The use of artificial intelligence (AI) techniques, specifically machine learning (ML), in predictive maintenance is revolutionizing asset management by analyzing the current operational conditions to anticipate equipment failures and optimize maintenance schedules, leading to enhanced operational efficiency.
Specifically, AI tools can identify small efficiency reductions that indicate maintenance needs by evaluating the current equipment performance against the baseline data. This article discusses the importance and applications of AI in predictive maintenance.
Image credit: MONOPOLY919/Shutterstock
Importance of AI in Predictive Maintenance
Predictive maintenance primarily involves using large volumes of data to address and anticipate potential issues in assets before they lead to breakdowns in processes, operations, systems, or services. AI can be used in predictive maintenance to predict equipment problems/failures to increase operational reliability accurately.
The technology can combine historical performance data, real-time analytics, and engineering specs to create condition-based, user-specific alerts and alarms to allow users to quickly resolve a potential issue to prevent or significantly reduce unplanned downtime and avoid expensive repairs/servicing, leading to greater productivity.
Moreover, the stress that leads to performance problems during the design process can be simulated in a better manner using data obtained using AI-enabled processes to ensure that machines can effectively endure real-world conditions and determine the predictive alert and alarm points.
Predictive maintenance models using AI can assess several variables that represent the current status of an asset and make usage trend-based predictions to eliminate production losses. Predictive maintenance can be planned based on workers’ schedules when AI is used to predict equipment failures/issues and improve productivity.
Increased asset utilization, fewer productivity lags, and maximized uptime can be achieved when workers do not spend a significant amount of time addressing an unexpected malfunction. Moreover, the safety of workers and service technicians can be improved by not sending them to hazardous situations, as AI can precisely predict the malfunction/breakdown of equipment. These predictions can ensure that workers remain safe from equipment/machines that are predicted to experience malfunctions and allow service technicians to resolve the identified issues safely.
AI Techniques in Predictive Maintenance
Several AI methods can be employed for predictive maintenance, including artificial neural network (ANN), support vector machine (SVM), decision tree (DT), random forest (RF), logistic regression (LR), extreme gradient boosted trees (XGBoost), gradient boosting machines (GBM), linear regression, and symbolic regression (SR).
ANN: ANN models are used extensively for predictive maintenance because they can learn from examples and effectively analyze fuzzy, nonlinear, and random data. In a study, a trained ANN model was used to classify the tool state of a computer numerical control (CNC) milling machine using acceleration data.
A programmable prototyping platform with built-in sensors was utilized to monitor the tool wear and the ANN model performance was compared with other models. In the results, the ANN model demonstrated a better performance compared to the performance of k-nearest neighbors (KNN) and SVM models.
In another study, a methodology was proposed to process and convert the vibration measurement data collected from a vibration system simulating a motor, and a dataset was built to train and test an ANN model that can predict future equipment conditions, specifically potential failures.
Multilayer perceptron (MLP) methodology was utilized to perform the prediction task owing to its easier implementation with a good generalization index, and the proposed ANN model was then compared with other ML techniques, including RF, regression tree (RT), and SVM-based on efficiency and root mean square error (RMSE) performance index value. The results demonstrated that ANN can outperform other models in long-term and medium-term prediction. However, the performance of RT, RF, and ANN was similar in short-term predictions.
A bench top test-rig was designed and developed in a study to investigate the time domain vibration signatures of multiple key components in wind turbines. Faulty and healthy condition vibration signatures of the critical components were acquired, and then an ANN model was developed and applied to classify the healthy and faulty state features. The developed ANN model displayed 92.6% classification efficiency.
In another study, neural networks (NN) and physics-based models were developed to monitor the starter degradation of the auxiliary power unit (APU). A generic modeling technique was used to address the limitations of the lack of component characteristics, and a comparative analysis between forward-feed and back-spreading NN models was performed.
Both models were applied under deteriorated and nominal conditions and their capabilities were validated. The results showed that more consistent outcomes can be realized for cases with degraded starters using the physics-based approach, while the NN model displayed better performance with starters in healthy conditions.
SVM: SVM can perform pattern recognition, classification, and regression analysis, and is commonly used in predictive maintenance of industrial equipment to identify a specific status based on the acquired signal.
In a study, a data-driven prognostics and diagnostics framework was proposed for machines to reduce maintenance costs and increase efficiency, and real data from vending machines was used to validate the proposed framework for three classifiers including SVM, RF, and GBM.
Additionally, two models were developed for predictive maintenance, with one model for two-stage prognostics and the other for diagnostics. Results showed that the developed SVM model can be effectively used for diagnosis and prognostics monitoring of complex vending machines. Specifically, the prognostics model outperformed the one-stage conventional prediction models.
SVM and ANN algorithms were used in a study to develop gauge degradation measurement prediction for two rail track types, including curved and straight segments. The coefficient of determination and mean squared error were used for the performance evaluation of the models.
Although both the SVM and ANN models provided slightly similar and satisfactory outcomes, the ANN model's performance in predicting the gauge deviation of straight segments was better than the SVM models, while the SVM model was better in predicting the gauge deviation of curved segments.
RF: RF is an ensemble learning algorithm consisting of several DT classifiers that can manage high-dimensional data without choosing a feature and can be implemented in a simple manner. In a study, the downtime of a printing machine was forecasted based on real-time predictions of imminent failures. Unstructured historical machine data was utilized to train the ML classification algorithms, including LR, XGBoost, and RF, to predict machine failures.
Several metrics, such as the area under the receiver operating characteristic curve (AUC) and precision-recall curve (PRC), were analyzed to evaluate the fitness of the ML models. All ML algorithms demonstrated a similar performance based on the receiver operating characteristic curve (ROC), while XGBoost and RF performed better than LR based on decision thresholds.
In another study, ML algorithms, including SR, linear regression, and RF, were used to model the condition of healthy industrial machinery. A methodology was proposed to detect and predict drifting behavior/concept drifts in continuous data streams.
Additionally, a real-world case study on industrial radial fans was presented. The synthetic data results showed that both concept drift prediction and detection were highly successful. A predictive model using the term frequency-inverse document frequency (TF-IDF) and RF was proposed in a study that can forecast high-sensitivity faults in advance by analyzing the historical aircraft maintenance system data.
TF-IDF was employed to extract features from the raw data in the past consecutive flights, and various priorities were considered in classifying the faults using the RF model. The ROC curve was adopted as a performance metric due to the use of a highly imbalanced dataset. The proposed RF model attained a false positive rate of 0.13% and a true positive rate of 66.67% for the testing dataset.
Other Techniques: Convolutional neural networks (CNN) can be used for predictive maintenance of photovoltaic panels by monitoring the operation of panels. CNN coupled with deep neural network (DNN) can also be used for infrared thermal image-based machine health monitoring to detect different conditions of rotating machinery.
Reference and Further Reading
Çınar, Z. M., Abdussalam Nuhu, A., Zeeshan, Q., Korhan, O., Asmael, M., Safaei, B. (2019). Machine Learning in Predictive Maintenance towards Sustainable Smart Manufacturing in Industry 4.0. Sustainability, 12(19), 8211. https://doi.org/10.3390/su12198211
The Role of AI in Predictive Maintenance [Online] (Accessed on 04 November 2023)
Himes, E. (2023). What Is AI in Predictive Maintenance? [Online] (Accessed on 04 November 2023)
Predictive Maintenance [Online] (Accessed on 04 November 2023)