A Comparative Study of Classification and Clustering Methods for Data Analysis in Digital Transformation and IoT Systems

This studyemploys classification and clustering methodologies on datasets derived from digital transformation and Internet of Things (IoT) initiatives within the cable and automotive sectors. The analytical procedures are conducted utilizing the KNIME platform, employing Support Vector Machines (SVM) and K-Means algorithms. The results indicate that SVM exhibits superior accuracy rates compared to K-Means within both industries. The data collection methodology facilitated by the Mert Software IoT platform is identified as reliable and efficacious. The primary objective of this article is to augment decision-making precision in digital transformation software and contribute to the scholarly discourse within this domain.

Keywords: Machine Learning, Classification, ClusterAnalysis, Industry 4.0.

Digital transformation has emerged as a significant topic in factories in the world over the recent years. The concept of the Internet of Things (IoT) represents all entities connected to the internet through a network [1].  Digital transformation begins by strategically determining the items that will undergo production on assembly lines within a factory. Subsequently, it encompasses the identification of the staff scheduled for duty during production, specifying their cycle times, and recognizing instances of downtime in the course of the process. Digital transformation assumes a pivotal role in these dimensions, given its paramount importance for industry economies. These changes are imperative for the future and have garnered increased prominence in light of technological advancements. Detecting downtime during production is a critical factor that directly affects manufacturing. When a machine fails, this directly impacts the whole factory, and the length of duration is often unpredictable. The software solutions aim to provide real-time notifications to reduce downtime. The ultimate goal is to increase production quantities. In pursuit of this goal, the data collected by Mert Software has been transformed into knowledge discovery toolthrough classification and clustering analyses. Clustering analysis is used to classify data   based on similarities, particularly for data with an unknown number of groups and unclassified data. It is a technique that groups data into discrete clusters based on similarities with respect to units or variables [2]. This study concentrates on two distinct industries: the automotive industryand the cable production industry.The rationale behind choosing two disparate manufacturing sectors lies in the distinct operational characteristics inherent to each. Within the cable production industry, a singular product is manufactured sequentially utilizing a solitary machine. In contrast, the automotive industry permits the simultaneous production of multiple products. A pivotal demarcation lies in the StockId field within the data  sets, exhibiting variation for each record. This study endeavors to assess the efficacy of identical algorithms by applying them to these sectors characterized by disparate production methods.

Klein M., in her article titled "Scenarios of Digital Transformation in Enterprises -A Conceptual Model Proposal," underscores the imperative for businesses to formulate comprehensive digital transformation strategies encompassing processes, business models, customer relationships, and management. Within her study, she delineates the trailblazers of diverse digital transformation methodologies, foreseeing their role in aiding businesses in the development of robust digital transformation strategies [3]. In their article titled "Formation of Digital Factories with Production Tracking Systems and Conceptual Data Analysis," published in 2022, Mustafa and Halil discussed the significance of production tracking, including planning on workstations, raw material inputs, inventory tracking, quality control processes, and maintenance processes, as well as the importance of Key Performance Indicators , (KPIs)  for businesses. Their study delved into the relationship between Manufacturing Execution Systems , (MES) and Enterprise Resource Planning (ERP) systems, highlighting the advantages of digitizing every moment within businesses through these processes. They found that this digitization would significantly reduce production costs, increase labor productivity, decrease production expenses, and facilitate inventory tracking [4]. 

Altuntaş, conducted a quantitative study involving 114 corporate executivesin the article titled "The Impact of Digital Transformation Practices on Corporate Brand Value". The study revealed that 89% of the executives had engaged in digital transformation initiatives, and 34% reported increased revenue and improved customer relationships as a result. Altuntaşargued that the Industry 4.0 revolution would yield positive outcomes for every sector. Furthermore, the author anticipated that digital investmen ts, particularly those enabling brands in the fast-moving consumer goods and retail sectors to offer personalized services to customers, would strengthen brand loyalty andcontribute to an increase in brand value [5]. In the article titled "The Impact of the Internet of Things (IoT) Technology on Businesses in the Digital Transformation Process," published by Çarkin 2020, the author conducted a study to understand the influence of the Internet of Things (IoT) technology on businesses and to assess the existing literature on the subject. The study utilized content analysis as its methodology and reviewed data from the Web of Science    (WoS), Ulakbim, and DergiPark data  bases. Within the context of the digital transformation process, Çarkprovided recommendations for preparing and adapting individuals, organizations, and society as a whole to Industry 4.0 technologies. These recommendations aimed to facilitate a healthy transition and adaptation to the digital transformation era [1]. 

Gürkan's thesis, titled "Development of an Intelligent Factory Management and Information System Supported by Industry 4.0 and Digital Transformation Technologies," introduces a novel approach[6]. The primary goal is to establish a high-tech automation system infrastructure in industry while minimizing human interventi on. This approach aims to reduce error rates during the production phase to a minimum and produce high-quality products. It also focuses on enabling factories, especially in terms of production performance, to self-optimize via network connectivity. The objective is to maximize production speed in factories and minimize production costs. In the course of this research, Gürkandeveloped anAndroid-based IoT application and implemented it, using electronic cards, specifically in smart marble factories, particularly in the marble drying phase. As a result of this study, several advantages were realized, including the ability to facilitate planned production, contribute to high-quality production, accelerate mass production processes, and minimize production losses [6].Kaynarand his colleagues, in their 2016 article titled "Sentiment Analysis with Machine Learning Methods," conducted sentiment analysis on datasets containing movie reviews from IMDB using classification analyses, specifically the Support Vector Machines , (SVM) algorithm and Multilayer Perceptrons ,    (MLP) algorithms. The study revealed that the SVM analysis had a significantly higher accuracy rate compared to other methods [7].

KNIME is a platform that processes data and enables reporting through relationshi ps between nodes. Being open source, it is open to further development by programmerswho can add additional features to the system. It is generally used in data analysis applications in business intelligence processes. It has various components, and these components are called nodes. Analyses are performed through nodes without the need for coding. The outputs of each node can be viewed and interpreted separately. It is widely used in research related to data analytics. With its powerful features and open system architecture, it is becoming increasingly popular [8].

Classification refers to the process of appropriately distributing data into predefin ed classes within a dataset. Classification algorithms are used to learn the characteristics of classes from training data and to predict to which class incoming test data belongs. Among classification algorithms, the most commonly used ones in the literature include Naive Bayes, Artificial Neural Networks (ANN), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and the KStar algorithm. In this study,SVM was employed. SVM is generally usedfor pattern recognition and classification problems. It trains a support vector classifier using a multi-term kernel. It normalizes all attributes with predefined data [9]. The SVMalgorithm is based on Lagrange multiplier equations. Its primary objective is to find the optimal separating hyperplane that best divides data points into different classes. Support Vector Machines accommodate two scenarios: linear and non-linear. Linear SVM is applied to problems that are linearly separable. The structure of linear SVM is illustrated in Figure 1(a).

When there is no linear separability, moving the data to a higher-dimensional space can be considered as a solution. This is the fundamental theory behind non-linear Support Vector Machines. SVM achieves these transformations using kernel functions. In this case, a dataset of dimension 𝑎is transformed into a new dataset of dimension 𝑏, where 𝑏> 𝑎. The structure of this SVM is presented in Figure 1(b). There are numerous kernel functions developed for SVM, and in this study, a radial basis kernel function is utilized. The equation of this function is presented in Equation 1.[13]

Cluster analysis is a method that groups units under investigation in a research study into specific categories based on their similarities, enabling classification, revealing common characteristics of units, and making general descriptions related to these classes [10]. In cluster analysis, grouping is done based onsimilarities and differences. The inputs involve similarity measures or the necessity of calculating which similarities can be applied to the data. Depending on the purpose and field of use, the objectives of cluster analysis are as follows:

1.Identifying the correct types.

2.Building models.

3.Predictions based on groups.

4.Hypothesis testing.

5.Data exploration.

6.Hypothesis generation.

7.Data reduction.

In cluster analysis, distances are calculated between rows of the data matrix. In the formulas, "i" and "j" represent rows of the data matrix, "k" represents columns, "x_ik" represents data in the "i"th row and "k"th column, and "p" represents the total number of variables.The K-Means clustering method is an effective technique for evaluating many commercial datasets due to its efficiency in clustering large datasets. Being widely used in practical applications for over fifty years, K-Means has become the preferred clustering method in this study for several reasons, such as its speed in comparison to hierarchical clustering methods when there is an expectation of forming a low number of clusters, and its ease of implementation.

The K-Means clustering method is a simple yet effective algorithm for creating clusters based on the available data. The application steps for this method can be outlined as follows [14]:

Step 1: Determination of the number of clusters to which the dataset needs to be partitioned.

Step 2: Random assignment of initial cluster centers for k clusters.

Step 3: Identification of the nearest cluster center for each cluster data point.

Step 4: Calculation of the cluster centroid for k clusters and updating the position of each cluster center based on the new centroid value.

Step 5: Iteration of the processes between steps 3 and 5 until convergence or termination is achieved.In the third step, the distance from each cluster data point to the nearest cluster center is determined using the Euclidean Distance formula:

In manufacturing facilities where production processes occur, signals generated by machinery, including production and stoppage signals, are conveyed to operator panels through an electronic card known as IOCARD. The IOCARD serves the function of converting signals originating from the machinery into readabledata strings, which are subsequently transmitted to the panel software. The panel program, developed in C#, undertakes the reception of this data and parses it based on station-specific definitio ns. For instance, upon receipt of a signal indicating an increase in production count, the program increments the production accordingly. Similarly, if a signal denoting operational status is received, the program sets the status to 'operational.' In the event of a stoppage signal, the system state is modified to a predetermined 'stop' status. Notably, the system accommodates the establishment of approximately 40 distinct definitions.In this study, the accuracy of the production data collected by Mert Software IoT platformwas used for classification and clustering. This study was conducted using datasets collected in the automotive and cable industries. In Table 1, an example dataset used in the cable industryis provided. In Table 2, an example dataset used in the automotive industryis provided. Table 3 explains the variables, and Table 4 specifies the data types.

The process of collecting data from production and reporting the classified and clustered analyses of the collected data is as follows:

-Determination of points to receive signals from the machine.

-Installation of SQL SERVER for recording data collected from the machine. Activation of the system.

-Analysis of the collected data. In this phase, the system conducts initial checks before system analyses, verifying the existence of shift definitions. If a shift definition exists, the system initiates operations and performs clustering analysis for Personnel, Production, and the Manufactured product. The system operates synchronously, conducting Classification analysis for Production and Time.

-Following these analyses on the raw data, the results are utilized in reporting and dashboard tools for faster and analyzed data presentation.

4.1Confusion Matrix

It is a metric used to evaluate the performance of an algorithm in classification problems. The confusion matrix visualizes the number of correct and incorrect classifications by comparing the actual class labels with the predicted class labels. The confusion matrix is typically presented in the form of a 2x2 table, as shown in Figure 3, but it can be larger for multi-class classification problems. The 2x2 confusion matrix includes the following four elements:

True Positive (TP): The number of true positives. TP increases when the actual class label is positive, and the predicted class label is also positive.True Negative (TN): The number of true negatives. TN increases when the actual class label is negative, and the predicted class label is also negative.False Positive (FP): The number of false positives. FP increases when the actual class label is negative, but the predicted class label is positive. It is also known as a false alarm.False Negative (FN): The number of false negatives. FN increases when the actual class label is positive, but the predicted class label is negative. It represents the cases that were missed.

Confusion matrix is used to calculate various performance metrics using these four elements. Among these metrics, accuracy, precision, recall, and F1 score values are included. The confusion matrix is an important tool for understanding the performance of classification algorithms, assessing the strength of the model, and understanding classification errors [11].4.2Performance MetricsTo obtain reliable accuracy results, some measurements are made using the values in the confusion matrix. These measurements are achieved with the accuracy, precision, recall (sensitivity), F1 score, and specificity formulas [11]

Accuracy represents the accuracy value, which is the ratio of correct classifications to the total number of classifications.

Precision gives the ratio of correctly classified data to all the positives. Here is the formula for precision:

Sensitivity provides the ratio of data correctly classified as positive to the actual positive data. Here is the formula for sensitivity (recall):

F1 Score is a value calculated by taking the harmonic mean of precision and recall values. Here is the formula for calculating the F1 Score:

Specificity is a value calculated by taking the ratio of data correctly classified as negative to the actual negative data. Here is how you can calculate specificity:

5.Results

5.1 KNIME Workflow

The workflow developed on the KNIME platform is illustrated in Figure 4. Initially, data is read using the "Excel Reader" node, and then irrelevant elements in the dataset are removed using the "Column Filter" node. Elements such as primary key fields like "CompanyId" and "ID" are among those removed from the dataset. To facilitate learning, the dataset is divided into 80% for training and 20% for testing using the "Partitioning" node. Data is normalized for classification analysis. The "Normalizer" node is used to normalize the dataset, applying min-max normalization to the fields "PstopId," "PID," and "QTY," scaling values to a range between 0 and 1.Subsequently, the dataset is split for training and testing through the "Partitioning" node. The training set provides insights into how well the model explains information in the target variable, while the test set indicates the model's performance with unseen observations. In the modeling phase, learning and prediction nodes for Artificial Neural Networks are added to the model.Following the Normalizer node, for classification analysis, the data is linked to the "SVM Learner" node. This node selects the "Status" field for classification. Then, the output of the Normalizer node and the SVM Learner node are combined in the "SVM Predictor" node. This node aligns the learned data from the SVM Learner with the test data input. Subsequently, it connects to the "Scorer (JavaScript)" node for evaluation. Simultaneously, for clustering analysis, the "k-Means" node is employed, receiving input from the Partitioning node based on the fields "QTY" and "QTY2." This node generates two classes with a defined iteration count of 90. The output from the k-Means node, the Clustering Model, is linked as input to the "Cluster Assigner" node. Following the clustering analysis, to interpret the cluster names as "Working" and "Stopping," the "String Replacer" node is connected. This node replaces the name of cluster_0 with "Working" and cluster_1 with "Stopping." Subsequently, it is connected again to another "String Replacer" node, where cluster_1 is named "Stopping." Finally, the workflow connects to the "Scorer (JavaScript)" node and the "Scorer" node forevaluation.

After the model completes its job, the Scorer node produces results, and as shown in Figure 5, an Overall Accuracy of 74,14% and an Overall Error of 25,86% were obtained.

In this study, an attempt has been made to classify and cluster the production, downtime, and staffdata of companies involved in digital transformation projects by Mert Software. Machine Learning methods such as classification and clustering analyses were employed. In the data sets we used in this study, the SVM algorithm produced more successful results than the K-means algorithm in both the cable industry andthe automotive industry.The study revealed that the method employed by Mert Software for digital transformation projects achieves high accuracy rates.

A major problem with industrial data is that sometimes operators make decisions that can lead to incorrect feedback. This affects the accuracy of the data and causes problems. In this study, we selected examples from factories where signals about production and downtime are collected automatically so that operator errors do not affect data quality. We believe that this study can contribute to the decision-making accuracy of digital transformation software and contribute to scientific research in this field