Project topics and materials
Get Complete Project Material File(s) Now! »
Related Work
There are several studies published previously on the detection of faults in gearbox and the rotatory machinery by several groups using multiple techniques, as summarized below briefly.
F. Ribeiro et al. have used non-machine learning techniques such as similarity-based models (SBM) to automatically classify the faults in rotatory machinery [4]. As a result, they classify the faults with an accuracy of 96.43 percent.
In another study by A. Alzghoul et. al, the authors classified the rotatory faults with the accuracy of 97.1 percent using Artificial Neural Network (ANN) [5]. Like our study, MAFAULDA [6] machine fault database was used in both studies [4, 5].
Similarly, signal processing-based preprocessing algorithms and neural networks has been used to classify the gearboxes faults in another study by W.J. Staszewski et.al [7]. These models detect and classify the gearbox faults without any errors.
Zhang Qiang et.al has shown to use self-organizing map-based fault models to detect the gearbox faults with an accuracy of 95 percent [8].
With these emerging techniques and methodologies, there are still several challenges such as as computing resources and programming methods as discussed in detail in one of the study by S. R. Saufi et al. in 2019 [29]. In this study, they highlighted the challenges of machinery fault detection using deep learning. The main challenges of implementing a deep learning-based system for machinery fault prediction required high performance resources such as a GPU-based system [29].
Another challenge while performing this type of studies is at its architecture level to train the DL model. Selection of activation function and training the model required prior knowledge. Now a days different types of programming tools are using while implementing this type of system. Each programming environment have different coding styles. It might affect the fault diagnostic performance of the model. To build the DL model required huge amount of historical data to train and test the system [29].
In a more recent study published in 2021 by S. Ayva, comparative analysis and evaluation of several ML algorithms was performed by Serkan Ayva et al.[30] .Their results showed that random forest (RF) outperformed all the other algorithms studied. This enabled them to incorporate the best performing machine-learning model into the production system in the factory [30].
Industrial Maintenance and Machine Learning
Maintenance
The maintenance cost in many industries is higher than operational and production costs due to premature equipment failure [9]. The profitability of any industry generally depends on the maintenance process.
Normally maintenance in industries happens when the equipment reaches a certain age or stops working [10]. It is good to do scheduled maintenance, but it doesn’t provide any information about the equipment’s health in the future. To optimize the production lines and equipment reliability, different types of maintenance can be performed based on the resource. The most common types of industrial maintenance are Figure 3.1
1. Reactive Maintenance
2. Preventive Maintenance
3. Predictive Maintenance
Reactive Maintenance
In this approach, maintenance can be performed when components or machinery have a problem or stop working. Normally maintenance will perform after the equipment failure as shown in Figure 3.2. Although the component or machine is used full lifespan, drawbacks of this approach are
● Unscheduled maintenance
● Downtime is increased
Preventive Maintenance
In this approach, the machine or component is replaced in advance before it fails. It helps to avoid unscheduled maintenance. The maintenance will perform during the regular interval as shown in Figure 3.3. The drawback of this approach is [11,12,13]
● The component or machine is not fully utilized
● Over maintenance is performed
The drawbacks of regular maintenance are
● Breakdown time is increased
● Productivity is reduced due to regular maintenance
● Over maintenance of some equipment or machinery
● Operation cost is an increase
● The life span of a machine is decreased
● More skilled labor is needed to maintain the equipment
Predictive Maintenance
It predicts the fault and performs the maintenance on the machine or equipment before the fault or failure happens as shown in the Figure 3.4 . Only the components or machines can replace which is going to fail soon. It extends the life span of the equipment. There are several advantages of predictive maintenance [13,14,15] such as,
● It can reduce the unplanned downtime
● It can help to identify fault or equipment health by condition monitoring to avoid costly equipment failure
● It decreased the planned downtime by reducing inspection and premature repair Predictive maintenance system is an IoT based system. The drawback of this approach is the initial cost to build such a system is very high.
Figure 3.4: Predictive maintenance overview
Machine Learning (ML)
IoT and cloud computing make machine learning possible in manufacturing and other industries. Now it is much easier to get the data from the industrial equipment with IoT devices. These data from the industrial equipment will help us to build the ML models to predict the faults. ML transforms some of the tasks to a machine that was previously not possible with humans [16].
Types of Machine Learning
The ML is of three types
● Supervised Learning
● Unsupervised Learning
● Reinforcement Learning (RL)
1. Supervised Learning
Supervised learning techniques are easy to understand and implement. Labeled data is provided to the ML models [17,18]. It means both training and validation data are labeled. The training datasets comprise both inputs and target outputs in supervised learning as shown in Figure 3.5, which allow the model to learn and improve over time. When the model is fully trained it will predict the new or unseen data with a good label. It can be used for both classification and regression problems. The algorithms in supervised learning are decision trees, random forest, support vector machine, navies byes, linear regression, logistic regression, etc.
Unsupervised Learning
In this approach the user does not need to provide the label data to the model, it works with unlabeled data [19]. It allows the model to detect patterns and information on its own Figure 2.6. It is useful to find the unknown patterns in the data. The algorithms in unsupervised learning are clustering, K- Nearest Neighbors (KNN), anomaly detection, Principal Component Analysis (PCA), etc.
Reinforcement Learning
RL is a type of ML and does not require a lot of training data. Instead of environments are given to the RL models, the agent learns from its environment by trial and error to achieve goals and get rewards Figure 3.7.
Dataset and Faults
Data is the core component of any ML/DL model. Quality data is required to perform these models efficiently. The performance of the ML/DL model can improve by integrating more data into the ML/DL system. The data can be of many forms, but the ML model mainly rely on
● Numerical data
● Text data
● Categorical data
● Time series data
Experimental Setup
Spectraquest provides different types of simulators for training and studying industrial machine behaviors. These simulators accelerate learning and help to understand the different types of fault in industrial machinery [20]. The data we used to train and test the ML model was taken from these simulators
● SpectraQuest’s Gearbox Fault Diagnostics Simulator
● SpectraQuest’s Machinery Fault Simulator
Gearbox Dataset
The gearbox dataset used in this study is publicly available at OpenEi [21]. The data was recorded by OpenEi [21] with the four vibration sensors placed in different directions on spectra quests gearbox fault diagnostics simulator [20]. The dataset is recorded with a different load from 0 to 90 percent and contains information about the health conditions of the gearbox based on the vibrational sensors reading. Gearbox dataset describes only two states of gearbox such as
• Normal
• Broken teeth
Machinery Fault database
The data from spectraQuest Machinery Fault Simulator (MFS) are collected by sensors and stored in the machinery fault database [6]. The database contains 1951 multivariate time series data comprised of six different simulated states such as
● Normal
● Horizontal misalignment
● Vertical misalignment
● Imbalance faults
● Underhang bearing fault
● Outer bearing faults
The rotatory machinery faults database contains the following percentage of each category of data as shown in Figure 4.1.
The rotatory machinery database contains the least amount of class normal data and maximum class underhang bearing faults data. The summary of the measurements is shown in Figure 4.2
Rotatory machine states
The data stored in the machinery fault database is acquired with the help of six accelerometers, a microphone, and a tachometer attached to the machine fault simulator [4]. It contains a total of 1951 scenarios as shown in Figure 4.2. The data describe the normal and five faulty states of the rotatory machine.
Normal
The normal sequence means without any fault. The 49 measurements of the normal sequence were used in this study as shown in Figure 4.2. These sequences have been recorded with fixed rotation speed (range 737-3686 rpm) [6].
Imbalance
The total number of imbalance faults was 333 measurements [6]. The data was recorded with the load values (6g to 35g) as shown in Table 4.1
Table of contents :
Acknowledgments
List of Abbreviations
1 Introduction
1.1 Motivation
1.2 Background
1.3 Problems Definition
1.4 Proposed Solutions
2. Related Work
3 Industrial Maintenance and Machine Learning
3.1 Maintenance
3.1.1 Reactive Maintenance
3.1.2 Preventive Maintenance
3.1.3 Predictive Maintenance
3.2 Machine Learning (ML)
3.2.1 Types of Machine Learning
4 Dataset and Faults
4.1 Experimental Setup
4.2 Gearbox Dataset
4.3 Machinery Fault database
4.4 Rotatory machine states
4.4.1 Normal
4.4.2 Imbalance
4.4.3 Horizontal misalignment
4.4.4 Vertical misalignment
4.4.5 Underhang bearing fault
4.4.6 Overhang bearing fault
5 Methods
5.1 Raw data / Sensors reading
5.2 Preprocessing
5.2.1 Standard Deviation
5.3 Machine Learning Pipeline
5.3.1 Decision tree
5.3.2 Random Forest
5.3.3 Ada-boost (Adaptive Boosting)
5.4 Deep Neural Network (DNN) Pipeline
5.4.1 Activation Function
5.5 Performance Evaluation
5.5.1 Confusion matrix
5.5.2 Accuracy
5.5.3 Error Rate (ERR)
5.5.4 True Positive Rate (TPR)
5.5.5 False Positive Rate (FPR)
5.5.6 Precision
5.5.7 F1-Score
5.5.8 Mean Squared Error (MSE)
5.5.9 AUC Score
5.5.10 ROC Curve
6 Results
6.1 Gearbox Fault Prediction
6.1.1 Performance Evaluation on raw data
6.1.2 Performance Evaluation of normalized data
6.2 Machinery Fault Prediction
6.2.1 Performance evaluation of ML model in MFP dataset
6.2.2 Performance evaluation of DNN model in MFP dataset
7 Discussion
7.1 Gearbox Fault Prediction
7.2 Machinery Fault Prediction
8 Conclusions
