Most automobile repair and maintenance apps utilize the rule base and case base reasoning methodologies in

their implementations. These two methodologies have their strengths and limitations, some of these limitations

can be overcome by the Convolutional Neural Networks (CNN), a specialized subset of deep learning that has

with an accuracy of 92%, Precision of 91% and Recall of 90%. It is recommended that an integration of case-

based reasoning, Fuzzy logic reasoning and CNN be undertaken to improve on the results and a high resolution

camera be used to capture the images of the damaged parts for a better input to the CNN model.

Keywords: CNN, Automobile, Maintenance, Repair. Akwa Ibom.

Maintenance of an automobile requires (i) the identification of the problem or faulty part of the vehicle, (ii)

adjustment of the vehicle mechanism that may be responsible for the fault including cleaning, soldering,

lubrication and (iii) replacement of the faulty part with a new functional part. All these require the observations

and monitoring of the performance of vehicle while in motion. The observation and monitoring involve senses

of sighting, touching, smelling, hearing and perception. The knowledge of the type of sound, smell, behavior of

of the faulty part (Obot and Obike, 2025). This can be realized through a deep learning model, the Convolutional

Neural Networks (CNN). It has been used in many applications successfully to detect and extract cardinal

features of an object for classification with little or no human supervision. They are suitable for image, audio,

video and text-based datasets.

In this study, the damaged or faulty parts of a vehicle of is captured and represented as an image dataset whose

nodes are represented as pixel (picture element) and back propagation is applied in automatic and adaptive

learning. With a data repository of 998 cases including images, CNN are trained to refine solutions based on

similarities. The study aims at classifying and detecting a fault in a vehicle that helps its users diagnose a

malfunctioned vehicle to assist them in maintenance and repair of the vehicle. The specific objectives include

to; (i) acquire the images of faulty parts of vehicles from a workshop operated by AKTC (ii) use different variants

The automation of automobile maintenance and repair has been implemented in various approaches, including

rule-based expert systems, neural networks, fuzzy logic, case-based reasoning (CBR), and hybrid AI

frameworks. According to Ucar et al., (2024), the sequence of steps taken to undergo an AI-based predictive

maintenance include; sensing, data pre-processing, algorithms, modeling, communication, integration, user

interfacing, and reporting. Fernades et al. (2023), lists diagnosis of faults in automobile to include faults detection

and identification. Alkoty et al., (2018) employed the methodology of developing an expert system to design a

system that diagnose and fix regular car breakdowns. The system is used to train students of industrial technical

education the procedures and processes of repairing and maintenance of cars that have regular reoccurring faults.

Sandoval-Pillajo et al. (2019) also developed the same methodology to design automobile repair system.

Rahman et al. (2018) implemented an expert system that uses CBR to determine solutions to mechanical failures

CBR's main challenges and recommend combining CBR with other AI methods to enhance diagnosis in

intricately layered technical systems. These improvements would be of great use in the automotive systems since

their failure mechanisms are multifaceted and contextually dependent. In the analysis of vehicle and machinery

malfunctions, CBR has been used to pull fault cases and recommend fixes. Chen et al. (2022) highlighted the

use of CBR in automated fault diagnosis systems by classifying intricate fault patterns. This involved the

construction of semantic case matching, coupled with an entropy-based approach to attribute weighting. Other

studies, such as Zeng et al. (2025), explored the use of hybrid methods in equipment health management that

integrate fuzzy association rules and CBR. In these studies, fuzzy logic is used to understand uncertain

relationships between features, while CBR retrieves historical fault cases and recommends maintenance actions.

Obot and Obike (2024) also integrated CBR with fuzzy logic to develop an automobile maintenance and repair

hard threshold rules (Saatchi, 2024). In (Rojek et al. 2023), 533 cases of input data are used to demonstrate how

fuzzy logic could be used to reduce the problem of ambiguity of information in automobile repairs and

maintenance. Recent automotive fault detection systems increasingly leverage fuzzy logic to reflect real-world

ambiguities in sensor readings and maintenance heuristics. A fuzzy logic-based automobile fault detection

model by Kizito et al. (2024) used the Mamdani algorithm, which demonstrated improved detection accuracy

compared to crisp rule systems, highlighting fuzzy logic’s suitability for handling uncertain operational

conditions in vehicles. Akazue et al. (2024), developed an intelligent fuzzy logic system for vehicle diagnosis

with symptom-based rules, yielding acceptable accuracy and precision metrics for initial fault classification,

although system generalizability remains constrained by the underlying rule base. Broader surveys of fault

diagnosis methods in vehicle systems suggest that fuzzy and neuro-fuzzy methods have significant strengths in

In the real life scenario, it is important that the repairer observes, inspects and feel the impact of a damaged

vehicle as part of the diagnostic procedures. This helps them to assess the extent of damage to be able to proffer

a solution to the problem. Rule base, case base, fuzzy logic and the other techniques discussed so far lack the

successful, especially when large datasets are present (Jia & Li, 2023). This is primarily due to the ability of

CNNs to automatically capture features in a hierarchical way, which is how they function. Brito (2023), used

CNNs to recognize damaged parts of vehicles and integrate these models into inspection applications, focusing

on exact pattern recognition for visible physical damage. Expanded deep learning research shows that CNN

architectures can significantly outperform traditional methods in automotive fault classification tasks, achieving

higher accuracy and lower false positives in sensor and imaging applications, which is critical for automated

maintenance systems that rely on consistent diagnostic outputs (Siddique et al., 2025). Adelusi and Mike (2024)

developed a CNN-based automatic damage system and asses the performance with the traditional machine

learning methods. A dataset of images of vehicle with different types of severities of damages captured under

different conditions such as lighting, angle and weather. However, CNN not detect fine-grained severity of

approaches perform better than rule-based and statistical-based approaches by 15-25% accuracy. Min et al.

(2025) carried out a non-invasive diagnosis of faults in electric motors via CNN for fault classification and

compared the results obtained with the conventional methods. CNN approach achieved 99.76% accuracy.

Chen et al. (2025), the survey to examine the recent advances in predicting methods for vehicle service parts and

predictive maintenance, different models including SVM, LSTM and CNN were considered in the review using

statistical, data-driven, digital-twin and stochastic approaches. Results show that significant progress is made in

data-driven and digital-twin approaches. Van Ruitenbeck and Bhulai (2022) developed a damage detection

model to locate and classify learning algorithms to train 10,000 damage images. The system was able to

accurately detect small damages with various conditions and the performance evaluation carried out show an

appreciable comparison with that of domain experts. Hybrid deep learning models that combine CNNs with

fusion of architectures that utilise deep learning for the extraction of features and fuzzy inference coupled with

automation problem solving and employment of the methods for automated vehicle systems operations. It is in

such studies that the handling of uncertainty coupled with exemplary results for classifying the performances,

was evident. The integration and contemporary advancements in intelligent diagnostic systems emphasize the

use of hybrid models that synergize multiple artificial intelligence (AI) approaches. The recent literature

concerning AI-based vehicle fault diagnosis shows that the fusion of machine learning, deep learning, fuzzy

logic, and reasoning systems yields more precise, holistic, and scalable diagnostics, especially for the impending

generation of autonomous and connected vehicles (Hossain et al., 2024). These fusions and integrations come

From the reviewed studies, it is evident that traditional rule-based and fuzzy logic approaches are limited in their

ability to process visual fault data effectively. Although deep learning approaches have demonstrated strong

performance in image classification tasks, many studies rely on simulated or publicly available datasets. This

study addresses this limitation by utilizing real-world automobile fault images collected directly from workshop

To develop a robust dataset for automobile fault detection, images of faulty vehicle components were collected

directly from automobile repair workshops. The data collection was carried out in collaboration with experienced

mechanics who assisted in identifying defective parts during routine maintenance and repair activities. The goal

These classes were selected because they represent frequently occurring faults observed during workshop

maintenance operations. Figure 2 shows the distribution of collected workshop images across different fault

categories.

The fully connected layers perform the final classification of the extracted features into the predefined fault

categories. A softmax activation function is used in the output layer to generate probability scores for each fault

class. The model architecture is presented in Table 4.

The architecture of the proposed convolutional neural network for automobile fault detection showing image

acquisition, preprocessing, hierarchical feature extraction using convolution and pooling layers, and final fault

Output: Predicted Fault Category

1. Load labeled workshop image dataset

2. Preprocess images (resize, normalize)

3. Split dataset into training, validation, and testing sets

4. Initialize CNN model parameters

5. For each training epoch:

 Feed training images into CNN

 Extract features using convolution layers

 Apply pooling layers to reduce feature dimensions

 Pass features through fully connected layer

 Compute classification probabilities using Softmax

During training, the CNN model learns to associate specific visual patterns with particular fault categories. The

training process involves feeding labeled images into the network and adjusting the network weights using

and Python-based deep learning libraries implemented using TensorFlow/Keras framework.

dataset. The trained model demonstrated the ability to accurately classify images of faulty vehicle components

The experimental results show that the convolutional neural network successfully learned visual features

associated with different types of vehicle faults. For instance, the model was able to distinguish between radiator

fan damage and broken fan belts based on the structural patterns observed in the images. Similarly, Figure 4

shows spark plug carbon deposits that were accurately identified by the CNN model due to their distinctive

visual characteristics.

29. Ucar, A., Karakose, M., Kırımça, N. (2024), Artificial Intelligence for Predictive Maintenance

30. Yan, A., & Cheng, Z. (2024). A Review of the Development and Future Challenges of Case-Based