A stacking ensemble classification model for determining the state of nitrogen-filled car tires

1 Introduction

The tire is a flexible rubber component encircling the rim of a wheel that plays a crucial role in effectively maintaining the comfort and safety of the occupants in an automobile. The tire is considered one of the crucial parts in an automobile that influences a number of essential tasks like traction, road shock absorption, vehicle handling characteristics, proper road contact, and safety. The traction, stability and grip provided by the tires can assist the vehicles to be operated at a range of speeds on various types of road surfaces and in a variety of weather conditions. The weight of the car and its occupants is supported entirely by the tires, which also provide shock and vibration absorption for a smoother and more comfortable ride. Tires are manufactured in several variants to adapt to different weather conditions, terrains, and performance requirements. Additionally, tires are essential for braking, acceleration, cornering, and handling, all of which have an impact on the vehicle's overall performance and safety [1]. To ensure the proper functioning of tires, maintaining the specific tire pressure and tread depth is essential for conserving the stability of the vehicle and improving maneuverability, and braking performance [2]. Also, underinflation or overinflation of tires can significantly impact the longevity of tires and the performance of the vehicles. Climatic changes, thermal stresses, improper driving, varying road conditions, and continual operation of vehicles can influence the intrusion of faults in tires. Such fault scenarios create a concern for safety and degrade the comfort of the occupants. In recent times, vehicle manufacturers have designed and equipped a TPMS system that can monitor the pressure in tires through a wireless medium [3]. With the invention of such technology, various phenomena like tire life, vehicle dynamics, fuel efficiency, and performance of the vehicle have been enhanced [4]. TPMS is considered an utmost safety device; however, the gas utilized to inflate the tires is also considered essential to improve the vehicle characteristics. Of late, nitrogen gas is generally preferred over air for inflating tires due to the superior performance traits showcased [5]. Tires filled with nitrogen gas consist of pure nitrogen in comparison to air-filled tires, which contain a mix of oxygen, nitrogen, and other trace elements. Improved tire pressure retention, lesser tire gas leakage (larger nitrogen molecules), lower thermal expansion, and minimal impact on temperature variations are certain takeaways of using nitrogen gas in tires [6]. Despite the fact that nitrogen gas inflation is expensive and widely adopted for various applications, both nitrogen and air-filled tires require constant monitoring to enumerate elevated levels of safety [7].

Several research works have been carried out in recent years focussing on TPMS that are discussed as follows. TPMS can be categorized and carried out in two ways: indirect and direct TPMS. A recent study encompassed the advantages of both indirect and direct TPMS by formulating a hybrid TPMS system that measures the pressure of one wheel to indirectly estimate the other tire pressures. The phenomenon used for measurement was termed a three-index system [8]. Sachan and Iqbal developed a non-electronic stability program for vehicles to estimate the tire pressure of vehicles [9]. ResNet-50, a pre-trained model, was adopted by Muturatnam et al. to monitor the condition of nitrogen-filled tires [10]. In another study, the vibration-based assessment was performed by Robinson to study the condition of tires due to vibration generated from hammer impacts [11]. Tire stiffness and tire pressure were monitored by an adaptive Kalman filter through the measurement of unknown road roughness using a weight-based approach [12]. Fechtner et al. adopted piezo-resistive pressure sensors to develop an intelligent system for monitoring tire pressure through the mass of the vehicle [13]. Air leakage in commercial vehicle tires was determined by Huo et al. with wheel speed sensors that used radius and frequency methods [14]. A hybrid nanogenerator with a rotary cylinder was designed by He et al. to develop a self-powered TPMS [15]. Jin et al. performed a frequency analysis on tire vibrations that incorporates road roughness compensation, vibration disturbance elimination, and resonant frequency estimation for reliable and accurate monitoring of tire pressure [16].

The evolution of artificial intelligence (AI) has defined a significant growth in the adaptation of machine learning in numerous applications. Some of the renowned applications include robotics, fault diagnosis, speech recognition, object detection, surveillance, text classification, and medical imaging. Accurate results, swift operation, and minimal time have been the significant advantages of using machine learning in fault diagnosis studies. The algorithms in machine learning display robust capabilities that can be applied for both regression and classification tasks. Table 1 gives an overview of applying machine learning algorithms in mechanical fault-finding systems. It shows how different algorithms enhance the efficiency and accuracy of monitoring and diagnosing mechanical conditions.

The literature survey mentioned above provides some valuable research gaps that are described as follows:

• Numerous literature were oriented toward the analysis of direct TPMS systems while very minimal works have been performed using indirect TPMS. Additionally, direct TPMS is claimed to be a non-cost-effective solution.

• Studies related to the usage of nitrogen-filled tires were very minimal, with few tire conditions.

• Vibration-based assessments in TPMS have limitations, especially concerning the underexplored ARMA, statistical, and histogram features.

• The classification of diverse tire conditions using machine learning represents a field that is in its initial stages of development.

• Single (or individual) classifier performance was assessed widely. However, a stacking-based ensemble approach was not attempted.

Sensor errors in TPMS pose a significant problem due to false alarms or failure to identify real tire pressure. These errors are brought on by variables like temperature fluctuations and electromagnetic interference. Additionally, the TPMS sensor battery life may be constrained, necessitating routine maintenance and replacement, which may be inconvenient for car owners. Additionally, non-standard tire sizes or odd tire shapes may be difficult for TPMS to handle, impairing their capacity to precisely monitor tire pressure. Some TPMS signals may be open to hacking or malicious attacks, which could jeopardize vehicle security. Further, the complexity of TPMS systems may result in higher repair costs and present numerous technical difficulties for both car owners and mechanics. Addressing these issues is very crucial to ensure the reliable and effective operation of TPMS for enhanced vehicle safety and performance [24]. Hence, there is a noticeable research gap in the development of a thorough TPMS system adept at accurately determining tire pressure using indirect TPMS and machine learning techniques.

Considering the aforementioned concerns portrayed, the present study introduces a stacking ensemble classification model that combines the predictive abilities of various classifiers that can considerably improve the performance of monitoring tire conditions [25]. This method, which has been shown to be effective in a number of machine learning applications, can increase the accuracy and robustness of the identification of tire pressure anomalies. The ensemble can capture complex correlations within tire pressure data, thereby improving the system’s capability to precisely identify abnormal conditions. The technique can capture complex relationships within data by fusing the strengths of several base classifiers, improving accuracy and robustness. Owing to the aforementioned benefit, stacking is especially helpful for jobs where individual classifiers may struggle, such as when working with noisy data or severely unbalanced datasets. Additionally, stacking ensembles offer an adaptable framework for model combinations that enables the integration of different types of base classifiers. A more thorough study of the feature space is frequently the result of this diversity, which may improve generalization and predictive abilities. By updating or replacing base models, stacking models may adapt to dynamic and changing situations, improving performance over time. A stacking ensemble model gives the TPMS the ability to make decisions in real time while continuously learning, which lowers safety risks and improves overall vehicle performance. The present study consists of various technical contributions that are listed as follows.

• Vibration signals from nitrogen-filled tires were monitored using a cost-effective MEMS accelerometer sensor, focusing on four conditions: high pressure, normal pressure, idle, and puncture. Three types of features – statistical, histogram, and autoregressive moving average (ARMA) – were extracted, and the most significant features were selected using a J48 decision tree algorithm.

• Training and test datasets were created in an 80–20% ratio to evaluate the performance of various machine learning algorithms, including sequential minimal optimization (SMO), instance-based k-nearest neighbor (IBK), logistic model tree (LMT), random forest (RF), multilayer perceptron (MLP), J48, and naive Bayes (NB).

• A stacking-based machine learning approach was applied, combining the top five performing classifiers in various ensembles of two, three, four, and five classifiers. Predictions from the individual classifiers were aggregated in these ensembles and classified using a meta-classifier, aiming to leverage the strengths of each classifier for improved overall prediction accuracy.

• The study demonstrated that integrating multiple machine-learning techniques and tuning hyperparameters could significantly enhance the accuracy of tire pressure monitoring systems. The findings highlighted the potential of stacking-based ensembles in providing robust and reliable predictions, paving the way for advanced applications in automotive safety and maintenance. Ongoing research is crucial to address challenges related to data availability and model generalization, ultimately aiming to improve vehicle safety, performance, and operational efficiency.

2 Experimental setup

The present study for tire condition monitoring was carried out on a front wheel drive vehicle with nitrogen-filled tires. The reason for selecting a front wheel drive vehicle was aimed at the reduction of vibration interference from the engine and various moving parts in the car. The vibration signals from the vehicle were collected using a tri-axial waterproof MMA7361L MEMS accelerometer. The sensor was mounted on the rear axle closer to the left wheel, as represented in ref. [7]. The vibration signals were collected from the sensor, which had a frequency range of about 400 Hz and a sensitivity of 206 mV/g. The overall experiment was carried out in a real-world environment while the car operated on a smooth road on Indian highways. The car was equipped with a data acquisition (DAQ) unit that was connected to the accelerometer through shielded cords (to reduce interference of signals) and a laptop.

3 Experimental procedure

The initial phase of experimentation was started with a perfectly balanced tire consisting of an existing weight of 40 g. No additional weight was added to the tire under experimentation. The vibration data collection was performed for four different tire conditions at varying vehicle speeds between 10 and 90 kmph (excluding idle conditions). The various conditions considered along with the tire pressures are detailed as follows: idle condition (31 psi at speeds below 10 kmph), high (40 psi), puncture (19 psi), and normal (31 psi). Furthermore, the details of vibration data collected, along with feature extraction and feature selection methods, are detailed below:

• Number of signals collected for each condition – 60

• Total signals collected – 240

• Features extracted – statistical, histogram, and ARMA

• Feature selection – J48 decision tree

The features selected using the J48 algorithm were passed as input to various classifiers to determine the performance, which was further utilized to develop combinations of classifiers for the stacking approach.

4 Methodology

The complete workflow or overview of the proposed methodology is depicted in Figure 1. Initially, a low-cost MEMS sensor was used to collect vibration signals from nitrogen-filled tires on the car’s rear axle across four conditions: high pressure, normal pressure, punctured, and idle. Further, ARMA, histogram, and statistical features were extracted from the collected vibration signals. Feature selection was performed with the aid of the J48 decision tree algorithm to select the most significant features that can help improve the learning capability and optimize the computational time of the machine learning model. The training and test datasets were created with a ratio of 80:20 using the selected features. Several base classifiers like SMO, RF, J48, IBK, MLP, NB, and LMT were trained initially. The best-performing classifiers (top five) were combined to form two, three, four, and five classifier combinations for the application of stacking. Meta classifier was chosen to be the top performing individual classifier for the specific feature set. The obtained results were recorded and analyzed to verify the effectiveness of the stacking strategy.

Figure 1

A visual representation illustrating the sequence of steps in the present investigation. Source: Created by the authors.

5 Results and discussion

This section aims to evaluate and compare the performance of individual base classifiers and the stacking strategy applied to monitor nitrogen-filled tire conditions. The study utilized an 80% training and 20% testing data division. Performance was assessed using metrics like training, validation, and test accuracies. Validation accuracy was determined using a ten-fold cross-validation approach, where the training data were split into ten equal parts, nine for training and one for testing in each cycle. Detailed results from this evaluation process are presented in the next section.

5.2 Performance evaluation of stacking classifiers

The stacking algorithm was adopted in the current study for the combination of different classifiers. The ensemble of classifiers was broadly categorized into two, three, four, and five classifiers. The meta-classifiers adopted in the study are KNN (for statistical and ARMA) and RF (for histogram). The possible combinations of the classifiers adopted in the study are detailed below.

• Two classifier ensemble – 10 combinations (KNN + RF, KNN + J48, KNN + MLP, KNN + NB, RF + J48, RF + MLP, RF + NB, J48 + MLP, J48 + NB, MLP + NB)

• Three classifier ensemble – 10 combinations (KNN + RF + J48, KNN + RF + MLP, KNN + RF + NB, KNN + J48 + MLP, KNN + J48 + NB, KNN + MLP + NB, RF + J48 + MLP, RF + J48 + NB, RF + MLP + NB, J48 + MLP + NB)

• Four classifier ensemble – 5 combinations (KNN + RF + J48 + MLP, KNN + J48 + MLP + NB, KNN + RF + MLP + NB, KNN + RF + J48 + NB, RF + J48 + MLP + NB)

• Five classifier ensemble – 1 combination (KNN + RF + J48 + MLP + NB)

The combinations mentioned above were used in the stacking approach, and the best-performing classifier combinations were identified. The performance of every individual stacking ensemble classifier is presented in Tables 3–6.

Table 3

Accuracy of the two classifiers stacking for the statistical, histogram, and ARMA features

Feature type	Meta classifier	Two classifiers	Training accuracy (%)	Validation accuracy (%)	Testing accuracy (%)
Statistical	KNN	RF + J48	97.92	90.10	89.58
		KNN + J48	96.35	88.02	85.42
		J48 + NB	91.15	88.54	83.33
		RF + NB	95.31	92.71	81.25
		KNN + RF	93.23	90.10	79.17
		RF + MLP	95.83	89.06	79.17
		MLP + NB	90.10	86.46	79.17
		KNN + NB	92.19	80.73	75.00
		J48 + MLP	94.79	85.42	72.92
		KNN + MLP	92.19	81.25	70.83
Histogram	RF	RF + MLP	97.92	92.71	85.42
		KNN + J48	96.35	93.75	85.42
		RF + KNN	94.27	92.71	83.33
		RF + J48	95.31	91.67	81.25
		KNN + MLP	90.10	89.58	81.25
		J48 + MLP	97.40	93.23	81.25
		RF + LMT	95.83	91.15	79.17
		J48 + LMT	95.83	90.63	79.17
		KNN + LMT	95.31	92.19	77.08
		MLP + LMT	92.19	92.19	75.00
ARMA	KNN	KNN + J48	96.35	93.75	93.75
		KNN + LR	94.79	89.06	89.06
		KNN + MLP	90.10	89.58	81.25
		J48 + MLP	97.40	93.23	81.25
		J48 + LR	95.31	90.63	79.17
		J48 + LMT	95.83	90.63	79.17
		KNN + LMT	95.31	92.19	77.08
		MLP + LR	95.83	93.23	77.08
		LR + LMT	95.31	93.75	77.08
		MLP + LMT	92.19	92.19	75.00

Table 4

Accuracy of the three classifiers stacking for the statistical, histogram, and ARMA features

Feature type	Meta classifier	Three classifiers	Training accuracy (%)	Validation accuracy (%)	Testing accuracy (%)
Statistical	KNN	RF + J48 + NB	94.79	89.58	89.58
		KNN + J48 + NB	95.31	86.46	87.50
		KNN + RF + J48	97.40	89.06	85.42
		KNN + MLP + NB	94.79	85.94	83.33
		RF + MLP + NB	93.75	92.19	83.33
		KNN + RF + NB	95.83	91.67	79.17
		KNN + J48 + MLP	95.31	86.98	79.17
		RF + J48 + MLP	96.35	88.54	79.17
		KNN + RF + MLP	94.79	89.06	77.08
		J48 + MLP + NB	94.27	87.50	77.08
Histogram	RF	RF + KNN + MLP	97.40	92.19	83.33
		RF + J48 + MLP	97.92	92.71	83.33
		RF + MLP + LMT	94.79	92.71	83.33
		J48 + MLP + LMT	95.31	94.27	83.33
		KNN + J48 + MLP	98.44	92.71	81.25
		KNN + J48 + LMT	96.88	90.63	81.25
		RF + KNN + J48	95.83	92.71	79.17
		RF + J48 + LMT	96.88	89.58	79.17
		RF + KNN + LMT	95.83	93.23	77.08
		KNN + MLP + LMT	91.67	92.19	70.83
ARMA	KNN	J48 + MLP + LMT	95.31	94.27	83.33
		KNN + J48 + MLP	98.44	92.71	81.25
		KNN + J48 + LR	95.83	91.15	81.25
		KNN + J48 + LMT	96.88	90.63	81.25
		J48 + MLP + LR	97.92	92.19	79.17
		KNN + LR + LMT	97.40	91.67	77.08
		J48 + LR + LMT	95.31	92.71	77.08
		MLP + LR + LMT	92.71	93.23	77.08
		KNN + MLP + LR	96.88	91.15	72.92
		KNN + MLP + LMT	91.67	92.19	70.83

Table 5

Accuracy of the four classifier stacking for the statistical, histogram, and ARMA features

Feature type	Meta classifier	Four classifiers	Training accuracy (%)	Validation accuracy (%)	Testing accuracy (%)
Statistical	KNN	KNN + RF + J48 + NB	95.31	89.58	89.58
		KNN + J48 + MLP + NB	95.83	89.06	81.25
		KNN + RF + MLP + NB	94.27	91.15	81.25
		RF + J48 + MLP + NB	95.31	89.06	81.25
		KNN + RF + J48 + MLP	95.31	88.54	72.92
Histogram	RF	RF + KNN + J48 + MLP	97.92	92.71	83.33
		RF + J48 + MLP + LMT	95.83	91.67	83.33
		RF + KNN + MLP + LMT	94.79	92.71	79.17
		RF + KNN + J48 + LMT	96.88	91.67	77.08
		KNN + J48 + MLP + LMT	95.31	93.75	75.00
ARMA	KNN	KNN + J48 + MLP + LR	97.40	91.67	81.25
		J48 + MLP + LR + LMT	95.83	93.23	79.17
		KNN + J48 + LR + LMT	96.35	93.23	77.08
		KNN + J48 + MLP + LMT	95.31	93.75	75.00
		KNN + MLP + LR + LMT	93.75	92.19	70.83

Table 6

Accuracy of the five classifiers stacking for the statistical, histogram, and ARMA features

Feature type	Meta classifier	Five classifiers	Training accuracy (%)	Validation accuracy (%)	Testing accuracy (%)
Statistical	KNN	KNN + RF + J48 + MLP + NB	94.79	89.06	81.25
Histogram	RF	RF + KNN + J48 + MLP + LMT	96.35	92.70	77.08
ARMA	KNN	KNN + J48 + MLP + LR + LMT	95.83	93.75	79.16

The results are summarized in Table 7 for comparison and visualization. Additionally, performance metrics are presented in Table 8. Table 7 shows that the highest classification accuracies for individual features – statistical, histogram, and ARMA – were 89.58% (KNN), 75.00% (RF), and 91.66% (KNN), respectively. Applying a stacking-based ensemble approach significantly improved the accuracy for two-class classifiers using histogram and ARMA features, achieving 85.42% (RF + MLP) and 93.75% (KNN + J48), respectively. However, there was no improvement observed for statistical features with the stacking method. Confusion matrices supporting these findings are presented in Figure 2(a) and (b).

Table 7

Performance comparison of best-in-class classifiers for the statistical, histogram, and ARMA features

Feature type	Meta classifier	Classifier	Training accuracy (%)	Validation accuracy (%)	Testing accuracy (%)
Statistical		KNN	100.00	83.85	89.58
Histogram		RF	100.00	93.75	75.00
ARMA		KNN	100.00	92.71	91.66
Statistical	KNN	RF + J48	97.92	90.10	89.58
Histogram	RF	RF + MLP	97.92	92.71	85.42
ARMA	KNN	KNN + J48	96.35	93.75	93.75

Table 8

Additional performance metrics of best-in-class classifiers for the statistical, histogram, and ARMA features

Feature type	Meta classifier	Classifier	Precision	Recall	F-measure	MCC	ROC	PRC
Statistical		KNN	0.902	0.896	0.895	0.864	0.931	0.835
Histogram		RF	0.750	0.750	0.697	0.663	0.920	0.791
ARMA		KNN	0.920	0.917	0.917	0.890	0.944	0.868
Statistical	KNN	RF + J48	0.897	0.896	0.896	0.862	0.971	0.894
Histogram	RF	RF + MLP	0.863	0.854	0.853	0.810	0.964	0.883
ARMA	KNN	KNN + J48	0.940	0.937	0.938	0.918	0.937	0.938

Figure 2

Confusion matrix of (a) histogram features (RF + MLP) and (b) ARMA features (KNN + J48). Source: Created by the authors.

6 Conclusion

This study presents a novel stacking ensemble model to enhance TPMS for nitrogen-filled tires using cost-effective MEMS accelerometers to capture vibration signals under varying tire conditions. By employing ARMA, histogram, and statistical features combined with the J48 decision tree for feature selection, the stacking ensemble achieved high classification accuracy, particularly for ARMA (93.75%) and histogram (85.42%) features. These results demonstrate that the stacking approach significantly outperforms individual classifiers, particularly in identifying subtle patterns in vibration data that correlate with tire pressure conditions. The findings underscore the potential of ensemble methods to improve real-time TPMS performance, providing an accurate and reliable tool for maintaining vehicle safety. This approach not only enhances classification accuracy but also presents a practical and resource-efficient solution suited to real-world applications. Future research could explore additional ensemble configurations, real-time implementation on embedded systems, and the integration of more diverse data sources to improve generalization and robustness under various driving conditions. Ultimately, this study contributes to advancing TPMS technology, paving the way for safer and more efficient vehicle maintenance and monitoring systems.