Abstract

Rotating machinery is fundamental to a wide range of industrial applications, making its reliability and performance critical for operational efficiency. This research focuses on condition monitoring of such machinery through vibration analysis integrated with predictive maintenance strategies powered by machine learning. Traditional maintenance approaches often rely on scheduled inspections or reactive responses to failures, leading to increased downtime and operational costs. In contrast, this study explores the implementation of advanced vibration signal processing techniques combined with machine learning algorithms to enable early fault detection and prognosis. Key features are extracted from time-domain and frequency-domain vibration data, which are then used to train supervised and unsupervised learning models for anomaly detection, classification of fault types, and prediction of remaining useful life (RUL). The proposed approach demonstrates significant improvements in accuracy and timeliness of fault detection compared to conventional methods. Results indicate that integrating intelligent data-driven techniques with vibration-based monitoring systems can enhance the reliability, safety, and cost-effectiveness of maintenance practices for rotating machinery. This research contributes toward developing robust, scalable predictive maintenance frameworks essential for modern Industry 4.0 environments.

Introduction

Rotating machinery, such as motors, turbines, compressors, and pumps, forms the backbone of many industrial systems. The reliability and operational efficiency of these machines are critical, as unexpected failures can result in substantial downtime, production loss, and costly repairs. To mitigate these risks, condition monitoring has emerged as a vital practice, offering real-time insights into the health of machinery. Among various techniques, vibration analysis has proven to be one of the most effective and widely adopted methods for diagnosing mechanical faults such as unbalance, misalignment, bearing defects, and gear failures.

Traditional maintenance strategies, such as reactive or preventive maintenance, often fall short in optimizing machine life and minimizing operational disruptions. Predictive maintenance, on the other hand, leverages condition monitoring data to anticipate faults before they escalate, enabling timely and targeted interventions. Recent advancements in machine learning (ML) have further revolutionized predictive maintenance by enabling systems to learn from historical and real-time data, identify subtle patterns, and make accurate predictions regarding equipment health and remaining useful life (RUL).

This research aims to investigate the integration of vibration analysis with machine learning-based predictive maintenance frameworks for rotating machinery. By extracting meaningful features from vibration signals and applying ML algorithms for classification and prognosis, the study seeks to enhance fault detection accuracy, reduce false alarms, and improve maintenance planning. The outcome is expected to contribute toward more intelligent, data-driven maintenance systems that align with the principles of Industry 4.0 and smart manufacturing.

• I. LITERATURE REVIEW

The condition monitoring of gearboxes in rotating machinery has seen significant advancements in recent years, particularly through the integration of vibration analysis with machine learning techniques such as convolutional neural networks (CNNs) and wavelet transforms. Gearboxes are susceptible to a range of mechanical faults—gear tooth cracks, wear, misalignment—that manifest as changes in vibration signals. Traditional methods of analyzing these signals often rely on Fourier and statistical techniques; however, recent trends emphasize deep learning models that automatically learn fault-relevant features from raw or transformed vibration data.CNNs have emerged as a powerful tool for gearbox fault diagnosis due to their ability to learn spatial hierarchies from structured inputs such as time–frequency representations. Zare et al. developed a 2D-CNN model that classified healthy and faulty conditions of wind turbine gearboxes using texture images derived from time-domain vibration data, achieving 99.76% accuracy [2].

Similarly, Gubernatorov and Gavrilenkov utilized CNNs on frequency-domain features from vibration spectra collected during CNC milling operations. Their model was deployed in an edge-to-cloud architecture and demonstrated high fault detection accuracy, reflecting the trend of using CNNs in industrially relevant environments [3]. These studies emphasize the effectiveness of CNNs in learning discriminative patterns from image-like input representations of vibration signals.

Wavelet transforms have frequently been employed in conjunction with CNNs to enhance diagnostic performance. Lupea and Lupea applied continuous wavelet transforms (CWT) to triaxial gearbox vibration signals to generate time–frequency scalograms. These were fed into CNNs that classified four distinct health conditions with over 99% accuracy. The wavelet preprocessing step allowed better localization of gear mesh frequencies, which were crucial for accurate fault identification [4]. Nguyen et al. introduced a localized adaptive denoising technique (LADT) followed by discrete wavelet decomposition to create multi-scale vibration images. These were input into a deep CNN that successfully diagnosed multiple tooth faults under variable-speed conditions, achieving near-perfect classification results [5]. These studies illustrate that wavelet-based preprocessing not only improves signal clarity but also enhances CNN feature extraction capabilities.

Further improvements have been made through hybrid CNN architectures and the incorporation of attention mechanisms. Cheng et al. proposed a lightweight 1D-CNN model augmented with an efficient channel attention module (LECA) and combined it with transfer learning techniques. Their approach involved extracting wavelet-based features from vibration signals, which were then used to fine-tune a shallow CNN trained on a source domain. This method yielded robust results even with limited training data and under varying operating conditions [1]. Xu et al. contributed a denoising-focused approach using adaptive wavelet thresholding followed by a CNN-LSTM hybrid model. Their framework achieved 100% accuracy in clean environments and 99.97% accuracy under -4 dB signal-to-noise ratio (SNR), highlighting the effectiveness of wavelet-based denoising in noise-prone settings [6].

Despite the high performance demonstrated in controlled environments, several challenges remain in translating these methods to real-world applications. A major limitation is the scarcity of labeled fault data, especially for early-stage or compound failures. Most models are trained on laboratory datasets, which may not generalize well to field conditions due to variations in load, speed, and environmental noise [7]. Moreover, while CNNs excel at classification, their interpretability remains limited, making it difficult to understand the physical basis of the learned features. There is also a lack of standardized public datasets for gearbox diagnostics, in contrast to more established benchmarks in bearing fault detection.

In summary, recent studies underscore the effectiveness of CNNs, especially when combined with wavelet-based signal processing, for condition monitoring of gearboxes. Wavelet transforms contribute significantly by enhancing fault feature visibility and reducing signal noise, while CNNs automate the process of fault classification. Trends in research are moving toward lightweight, hybrid architectures with attention mechanisms and transfer learning to address data limitations. However, challenges such as generalization under variable conditions, explainability, and data availability need to be addressed to facilitate real-world deployment in predictive maintenance systems.

• II. METHODOLOGY

This study presents a framework for condition monitoring of rotating machinery, specifically gearboxes, using vibration signal acquisition, wavelet-based preprocessing, and deep learning-based classification using convolutional neural networks (CNNs). The methodology comprises the following key components: hardware setup for data acquisition, software setup for signal processing and model development, and the step-by-step workflow for predictive maintenance.

1. Hardware Setup

The experimental setup consists of a gearbox fault simulation test rig that includes:

• Electric Motor: A 3-phase AC induction motor is used to drive the rotating system at variable speeds.
• Gearbox: A spur or helical gearbox is coupled to the motor shaft. Fault conditions such as gear tooth cracks and wear are manually induced.
• Load Mechanism: A magnetic powder brake or a mechanical brake applies variable load conditions to simulate real operational stress.
• Vibration Sensors: High-sensitivity piezoelectric accelerometers (e.g., PCB Piezotronics, ±50g range) are mounted on the gearbox housing in horizontal, vertical, and axial directions to capture vibration signals.
• Data Acquisition System (DAQ): A 16-bit NI USB DAQ card (e.g., NI USB-4431) is used to collect data at a sampling rate of 20 kHz per channel. The DAQ is interfaced with a PC for real-time signal acquisition.
• Signal Conditioning: Low-noise amplifiers and anti-aliasing filters are used before digitization to ensure signal integrity.

1. Software Setup

The software setup integrates signal pre-processing, feature extraction, and fault classification. The primary tools and libraries include:

1. Python Environment

Python 3.10 serves as the primary development environment due to its extensive support for scientific computing, machine learning, and signal processing libraries. The open-source nature of Python also allows seamless integration of multiple tools, making it highly suitable for developing vibration-based condition monitoring systems.

1. Signal Preprocessing

Vibration signals obtained from the accelerometers are non-stationary and often noisy. To enhance the feature representation, signal preprocessing is carried out using the PyWavelets library. Both continuous wavelet transform (CWT) and discrete wavelet transform (DWT) are applied to convert the raw time-domain signals into time-frequency representations. These are then transformed into 2D scalograms and multi-scale wavelet images that are more suitable for input into convolutional neural networks (CNNs).

1. Deep Learning Framework

The machine learning backbone of the system is built using TensorFlow and Keras. A custom CNN model is developed to classify different fault conditions based on the wavelet-transformed images. To address the issue of class imbalance and limited fault data, data augmentation techniques such as random rotation, noise injection, and flipping are employed during training. This ensures the robustness and generalizability of the model.

1. Visualization

To facilitate understanding and interpretation of both raw and processed data, visualization tools like Matplotlib and Seaborn are utilized. These tools are used to plot time-domain waveforms, frequency spectra, wavelet scalograms, and model training metrics such as loss curves and accuracy trends. These visualizations help validate the quality of pre-processing and monitor model performance during training.

1. Evaluation Metrics

A set of standard classification metrics is used to evaluate the performance of the CNN model. These include accuracy, precision, recall, F1-score, and confusion matrices. These metrics provide a comprehensive understanding of the model's ability to correctly classify both faulty and healthy conditions across multiple test scenarios.

1. Deployment and Testing

For practical deployment in an industrial or embedded environment, the trained CNN model is converted to the TensorFlow Lite format. This optimized model is then tested on a Raspberry Pi 4 device equipped with 8GB RAM, which acts as a compact edge-computing platform. The Raspberry Pi interfaces with the real-time data acquisition (DAQ) system, allowing for live classification and fault monitoring in a resource-constrained setting.

1. Workflow

The end-to-end methodology follows this sequence:

• Data Collection: Raw vibration signals are collected from the gearbox under various operating conditions (healthy, chipped tooth, crack, wear).
• Wavelet Transform: CWT is applied to transform 1D time-domain signals into 2D scalograms.
• Feature Learning: The scalograms are input into a 2D CNN for automatic feature extraction and fault classification.
• Model Training: The CNN is trained on 70% of the data with the rest used for validation and testing. Data augmentation (flipping, Gaussian noise) is applied during training.
• Prediction and Monitoring: The trained model is deployed to classify incoming real-time signals into predefined health states, enabling predictive maintenance decisions.

IV. RESULTS AND DISCUSSIONS

The proposed methodology was evaluated using vibration data collected under four different gearbox conditions: healthy, worn tooth, cracked tooth, and broken tooth. After preprocessing with wavelet transforms and training the CNN model on the generated scalograms, the system demonstrated high diagnostic accuracy and robustness.

1. Classification Performance

The CNN model trained on the continuous wavelet transform (CWT) scalograms achieved an average as in fig. 1classification accuracy of 99.13% on the test dataset. The confusion matrix revealed that the model was able to distinguish between the four fault categories with minimal misclassifications. Precision and recall values for each class exceeded 98.5%, indicating that the model not only correctly identified faults but also maintained a low false-positive rate.

To assess robustness, additional experiments were conducted by introducing Gaussian white noise (SNR = 10 dB) into the test signals. Despite the noise, the model maintained a high accuracy of 97.4%, suggesting strong resilience to signal corruption and environmental variability—a critical requirement for real-world deployment.

1. Impact of Wavelet Transform

Wavelet-based preprocessing significantly improved the quality of input data by enhancing localized fault features. Compared to raw time-domain signals, wavelet scalograms captured transient anomalies more effectively, which contributed to better CNN feature learning. A control experiment using FFT-based spectrograms instead of wavelet scalograms resulted in a lower accuracy of 93.8%, reinforcing the value of wavelet techniques in vibration-based diagnostics.

Fig. 1 Confusion matrix of CNN classifier

1. Model Efficiency and Edge Deployment

The final trained model had approximately 1.2 million parameters and required 6.2 MB of storage in TensorFlow Lite format. When deployed on a Raspberry Pi 4, the model performed real-time classification with an average inference time of 42 ms, which is suitable for live monitoring applications. Resource usage remained within acceptable limits, with CPU utilization averaging 58% during peak operation.

This efficiency confirms that the proposed architecture is not only accurate but also practical for embedded predictive maintenance systems as in table 1, especially in remote or decentralized industrial environments.

Table 1: Model Comparision Performance

1. Limitations and Considerations

While the results are promising, a few limitations need to be addressed. The model was trained and tested on a controlled dataset with clearly labeled fault conditions. In real-world scenarios, fault progression is gradual and can involve compound failures. Additionally, real field data may exhibit irregular sampling rates, signal dropouts, and unknown noise sources. Thus, future work should incorporate transfer learning and semi-supervised approaches to adapt the model to unlabeled or partially labeled real-world data.

Conclusion

This study demonstrates the effective integration of vibration analysis and machine learning techniques for condition monitoring and predictive maintenance of rotating machinery. By leveraging continuous wavelet transform pre-processing and convolutional neural networks, the proposed approach achieved superior fault classification accuracy of 99.13%, outperforming traditional methods such as FFT-based models. The robustness of the system was further validated under noisy conditions, highlighting its potential for reliable real-time deployment on resource-constrained edge devices like the Raspberry Pi 4. These results affirm that advanced signal processing combined with intelligent data-driven models can significantly enhance early fault detection, reduce unplanned downtime, and optimize maintenance scheduling in industrial environments. However, challenges remain in adapting the model to complex, evolving fault patterns and real-world operational variability. Future work focusing on transfer learning and semi-supervised techniques will be essential to extend the applicability of this framework to diverse and unlabelled datasets, thereby supporting scalable and resilient Industry 4.0 predictive maintenance solutions.

References

[1] Cheng, Xuemin& Dou, Shuihai& Du, Yanping & Wang, Zhaohua. (2024). Gearbox fault diagnosis method based on lightweight channel attention mechanism and transfer learning. Scientific Reports. 14. 10.1038/s41598-023-50826-6. [2] S. Zare, M. Ayati, M. R. Ha’iriYazdi, and A. KabirAnaraki, “Convolutional neural networks for wind turbine gearbox health monitoring,” Energy Equipment & Systems, vol. 10, no. 1, pp. 73–82, 2022. [3] K. Gubernatorov and S. Gavrilenkov, “CNC machine gearbox fault detection with convolutional neural network,” World Journal of Advanced Engineering Technology and Sciences, vol. 11, no. 2, pp. 379–388, 2024. [4] I. Lupea and M. Lupea, “Continuous wavelet transform and CNN for fault detection in a helical gearbox,” Applied Sciences, vol. 15, no. 2, p. 950, 2025. [5] C. D. Nguyen, Z. Ahmad, and J. Kim, “Gearbox fault identification framework based on novel localized adaptive denoising technique, wavelet-based vibration imaging, and deep convolutional neural network,” Applied Sciences, vol. 11, no. 16, p. 7575, 2021. [6] H. Xu, C. Cai, Y. Chi, and N. Zhang, “Fault diagnosis of gearbox based on adaptive wavelet denoising and convolution neural network,” Advances in Mechanical Engineering, vol. 15, no. 2, 16878132231157186, 2023. [7] M. Tiboni, C. Remino, R. Bussola, and C. Amici, “A review on vibration-based condition monitoring of rotating machinery,” Applied Sciences, vol. 12, no. 3, p. 972, 2022.

Copyright

Copyright © 2025 Jayant Kulkarni, Sanika Patankar, Ananya Sharma, Atharva Ansingkar, Sarvesh Bajare, Sarthak Biyani. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.