Wafer-transfer robots are critical components in the semiconductor manufacturing process and are responsible for the precise and automated handling of semiconductor wafers at various stages of fabrication. The accuracy and reliability of these robots are paramount because any mechanical deviation or failure can lead to significant defects in the wafers, resulting in costly production delays and reduced yields [1]. One of the key mechanical elements in these robots is the Z-axis belt, which controls the vertical movement essential for the correct positioning of the wafers. This belt experiences wear and tension degradation over time due to continuous operation, which can lead to misalignment, vibration, and, ultimately, equipment failure if improperly maintained. Current maintenance strategies for the Z-axis belts in WTRs typically rely on traditional methods, including periodic inspections and preventive measures. These approaches effectively reduce the risk of unexpected failures by ensuring regular system tests. Although traditional maintenance supports reliability, it may occasionally lead to planned downtimes and associated costs that can be further optimized. In addition, following fixed schedules can occasionally allow belt wear to progress between inspections, potentially leading to system interruptions. Continuous monitoring or predictive maintenance solutions can enhance precision and efficiency [2]. Therefore, there is a critical need for an intelligent maintenance approach that can predict the Z-axis belt wear and tension degradation in real time, enabling targeted interventions before issues arise. Artificial intelligence (AI) has recently gained significant traction in numerous domains, including computer vision, biomedical engineering, autonomous systems, and industrial digitalization [3,4]. Machine learning (ML) and deep learning (DL) approaches have demonstrated immense potential in prognostic and health management (PHM) systems, which are employed to monitor and predict the condition of various rotating machinery components such as those found in power generation systems, gearboxes [5–8], bearings, and motor [9]. Notable data-driven methods [10,11] include backpropagation neural networks [12], support vector machines [13], and artificial neural networks [14], which are effective in enhancing fault diagnosis processes [15]. Among the various algorithms used for data-driven fault detection, convolutional neural networks (CNNs) [16–18] are currently the most popular because they automatically extract and learn features from complex data. However, this study overlooks CNNs because of the limited availability of data, which is crucial for effectively training deep-learning models. In cases where data scarcity is an issue, traditional ML algorithms are often more suitable because they require less data to achieve accurate predictions. Moreover, ML techniques such as artificial neural networks, decision tree methods, hybrid models, and latent variable models have been employed in studies focusing on mechanical fault diagnosis and prognosis in real industrial manufacturing use cases [19,20]. While many studies focus on fault diagnosis for individual mechanical components like motors and gears [21–25], limited research addresses diagnostics at the system level in robotic applications, particularly semiconductor wafer transfer robots (WTRs). This study addresses that gap by focusing on fault prediction for Z-axis belt wear within the WTR system, marking a novel approach in predictive maintenance for robotic systems in high-precision manufacturing. By examining the Z-axis belt as part of a larger robotic subsystem, this research offers a new perspective on the semiconductor industry’s fault detection and maintenance strategies. Despite the limited focus on Z-axis belt wear in prior research, industry experts identify belt tension degradation as a primary cause of malfunctions in WTRs, often leading to issues with the Z-axis motor. Such wear can cause misalignment and other operational disruptions, emphasizing the need for real-time monitoring and maintenance of the Z-axis belt. This study fills this gap by providing critical insights into belt wear and degradation to enhance WTR reliability and longevity in high-precision environments.

This work proposes a novel fault prediction approach for Z-axis belt tension wear using data from a wireless acceleration sensor. For the first time, mechanical faults related to Z-axis belt wear were predicted through sensor-driven data collection, utilizing an industrial robotic system operating under realistic conditions to ensure applicability to practical industrial scenarios. Acceleration data were wirelessly collected from sensors during various robotic movements, with faults classified using ML algorithms. The model was validated using unseen data under multiple operating conditions (low, high, and variable speeds), demonstrating high predictive accuracy across scenarios. This study develops a predictive ML model based on real-time sensor data to prevent unexpected mechanical failures and optimize maintenance schedules. This research advances semiconductor manufacturing by introducing an ML model to monitor and predict wear and tension loss in WTR Z-axis belts. Early detection of tension loss enables manufacturers to reduce unplanned downtime, lower repair costs, and improve WTR performance and efficiency overall.

The remainder of this paper is structured as follows: Section 2 provides a comprehensive review of the existing literature on wafer transfer systems, particularly emphasizing the use of wafer WTRs in the semiconductor industry. Section 3 describes the dataset used in this study and outlines the methodology employed to develop the predictive model. Section 4 focuses on signal analysis and the application of ML to predict the wear in the Z-axis belt. Section 5 presents the results of the model evaluation, including its accuracy and potential implications for optimizing maintenance strategies. Finally, Section 6 concludes the study with a discussion of the findings and offers recommendations for future research.

Fig. 1 illustrates the comprehensive data acquisition and analysis process for monitoring the wafer WTRs. The experimental setup involved monitoring the z-axis belt tension wear, simulating various operational conditions, and artificially adjusting the belt tension. During the data acquisition phase, data was collected using an acceleration sensor and communicated wirelessly to the DAQ PC for further analysis. The signal processing stage included data preprocessing, visualization, and dataset creation. Finally, the performance evaluation used ML models to analyze the results through accuracy scores and receiver operating characteristic (ROC) curves to assess the efficiency and predictive capabilities of the system.

WTRs play a critical role in semiconductor manufacturing by ensuring precise handling and transporting of wafers through crucial processes such as etching, deposition, and inspection. These robots can function in environments where precision and cleanliness are paramount because even the slightest handling errors can result in significant defects in semiconductor devices. The demand for increased throughput in semiconductor production has surged, requiring faster and more efficient wafer transfer. WTRs have evolved to meet these standards as wafers become increasingly precise and fragile. Precision is crucial because even minor impacts or misalignments can lead to wafer damage, resulting in costly delays and lower yields. To address the need for both speed and accuracy, modern WTRs are designed to handle wafers at high speeds with rapid acceleration and deceleration while maintaining precise positioning, ensuring increased productivity without compromising the delicate nature of the wafers. Figure 2 shows the NT series NT420 Kawasaki wafer-transfer robot used in this study.

A wafer transfer robot consists of a stable base, an arm structure for horizontal and vertical movement, and an end effector, typically equipped with a vacuum-based gripper that securely holds and releases wafers. The robot’s movements were driven by a combination of motors and belts, with the Z-axis belt playing a critical role in controlling the vertical position of the end effector. This Z-axis belt’s accurate functioning(operation and positioning) is crucial because it ensures the precise lifting and lowering of wafers. However, over time, the Z-axis belt is subjected to mechanical stress, which can lead to wear and loss of tension, potentially compromising the robot’s accuracy and reliability. Maintaining the Z-axis belt tension is challenging because the wear and degradation of the belt are influenced by various factors, including the operational environment, frequency of movements, and loads applied during wafer handling. Traditional maintenance practices typically involve scheduled inspections and preemptive belt replacement based on estimated wear patterns. However, these methods are not consistently effective because unexpected failures can occur between maintenance intervals, leading to unplanned downtimes and increased operational costs. Moreover, unnecessary belt replacement can further contribute to inefficiency. Given these challenges, there is growing interest in predictive maintenance strategies that aim to monitor the Z-axis belt condition in real time and predict when maintenance is required. By accurately forecasting belt wear and tension degradation, predictive maintenance can help prevent unexpected failures and optimize maintenance schedules, ultimately improving operational efficiency and reducing costs. Integrating ML into wafer transfer robotic systems offers a potential solution for achieving effective predictive maintenance. By continuously analyzing the acceleration sensor data from the robot’s operations, machine-learning models can detect subtle changes in the robot’s operation, which may indicate the onset of belt wear or tension loss. This data-driven approach enables precise and timely maintenance actions, ensuring the robot’s reliability and extending its components’ lives.

Table 1 lists the specifications of the Kawasaki NT420 model used in this study and details its vital performance characteristics. It outlines the basic structure and motion range, including the rotational movements of the θ1 and θ2 axes, which correspond to the first and second arms, respectively, and the H1 and H2 axes, representing the upper and lower blades. The Z-axis indicates the up-and-down motion. Additionally, the table specifies the robot’s position repeatability, ensuring a high precision level within ±0.1 mm and its maximum reach of 1,280 mm, making the NT420 ideal for precise and efficient wafer handling in semiconductor manufacturing environments [26].

ML-based fault prediction in industrial systems applies sophisticated algorithms to analyze real-time sensor data and forecast failures in advance. For example, ML models monitor operational parameters, such as belt tension, acceleration, and vibration, to assess the health of WTRs. These models compare live sensor readings with learned patterns from normal operating conditions to detect anomalies such as wear or mechanical issues, which might indicate potential faults. For instance, if the belt tension in a WTR drops from the standard 350 Hz to lower levels like 300 or 230 Hz, the system flags this deviation as an early sign of degradation. ML models are trained on extensive datasets encompassing both normal and faulty scenarios, which enables them to recognize subtle changes in performance that may precede mechanical failures. By providing real-time predictions, these models allow preventive maintenance, reduce unexpected downtime, lower maintenance costs, and ensure that robots operate efficiently in high-precision environments such as semiconductor manufacturing.

Machine learning classifiers were used to predict the Z-axis belt tension wear in WTR to identify normal and abnormal operating conditions based on sensor data collected under varying belt tension and operating speeds. The WTR was operated under silicon (Si) and aluminum (Al) wafers to ensure robustness in the model. Standard conditions were defined as a belt tension of 350 Hz, whereas abnormal conditions were simulated by progressively reducing the belt tension to 300, 280, or 230 Hz. These changes in the belt tension were designed to mimic wear and degradation, which could affect the mechanical stability and precision of the robot.

Accuracy is a widely used metric for assessing the performance of ML algorithms. It quantifies the proportion of correct predictions made by a model on the test data and is expressed as a percentage. This metric was calculated by dividing the number of correct predictions (True Positives and True Negatives) by the total number of predictions. The accuracy [31] of a classifier is typically evaluated using the following formula:

True positives (TP) and True Negatives (TN) represent instances in which the model correctly predicts the positive and negative classes. False Positives (FP) and False Negatives (FN) indicate cases where the model incorrectly predicts the positive and negative classes. Accuracy provides insight into the ability of the model to correctly classify instances across both positive and negative classes, reflecting a balance between true and false predictions. Sensitivity is also called recall and measures the ratio of accurate optimistic predictions to the total number of positive instances in a dataset. This is a crucial metric, mainly when the cost of missing positive instances (false negatives) is significant. Sensitivity is mathematically defined as follows:

Specificity evaluates the accuracy of optimistic predictions generated by a classification model. This quantified the ratio of true-positive predictions to the total number of instances predicted as positive by the model. Mathematically, the specificity is expressed as

The F1 score is a metric used in ML that combines precision and recall into a single value. It is beneficial in situations with uneven class distribution or when false positives and negatives have different implications. The F1 scores were calculated using the following formula:

The evaluated classifiers included KNN, Logistic Regression, Naive Bayes, random forest, and decision Tree, each selected for its potential to classify normal and abnormal conditions based on sensor data effectively. The performance of these classifiers was evaluated using key metrics, such as accuracy, sensitivity (recall), specificity, and F1-score, and the results are summarized in Table 4.

The results indicated a significant variation in performance across the classifiers. The KNN is highly effective in capturing the nonlinear relationships between the belt tension, wafer material, and speed. It achieved impressive sensitivity and specificity values (99.80 and 99.93%, respectively) with a high accuracy of 99.78 and an F1-score of 99.80%, indicating its strength in accurately identifying normal and abnormal belt tension states. While performing well in binary classification tasks, logistic regression was less successful in detecting abnormal wear patterns than KNN and Random Forest, achieving an accuracy of 84.19 and an F1-score of 84.23%. Its sensitivity was also low (85.76%), suggesting it is better suited for identifying normal conditions than subtle degradation patterns in belt tension. Naive Bayes struggled with this dataset because it assumed feature independence, which may not hold in real-world interactions between belt tension, wafer type, and speed. It underperformed with an accuracy of 67.78, a sensitivity of 70.94, and an F1-score of 66.86%, making it less effective at distinguishing between normal and abnormal wear conditions. However, the decision Tree performed exceptionally well, with an accuracy of 99.79, sensitivity of 99.8%, and specificity of 99.93%. Its ability to model nonlinear relationships allows it to detect wear thresholds accurately, providing valuable insights into belt tension wear. Random forest emerged as the top-performing model, slightly outperforming KNN and decision tree, with an accuracy of 99.80, sensitivity of 99.82, and specificity of 99.93%. Its ensemble nature helped mitigate overfitting and effectively handled the noisy sensor data. Its high F1-score of 99.80% highlights its balanced performance across all metrics, making random forest the most reliable choice for detecting Z-axis belt wear.

Following the classifier evaluation, Fig. 7 compares the training and test accuracies of the various ML classifiers used in this study for fault prediction in the Z-axis belt tension wear. The Logistic Regression model demonstrated equal training and testing accuracies of 84%, suggesting that the model generalized well to the unseen data. However, it performed poorly than the other models because of its linear nature, which may have limited its ability to capture the nonlinear relationships between the belt tension, speed, and wafer type.

In contrast, the decision tree, random forest, and KNN methods exhibited perfect training and testing accuracy (100%), indicating exceptional performance. This result highlights the ability of these models to effectively capture both normal and abnormal wear conditions in the belt, mainly when nonlinear interactions exist between variables such as belt tension and operational speed. The random forest ensemble method likely contributed to its ability to handle noisy data and avoid overfitting, whereas the KNN proximity-based approach helped detect nuanced patterns in the dataset. In contrast, naïve Bayes, which assumes feature independence, achieved an accuracy of 68% on both the training and testing sets. This relatively lower accuracy compared to other models indicates that the feature dependencies in the dataset (such as the relationship between speed, belt tension, and wafer material) reduced the Naive Bayes’ predictive capabilities, as the assumption of feature independence did not hold in this context.

The figure demonstrates that the decision tree, random forest, and KNN are better suited for this fault prediction task. In contrast, logistic regression and naïve Bayes show limitations in handling the complex, nonlinear relationships present in the dataset. The consistent performance between the training and test sets suggests that the top-performing models generalized well to the new data, making them highly suitable for real-world applications in predicting belt tension wear in WTRs.

Fig. 8 illustrates the ROC (Receiver Operating Characteristic) curves for the Random Forest classifier applied to Z-axis belt wear prediction. The plot includes a central black diagonal line, known as the “line of no discrimination”, which represents the performance of a random classifier, where the true positive rate (TPR) equals the false positive rate (FPR) at all points, resulting in an area under the curve (AUC) of 0.5. However, the Random Forest model’s ROC curves for each class closely follow the top-left corner of the plot, achieving a perfect AUC of 1.0 across all classes (class 0, class 1, class 2, and class 3). This ideal AUC score of 1.0 for each class indicates flawless classification performance, meaning the model effectively distinguishes between normal and abnormal belt wear conditions without error. The ROC curves, which plot the TPR against the FPR, confirm the model’s high sensitivity and specificity, demonstrating the Random Forest classifier’s exceptional capability for accurately monitoring belt wear conditions.

Random forest consistently outperformed the other models, including KNN, logistic regression, and Naive Bayes, despite needing further validation across all key metrics. Its ensemble approach and ability to model nonlinear relationships between belt tension, speed, and wafer material make it highly effective for fault detection. The balanced performance of the RF model in terms of accuracy, precision, recall, and F1 scores (as seen in earlier evaluations) further justifies its selection as an optimized model for this system. This robust performance ensures the model can accurately predict the Z-axis belt wear under various operational conditions, which is crucial for predictive maintenance in WTRs. By choosing a random forest as the final model, the system benefits from reliable early detection of belt wear, which helps prevent costly downtime and ensures efficient operation. Nevertheless, given the risk of overfitting, additional steps such as cross-validation with different datasets and real-world testing are recommended to confirm the generalizability and robustness of the model in diverse operating environments. This validation is essential to ensuring the long-term reliability and capacity of the system to handle new, unseen data in real-world scenarios effectively.

In conclusion, this study explored the application of ML to predictive maintenance in WTR, focusing on detecting Z-axis belt-tension wear. The robots’ high-frequency acceleration sensor data (local X, Y, Z, and global Z axes) were utilized to create a data-driven approach that improved traditional fault-detection methods by offering early and accurate predictions of belt degradation. Various machine-learning classifiers, including logistic regression, naive Bayes, random forest, and KNN, have been implemented and evaluated. Among these, RF, decision tree, and KNN exhibited superior performances, with RF achieving an impressive accuracy of 99.8%, demonstrating its capability to outperform conventional techniques. This high accuracy establishes Random Forest as this study’s most optimized model for predicting belt wear. The ROC (Receiver Operating Characteristic) curve for the Random Forest model displays an impressive 100%, indicating the model’s exceptional ability to distinguish between different levels of degradation in the Z-axis belt wear. This perfect score reflects Random Forest’s effectiveness in accurately classifying belt conditions, making it a powerful tool for early detection and maintenance scheduling in real-world industrial applications. This study confirmed the value of ML in proactive maintenance, helping to identify early signs of wear before they result in costly downtime. By applying ML to WTR systems, this study emphasizes the importance of early detection in optimizing system uptime and reducing maintenance costs. Moreover, integrating wireless data acquisition enhances the system’s flexibility, reducing manual intervention and operational overhead. This study provides a strong foundation for future advancements, suggesting that ML models can be extended to monitor other critical components within WTR systems. Future research should focus on real-world deployment and cross-validation techniques to ensure these models generalize well across different environments. This study is foundational for using predictive maintenance to extend the lifespan of industrial robots, particularly in semiconductor manufacturing, where minimizing downtime is crucial. These findings underscore the broad potential of data-driven fault detection for improving reliability engineering across industries.