulation model uses a computer simulation platform for creating a model with the property of the actual model which can be used for data generation. An engineer can model faulty components or a variety of failure scenarios by adjusting temperatures, flow rates, or vibrations or adding a sudden fault in the system.
These faults containing model can be simulation and results containing failure data can be labeled and stored processed for further analysis. Many industries have used this approach for creating virtual factory, machine health testing applications. (Klein & Bergmann 2018, p. 5.)
Synthetical data generation based on a simplified physical model uses a similar approach but instead of using a virtual model, it uses a simplified physical-based model. The models can be replicated in two different ways. One using actual components which are used in real machines and the other using not real components. The benefit of the second method is the remarkably low cost of constructing such a model. Lego Mindstorms and Fischertechnik (FT) provides such construction of the model at a low cost. (Klein & Bergmann 2018, p. 6.)