Gearbox Fault Diagnosis Using Industrial Machine Learning Techniques ^†

Abstract

This paper highlights the need for precise and reliable diagnostic methods for early fault detection in gearbox systems, something critical for industrial maintenance. Advances in machine learning (ML) and image processing have opened new avenues for diagnosis. This study explores ML techniques, particularly edge detection and maximized pooling, with the Inverse Distance Weighting method, for diagnosing gearbox faults from vibration signal images. Using the ODYSSEE-A Eye platform, a model was developed that achieved 96% accuracy in identifying faults from a 500-sample dataset. The research results promote further investigation and progress in this area, indicating specific possible directions for further research.

1. Introduction

Gearbox systems play a central role in numerous industrial operations, as they guarantee high-performance power and motion transmission through machinery. Nonetheless, these systems are also susceptible to different types of operational faults due to stresses, as well as material and environmental wear. With the failure to identify and resolve such faults, gearboxes can result in severe machinery failures, which are often costly in terms of downtime and repair work [1]. Therefore, the innovation and development of accurate and efficient diagnostic tools for the early detection of gearbox faults remains an imperative area of research in industrial maintenance. In recent years, machine learning (ML) and image-processing capabilities have provided new opportunities for fault-detection methodologies [2]. These tools facilitate the processing of advanced data forms, such as vibration signal images, to detect early signs of wear and faults in machinery [3]. Such tools are more efficient than other diagnostic approaches that rely on human judgment and may involve shutting down equipment. Therefore, this research aims to explore the application of ML techniques, including edge detection, and maximized pooling, in combination with the Inverse Distance Weighting (InvD) method, to detect gearbox failures based on vibration signal images [4,5]. More specifically, we strive to create a model within the ODYSSEE-A Eye 2024.1. software to illustrate how these specialized ML and image-processing tools can help efficiently analyze vibration signals and identify gearbox faults in their early stages [6]. In that regard, this paper provides an analysis of the methodology of data collection, image processing, and diagnosis. The rationale behind using image format data is originated from industrial applications. In some circumstances, the numerical data extraction from vibration analysis and measurement software tools is circumstantial or requires high-level license models. The image format, however, is always available from the graphical user interface of these software tools.

2. Methodology

2.1. Data Collection

The present research does not include actual vibration data collection but relies on an existing database to create the operational dataset. The dataset was obtained from an external database containing vibration signals collected from gearboxes monitored under five different conditions [7]. These vibration signals are information-rich signals showing the operating state or condition of a gearbox system, offering the flexibility of assessing the performance and health of a system through numerous methods [8]. The dataset was obtained from the gearbox assembled using a two-stage gear stage. Figure 1, which is edited by us based on [7,8], shows this setup, where the input shaft’s speed and gearbox vibration signals were measured using a tachometer and accelerometer, respectively.

The data are available at a 20 kHz sampling rate to have detailed data on the dynamic response vibrations of the gearbox [8]. However, the time resolution of the pictures is less than 20 kHz: there are about 800 datapoints per revolution. With a total of 500 time-histories obtained from the database, the five gear conditions mentioned above were used in our research: (a) healthy, (b) missing tooth, (c) crack, (d) spalling, and (e) chipping tip. Note that all samples of condition e) were pooled and reduced to see if the recognition of chipping tip faults is still reliable. This created a comprehensive dataset that is exploited to efficiently train and test our model’s fault-detection capabilities. The proportionality of the samples created an almost perfect balance that helped to realize the required balance in the learning set. In addition, the five different gear conditions, each identified by 100 samples, provided the data required to realize the faults. Also note that a wide array of samples presented diversity in the vibration signal, which ultimately strengthens our diagnosis results. The database includes images, which are an alternative of numerical data in some vibration measurement software tools. Figure 2 shows vibration signals for each set of the gear condition. Ultimately, the graphically interpreted vibration data form the basis of our research, from which we make an identification, mainly based on the identification features of each fault.

2.2. Image-Processing Technique

For image data processing, we used the Odyssee A-eye Manager software [9], and for the above processing techniques, namely edge detection and maximized pooling, we used the Edge Detection tool with Max Pooling. The image formats of the vibration signals processed using the described techniques are 875 × 656 pixels in a JPG format. Such parameters of the processed images give the necessary detailed images for further data analysis. Edge detection is an image-processing technique that helps highlight the most important contours and boundaries of an object [10], providing the ability to identify specific structures such as vibration waves and amplitudes. Such an image-processing method significantly assists in analyzing an image-based measurement of vibration signals because it enables the ability to measure and extract structures on which the fault indicators are based. Maximized pooling, or, as it is also called, max pooling, is an additional image-processing stage that allows for a reduction in the image’s size and, at the same time, maintains the most important features [11]. In our method, after determining the edges of the image, only the maximum values from the corresponding areas are included, thus maintaining the most important features such as the vibration amplitude of the waveform. These two image-processing techniques were sufficient to identify and measure key features within the vibration signal images. The described techniques are essential when using the ML model since they allow the extraction of important features and landmarks to identify the gearbox failure type.

2.3. Machine Learning Model

We applied a model within the ODYSSEE-A Eye platform, to forecast potential faults in gearbox gears based on vibration signal analysis [12]. The model’s primary function should be to identify the potential for failure early on, allowing for its prevention and elimination. The model operates by analyzing descriptive variables (X), derived from time-domain vibration signals, to provide outcomes (Y), which are categorized into the following fault types: (a) healthy, (b) missing tooth, (c) root crack, (d) spalling, and (e) chipping tip, as illustrated in Figure 3.

The model follows a structured training and assessment protocol, where the total sample size is 500. This includes 100 samples distributed equally across five gear conditions. Both the testing and learning sets incorporate these samples, with a 10% allocation for testing and the remaining 90% for learning, as can be seen in Figure 4.

The chosen ML method within the A-Eye Manager, a general-purpose GUI tool for the execution of explicitly defined customizations, is based on direct interpolation using InvD. The InvD method, a form of multivariate interpolation, calculates the values at unknown points by taking a weighted average of the values at known points, with weights determined by the inverse of the distance to each known neighboring point [13,14]. For this application, the hyperparameters were optimally set to use seven closest neighbors and a shape of interpolation curvature (power) value of p = −10.

This configuration is specifically tailored to the unique requirements of identifying gearbox faults from vibration signal images shown in Figure 5. The significance of InvD in this context lies in its capability to estimate and interpolate values across multiple dimensions, leveraging the proximity and similarity of known data points. By assigning greater weight to the seven closest neighbors of a query point—a specific vibration signal pattern—the model efficiently identifies the corresponding fault type. This balance between avoiding oversaturation with data from dense clusters and maintaining a generalizing approach ensures sensitivity to minor differences in vibration signals, which are crucial for distinguishing various gearbox faults.

Thus, the model adeptly navigates the challenge of identifying even subtle variations in the vibration signals that indicate different gearbox conditions, providing a theoretically sound and structured methodology for differentiating between vibrations under various operational conditions.

3. Results and Discussion

The results contain valuable data about the precision and reliability of gearbox fault classification. Figure 6 presents an interface for an ML application designed for image-based fault detection. Specifically, out of 50 test data samples, which included 10 samples of each of the five fault conditions, the model generated 48 reliable results. This yields 96%, which can be regarded as a high accuracy rate for the area of ML and industrial diagnostics. One of the two incorrect results turned out to be equally distributed between the two samples identified as healthy gearbox samples. In terms of the amplitude peaks, those identified for the fault and the faultless cases were slightly higher than the average peak of the healthy gearbox sample, and the model made a slight mistake in identifying them. This reveals the need to refine the ML model further, make it more complicated, and be sensitive to even the most subtle deviations from the patterns of healthy gearboxes.

The high (96%) accuracy evidenced the relevance of the model for early gearbox fault detection. The feasibility of the use of the INVD method, edge detection, and pooling maximum is confirmed for the identification of fault types based on vibration parameters. Nonetheless, the model requires further development to improve its performance and achieve an even higher classification accuracy when the variation in the signal becomes less distinct. Learning from the classification in these two cases may become the key to this development and increase the validity and accuracy of the final classification results. Furthermore, the fact that the model performed best for the machine statuses, which indicated a serious physical fault condition such as missing teeth or significant magnitude cracks, shows that it works the most reliably with the larger, more obvious signs of failure. Note that the teaching and testing data contain a healthy set and sets for each fault, with a similar number of samples. However, when the ML model is employed in real circumstances, there are very few faulty transmissions expected. Still, they will be identified.

4. Conclusions

The main goal of this work is the extraction of the fault categorization features from the data. The presented concept is that feature categorization is delegated to the industrially developed ML model. This paper has presented the research and results obtained from the work, which displays the enormous possibilities of ML methods, namely edge detection and maximized pooling with the help of the INVD method, in the field of the diagnostics of gearbox faults using vibration signal imaging. The model was first created with the help of the ODYSSEE-A Eye and then tracked and diagnosed with a 96% accuracy the gearbox fault states from the 500-sample dataset in the real experiment. Although, this dataset is smaller than that used in [7,8], the reliability of the classification results for chipping tip failure was good. The performance of the model once again confirmed the ability of the proposed method to find and diagnose gearbox faults with great accuracy early on. The high performance of the model suggests not only its practical significance but also the importance of applying various machine and image-processing methods in improving the reliability of mechanical systems. The high fault-detection rate of the gearbox faults states before their critical degeneration will reduce downtime and the maintenance costs incurred, thereby improving plant-level operational efficiency.

There are, however, unexplored possibilities for further improvement in this study. An in-depth analysis of wrong fault classifications is one of them, since it elucidates the possibilities for critical enhancement in the areas of faulty evaluation, specifically as regards more subtle fault characteristics. These comprise the increase in the size of the dataset and consideration of a variety of gearbox fault conditions, the refinement of image-processing methods to properly track the more subtle fault features, and the application of additional machine learning algorithms to enhance the model’s accuracy and robustness. To sum it up, the research paper has demonstrated the appropriateness, viability, and efficiency of identifying gearbox faults from vibration imaging signals. The paper has opened new perspectives for researching and applying automatic maintenance, something which entails the least amount of operational disruptions and maximizes the useful life of critical equipment. The resultant applications of these modeling methods imply changing maintenance strategies with high-level ML tools, proactivity, and cost-efficiency in platforms of different industries.