A Hybrid Method for Technical Condition Prediction Based on AI as an Element for Reducing Supply Chain Disruptions

Abstract

In the field of transport, and more precisely in supply chains, if any of the vehicle components are damaged, it may cause delays in the delivery of goods. Eliminating undesirable damage to the means of transport through the possibility of predicting technical conditions and a state of failure may increase the reliability of the entire supply chain. From the aspect of sustainability, the issue of reducing the number of failures also makes it possible to reduce supply chain disturbances, to reduce costs associated with delays, and to reduce the materials needed for the repair of the means of transport, since, in this case, the costs only relate to the replaced elements before their damage. Thus, it is impossible for more serious damage to occur. Often, failure of one item causes damage to others, which generates unnecessary costs and increases the amount of waste due to the number of damaged items. This article provides an author’s method of technical condition prediction; by applying the method, it would be possible to develop recommended maintenance activities for key elements related to the safety and reliability of transport. The combination of at least two artificial intelligence methods allows us to achieve very good prediction results thanks to the possibility of individual adjustments of weights between the methods used. Such predictive maintenance methods can be successfully used to ensure sustainable development in supply chains.

1. Introduction

In technical diagnostics, significant emphasis is placed on methods that, apart from the detection of damage and classification of technical conditions, also enable their prediction. Advanced systems for continuous diagnostics of devices enable the monitoring of essential features of these devices in real time (i.e., condition monitoring systems). Such systems consist of measuring sensors, usually connected to a measuring computer or a control module, which make it possible to constantly record the value of diagnostic features of the examined technical object. The ongoing monitoring of the technical condition of an object, then enables the use of methods of predicting the value of the examined features [1]. This, in turn, makes it possible to implement a predictive management model. This optimal management strategy is called predictive maintenance.

Therefore, predictive maintenance is based on the extrapolation of acquired data during measurements and setting boundaries for key diagnostic features. Artificial intelligence tools are increasingly being used to extrapolate the obtained data. These are tools that use such areas of knowledge as artificial neural networks, expert systems, fuzzy logic reasoning, and genetic algorithms. Nowadays, thanks to the development of computer science, measurement techniques, and automation, it is possible to build automatic diagnostic systems in which formulation of the assessment of the technical conditions of objects, detection and localization of malfunctions, as well as forecasting the period of correct operation of these objects can take place without or with minimal human participation [2,3].

Artificial neural networks turn out to be useful in the process of identifying failures. Their proper application also enables forecasting of the processes and phenomena that occur in technical objects and the interactions between these objects.

Predictive maintenance usually involves real-time monitoring of the relevant diagnostic features. In the case of vehicle diagnostics, predictive maintenance can be successfully used in monitoring systems and when measurements of diagnostic features are performed only during technical inspections. In this case, the use of predictive maintenance methods seems to be less effective than in the case of constant monitoring of diagnostic features. However, with the help of monthly surveys, it is possible to obtain sufficiently accurate prediction results, which are presented in the research part of this paper.

In the case of individual elements of technical objects, many diagnostic features are measured during a standard technical inspection. Therefore, many factors (diagnostic features) can be identified that affect the technical conditions of such objects and the need to replace elements. These factors are often interrelated, and the ability to indirectly measure selected features also allows us to take into consideration those features where the measurements are more complicated. For example, information on the quality of the railway infrastructure cannot be obtained directly, and the quality of the infrastructure has a huge impact on the wear and tear process of individual components of the rail vehicle. However, such information can be obtained indirectly, e.g., based on an analysis of the distance traveled by the vehicle, the service intervals, wear intensity of the tested element, weather conditions, and vehicle speed. The use of artificial intelligence methods makes it possible to easily consider factors that could not be directly taken into consideration. In addition, determining the technical condition of a technical object as well as identifying the reasons for its replacement and potential causes of damage can be treated as a problem of pattern recognition and classification. For this purpose, artificial intelligence methods are the best solution due to the possibility of analyzing very large datasets related to each other indirectly.

In Section 3 (Methods of Artificial Intelligence), we describe selected artificial intelligence methods that can be used to detect damage and classify the technical condition of the technical object. These methods can be successfully used in the failure prediction method developed by the authors, which has been tested on a selected vehicle component.

2. State of the Art

Problems of predictive maintenance usage are presented in several publications. They often consider the use of artificial neural networks [4,5,6,7,8,9,10,11], fuzzy logic [12,13,14,15], genetic algorithms [16,17,18], or expert systems [19,20,21,22].

In the field of Industry 4.0 and IoT, one of the sectors in which research is particularly active is the area of predictive maintenance (PdM), the purpose of which is to improve the industrial production process [23,24]. Predictive maintenance is fundamental for modern industries, to improve physical assets availability, for decision making, and to rationalize costs [25]. The need to fulfil the requirements of present-day customers has led to the development of IoT for smart manufacturing and smart logistics [26]. This type of Industry 4.0 method can also be used for sustainable management of maintenance resources [27]. The use of artificial intelligence is popular in the area of predictive maintenance. In [28], the authors investigated the health of an engine through experimental observation using an artificial neural network (ANN). An ANN can also be used for early fault detection [29,30]. Another interesting publication showed how to use deep learning with LPC and wavelet algorithms for driving fault diagnosis [31]. Also, for the dynamic maintenance strategy, genetic algorithms can be used. In [32] the authors presented the use of an ANN and machine learning for predicting ship maintenance and repair labor. In [33], the authors presented this type of strategy. They considered the structure of multicomponent systems; the maintenance strategy was determined according to the importance of the components. In their research, the authors proved that the strategy could minimize the expected depreciation cost of the system and the system could be divided into optimal groups that met the economic requirements. Also, different researchers [33] have used predicting methods based on clustering and autoencoders for fault detection and diagnosis of industrial processes. To ensure a reliable maintenance strategy, establishing a plan for the entire maintenance process for the assets, from preliminary asset confirmation through to entering the master data in the management system, is necessary [34]. ANN-based simulation models integrated into decision support systems (DSSs) significantly improve decision making in complex scenarios requiring quick responses [35].

Machine learning (ML) is a branch of the science of artificial intelligence (AI). There are two basic types of machine learning (Figure 1). On the one hand, for supervised learning, classification and regression algorithms are used. The developed predictive model, in this case, is built based on input and output data.

Unsupervised learning, on the other hand, is based on cluster analysis, i.e., clusterization. In this case, the grouping and interpretation of data are based only on the input data.

2.1. Selected Classification and Regression Algorithms

Due to the fact that this work uses supervised machine learning, selected classification and regression algorithms are presented below. The prediction and classification algorithms are the two main types of prediction algorithms, where classification is used to predict discrete values and regression is used to estimate continuous or ordered values. In the case of classification algorithms, several basic algorithms presented below can be distinguished [36].

• Naive Bayes classifier

The naïve Bayes classifier is a two-class and multiclass classification algorithm. The “naivety” of this classifier lies in the fact that it assumes the independence of individual variables, which is rarely the case in reality. Its operation is based on calculating the probability of events. Therefore, it is classified as one of the simplest algorithms [9]:

$$P(c_i|x)=\frac{P\left(c_i\right)P\left({x|c}_i\right)}{P\left(x\right)}$$

(1)

where $P\left(c_i\right|x)$ is the probability that the example x belongs to a category $c_i$, $P\left(c_i\right)$ is the probability of the $c_i$ category occurrence in general, $P\left({x|c}_i\right)$ is the probability of the x example occurrence provided the $c_i$ category occurs, $P\left(x\right)$ is the probability of the example x occurring.

Despite its simplicity and independence of variables, this classifier works very well when the amount of data available for training the model is limited. The Bayes classifier is also a good choice when CPU and memory resources are constrained. Since this classifier is very simple, it does not tend to overflow and can be trained very quickly.

• Nearest neighbors algorithm

The nearest neighbors algorithm categorizes data points based on their distance to other points in the training dataset. It is one of the nonparametric regression algorithms used to predict the values of random variables. This makes it a simple yet effective way to classify data. Therefore, the assumption of this algorithm is as follows: similar problems have similar solutions.

In two-dimensional Euclidean space, the distance can be calculated from the Pythagorean theorem. However, since we rarely analyze a dataset with only two features (two-dimensional), by generalizing the problem of distance to n-dimensional space, the distance function can be defined with the following metric:

$$d=\sqrt{\sum _{i=1}^n{\left(x_i-y_i\right)}^2}$$

(2)

This is the most common way to calculate distance, but not the only one. In general, the measure of closeness can be calculated by using methods such as the squared Euclidean measure method, the Manhattan method, and the Chebyshev method. The dependent and independent variables can both be continuous and discrete (categorized).

• Random forests

The random forest algorithm is used for two-class and multiclass classification (Figure 2). Its operation, as the name suggests, is based on classification, which uses a group of decision trees.

The random forest algorithm is an improvement of the decision tree algorithm due to the application of the so-called bootstrap aggregation otherwise known as bagging. It consists of creating a few training subsets from the original dataset. This choice is made by random selection. Some of the observations that have not been selected for the training set automatically act as the validation set. Due to this approach, there is no need to divide the data into training and test sets. In the next step, explanatory variables are selected for each of the subsets. Then, the construction of decision trees for each of the subsets follows. The final result is selected by voting; the value that has been predicted by the largest number of individual decision trees is selected.

• Neural networks

A neural network algorithm can be used in both classification and regression problems. In the case of classification, neural networks are used for two-class and multiclass classification. The scheme of operation is presented using a directed graph. There are neurons in the network with any number of inputs and outputs. In successive layers, operations on variables are performed until the result value is reached at the end of the graph.

Figure 3 shows the structure of the neuron. The mathematical formula describing such a neuron is as follows [37]:

$$a=f\left(wp+b\right)=y$$

(3)

In the case of artificial neural networks, the values of the input weights for individual neurons as well as the values of the bias inputs are modified during the learning process, where a is the output value from the neuron, p is the neuron input value, f is the activation function, w is the neuron weight value, b is the bias input value (threshold), y is the output value from the neural network.

• Support vector machine

The support vector machine algorithm works on the principle of an abstract machine model. Classification is performed by using nonlinear decision boundaries. The decision-making space is divided by building boundaries that separate objects of different class affiliations. When determining partition hypersurfaces, this algorithm takes into consideration only the data that are located near the class boundary. Since there can be a very large number of functions separating the two classes, the algorithm for computing the hyperplane aims to obtain the largest margin possible. On this basis, the support vectors are determined on which the separating hypersurface is spread. An example showing the classification of data to appropriate classes is shown in Figure 4.

The distance of a point from a plane is given by the formula:

$$\mathrm{d}=\frac{\left|w_1{x}_1+w_2{x}_2+\dots +w_n{x}_n+w_0\right|}{\sqrt{w_1^2+w_2^2+\dots +w_n^2}}=\frac{\left|w^{T}x+w_0\right|}{\left|\left|w\right|\right|}$$

(4)

The vector of the coefficients w = [w1, w2, …, wn]^T is the normal vector of the plane (it is perpendicular to the plane), while ‖w‖ is the length of this vector. Then, it is established that the distance between each point xi and the discriminant plane is not less than a certain value of M called the margin.

The final decision rule takes the following form:

$$\mathrm{f}\left(\mathrm{x}\right)=\mathrm{s}\mathrm{g}\mathrm{n}\left({\mathbf{w}}^{\mathrm{T}}\mathit{x}+w_0\right)=sgn\left(\sum _{i\in \mathsf{\Omega }}{\alpha }_i{y}_i\left({\mathit{x}}_i·\mathit{x}\right)+w_0\right)$$

(5)

where sgn (z) = 1, when z > 0 and then the point belongs to class (+1); sgn (z) = −1, when z < 0 and then the point belongs to class (−1); (x_i · x) is the scalar product of two vectors.

The machine learning methods can be successfully used to develop models used in the method described below.

• Linear regression

Linear regression is typically a first-choice algorithm when predicting continuous values. It assumes the existence of a relationship between model variables and predictors. The fitted regression line or curve, therefore, represents the estimated value of the expected output variable, i.e., the predictor. Thus, appropriately selected parameters constitute a solution to the prediction problem. These parameters are selected in such a way that the distances from the straight line to the observation are as short as possible. The least squares estimation method is usually used for this purpose. The smaller the distance between the real and theoretical values, the better the model, and the estimators of the model parameters minimize the sum of the distances between $y_i$ and $\widehat{y_i}$:

$$\sum _{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2=\sum _{i=1}^n{e}_i^2$$

(6)

• Decision trees

This algorithm is often referred to as the classification tree (Figure 5). It is based on tree-shaped structures. Nodes contain conditional tests for attributes, and leaves contain categories or probabilities of categories. The decision tree is built in an iterative way from the root to the leaves. Attributes are selected by the feature selection algorithm in such a sequence as to maximize the information gained from a given node.

The model-building process ends when all the leaves are in the same class or when there are no classes to divide. In this case, the algorithm may be overfitted; therefore, the process of building such a tree can be limited to a specific depth. To avoid overfitting, apart from limiting the learning depth, the number of branch divisions can also be limited, and the so-called reinforcement consists of a random generation of new training strings, which are used to train subsequent versions of the classifier.

2.2. Artificial Intelligence in Science and Industry

Several prevalent AI techniques are being used in supply chain management (SCM). For example, predictive analytics involves using historical data and machine learning algorithms to forecast future demand, inventory levels, and supply chain disruptions. This helps with optimizing inventory management and improving demand forecasting accuracy. Machine learning (ML) techniques are used for various SCM tasks such as demand forecasting, route optimization, and quality control. ML models can analyze large datasets to identify patterns and to make data-driven decisions. Computer vision technologies can be applied to automate quality control processes by inspecting products and identifying defects. They can also help with inventory tracking using visual data. Blockchain technology can be used for enhanced transparency and traceability in supply chains. It helps to verify the authenticity of products and to ensure that goods are not tampered with during transit. AI can be used to assess supplier risk by analyzing various factors like financial stability, geopolitical events, and natural disasters, allowing organizations to proactively manage supply chain disruptions. AI techniques can be used to optimize inventory levels by considering factors like demand variability, lead times, and carrying costs, ensuring that the right amount of stock is available without excess.

AI can analyze vast amounts of data to monitor and measure environmental and social sustainability metrics across a supply chain. This includes tracking carbon emissions, water usage, waste generation, and labor conditions. By providing real-time insights, AI helps organizations to identify areas where sustainability goals are not being met and to take corrective actions. AI can predict and identify potential environmental and social risks in a supply chain, such as natural disasters, regulatory changes, and labor issues. This early warning system allows companies to proactively address these risks, reduce disruptions, and ensure the well-being of workers and the environment.

AI plays a crucial role in promoting sustainability in supply chains by providing data-driven insights, optimizing processes, ensuring transparency, and helping organizations to make informed decisions. By addressing environmental and social challenges, AI contributes to the creation of more resilient and responsible supply chains that are better equipped to handle the threats posed by climate change, resource scarcity, and social issues.

Using predictive maintenance with AI methods to manage vehicle components in a supply chain can offer several key benefits and implications for reducing disruptions, costs, and waste while ensuring sustainability:

• Reduced Downtime and Disruptions. Predictive maintenance helps in identifying potential issues with vehicle components before they fail. This proactive approach minimizes unexpected breakdowns and downtime, ensuring that goods are delivered on time.
• Improved Resource Allocation. Organizations can allocate resources more efficiently by focusing on components that require immediate attention. This ensures that maintenance efforts are directed where they are most needed.
• Reduced Waste. By minimizing the need for premature component replacements, predictive maintenance reduces waste, including discarded vehicle parts, which is environmentally harmful.
• Resource Conservation. Extending the lifespan of vehicle components reduces the need for raw materials and energy required for manufacturing new components, thus contributing to resource conservation.
• Emissions Reduction. Predictive maintenance reduces the frequency of vehicle breakdowns and the associated emissions from idling vehicles waiting for repairs or towing services.
• Resilience and Risk Mitigation. Predictive maintenance helps to ensure that vehicles are in optimal condition, reducing the risk of disruptions due to breakdowns or accidents. This improves the overall resilience of a supply chain.
• Reputation and Customer Satisfaction. Sustainable practices including reliable delivery services due to effective predictive maintenance enhance a company’s reputation and customer satisfaction.

The combination of AI methods for predictive maintenance has applications that extend far beyond supply chain management. These principles that leverage data and AI algorithms to anticipate and prevent equipment failures are adaptable to a wide range of industries and sectors. For example, in manufacturing, AI-based predictive maintenance can enhance equipment reliability, reduce downtime, and optimize resource utilization. In healthcare, it can ensure the reliable operation of medical devices, minimizing equipment downtime and improving patient care. The energy sector can benefit from improved power plant performance and grid management, while aviation can achieve enhanced safety and aircraft availability. However, the adoption of AI-based predictive maintenance is not without its challenges. It often involves data integration, security and privacy concerns, and compliance with industry-specific regulations. Nevertheless, the potential benefits, including increased reliability, cost savings, and improved efficiency, make it a promising approach for various industries as technology continues to advance.

The implementation of predictive maintenance with AI in a supply chain offers significant benefits but also comes with the following limitations and potential risks that need to be carefully addressed to ensure successful adoption and integration:

• Data Quality and Availability. Predictive maintenance relies on high-quality and timely data from sensors and equipment. If the data are inaccurate or unavailable, the effectiveness of the AI models can be compromised.
• Cost of Implementation. The upfront cost of implementing AI-driven predictive maintenance systems, including sensors, data infrastructure, and AI software, can be substantial. Smaller businesses may find it challenging to make this initial investment.
• Integration Complexity. Integrating AI systems with existing supply chain management and maintenance systems can be complex. Compatibility and data sharing between systems may require significant effort and resources.
• Skill and Training. Organizations need skilled personnel to develop, operate, and maintain AI-based predictive maintenance systems. A shortage of talent and the need for continuous training can be obstacles.
• Maintenance of AI Systems. AI models require continuous maintenance and updates to remain effective. Failure to maintain AI systems can lead to performance degradation.
• False Positives/Negatives. Overreliance on AI predictions may lead to false alarms or missed maintenance opportunities. Organizations must carefully validate AI-generated alerts.
• Security Concerns. AI systems can be vulnerable to cyberattacks. Protecting sensitive maintenance data and ensuring the security of AI systems is essential.
• Privacy Issues. Collecting and analyzing data related to equipment and maintenance may raise privacy concerns, especially when personal data are involved such as employee information.
• Regulatory Compliance. In some industries, predictive maintenance systems must comply with specific regulations and standards. Failure to do so can result in legal and regulatory penalties.
• Dependency on Technology. Overreliance on AI systems can make organizations vulnerable to disruptions if the technology fails or experiences downtime.

AI-based predictive maintenance holds great promise for industries, offering benefits such as reduced downtime, increased productivity, and enhanced equipment reliability. Research efforts should focus on refining AI algorithms, data analytics, and sensor technologies to drive further advancements. This technology also aligns with sustainability goals by conserving energy and reducing environmental impact, making it a win–win situation for both industry and research.

2.3. The Previous Predictive Models Tested Separately

In the area of analyzing and predicting the state of a technical object on an example of pantograph’s sliding strips, previous research has been published in [38,39], and the following predictive models have been described, developed based on real data from maintenance systems:

• Model based on the decision trees method.
• Model based on artificial neural networks.

The paper by [38] described a predictive model based on decision trees. To predict the technical condition of the object, a three-valued prediction method based on machine learning and failure analysis was developed. The results from the developed model were in a discrete form. Therefore, the anticipated technical conditions can be described as follows:

$$\begin{gathered} S_1⟺\left[y_{13}=1\right] \\ {S}_2⟺\left[y_{13}=2\right] \\ {S}_3⟺\left[y_{13}=3\right] \end{gathered}$$

(7)

Based on the results, it can be determined that the correctness of the classification was about 81%. However, correct classification of Class 2 is the most important. Class 2 indicates that the technical object will fail soon (which defines Class 3). In the confusion matrix, 49.6% of cases in Class 2 were identified as Class 1 (class 1—possibility of further use). In practice, it means that about half of cases can be identified before failure (i.e., reach Class 3—no use, it is necessary to replace the object).

In a second paper by [39], a predicting method based on an ANN was developed. Due to the large number of conducted studies, it was decided to include only representative models based on artificial neural networks. The most representative models are presented in

Table 1. List of selected predictive models and parameters.

Model No.	Type of Learning Method			Model Parameters
1	Feed-forward artificial neural network with backpropagation	Activation function: TANSIG	Learning algorithm: TRAINLM	Number of hidden layers: 5 (14-14-14-14-14-3)
2				Number of hidden layers: 5 (12-12-12-12-12-3)
3				Number of hidden layers: 5 (12-12-12-12-12-3)
12				Number of hidden layers: 1 (10-3)
4		Activation function: TANSIG/PURELIN	Learning algorithm: TRAINLM	Number of hidden layers: 1 (12-3)
5				Number of hidden layers: 1 (6-3)
6				Number of hidden layers: 1 (10-3)
7			Learning algorithm: TRAINBR	Number of hidden layers: 1 (10-3)
8			Learning algorithm: TRAINLM/TRAINBR	Number of hidden layers: 1 (10-3)
9	Feed forward artificial neural network with backpropagation distributed time-delay	Activation function: TANSIG	Learning algorithm: TRAINBR	Number of hidden layers: 1 (10-3)
10				Number of hidden layers: 1 (10-3)
11			Learning algorithm: TRAINCGB	Number of hidden layers: 1 (10-3)

.

The selected model can be described as follows:

$$\mathrm{a}2=\mathrm{f}2(\mathrm{I}\mathrm{W}\mathrm{2,1}\mathrm{f}1\left(\mathrm{I}\mathrm{W}\mathrm{1,1}\mathrm{p}+\mathrm{b}1\right)+\mathrm{b}2)$$

(8)

Since the simulation results were continuous, to determine the technical condition, it was necessary to define the ranges of values belonging to a given technical condition. For example, for the model giving the best results, it was defined as follows:

$$\begin{gathered} S_1⟺\left[0\le y_{12}<1.25\right] \\ {S}_2⟺\left[1.25\le y_{12}\le 2.5\right] \\ {S}_3⟺\left[2.5<y_{12}\le 3.5\right] \end{gathered}$$

(9)

where

• $S_1$ is the first technical condition, i.e., able to further use;
• $S_2$ is the second technical condition, i.e., the limited ability of further use, it will be necessary to replace the sliding strip for the next inspection;
• $S_3$ is the third technical condition, i.e., not able to further use, it is necessary to replace the sliding strip/pantograph;
• $y_{12}$ is the value obtained during the prediction based on use of the ANN predictive model.

The simulation results based on the ANN allowed us to obtain the correctness of the classification of State 2 at the level of 85.1%.

3. Prediction Hybrid Method

Traditional predictive methods, such as statistical models, may not always effectively handle the complex relationships between data and the technical state of a TO. Therefore, machine learning methods are increasingly being employed as they can uncover hidden patterns and dependencies in the data.

Machine learning methods used in a predictive model can be based on classification and regression methods. Classification allows for identifying faults and planning maintenance, with example methods being decision trees, SVM, k-NN, random forests, and neural networks. Regression predicts numerical values, enabling optimal resource management. Example regression methods include linear regression, logistic regression, SVM, random forests, and neural networks. These methods support precise forecasts and appropriate corrective or maintenance actions to ensure the reliability and efficiency of technical objects.

To increase the effectiveness of predicting the technical state of a TO, a hybrid model has been developed, combining at least two predictive models. Such an approach allows for leveraging the advantages of both methods and achieving synergy, resulting in more accurate forecasts.

The benefits of employing a hybrid model are numerous. Firstly, by combining different predictive methods, it is possible to incorporate different aspects and perspectives in a data analysis, leading to higher quality predictions. Secondly, a hybrid model may be more resilient to errors and disturbances, as errors from one model can be compensated for by the other. Additionally, the use of a hybrid model can contribute to cost reduction in repairs and maintenance through better planning and optimal resource utilization.

This section presents a detailed description and analysis of the hybrid method for predicting the technical state of a TO. The individual stages of the developed method are discussed and the results of experiments and comparisons with other approaches are presented. The results confirm the effectiveness of the hybrid method and its advantages over individual predictive methods.

The research hypothesis was defined as follows:

“Combining two different AI methods can lead to achieving predictive results that more accurately correspond to real outcomes than using each predictive method separately.”

In our research, we use the MATLAB software (version 23.2) for simulation, and Tableau for the data analysis and results visualization. Simulations allow the modeling of real-life situations or processes, enabling the testing of various scenarios and preparation for potential events. Simulation studies facilitate conducting virtual experiments in a controlled and safe environment, without the need for engaging physical resources or costly field experiments. In addition, simulations enable the verification and validation of theoretical assumptions and the comparison of model results with reality. The MATLAB software, when applied, provides tools for rapid prototyping and iterative testing of ideas and algorithms.

The use of AI in the presented research is justified by the fact that AI can tailor the diagnostic process to specific devices or systems, enabling more precise results and an approach to maintenance. AI enables continuous real-time monitoring of the technical condition of devices, allowing for real-time response to changes. The elimination or reduction of human errors in the diagnostic process improves not only accuracy but also safety.

3.1. Technical State Prediction Hybrid Model

To increase the efficiency of predicting the technical state of a technical object (TO), a hybrid model was created. The presented methodology enables the use of two different AI methods; in the model, two high-performance methods should be used for prediction.

The three-valued method of assessing the technical condition of a technical object consists of six stages. In this work, the concept of a stage means a set of interconnected activities aimed at fulfilling a specific task (presented in the heading for each stage). Steps can be repeated in a loop in a way that results from the connections between them.

In the developed method, we can distinguish the following stages:

• Training a predictive model based on data collected during reviews in Level 2.
• Prediction of the technical state of a technical object using the developed hybrid model for reviews at Level 2.
• Recommended maintenance activities during technical review at Level 2.
• Training the predictive model based on data collected during reviews at Level 1.
• Prediction of the technical state using the developed prediction model for review at Level 1.
• Recommended maintenance activities during review at Level 1.

The hybrid model that has been developed combines the advantages of two different prediction methods presented in Section 2.3. For hybrid method testing, the decision tree was used as AI Method 1 and the artificial neural network as AI Method 2. The decision tree is used for data analysis and decision making based on rules, while the artificial neural network is an effective tool for learning from a training set and making predictions based on previously acquired knowledge. Combining these two methods allows for synergy and increased accuracy in predicting the technical state of the TO and clearly illustrates the operation and advantages of using a hybrid model.

3.1.1. Stage I—Training a Predictive Model Based on Data Collected during Reviews in Level 2

In Stage I, based on measurement data obtained during the technical review at Level 2, as well as the rules for identifying the reasons for technical object replacement and the current technical condition, a set of input (

Table 2. Input data for the prediction model.

Type of Input Data		Input Data Structure Number
		1	2	3
The number of input data sin		14	12	10
1	Review number	X	X	X
2	New measuring cycle	X	X	X
3	The number of days since the exchange	X	X	X
4	Quarter of the year	X	X	X
5	Average temperature in the month (°C)	X	X
6	Average wind speed for the month (km/h)	X	X
7	Total rainfall for the month (mm)	X	X
8	Pantograph type	X	X	X
9	Front/rear pantograph	X	X	X
10	The difference in thickness of the strip N1 between inspections	X	X	X
11	The difference in thickness of the strip N2 between inspections	X	X	X
12	Sliding strip thickness N1	X
13	Sliding strip thickness N2	X
14	Reason for replacement during the previous measurement	X	X
15	Earlier technical condition			X
16	Reason for replacement			X

) and output data was developed to teach predictive models (and thus a hybrid model). The result of Stage I is the state S = {1, 2, 3}, the value obtained during the prediction of the technical condition for the hybrid model.

For the technical review, the following levels were determined:

• Level 1, short interval reviews, about 2–3 days;
• Level 2, medium interval reviews, about one week;
• Level 3 and more, long-term reviews, over one month are not important for this method.

In Figure 6, the training process of a predictive model is presented.

The developed training data were used in the hybrid model to predict the technical state in the form of S variable.

3.1.2. Stage II—Prediction of the Technical State of a Technical Object Using the Developed Hybrid Model for Reviews at Level 2

In the process of predicting the technical state of the technical object, the same hybrid model is used as in Stage I. In Stage II, work begins with data that were analyzed and imputed during the review at Level 2.

A total of about 750 service cards from one of the main polish rail carriers with over 1500 measurements of current collectors made over 2 years were analyzed. The measurements concerned 62 locomotives, types EP09 and EU07. The data were preprocessed; the paper service cards were transformed to digital and the reason for the changing pantograph parts was presented in previous research [38]. These cards show that during the P2/P3 maintenance (at the P2/P3 “maintenance level” according to DSU), 8.3% of the current collectors were replaced in full (127 cases), while in 273 cases, the sliding block was replaced, which accounts for as much as 17.8% of all cases. The analysis of the collected data shows that in the enterprise, the most serviced were vehicles with AKP-4E collectors (56.6%), and vehicles with 5ZL collectors constituted 29.9%. It follows that 86.5% of the serviced current collectors were four-arm collectors and only 13.5% were single-arm DSA 150 collectors. The input data structure is presented below.

Table 3. List of selected predictive models and results.

Model No.	Type of Learning Method	Input/Predictors Regarding Table 1	Output/Response Regards to Table 2	Model Parameters
1	(1) ANN F-T-Lm	1	1	Number of hidden layers: 5 (14-14-14-14-14-3)
2	(1) ANN F-T-Lm	2	1	Number of hidden layers: 5 (12-12-12-12-12-3)
3	(1) ANN F-T-Lm	2	1	Number of hidden layers: 5 (12-12-12-12-12-3)
4	(2) ANN F-TP-Lm	2	1	Number of hidden layers: 1 (12-3)
5	(3) ANN F-TP-Lm	2	1	Number of hidden layers: 1 (6-3)
6	(2) ANN F-TP-Lm	3	1	Number of hidden layers: 1 (10-3)
7	(3) ANN F-TP-Br	3	1	Number of hidden layers: 1 (10-3)
8	(4) ANN F-TP-Lm/Br	3	1	Number of hidden layers: 1 (10-3)
9	(5) ANN Ft-T-Br	3	1	Number of hidden layers: 1 (10-3)
10	(5) ANN Ft-T-Br	3	1	Number of hidden layers: 1 (10-3)
11	(6) ANN Ft-T-C	3	2	Number of hidden layers: 1 (10-3)
12	(2) ANN F-TP-Lm	3	2	Number of hidden layers: 1 (10-3)

lists 12 predictive models.

In Figure 7, the initial conditions specifying the technical condition of the technical object and the type of technical reviews are presented. The technical condition in the initial conditions is assumed to be the state S = 1 and the type of technical review at Level 2. In the case of the next iteration, i.e., during the next review at Level 2 for the same technical object, the database is supplemented with subsequent values, e.g., wear condition, as well as the automatic addition of recommended maintenance activities and possible changes in the technical state and type of technical review resulting from the analysis for the previous review.

3.1.3. Stage III, Recommended Maintenance Activities during the Technical Review at Level 2

Recommended maintenance activities are determined based on the technical state resulting from the prediction process (Figure 8). If the state S = 3 was identified, a replacement of the technical object is required. In this situation, the next review of the same object will take place during the review at Level 2. In the case of an unserviceable state (S = 3), the reason for replacement is also determined. The reason for the replacement was determined during data analysis based on decision rules, which are determined individually for each type of technical object. The need to identify the reason for the replacement results from the fact that it was not recorded during the reviews. In the proposed method, if a condition of unserviceability is identified, it will be necessary to record the reason for replacement. This will enable verification of the correctness and adjustment of the developed rules for identifying the reasons for replacement, and thus predictive models.

In the case of identification of the “able to further use” state (S = 1) at review n, the information obtained at this stage is added to the next review (n + 1), i.e., information about the technical state of review n. The type of review does not change then, so, i.e., review (n + 1) is also a review at Level 2.

In the case of identifying the state of limited possibility of further use (S = 2) at review n, it is necessary to increase the frequency of measurements. Therefore, the interval of review should be shortened to Level 1. The next inspection (n + 1) for the same technical object will be in such a case a review at Level 1, which will be carried out by Stage IV.

3.1.4. Stage IV—Training the Predictive Model Based on Data Collected during Reviews at Level 1

Stage IV takes place during the review at Level 1 (n + 1), i.e., only in the case of identifying the state of limited possibility of further use (S = 2) at the review with the identifier n. If such data are not collected by the service, it will be necessary to collect sufficient data for the review at Level 1, shortening the measurement interval to Level 1.

In this method, it was assumed that the learning process of the predictive model for the review at Level 1 will finish when data from at least 100 review cases (P1n) has been collected. This will enable verification of the predictive model based in its original form on the developed hybrid model. The result of this stage will be the value S = 1, obtained during the prediction of the technical state during the review at Level 1. Stage IV is presented in Figure 9.

3.1.5. Stage V—Prediction of the Technical State Using the Developed Prediction Model for Review at Level 1

Stage V is presented in Figure 10 and concerns prediction of the technical state of the technical object during the review at Level 1. This stage starts when data from at least 100 cases are collected. Then, using the already tested predictive model for data from review at Level 1, it is possible to identify the technical state. For the shortened interval to Level 1, only the state of limited ability of further use (S = 2) and not able to further use (S = 3) will be identified.

The technical state is determined by the value obtained from the prediction of the technical state for the Level 1 review and the use of rules defining the ranges of values for individual technical states.

3.1.6. Stage VI—Recommended Maintenance Activities during Review at Level 1

In stage VI, as schematically shown in Figure 11, recommended maintenance activities are determined based on the technical state identified during review at Level 1. In the case of identifying S = 2, i.e., limited ability of further use, it is possible to continue work using the technical object. If S = 3, i.e., not able to further use, was identified, it is necessary to replace the technical object with a new one. Then, as in the case of identifying the state S = 3, it is also necessary to determine the cause of replacement. This will help to correct the rules for identifying reasons for replacement. At this point, the identification of the object’s technical condition is completed.

3.2. Hybrid Method Testing

In the hybrid method, the decision tree model was used as AI Method 1, and the artificial neural networks as AI Method 2. The hybrid method was tested on data obtained from monthly reviews at Level 2 technical reviews. About 900 training cases were used for this purpose, 80% of which were used for the model training process, while 20% were used for the simulation process. Only the input data were given then, and the results were compared with the actual values obtained during technical inspections.

The AI Method 1 uses the decision tree method, which consists of decision nodes and leaf nodes. Each decision node represents the optimal splitting decision based on xi values. The decision tree is constructed by recursively splitting the root nodes to the left and right until the maximum depth is reached. In each prediction step, the value of one predictor (variable ×1 to ×10) is checked, and the Gini diversity index is calculated. In the proposed model, the maximum number of splits was 100, the Gini diversity index was used as the splitting criterion, and there were no surrogate decision splits. The Gini diversity index calculates the probability of a specific feature being misclassified when randomly selected. If all elements are assigned to a single class, it is considered pure. The Gini diversity index ranges from 0 to 1, where 0 indicates classification purity, meaning all elements belong to a specified class or only one class exists. Additionally, a value of 1 indicates a random distribution of elements across different classes. This model was presented and tested on the predictive model described in the article by [38].

The AI Method 2 uses the artificial neural network method. Selected predictive models have been presented and tested in the article by [39].

Table 4. Summary of the results of selected predictive models.

Model No.	Method Type (acc. to Table 4)	Input (acc. to Table 1)	Output (acc. to Table 2)	Training		Simulation
				MSE	R	The Correctness of the Classification of All Technical Conditions	The Correctness of Classification of the Second Condition ${\mathit{S}}_2$
1	(1) ANN F-T-Lm	1	1	0.12497	0.75943	59.8	4.3
2	(1) ANN F-T-Lm	2	1	0.13384	0.67901	41.2	31.9
3	(1) ANN F-T-Lm	2	1	0.11970	0.71918	48.1	10.6
4	(2) ANN F-TP-Lm	2	1	0.11913	0.72838	53.2	17.0
5	(2) ANN F-TP-Lm	2	1	0.11669	0.71501	49.3	12.8
6	(2) ANN F-TP-Lm	3	1	0.069108	0.84538	76.5	38.3
7	(3) ANN F-TP-Br	3	1	0.043195	0.87944	76.2	38.3
8	(4) ANN F-TP-Lm/Br	3	1	0.038862	0.88206	78.1	42.6
9	(5) ANN Ft-T-Br	3	1	0.014731	0.9222	78.6	61.7
10	(5) ANN Ft-T-Br	3	1	0.020364	0.92088	82.0	61.7
11	(6) ANN Ft-T-C	3	2	0.064105	0.8938	81.5	80.9
12	(2) ANN F-TP-Lm	3	2	0.15974	0.90966	82.5	85.1

The designations presented in Table 5 are MSE—mean square error and R—Pearson correlation coefficient.

presents a summary of results of selected ANN predictive models. As a loss function, a MSE function was used as follows:

$$loss=\frac{1}{2N}\sum _{i=1}^M{\left(X_i-T_i\right)}^2$$

(10)

where X_i is the network prediction, T_i is the target value, M is the total number of responses in X (across all observations), and N is the total number of observations in X.

The best result was achieved for Model Number 12. In this model, a feed-forward network with backpropagation was used. One hidden layer with 10 neurons was applied. The tangential activation function (TANSIG) was used, as well as the linear activation function for the output layer (PURELIN). The Levenberg–Marquardt algorithm (TRAINLM) was used as the learning algorithm.

The best result was achieved for 12 epochs, with MSE = 0.15974, R = 0.90966, gradient = 0.1293, and Mu = 0.0001. For training, the accuracy was 0.91632, and for testing, the accuracy was 0.94019.

In both tested AI methods, i.e., AI Method 1 (decision tree) and AI Method 2 (ANN model) the same structure of input and output data was used, but to develop a hybrid method it is not necessary to have the same structure of such data. For the hybrid method, the data structure which is the best for each AI method should be used. It should also be noted that the developed method is a classification method, and the results are presented in a discrete form, while the input data can be both discrete and continuous.

In the hybrid method, by determining the ranges of limit values defining the expected technical state, it was possible to determine the correctness of the classification of technical conditions. For this purpose, the percentage share of S values correctly classified to technical state was calculated. Thus, the correctness of the classification means the number of cases classified to the same technical condition as occurred in reality.

Table 5. The correctness of the classification of technical state for the presented model.

Technical State	${\mathit{S}}_1$	${\mathit{S}}_2$	${\mathit{S}}_3$
Correct classification [%]	61.5	93.6	100.0

presents the correctness of the classification of technical conditions obtained by using the presented AI models and the hybrid method.

To better illustrate the advantage of the hybrid method, the comparison of the three methods is presented in Figure 12. For State S = 2 (the most important state), the correct classification for the developed hybrid method was obtained at the level of 93.6%.

Despite using a regression method that produces continuous results in Method 1, the results of the regression were transformed into a discrete form by adjusting appropriate value ranges for each class. This transformation allowed for the development of a hybrid model and an objective comparison of the results of all methods. The confusion matrix is a fundamental tool in machine learning evaluation, providing a summary of the performance of classification models. Therefore, in Figure 13, three confusion matrices are presented for three different methods, offering insights into their respective classification accuracies and error rates.

For the three class classification problem, the metrics were computed both globally (for all classes combined, i.e., global accuracy), and individually for each class, according to the Equations (11)–(14).

• Global Accuracy:

$$Accuracy=\frac{{\mathrm{T}\mathrm{P}}_{\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{e}\mathrm{s}}+{\mathrm{T}\mathrm{N}}_{\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{e}\mathrm{s}}}{{\mathrm{T}\mathrm{P}}_{\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{e}\mathrm{s}}+{\mathrm{T}\mathrm{N}}_{\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{e}\mathrm{s}}+{FP}_{\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{e}\mathrm{s}}+{\mathrm{F}\mathrm{N}}_{\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{e}\mathrm{s}}}$$

(11)

To determine the accuracy for individual classes, metrics such as precision, recall, and the F1 score were utilized. These metrics provide information about the quality of classification in the context of individual classes. For each class (i), these coefficients were calculated using TP, FP, and FN specific to that class:

$${Precision}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}=\frac{{\mathrm{T}\mathrm{P}}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}}{{\mathrm{T}\mathrm{P}}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}+{FP}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}}$$

(12)

$${Recall}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}=\frac{{\mathrm{T}\mathrm{P}}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}}{{\mathrm{T}\mathrm{P}}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}+{\mathrm{F}\mathrm{N}}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}}$$

(13)

$${F1Score}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}=\frac{2·{Precision}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}·{Recall}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}}{{Precision}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}+{Recall}_{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}i}}$$

(14)

where n is the number of classes, TP_i is the number of true positives for class i, TN_i is the number of true negatives for class i, FP_i is the number of false positives for class i, FN_i is the number of false negatives for class i.

Below, in

Table 6. The accuracy, precision, recall, and F1-score metrics for the AI Method 1.

	AI Method 1 (Decision Tree) Class—Technical Condition:
	s1	s2	s3
Precision:	0.892	0.521	09.79
Recall:	0.883	0.532	1.000
F1 Score:	0.887	0.526	0.989	All classes:
Accuracy:	0.846	0.849	0.997	0.897

,

Table 7. The accuracy, precision, recall, and F1-score metrics for AI Method 2.

	AI Method 2 (ANN) Class—Technical Condition:
	s1	s2	s3
Precision:	0.948	0.345	0.979
Recall:	0.624	0.851	1.000
F1 Score:	0.753	0.491	0.989	All classes:
Accuracy:	0.719	0.722	0.997	0.813

and

Table 8. The accuracy, precision, recall, and F1-score metrics for the hybrid method.

	Hybrid Method (Decision Tree + ANN) Class—Technical Condition:
	s1	s2	s3
Precision:	0.977	0.361	0.979
Recall:	0.615	0.936	1.000
F1 Score:	0.754	0.521	0.989	All classes:
Accuracy:	0.726	0.729	0.997	0.817

, indicators for each of the three methods are presented.

Particular attention should be given to the fact that despite a lower precision value for class s2 (State 2) in the hybrid method compared to the AI Method 1 (decision tree), the recall for s2 in the hybrid method is significantly higher at 0.936. In this specific case, this coefficient determines the quality of the predictive model. A recall of 0.936 for the hybrid method means that approximately 94% of actual s2 states were correctly classified as s2 during prediction. In other words, almost all s2 states were detected correctly. The low precision coefficient for s2 in the hybrid method, which is 0.361, is associated with about 38% of s1 states being incorrectly assigned as s2, without any incorrect classification of s3 states as s2. Such an outcome is satisfactory, especially when increasing the frequency of measurements is not significantly costly, and the costs of damaging the inspected element far exceed the costs of measurements. For example, if s2 were incorrectly recorded immediately after the P2 inspection when it was actually s1, according to the procedure, the frequency of P2 measurements would be increased (during the standard P1 measurement, which takes place every 2–3 days, an extended P2 measurement called “extended P1 inspection” would be carried out). In that case, the cost of inspections would be twice as high, and instead of 100 euros (cost of P1 per month), it would be 200 euros (cost of extended P1 per month). The cost of replacing the current collector component with a new one in the case of damage is at least 14 times higher than P1 and can range from 1400 euros to 7500 euros for the replacement of the entire pantograph. It should be noted that the aforementioned cost of the extended P1 inspection is the maximum cost that may occur. In the event that s1 is incorrectly classified as s2, for example, halfway between P2 inspections, this cost would only be higher by 50 euros, and in the discussed case, it would be 150 euros instead of 100 euros.

Undoubtedly, a better solution would be if all technical states were correctly classified. However, in the case of predicting the technical state of elements related to safety, it is better if a state of fitness is inaccurately classified as a state of incomplete fitness, and a state of incomplete fitness that can result in damage is fully correctly identified. In this way, safety can be significantly increased, which is particularly important in the context of transportation, rail vehicles, and many other areas where human safety may be compromised in the event of element damage.

4. Conclusions

The developed method enables identification of the technical conditions of technical objects (TOs). In the presented method, the key is to identify three technical conditions, not two as usual. Apart from the first technical condition, i.e., able to further use, and the third technical condition, i.e., not able to further use, it is important in the process of predicting the State 0 (often caused by damage) to determine the state of limited ability of further use (State 2). This condition can be determined based on archival data in technical inspections, where it is necessary to specify the last inspection during which the technical object was operational. Periodic technical inspections are used in the case of many technical objects, where, among others, we can distinguish vehicles and individual elements of various types of vehicles.

In response to the hypothesis presented in this article, based on the conducted research, it can be stated that combining two different AI methods can lead to achieving predictive results that more accurately correspond to real outcomes than using each predictive method separately. The hybrid method, i.e., combining ANN and decision trees, proved to yield better classification results for State 2 than each of the mentioned methods separately.

In the field of transport, and more precisely in supply chains, if any vehicle components are damaged, it can result in delays in the delivery of goods. The elimination of undesirable damage to the means of transport through the possibility of predicting the state of failure may increase the reliability of an entire supply chain. From the aspect of sustainability, the issue of reducing the number of failures also makes it possible to reduce supply chain disturbances, to reduce costs associated with delays, and to reduce the materials needed for the repair of the means of transport, since in this case the costs only relate to the replaced elements before their damage. Thus, it is impossible for more serious damage to occur. Often, failure of one item causes damage to others, which generates unnecessary costs and increases the amount of waste due to the number of damaged items.

The proposed predictive maintenance approach, which involves a hybrid method of technical condition prediction based on AI for reducing supply chain disruptions, aligns well with the principles of sustainability and contributes to the broader concept of sustainable development in supply chain management in several ways as follows:

• Resource Efficiency. By using AI to predict the technical condition of equipment and components, organizations can optimize maintenance activities. This reduces the consumption of spare parts, minimizes energy usage, and prolongs the lifespan of assets. These resource-efficient practices contribute to sustainability by conserving resources and reducing waste.
• Waste Reduction. Predictive maintenance minimizes unplanned breakdowns and equipment failures, reducing the need for emergency repairs and replacement parts. This leads to a reduction in the generation of waste, including discarded equipment and components, which aligns with intending to minimize environmental impact.
• Energy Savings. By ensuring that equipment is in optimal condition and operating efficiently, predictive maintenance reduces energy consumption. This leads to lower greenhouse gas emissions and contributes to efforts to combat climate change.
• Improved Supply Chain Resilience. Predictive maintenance enhances supply chain resilience by reducing the risk of disruptions caused by equipment failures. A more resilient supply chain is better equipped to handle unexpected challenges, such as natural disasters, and therefore, is more sustainable in the long run.
• Cost Savings. Predictive maintenance can lead to significant cost savings in terms of reduced maintenance expenses, fewer breakdown-related costs, and improved asset utilization. These financial benefits contribute to the economic pillar of sustainability.
• Transparency and Accountability. Implementing a data-driven predictive maintenance approach enhances transparency by providing visibility into equipment conditions and maintenance actions. This transparency can help organizations be more accountable for their sustainability goals and practices.
• Long-Term Planning. Predictive maintenance encourages long-term planning by considering the lifespan of equipment and components. This aligns with sustainable development principles, which emphasize the need to meet present needs without compromising the ability of future generations to meet their needs.
• Safety and Risk Mitigation. A well-maintained supply chain, thanks to predictive maintenance, is safer for employees, suppliers, and the public. Ensuring safety and mitigating risks is a key component of sustainable development.
• Circular Economy. Predictive maintenance supports the transition to a circular economy by extending the life of products and equipment. This reduces the need for extraction of new raw materials and promotes reuse and recycling.
• Customer Satisfaction. Sustainable practices, including reliable supply chain operations enabled by predictive maintenance, enhance customer satisfaction. Satisfied customers are more likely to support businesses committed to sustainability.

By applying the method proposed in this article, it would be possible to develop recommended maintenance activities for key elements related to the safety and reliability of transport. The combination of at least two artificial intelligence methods allowed us to achieve very good prediction results due to the possibility of individual adjustment of weights between the methods used. Such predictive maintenance methods can be successfully used to ensure sustainable development in a supply chain.

In summary, predictive maintenance with AI methods can significantly contribute to sustainability in a supply chain by reducing waste, conserving resources, minimizing disruptions, and improving the overall efficiency of vehicle components. It aligns with environmental and economic sustainability goals while ensuring the smooth operation of supply chain logistics. The conducted research shows that the hybrid method can obtain more accurate prediction results than each of the AI methods individually. The hybrid AI method can be successfully utilized as an element for reducing supply chain disruptions.