A
Simulation-Based Predictive Maintenance for Rotor Fault Diagnosis in Autonomous Robotic Systems Using Deep Learning Modelsc) ypzhang@mail.xjtu.edu.cn
Authors: Al Imran1,2,a), Changbiao Li1,2,b) and Yanpeng Zhang1,2,c)
1Key Laboratory for Physical Electronics and Devices of the Ministry of Education, Xi’an Jiaotong University, Xi’an, 710049, China
2Shaanxi Key Lab of Information Photonic Technique, Xi’an Jiaotong University, Xi’an, 710049, China
Corresponding Emails: a) emonemon45@gmail.com, b) cbli@mail.xjtu.edu.cn and
Abstract
Autonomous robotic systems, particularly those involved in industrial applications, rely heavily on the performance of their mechanical components, with rotors being central to their operation. Any fault in these systems, such as rotor misalignment, imbalance, or detachment, can lead to catastrophic failures, operational downtime, and financial losses. Predictive maintenance (PdM) strategies, based on fault diagnosis and remaining useful life (RUL) prediction, are crucial for mitigating these risks, extending the lifespan of robotic platforms, and enhancing operational efficiency. This paper presents a simulation-based approach to fault diagnosis and RUL prediction for rotor systems in autonomous robots. We propose a novel methodology using deep learning techniques—specifically Convolutional Neural Networks (CNN) for fault classification and Long Short-Term Memory (LSTM) networks for RUL prediction. The system is trained on synthetic rotor fault data generated from a robust simulation environment that models various fault conditions, including rotor drop-off, misalignment, and imbalance. The results show that the CNN model can classify fault types with an accuracy of 95%, while the LSTM model predicts RUL with a mean absolute error (MAE) of 5.6 hours, demonstrating the effectiveness of deep learning in enhancing predictive maintenance strategies for robotic systems. This approach shows promise for real-world applications in autonomous robotics by enabling early detection of faults and improving the reliability and safety of robotic operations.
Keywords:
Predictive Maintenance
Rotor Fault Diagnosis
Deep Learning
Autonomous Robotics
Remaining Useful Life
Convolutional Neural Networks
LSTM Networks
Vibration Analysis
Fault Classification
Fault Detection
1. Introduction:
Autonomous robotic systems are increasingly being deployed in complex, dynamic environments across industries, from manufacturing to healthcare. These systems often rely on mechanical components, such as rotors, which are essential for their movement and functionality. A failure in these components, whether due to misalignment, imbalance, or even complete rotor detachment, can compromise the entire system’s operation, leading to expensive downtime and potentially catastrophic consequences, underscoring the importance of real-time monitoring and predictive maintenance [1–4].
In the context of autonomous robots, it is critical to implement effective predictive maintenance strategies that can detect faults early and predict the remaining useful life (RUL) of critical components like rotors [5, 6]. Traditional fault diagnosis methods often rely on direct sensor measurements and heuristic approaches, which can be limited in their ability to handle the complexity and variability of real-world fault conditions [7, 8]. With advancements in artificial intelligence and machine learning—particularly deep learning—there has been growing interest in developing more robust and intelligent fault detection and RUL prediction models [9–13].
Among the most promising techniques, Convolutional Neural Networks (CNNs) have demonstrated strong capability in automatic feature extraction from vibration signals and time–frequency spectrograms, while Long Short-Term Memory (LSTM) networks have shown high accuracy in capturing temporal dependencies for RUL prediction [14–17]. These deep models outperform traditional methods such as Support Vector Machines (SVMs) and Decision Trees in terms of generalization and adaptability under varying fault severities [18].
In this paper, we propose a simulation-based framework that utilizes deep learning models, including CNN for fault classification and LSTM for RUL prediction. This framework addresses the challenges of rotor fault diagnosis by leveraging a simulated environment to generate realistic fault scenarios and train models capable of generalizing to real-world autonomous robotic systems [19–20].
The paper is organized as follows: Section 2 provides an overview of related work in predictive maintenance, fault diagnosis, and the application of deep learning techniques to rotor systems. Section 3 describes the simulation environment used to generate synthetic rotor fault data. Section 4 presents the methodology, including preprocessing steps, model architecture, and training procedure. Section 5 discusses the results of our experiments, comparing the performance of the deep learning models with baseline approaches. Section 6 offers a discussion of the implications of these findings and potential directions for future research. Finally, Section 7 concludes the paper and outlines the key takeaways.
2. Related Work:
The application of predictive maintenance (PdM) techniques in robotics has gained significant attention in recent years, particularly in the context of machinery health monitoring and fault detection. A number of studies have explored the use of vibration-based analysis to detect faults in rotating machinery, with approaches ranging from traditional signal processing techniques to more advanced machine learning methods.
A
One of the key challenges in predictive maintenance for autonomous robots is the variability and complexity of fault signatures, which can vary based on the system's operating conditions, the type of fault, and the presence of noise in the data. Early approaches to fault diagnosis relied on signal processing techniques such as Fast Fourier Transform (FFT), Short-Time Fourier Transform (STFT), and wavelet transforms to extract features from vibration signals. While these methods are useful in isolating fault-related frequency components, they often lack the capability to classify fault types and predict the remaining useful life effectively.With the rise of machine learning, researchers have started to explore more data-driven approaches. Support Vector Machines (SVMs), Random Forests, and k-Nearest Neighbors (k-NN) have been widely used for fault classification, while regression models have been applied to predict the remaining useful life (RUL) of machinery components. However, these methods often require careful feature engineering and may not capture the complex temporal relationships present in sequential sensor data.
Recent advancements in deep learning have led to the development of more powerful models for fault diagnosis and RUL prediction. Convolutional Neural Networks (CNNs) have shown great promise in extracting features from time-frequency representations of vibration data, such as spectrograms. Recurrent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) networks, are well-suited for handling sequential data and have been effectively used for predicting RUL. However, few studies have combined these techniques in the context of rotor fault diagnosis in autonomous robotic systems, and even fewer have leveraged simulation-based datasets to train and evaluate these models.
In this paper, we build upon these advancements by integrating both CNNs and LSTMs into a unified framework that addresses the unique challenges of rotor fault detection and RUL prediction in autonomous robots. The combination of these models, trained on synthetic data generated by a detailed rotor dynamics simulation, provides a robust solution that is both accurate and scalable for real-world applications.
3. Simulation Environment:
The simulation environment was developed using MATLAB and Simulink, which are widely used tools for modeling and simulation in engineering applications. The rotor fault scenarios were modeled based on the physical dynamics of rotating machinery, including the effects of misalignment, imbalance, and rotor drop-off. These scenarios were chosen because they represent common faults in rotor systems, which can cause significant damage if not detected early.
In the simulation, the rotor system was modeled as a dynamic system with a set of equations governing the motion of the rotor and the forces acting on it. Faults such as misalignment and imbalance were introduced by modifying the system’s mass distribution and geometric configuration. Rotor drop-off was simulated by reducing the effective mass of the rotor at certain points during the simulation, which caused sudden changes in vibration patterns.
For each fault scenario, the simulation generated time-series vibration data, which were then processed into time-frequency representations using Short-Time Fourier Transform (STFT). These spectrograms served as the input for the deep learning models. Additionally, noise was introduced into the simulation to replicate real-world conditions, where sensor readings are often subject to various types of interference. The dataset was generated to cover a wide range of fault severities, from minor misalignments to complete rotor detachment.
The simulated data provided a controlled environment for training the models, allowing us to explore how the models perform under different fault conditions and operating scenarios. A total of 100,000 data samples were generated, with 80% used for training, 10% for validation, and 10% for testing the models.
4. Methodology:
The methodology for fault diagnosis and RUL prediction involves several key steps: data preprocessing, feature extraction, model architecture design, and training.
Data Preprocessing:
The raw vibration data from the simulation were first subjected to noise filtering and normalization. The time-domain signals were converted into time-frequency spectrograms using the Short-Time Fourier Transform (STFT), which provided a more informative representation of the data for the deep learning models. The spectrograms were then resized to a fixed size to make them suitable for input into the CNN model.
Fault Classification using CNN:
The CNN model was designed to classify fault types based on the time-frequency representations of the vibration data. The architecture consisted of multiple convolutional layers followed by pooling layers to extract hierarchical features from the spectrograms. The output of the CNN was a set of classification probabilities for the three fault types: misalignment, imbalance, and rotor drop-off. The model was trained using a categorical cross-entropy loss function, and softmax activation was applied at the output layer.
RUL Prediction using LSTM:
For the RUL prediction task, we used an LSTM network, which is capable of learning long-term dependencies in sequential data. The LSTM model was trained to predict the remaining useful life of the rotor system based on the sequential vibration data. The input to the LSTM was a sequence of spectrograms, and the output was a continuous value representing the predicted RUL. The model was trained using mean squared error (MSE) as the loss function to minimize the prediction error.
Model Training:
Both models were trained using the Adam optimizer with a learning rate of 0.001. Training was performed over 100 epochs with a batch size of 32. The training set consisted of 80% of the generated data, with validation and testing sets covering 10% each. The models were evaluated on their classification accuracy and RUL prediction error, with the CNN model being assessed using accuracy metrics (precision, recall, F1-score) and the LSTM model using MAE and Root Mean Squared Error (RMSE).
5. Results and Discussion:
The proposed deep learning models—Convolutional Neural Networks (CNN) for fault classification and Long Short-Term Memory (LSTM) networks for Remaining Useful Life (RUL) prediction—were evaluated on the synthetic rotor fault data generated from the simulation environment. The results indicate that both models performed exceptionally well in their respective tasks, and a deeper analysis of the results highlights their potential impact on predictive maintenance in autonomous robotics.
Fault Classification Performance (CNN):
The CNN model demonstrated high performance in fault classification, achieving an overall classification accuracy of 95%. Figure 5 presents the confusion matrix for the fault classification task, showing that the model successfully classified misalignment, imbalance, and rotor drop-off with high accuracy. The confusion matrix highlights the strong distinction between misalignment and imbalance faults, which often present similar vibration signatures.
Fig. 1
Simulation Setup
Figure 1 illustrates the rotor fault simulation model created in MATLAB/Simulink. It shows how various fault scenarios—misalignment, imbalance, and rotor drop-off—are simulated within the rotor system. The diagram visualizes how the system is configured to simulate realistic faults, with adjustable parameters for each fault type. The rotor’s vibration data, generated by these faults, is collected through sensors strategically placed on the system. This data forms the foundation for training and testing the deep learning models used for fault classification and remaining useful life (RUL) prediction. The figure provides a clear picture of the simulation environment used to create synthetic data that mimics real-world fault conditions in robotic rotor systems.
Fig. 2
CNN Model Architecture
This Fig. 2 presents the architecture of the Convolutional Neural Network (CNN) model used for fault classification in the rotor system. It outlines the layers of the CNN and their function. The input layer receives vibration data in the form of time-frequency spectrograms. The convolutional layer applies convolution operations to the input data, extracting features that are characteristic of different fault types. The max pooling layer reduces the spatial dimensions of the feature maps generated by the convolutional layers, allowing the model to focus on the most important features while reducing computational complexity. After pooling, the extracted features are passed through fully connected layers, where neurons are connected to every neuron in the previous layer. This layer is responsible for making higher-level abstractions of the features. Finally, the output layer produces a classification result, indicating the type of fault (misalignment, imbalance, rotor drop-off). This CNN model is designed to classify rotor faults by learning patterns from vibration signals and their corresponding time-frequency spectrograms.
Fig. 3
Example Vibration Signals
figure 3 displays time-domain vibration signals for three different rotor fault conditions: misalignment, imbalance, and rotor drop-off. The top plot shows a clean, periodic vibration signal under normal conditions (misalignment), which can be characterized by smooth, periodic oscillations. The middle plot represents an imbalance fault, which introduces irregular, high-frequency noise into the signal due to uneven mass distribution. The bottom plot illustrates a rotor drop-off fault, where the vibration signal exhibits irregular, decaying patterns, signaling the detachment or severe damage to the rotor. These signals are crucial inputs for training machine learning models to detect faults and predict the system's remaining useful life (RUL). The figure visually distinguishes how each fault condition impacts the vibration signal, helping to guide the development of fault detection algorithms.
Fig. 4
Time-Frequency Spectrograms
This Fig. 4 presents time-frequency spectrograms generated from the vibration signals for each fault type using the Short-Time Fourier Transform (STFT). Spectrograms provide a time-based analysis of frequency content, which is essential for distinguishing different fault conditions. Each subplot displays the frequency spectrum of the corresponding time-domain signal shown in Fig. 3, but in the frequency domain, making it easier to identify the unique frequency patterns that occur during various faults. For the misalignment fault, the spectrogram shows relatively stable frequency components with minor fluctuations over time, indicating a periodic fault pattern. The imbalance fault results in a wider frequency spectrum, reflecting irregularities and high-frequency noise caused by uneven rotor mass. The rotor drop-off fault leads to sharp, transient spikes in frequency, signaling an abrupt fault like rotor detachment. By analyzing these spectrograms, the model can learn to distinguish the underlying faults and improve fault classification accuracy.
Fig. 5
CNN Fault Classification Performance
This figure shows the performance of the CNN model in classifying different fault types through a confusion matrix. The confusion matrix displays the number of correct and incorrect classifications for each fault type: misalignment, imbalance, and rotor drop-off. The matrix highlights that the model achieves high accuracy in distinguishing misalignment and imbalance faults, with few misclassifications. The rotor drop-off fault is also correctly identified, as its distinct vibration signature is easily recognizable. The model demonstrates strong performance across all fault types, with precision and recall values close to 100% for each class. This figure illustrates how the CNN model effectively learns to classify fault types based on time-frequency features, making it a powerful tool for predictive maintenance in autonomous robotic systems.
Fig. 6
LSTM RUL Prediction Performance
This Fig. 6 compares the predicted vs. actual Remaining Useful Life (RUL) for the rotor system using the Long Short-Term Memory (LSTM) model. The plot shows two curves: the blue curve represents the actual RUL of the rotor, which is based on the time to failure from the simulation, while the red curve represents the predicted RUL as estimated by the LSTM model. The closeness of the predicted RUL to the actual values demonstrates that the LSTM model performs well in estimating the time left before failure. The low Mean Absolute Error (MAE) of the model shows that it can reliably forecast the remaining life of the rotor, aiding in predictive maintenance by allowing maintenance actions to be scheduled before catastrophic failures occur. This figure highlights the capability of the LSTM model in tracking and predicting rotor health based on sequential vibration data.
A
Fig. 7
Comparison of Baseline Models
This figure compares the performance of the CNN, LSTM, Decision Tree, and SVM models for both fault classification and RUL prediction. Two metrics are compared: Fault Classification Accuracy and RUL Prediction Error. The CNN and LSTM models show significantly higher classification accuracy than Decision Trees and SVMs, with CNN achieving 95% accuracy. For RUL prediction, the LSTM model performs the best, with the lowest Mean Absolute Error (MAE), outperforming Decision Trees and SVM in this task. The bar chart visually emphasizes the superior performance of deep learning models (CNN and LSTM) compared to traditional machine learning algorithms (Decision Trees and SVM), underscoring the potential of deep learning for predictive maintenance in autonomous robotics.
7. Conclusion:
This paper presents a simulation-based framework for rotor fault diagnosis and remaining useful life prediction in robotic systems using deep learning models. The proposed approach, which combines Convolutional Neural Networks for fault classification and Long Short-Term Memory networks for RUL prediction, shows promising results in both tasks. The models were trained on synthetic data generated through a detailed rotor dynamics simulation, and the results demonstrate that deep learning can significantly enhance predictive maintenance capabilities in robotics. This work contributes to the advancement of autonomous robotic systems by improving fault detection and reducing downtime, and lays the groundwork for future studies that integrate real-world data and expand the scope of fault detection in robotic systems.
Acknowledgements
This work is sponsored by National Key R&D Program of China (2017YFA0303700); National Natural Science Foundation of China (11474228, 61605154, 11604256); Key Scientific and Technological Innovation Team of Shaanxi Province (2014KCT-10).
A
Author Contribution
Al Imran: Conceptualization, Writing—Original Draft, Visualization. Changbiao Li: Validation, Formal Analysis, Data Curation. Yanpeng Zhang: Investigation, Visualization, Writing—Review & Editing. All authors have approved the final version of the manuscript.
A
Funding
This work is funded by National Key R&D Program of China (2017YFA0303700); National Natural Science Foundation of China (11474228, 61605154, 11604256); Key Scientific and Technological Innovation Team of Shaanxi Province (2014KCT-10).
A
Data Availability
The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.
Declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
References:
1.
Lei Y, Li N, Guo L, et al. Machinery health prognostics: A systematic review from data acquisition to RUL prediction. Mech Syst Signal Process. 2018;104:799–834.
2.
Lee J, Davari Ardakani H, Yang S, Bagheri B. Industrial Big Data analytics and cyber-physical systems for future maintenance & service innovation. Procedia CIRP. 2015;38:3–7.
3.
Zhang L, Yang P, Sun X, Zhang H. Deep learning-based fault diagnosis in industrial systems: A review. IEEE Access. 2021;9:42370–90.
4.
Wang T, Qian K, Xu L. Deep learning for motor fault detection and diagnosis: A review. IEEE Trans Instrum Meas, 70, 2021.
5.
Ma J, Zhang X, Chen J. A hybrid deep learning model for fault diagnosis and RUL prediction of rotating machinery. Reliab Eng Syst Saf. 2021;209:107453.
6.
Li X, Zhang W, Ding Q. Deep learning-based remaining useful life estimation of bearings using multi-scale feature fusion. IEEE Trans Industr Electron. 2019;66(12):9639–48.
7.
Ahmad S, Qin SJ, Chen T. Process monitoring and fault diagnosis using statistical methods: A review. IEEE Trans Control Syst Technol. 2018;26(3):706–21.
8.
Yang G, Zhao J, Liu Y. A comparative study of machine learning approaches for bearing fault diagnosis. Measurement. 2021;182:109679.
9.
Xia F, Zhang L, Yu J. Recent advances in deep learning for industrial fault diagnosis: A survey. Neurocomputing. 2022;511:274–300.
10.
Malhi A, Gao R. PCA-based feature selection scheme for machine defect classification. IEEE Trans Instrum Meas. 2004;53(6):1517–25.
11.
Zhao M, Kang X, Shao X. Machine fault diagnosis using deep learning with time–frequency representation. Meas Sci Technol, 31, 2, 2020.
12.
Guo S, Yang T, Gao W, Zhang C. A novel fault diagnosis method for rotating machinery based on deep convolutional neural networks. Measurement. 2019;140:213–24.
13.
Chen Y, Chen Z, Li W. An interpretable deep learning framework for remaining useful life prediction of bearings. Reliab Eng Syst Saf. 2022;218:108123.
14.
Qin Y, Ding C, Chen J. Hybrid CNN-LSTM model for fault diagnosis of rotating machinery. IEEE Trans Industr Inf. 2022;18(7):4838–49.
15.
Li J, Lei Y, Lin J. A new data-driven method for bearing remaining useful life prediction based on deep learning. Mech Syst Signal Process. 2019;134:106400.
16.
Zhao Z, Yang R, Wang T. Attention-based convolutional recurrent networks for remaining useful life prediction of industrial equipment. Appl Soft Comput. 2021;113:107981.
17.
Zhang K, Wang X, Liu Y. Fault detection of mechanical systems using CNN and data augmentation. Sensors. 2020;20(13):3739.
18.
Singh P, Meena DS, Sharma AK. A comparative evaluation of machine learning algorithms for predictive maintenance in industrial applications. Measurement. 2022;194:111024.
19.
Gao R, Zhao J, Wang L. Simulation-based digital twin for fault diagnosis and prognosis of industrial robots. J Manuf Syst. 2021;58:49–57.
20.
Kang M, Kim S, Kim H. Deep transfer learning for remaining useful life prediction under variable operating conditions. Reliab Eng Syst Saf. 2023;229:108944.