Predictive maintenance is increasingly becoming crucial to ensure the functional safety and reliability of vehicles over their lifetime and reduce the ever-rising vehicle maintenance costs. Artificial intelligence (AI), specifically machine learning (ML), can be used effectively for the predictive maintenance of vehicles as most modern vehicles can provide substantial amounts of operating data. This article discusses a framework for automobile predictive maintenance using deep learning (DL) under Industry 4.0, the applications of AI techniques for predictive maintenance of vehicles, and the major challenges of using AI.
Image credit: Thx4Stock team/Shutterstock
The DL-based Framework
A framework has been recently designed using DL for automobile predictive maintenance that utilizes multisource data that is relevant to the automobile lifecycle. The framework has several stages, including the data collection stage, data mapping pre-processing and integration stage, cloud data transmission and storage stage, DL for automobile time-between-failure (TBF) modeling stage, and decision support for the predictive maintenance stage
Initially, the multi-source data is collected and uploaded to the cloud, and then the data is mapped, pre-processed, and integrated before it is utilized for modeling. DL, coupled with semi-supervised learning and reliability analysis, was used in the modeling stage to establish a TBF prediction model. Eventually, the predicted TBF of an automobile is utilized to optimize maintenance planning, job scheduling, and spare parts management.
AI for Vehicle Predictive Maintenance
Several ML techniques are used for the predictive maintenance of vehicles, including linear regression (LR), artificial neural network (ANN), Gaussian process regression (GPR), k-nearest neighbors (KNN), decision tree (DT), support vector machine (SVM), random forest (RF), multiple linear regression (MLR), GP, long short-term memory (LSTM), autoencoder (AE), merged-LSTM, convolutional autoencoder (CAE)+LSTM, error fusion of multiple SAEs (EFMSAE)-LSTM, convolutional neural network (CNN), and ensemble method.
LR and ANN, GPR, KNN, and MLR can be used for predictive maintenance of diesel engines fueled with biodiesel alcohol mixtures, remaining useful life (RUL) prediction for slow speed bearings, classification of vibration gravity to predict anomalies in electric inductive motors, and fatigue life evaluation of automotive coil springs, respectively.
DT, SVM, RF, and KNN can be utilized for monitoring and fault-predicting systems in vehicles, RF, SVM, ANN, and GP can be employed for fault diagnosis in turbocharged petrol engine systems, and ANN, SVM, LR, and GPR can be used for state of charge (SoC) estimation of lithium-ion battery for electric vehicles (EVs).
Specifically, ANN and GPR can effectively assist in designing the optimum battery management system for EVs based on SoC predictions. SVM can be utilized for fault diagnosis of vehicle suspensions and fault detection, identification, and prediction for autonomous vehicle controllers, while ANN + KNN can be used for vehicle health monitoring.
Additionally, LSTM+GPR, AE, Merged-LSTM, and LSTM/RF can be employed for RUL prediction for lithium-ion (Li-ion) batteries with reliable uncertainty management, prediction of upcoming failures in trucks, TBF prediction modeling based on multisource data, and heavy medium lead-acid battery prognosis, respectively.
Moreover, CNN can be employed for multi-sensor fault detection for autonomous vehicles, while LSTM can be used for the prediction of the remaining fatigue life of automotive suspension. CAE+LSTM and EFMSAE-LSTM can predict RUL for EVs and mechanical fault time series, respectively.
Major Applications of AI
Detection of Faults in Engines: In a study, a method based on an ensemble of Bayesian extreme learning machines (ELM) in a supervised classification setup was developed for the detection of engine faults focusing on simultaneous faults/ multiple single faults occurring concurrently.
Although the base classifiers were trained only on various single faults, the results from experiments performed on data from a vehicle displayed that the ensemble can detect both single and simultaneous faults. In another study, a novel variant of an ANN, designated as extension neural network type-1, was used for engine fault detection from vibration signals. The novel ANN variant was computationally more efficient compared to a standard ANN and could measure the distances between cluster centers and new data points of fault types.
A hybrid method based on ANNs, discrete wavelet transform, and vibration signature analysis using fast Fourier transform has been developed for effective combustion fault detection of a 12-cylinder 588 kW diesel engine. A DL with four CNN layers, two LSTM layers, and one softmax layer was used in a study to detect pre-ignition faults in turbocharged petrol engines using data obtained from electric control units (ECU) of a vehicle fleet. The approach identified faults with a 0.9 F1 score and was better compared to stand-alone CNNs, LSTMs, linear SVMs, logistic regression, and RFs.
In another study, one-class SVMs enhanced by Weibull calibration were employed to individually model fault classes. The study objective was to detect both unknown and known fault classes, and data from a combustion engine was used to evaluate the approach. Results demonstrated that the method can detect seven known fault types and previously unseen fault types.
An ANN trained using discriminative features extracted using wavelets has been used to classify faults in engines based on sound measurements. The approach can classify data successfully from an engine on a test bed into fault-free and any of 8 faults.
Detection of Faults and State-of-Health (SoH) of EV Batteries: An ML-based approach was proposed in a study for SoH estimation/capacity degradation for batteries of EVs. The objective was to identify two health indicators from the internal resistance of the battery using correlation analysis.
An ELM was trained and evaluated using a dataset generated using batteries with various levels of capacity loss. Comparative analysis of the ELM with a standard ANN displayed that ELM yields better accuracy, faster to train, and easier to use.
In another study, an approach was proposed for detecting external short circuits in batteries, a safety-relevant fault in EVs. In the initial step, two physical models were compared and their parameters were estimated using a genetic algorithm. Then, researchers utilized their domain knowledge to identify features that enable the identification of fault, and an RF classifier was eventually trained in a supervised manner. The approach was validated using real batteries in a lab environment. A hybrid model was developed in a study for fault detection of a lithium iron phosphate (LFP) power cell type used in EVs.
A k-means clustering algorithm was utilized to identify data groups with the same behavior, and then, various regression techniques, including support vector regression (SVR), ANN, and polynomial regression, were tested for each group. Results showed that the model could successfully classify all fault situations in the LFP power cell type. However, the model failed to distinguish between erroneous measurements.
Detection of Faults in EV Powertrain: In a study, researchers investigated data-driven fault diagnosis and detection for regenerative braking systems of hybrid electric vehicles (HEVs) by modeling the overall powertrain of a two-motor series and conducted fault injection experiments to simulate the most common system faults.
During simulations, 25 system state variables were monitored, including battery SoC, engine torque, and wheel speed. Multi-way partial least squares and multi-way principal component analysis were used to reduce the state space to minimize the computational costs.
Several algorithms, including KNN, partial least squares, a probabilistic ANN, and SVMs, were implemented based on the reduced data to classify the faults into 12 classes. Results displayed that SVM and KNN can classify faults, such as Motor1 current sensor fault, wheel inertia fault, and battery SOC fault, with up to 100% accuracy.
Full Vehicle Fault Detection: A data-driven fault detection framework using principal component analysis and unsupervised independent component analysis (ICA) was proposed in a study for data clustering and reduction based on feature distance and mutual information. The approach can be feasibly used for anomaly detection in automotive MAF sensors.
In conclusion, the use of AI techniques in the predictive maintenance of vehicles can effectively prevent costly breakdowns and optimize service schedules, leading to extended vehicle lifespans. However, the major challenges of using AI for vehicle predictive maintenance, including non- or insufficient availability of public real-world datasets, lack of labeled data, complexity of problem setting, and acceptance of ML-based maintenance, must be addressed to exploit the full capability of AI technologies.
References and Further Reading
Chen, C. (2020). Deep learning for automobile predictive maintenance under Industry 4.0 https://orca.cardiff.ac.uk/id/eprint/137968/
Arena, F., Collotta, M., Luca, L., Ruggieri, M., Termine, F. G. (2022). Predictive Maintenance in the Automotive Sector: A Literature Review. Mathematical and Computational Applications, 27(1), 2. https://doi.org/10.3390/mca27010002
Guo, J., Lao, Z., Hou, M., Li, C., Zhang, S. (2021). Mechanical fault time series prediction by using EFMSAE-LSTM neural network. Measurement, 173, 108566. https://doi.org/10.1016/j.measurement.2020.108566
Theissler, A., Pérez-Velázquez, J., Kettelgerdes, M., Elger, G. (2021). Predictive maintenance enabled by machine learning: Use cases and challenges in the automotive industry. Reliability Engineering & System Safety, 215, 107864. https://doi.org/10.1016/j.ress.2021.107864