Deep Learning Techniques for Fault Diagnosis in Interconnected Systems: A Comprehensive Review and Future Directions

Abstract

As systems in industry become increasingly interconnected and sophisticated, the task of fault detection and diagnosis becomes significantly more difficult. Predictive maintenance, in conjunction with sophisticated multimodal learning methods, has been found to be an effective solution for tackling such challenges. Presently, data are collected across numerous sources, ranging from sensors and operational variables to environmental variables, making it vital to combine these heterogeneous data for effective diagnostics. Advanced learning methods like deep learning, transfer learning, and hybrid models are tailored to processing and aggregating such disparate streams of data, thereby leading to higher diagnostic accuracy. This leads to more efficient and reliable predictive maintenance methods. This paper provides a comprehensive review of how various learning methods are applied to fault diagnosis in interconnected systems, particularly in predictive maintenance. It examines different approaches that integrate data across domains, evaluating how each contributes to improved fault detection and enhanced system reliability. Additionally, it addresses emerging research areas, such as real-time fault detection, innovative data fusion processes, and the increasing application in power grids, manufacturing, and the automation sector. This paper serves as a valuable resource for both researchers and practitioners, emphasizing the significant potential of multimodal learning in advancing fault diagnosis and predictive maintenance within increasingly interconnected and complex systems.

1. Introduction

As industrial systems grow in complexity and become more interconnected, the challenge of identifying and diagnosing faults has intensified. Traditional diagnostic methods which rely heavily on fixed models or expert-driven rules often fall short in real-world environments, especially when confronted with dynamic conditions and unpredictable variables. This limitation has paved the way for more adaptable, data-driven approaches. Among the most promising of these are machine learning (ML) and deep learning (DL), which possess the unique ability to learn directly from data. These technologies can automatically detect and classify faults without needing explicitly defined models [1,2,3,4,5,6,7]. Machine learning encompasses a wide range of techniques for fault detection. One commonly employed approach is supervised learning, which proves especially effective when large datasets with clearly labeled fault examples are accessible. With these datasets, ML algorithms can identify patterns and correlations between system behaviors and specific types of faults. Some widely adopted supervised learning techniques in this context include support vector machines (SVMs), decision trees, and various forms of neural networks [4,5,6,7,8].

Not all fault detection scenarios, however, provide access to labeled data. In such cases, unsupervised learning methods become invaluable. Techniques such as clustering and anomaly detection enable systems to identify unusual behaviors or deviations from normal patterns offering valuable insights even when clear fault labels are absent [3,4,6,7].

Interconnected systems are made up of different parts that work together and share data or resources to make things run more smoothly, efficiently, and flexibly. For example, in smart homes, devices like thermostats, lighting, and security systems are all connected, communicating with each other to keep things comfortable and reduce energy consumption. In the human body, systems like the nervous and circulatory networks collaborate to keep everything functioning properly. In cities, smart infrastructure such as water, electricity, and transportation networks work together to manage resources in a more effective and sustainable way. These examples show just how important interconnected systems are in shaping and improving our daily lives [9,10,11].

Deep learning has emerged as a powerful tool in fault diagnosis. Architectures such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are particularly well suited to handling time-series data, which are common in industrial settings. These models are capable of detecting subtle, time-dependent patterns that traditional techniques might miss [4,5,6,7,8]. This makes them especially valuable in diagnosing faults that develop progressively over time [5,8].

Another promising advancement is reinforcement learning (RL), in which an agent learns to make decisions by interacting with its environment and receiving feedback. In the context of predictive maintenance, RL can be utilized to determine the optimal times for maintenance, predict failures in advance, and optimize system performance ultimately reducing downtime and enhancing efficiency [5,7,8].

Among all the applications of ML in fault diagnosis, predictive maintenance (PdM) stands out as particularly impactful. PdM focuses on forecasting equipment failures to schedule maintenance just in time avoiding unnecessary interventions and minimizing unexpected breakdowns. By using continuous streams of sensor data, ML models can identify early warning signs of failure, allowing for proactive, rather than reactive, maintenance strategies [1,2,6,7]. This strategy is particularly beneficial in industries such as manufacturing, energy, and transportation, where unexpected downtime can be incredibly expensive [2,3,6,7].

Despite advancements, challenges still persist. One key obstacle is the need for large, high-quality datasets. In practice, industrial data are often noisy, incomplete, or imbalanced, making it harder for ML models to perform effectively. Furthermore, many existing models struggle to generalize to new, previously unseen fault conditions, which limits their reliability in rapidly changing environments [3,4,6,7].

Several emerging techniques show promise in overcoming these limitations. Transfer learning, few-shot learning, and semi-supervised learning can help lessen the reliance on large labeled datasets by utilizing knowledge from related tasks or integrating limited supervision with unsupervised learning strategies [3,4,6,7]. By integrating information from diverse sources including sensor outputs, environmental data, and maintenance records fault diagnosis systems can achieve greater precision and reliability, enhancing decision-making in complex and unpredictable industrial environments [3,4,5,7].

Ultimately, machine learning is transforming the detection and diagnosis of faults in industrial systems. While data quality and model generalization challenges persist, the potential of ML technologies remains vast. As these tools evolve, they are expected to enhance fault diagnosis and enable more cost-effective and efficient predictive maintenance strategies. The influence of ML across industries worldwide is set to grow as these innovations continue to develop [6,7,8].

2. Multimodal Learning Techniques for Fault Diagnosis

2.1. Deep Learning Approaches

Deep learning has truly revolutionized fault diagnosis by providing powerful, scalable solutions for detecting faults in complex systems. What sets deep learning apart is its ability to automatically learn patterns directly from raw data, making it highly effective, especially when handling unstructured data such as images, sounds, and time-series signals. Traditional manual extraction of features can be labor-intensive and prone to errors, but deep learning navigates these challenges effortlessly [12].

In fault diagnosis, convolutional neural networks (CNNs) are frequently the preferred option due to their proficiency in processing spatial data, such as images from thermal cameras or visual inspections. What’s particularly remarkable is that CNNs need minimal preprocessing [12]. Conversely, recurrent neural networks (RNNs), particularly long short-term memory (LSTM) networks, excel at handling sequential data, such as sensor readings over time, models like You Only Look Once (YOLO) and Single Shot Detector (SSD) have proven to be highly effective in detecting structural defects, surface cracks, and other anomalies in industrial visual data [13,14]. These models effectively identify temporal patterns that may signal an impending fault [15]. A typical deep learning model used in fault diagnosis might look like this:

$$\widehat{y}=f(X;\theta )$$

(1)

In Equation (1), $\widehat{y}$ stands for the predicted fault type (or output class), X represents the input data (such as sensor readings or images), and $\theta$ refers to the model parameters, which are adjusted during the training process.

To train the model, an optimization algorithm such as stochastic gradient descent (SGD) is employed to minimize the loss function:

$$L\left(\theta \right)={\displaystyle \frac{1}{N}}\sum _{i=1}^N\ell (f(X_i;\theta ),y_i)$$

(2)

Equation (2) represents the loss function $L\left(\theta \right)$, where ℓ is the chosen loss metric (such as cross-entropy), $X_i$ is the input for the i-th training example, $y_i$ is the corresponding true label, and N is the total number of samples in the dataset. The objective is to minimize this loss over the parameter space.

Although deep learning has demonstrated remarkable success in fault diagnosis, it is not without challenges. One of the primary obstacles is the requirement for large, labeled datasets to train these models effectively, which can be challenging in industrial settings where faults are often infrequent. Additionally, deep learning models can be computationally demanding, necessitating powerful hardware for both training and real-time implementation.

For example, in agricultural systems, especially in fish farming, deep learning is exerting a great impact on how we monitor and maintain fish health. By connecting various sensors and data sources, these systems can provide real-time updates on things like water quality, fish behavior, and overall health. For example, deep learning models like CNNs can analyze images of fish to spot early signs of illness or stress, while other models, like RNNs, can track changes in behavior or environmental conditions over time. These tools also use data like temperature, pH levels, and oxygen content to predict potential health problems or optimize feeding. This interconnected approach not only boosts efficiency and sustainability in fish farming but also helps catch issues early, reducing the chances of disease outbreaks and improving overall management.

Looking to the future, several exciting directions exist for enhancing deep learning in fault diagnosis. A key area is improving the interpretability of these models, where explainable AI techniques come into play [16]. Additionally, there is significant potential in enhancing these models’ abilities to manage smaller datasets. Techniques such as transfer learning and data augmentation can significantly enhance the performance of deep learning models, even when labeled data are limited [17].

2.2. Transfer Learning

Transfer learning is a powerful technique that enables a model trained on one task to be applied to a different, yet related, task. This approach is especially beneficial in fault diagnosis, particularly when labeled data are scarce or costly [18]. Rather than starting from scratch, transfer learning harnesses the knowledge acquired from one task (the source domain) and applies it to another task (the target domain).

For instance, consider a model trained to detect faults in one type of machine. With transfer learning, this model can adapt to diagnose faults in a different type of machine with minimal additional training. Mathematically, this can be expressed as follows:

$${\widehat{y}}_{target}=f(X_{target};{\theta }_{source})$$

(3)

In this equation, $X_{target}$ represents the data from the target domain, while ${\theta }_{source}$ signifies the parameters learned from the pre-trained model on the source domain.

A widely used method in transfer learning is fine-tuning, which involves adapting a pre-trained model by retraining only the final layers of the network on new data while leaving the earlier layers unchanged. This greatly reduces the time and computational resources required to train the model [19,20].

The advantages of employing transfer learning in fault diagnosis are evident: it reduces the necessity for extensively labeled datasets and speeds up the implementation of fault detection systems [18]. Nevertheless, the success of transfer learning significantly hinges on the degree of similarity between the source and target domains. Transfer learning may not produce optimal results if the two domains are too dissimilar. Numerous successful applications of transfer learning in fault diagnosis have been documented. For instance, in wind turbine systems, models trained on one type of turbine have been successfully utilized to diagnose faults in other turbine types with comparable structural characteristics [21,22].

2.3. Hybrid Models and Ensemble Learning

Hybrid models provide a compelling strategy for fault diagnosis by combining various machine learning techniques to capitalize on their individual strengths. For example, integrating decision trees with neural networks or support vector machines (SVMs) can lead to more accurate and robust fault detection systems [23,24]. The core idea behind a hybrid model can be mathematically expressed as follows:

$$\widehat{y}={\mathrm{Model}}_1\left(X\right)\oplus {\mathrm{Model}}_2\left(X\right)$$

(4)

Here, ⊕ denotes the method for integrating the outputs of the models this could involve simple averaging, majority voting, or weighted summation, depending on the task requirements.

A specific and widely used type of hybrid approach is ensemble learning. This methodology focuses on combining multiple base models to create a stronger, more generalizable predictive system [25]. Common ensemble methods include techniques like bagging, boosting, and stacking. A typical ensemble model can be represented as follows:

$${\widehat{y}}_{ensemble}={\displaystyle \frac{1}{M}}\sum _{m=1}^M{\widehat{y}}_m$$

(5)

In this equation, M denotes the number of base models, and ${\widehat{y}}_m$ is the prediction output of the m-th model.

The primary strength of ensemble learning is its capacity to minimize the prediction errors that individual models may generate, thereby improving the reliability of fault diagnosis systems [26]. For instance, in applications related to fault diagnosis for rotating machinery, ensemble methods have shown superior classification performance by merging predictions from models like decision trees, random forests, and neural networks [27].

Case studies. Ensemble learning in industrial applications highlights its effectiveness. In power grid systems, for example, multiple machine learning models are used collectively to assess the health status of transmission lines [28]. In manufacturing environments, hybrid models are similarly utilized to predict equipment failures, facilitating preemptive maintenance actions, and minimizing operational downtime [29].

As illustrated in Figure 1, the ensemble learning model combines predictions from different classifiers, improving fault detection accuracy, particularly in noisy environments.

Hybrid models and ensemble learning techniques offer a robust and flexible framework for enhancing fault diagnosis accuracy. These approaches can effectively manage the complex, noisy, and variable data typically encountered in industrial systems by strategically combining various algorithms [26,30].

3. Data Fusion and Multimodal Integration

Data fusion and multimodal integration are essential components of fault diagnosis systems, particularly when dealing with various data sources such as sensor readings, images, and textual information. These techniques improve diagnostic accuracy by merging information from different modalities and addressing the inherent uncertainty in real-time environments [31,32].

Multimodal data fusion has gained significant traction across various industries, especially in healthcare, where it enhances predictive performance and provides deeper insights by integrating data from diverse sources. Fusion strategies are typically categorized as early, intermediate, or late fusion each classification depending on the stage at which data integration takes place.

Early fusion merges raw inputs, delivering simplicity but may face noise and high dimensionality challenges. In contrast, late fusion compiles outputs from separate models, promoting modularity yet risks neglecting cross-modal relationships. Intermediate fusion acts between the input and decision phases, enabling advanced feature-level integration that generally harmonizes the advantages and drawbacks of the other two techniques. This section delves into these fusion strategies in depth, focusing on their implementation in healthcare applications where integrating multimodal data is essential for delivering accurate diagnoses, effective treatment plans, and comprehensive patient care.

• Early fusion: This method integrates data from multiple modalities at the input level, feeding them into a single unified model. Inputs may consist of raw signals, handcrafted features, or deep features. Common techniques for early fusion include concatenation, element-wise summation, multiplication (Hadamard product), and bilinear pooling (Kronecker product). One major advantage of early fusion is the simplicity of working with a single model; however, it assumes that the model can effectively handle all modalities. Synchronization is also necessary, which may pose challenges when data are captured at different time points. In healthcare, early fusion is applied in tasks such as combining ultrasound imagery for breast cancer diagnosis or merging imaging data with electronic medical records (EMRs) for applications like skin lesion classification and cervical dysplasia prediction. It is also used to integrate genomics with histology or radiology data for cancer classification, survival prediction, and treatment response analysis.
  As shown in Figure 2, early fusion integrates multiple inputs at the raw level, allowing a unified model to process them jointly. This approach is often challenged due to data synchronization issues.

• Late fusion: Also referred to as decision-level fusion, this approach trains separate models for each modality and then combines their predictions. Methods such as averaging, majority voting, or Bayesian inference are commonly used. Since synchronization is not required, different architectures can be applied independently to each modality. This is especially beneficial in scenarios involving heterogeneous or incomplete data. Furthermore, new modalities can be added without retraining the entire model. Late fusion is well suited to cases where modalities are loosely related. In healthcare, it is used in applications like combining MRI data with PSA blood tests for prostate cancer detection or integrating genomics with histology data for survival prediction.
  Figure 3 illustrates the late fusion technique, where separate models process each modality independently and then combine their results into a final decision. This approach is particularly effective when the modalities vary greatly from one another.

• Intermediate fusion: Positioned between early and late fusion, intermediate fusion combines features at various abstraction levels, enhancing the model’s ability to learn cross-modal relationships. Unlike early or late fusion, intermediate fusion allows the loss function to influence feature extraction, enabling each modality to improve its own representations in a multimodal setting. Fusion may occur simultaneously or progressively, starting with strongly correlated modalities before incorporating less-related data. Guided fusion can refine this process by letting one modality influence feature extraction in another (e.g., using genomics data to guide histology feature selection). Applications include lung cancer detection through PET and CT fusion, prostate cancer classification with MRI and ultrasound, and multi-omics cancer subtyping for survival analysis.
  Figure 4 illustrates how intermediate fusion strikes a balance between early and late fusion by combining features during the learning process. This helps capture deeper relationships between different modalities.

To make the different multimodal learning approaches easier to understand,

Table 1. A summary comparing key techniques used in multimodal fault diagnosis. The table provides an overview of each method’s main features, advantages, and drawbacks, making it easier to choose the right approach for practical applications.

Technique	Key Features	Advantages	Drawbacks
Deep learning	Automatically extracts complex features from raw data using models like CNNs, RNNs, and YOLO.	Performs well with complex data (images, time series) and needs little feature engineering.	Needs large labeled datasets, can be computationally expensive, and lacks interpretability.
Transfer learning	Transfers knowledge from a source task to a related target task, requiring minimal retraining.	Works well with limited labeled data and reduces training time.	Performance drops if the source and target tasks are too different.
Hybrid models	Combines different machine learning models (e.g., SVM and Neural Networks).	Improves robustness and combines the strengths of different models.	Designing and tuning can be complex, and training time may increase.
Ensemble learning	Combines predictions from multiple models (e.g., bagging, boosting, stacking).	Helps reduce overfitting and improves accuracy.	Can be computa tionally expensive and less transparent.
Fusion strategies	Integrates data from multiple sources at different levels (input, features, or decisions).	Leads to richer representations and more accurate diagnostics.	Requires careful data synchronization and choosing the right fusion method.

offers a comparative summary. It outlines the key features, advantages, and drawbacks of techniques like deep learning, transfer learning, hybrid models, ensemble methods, and fusion strategies that are commonly used in fault diagnosis. This comparison acts as a helpful guide for both researchers and industry professionals when choosing the best methods for their specific needs.

To help clarify the different multimodal learning approaches, Table 1 provides a comparative summary. It highlights the main features, benefits, and limitations of techniques such as deep learning, transfer learning, hybrid models, ensemble methods, and fusion strategies commonly used in fault diagnosis. This comparison serves as a useful reference for both researchers and industry professionals, helping them choose the most suitable methods for their specific applications.

3.1. Data Fusion Techniques

3.1.1. Understanding Data Fusion in Multimodal Learning

Data fusion involves combining information from multiple sources to generate a more accurate and complete understanding of a system. In fault diagnosis, this often means aggregating inputs from sensors that monitor different physical phenomena such as temperature, vibration, pressure, and acoustics. By merging these diverse signals, one can create a more holistic view of system health [33].

Multimodal learning enables models to process and learn from different types of data, such as visual, auditory, or sensor-based inputs. Data fusion is a foundational aspect of this approach, facilitating the integration of these varied inputs into a single learning framework. This not only allows the model to uncover relationships between modalities but also leads to more accurate predictions [34].

Formally, the fusion process can be expressed as a function that combines data from m distinct modalities:

$${\mathbf{D}}_{\mathrm{fusion}}=f({\mathbf{D}}_1,{\mathbf{D}}_2,\cdots ,{\mathbf{D}}_m)$$

(6)

Here, $f(·)$ represents the fusion function, a weighted combination, statistical method, or a learned function optimized through machine learning.

3.1.2. Methods for Integrating Diverse Data Sources

The method used for fusing data depends on the nature of the data and the complexity of the task. The primary levels of data fusion include the following:

• Low-level fusion (sensor level): This approach integrates raw sensor outputs before any feature extraction. It is the most direct form of fusion and often uses simple mathematical operations. A typical formulation is a weighted sum:

  $${\mathbf{D}}_{\mathrm{fusion}}=w_1{\mathbf{D}}_1+w_2{\mathbf{D}}_2+\cdots +w_m{\mathbf{D}}_m$$

  (7)

  where $w_1,w_2,\cdots ,w_m$ are weights assigned based on sensor reliability or data quality [35].
• Mid-level fusion (feature level): In this method, features are first extracted from each modality, then combined into a single set before classification. This is effective when different modalities provide complementary insights. For instance, combining thermal and vibration features enhances diagnostic coverage. Principal component analysis (PCA) is a common technique here:

  $${\mathbf{F}}_{\mathrm{fusion}}=\mathrm{PCA}({\mathbf{F}}_1,{\mathbf{F}}_2,\cdots ,{\mathbf{F}}_m)$$

  (8)

  where ${\mathbf{F}}_i$ represents features from the i-th modality [36].
• High-level fusion (decision level): This method combines outputs from separate models trained on individual modalities. Techniques such as weighted voting or ensemble averaging are used to generate the final decision:

  $${\mathbf{D}}_{\mathrm{final}}=\sum _{i=1}^m{w}_i·{\mathrm{Decision}}_i$$

  (9)

  where ${\mathrm{Decision}}_i$ is the output from the i-th classifier, and $w_i$ is its corresponding weight [37].

3.1.3. Challenges in Real-Time Data Fusion

Real-time data fusion presents several challenges that must be addressed to ensure accurate and timely fault detection:

• Latency: Low latency is crucial for detecting faults in real-time systems. However, the fusion process, particularly when handling large datasets from various modalities, can create delays. Minimizing this latency is essential, and it can be accomplished by optimizing the fusion algorithms and utilizing high-performance computing techniques [38].

  $${\mathrm{Latency}}_{\mathrm{total}}={\mathrm{Latency}}_{\mathrm{sensor}}+{\mathrm{Latency}}_{\mathrm{fusion}}+{\mathrm{Latency}}_{\mathrm{diagnosis}}$$

  (10)
  Each term denotes the duration allocated for sensor data collection, the fusion process, and fault diagnosis.
• Data synchronization: Data from various sensors may not always be synchronized, complicating the fusion process. Time-stamping, interpolation, and synchronization algorithms are essential to align the data streams before fusion can occur.
• Noisy data: Sensor data are often noisy in real-world environments, which can impact the accuracy of the fusion process. Filtering or denoising techniques are essential to ensuring that the data remain clean and reliable [39].
• Scalability: As the number of sensors or modalities increases, the complexity of the fusion system also grows. Efficient fusion algorithms are necessary to manage large datasets and ensure real-time processing capabilities.

To make research more accessible and reproducible, there are several widely used benchmark datasets and open-source tools in the field of fault diagnosis. For example, the Case Western Reserve University (CWRU) Bearing dataset is often used to test models for diagnosing issues in rotating machinery. The NASA Prognostics and Health Management (PHM) Data Challenge datasets are also common, especially in studies focused on aerospace and industrial health monitoring. Additionally, there are other useful datasets like the MIMII dataset, which is designed for detecting acoustic faults, and the SEU vibration dataset, which is helpful for vibration-based fault detection. As for tools, open-source frameworks like ensorFlow 2.12, PyTorch 1.13, and Scikit-learn 1.1.3. are incredibly popular. These tools support the building, training, and deployment of machine learning and deep learning models. They come with ready-to-use modules and GPU acceleration, and they benefit from strong community support, all of which help researchers create reliable, scalable fault diagnosis systems more efficiently.

3.2. Multimodal Data Representation

3.2.1. Approaches for Representing Data from Different Modalities

Representing multimodal data involves integrating information from various sources or modalities into a unified format. This cohesive representation allows models to learn complex relationships and patterns across data types. Several common approaches are used to represent multimodal data [40]:

• Feature concatenation: This approach combines the features from different modalities into a single vector, which the model then processes. It is particularly effective when the features from various modalities are similar in size and scale.
• Canonical correlation analysis (CCA): CCA is a statistical technique that investigates the relationship between two sets of variables. It identifies highly correlated linear combinations of features from each modality [41]. This method is ideal when the different modalities provide complementary information, helping to align them effectively.

  $${\mathbf{X}}_1·{\mathbf{a}}_1={\mathbf{X}}_2·{\mathbf{a}}_2$$

  (11)

  where ${\mathbf{X}}_1$ and ${\mathbf{X}}_2$ are data matrices from the two modalities, and ${\mathbf{a}}_1$ and ${\mathbf{a}}_2$ are the corresponding weight vectors.
• Multimodal autoencoders: These specialized neural networks are designed to compress data from different modalities into a lower-dimensional space [42]. Each modality is processed via a separate encoder, and the encoded outputs are combined before decoding. This allows the model to learn a shared latent representation of the multimodal data.

  $${\mathbf{z}}_{\mathrm{fusion}}=f_{\mathrm{fusion}}({\mathbf{z}}_1,{\mathbf{z}}_2,\cdots ,{\mathbf{z}}_m)$$

  (12)

  where ${\mathbf{z}}_i$ represents the encoded features from modality i, and ${\mathbf{z}}_{\mathrm{fusion}}$ is the resulting fused representation.

3.2.2. Strategies for Handling Complex, Noisy, and High-Dimensional Data

Navigating complex, noisy, and high-dimensional data presents one of the primary challenges in multimodal learning [43]. Here are several techniques to tackle these challenges effectively:

• Dimensionality reduction: Techniques such as principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and autoencoders are commonly used to reduce the dimensionality of data. By decreasing the number of features, these methods simplify the fusion process and mitigate challenges like the curse of dimensionality, which can hinder model performance [40].

  $${\mathbf{X}}_{\mathrm{reduced}}=\mathrm{PCA}\left(\mathbf{X}\right)$$

  (13)

  where $\mathbf{X}$ represents the original high-dimensional data, and ${\mathbf{X}}_{\mathrm{reduced}}$ is the smaller, reduced feature set.
• Data augmentation: This includes techniques such as adding synthetic noise and generating new data points to enhance the robustness of the multimodal system. These methods enable the system to manage noise and variability in real-world data more effectively.
• Noise filtering: to enhance the quality of sensor data before it is fused, noise filtering methods such as Kalman filters, wavelet transforms, or deep learning-based denoising autoencoders can be utilized to clean the data.

4. Applications of Multimodal Learning for Fault Detection Diagnosis

In recent years, multimodal learning which integrates various types of data, such as vibration, temperature, acoustic, and visual information has emerged as a crucial approach in the field of fault detection and diagnosis (FDD) [44]. By leveraging the complementary strengths of different data modalities, multimodal techniques significantly enhance diagnostic accuracy and overall system performance [45]. This section highlights various practical applications of multimodal learning in FDD, focusing on industries such as manufacturing, renewable energy, and automotive systems.

4.1. Applications in Manufacturing Systems

Maintaining high operational efficiency while minimizing equipment downtime is crucial for productivity and cost-effectiveness in the manufacturing sector. Fault detection and diagnosis (FDD) systems play a key role by identifying issues early, enabling timely maintenance and reducing the risk of unexpected failures. Recent advances in machine learning (ML) and deep learning (DL) have significantly enhanced the effectiveness of FDD systems, making them indispensable in modern manufacturing environments.

A major application of FDD in manufacturing is the condition monitoring of critical machinery, such as motors, conveyors, and pumps. These components are essential for continuous production, and any malfunction can result in costly delays. Choudhary et al. [46] proposed the use of convolutional neural networks (CNNs) for motor fault detection. By analyzing both vibration and acoustic sensor data collected from motors, the model effectively identified faults such as imbalance, misalignment, and bearing damage. This enabled early interventions and the implementation of predictive maintenance strategies.

FDD is also essential for monitoring robotic arms and automated production lines, which are frequently used in operations such as assembly, welding, and material handling. Failures in these systems can interrupt workflows and jeopardize product quality. Wang et al. [47] employed recurrent neural networks (RNNs) to detect anomalies in the behavior of robotic arms. Their approach examined time-series data from joint angle sensors, force sensors, and torque measurements to identify unusual motion patterns that indicate mechanical degradation or sensor errors (See

Table 2. Summarizes the applications of machine learning techniques in automotive systems.

Application	Paper	Year	Technique Used
Vehicle Fault Detection	Zhang et al. [48]	2023	Convolutional Neural Networks (CNNs)
Automotive System Monitoring	Hussain et al. [49]	2024	Long Short-Term Memory (LSTM)
Engine Fault Diagnosis	Wang et al. [50]	2022	Support Vector Machines (SVMs)
Brake System Failure Detection	Hsu et al. [51]	2024	Random Forest (RF)
Vehicle Battery Health Monitoring	Zhang et al. [52]	2020	Decision Trees (DTs)
Transmission Fault Detection	Abed et al. [53]	2020	K-Nearest Neighbors (KNN)
Vehicle Condition Monitoring	Vasan et al. [54]	2022	Deep Neural Networks (DNNs)
Hybrid Vehicle Fault Diagnosis	Xie et al. [55]	2021	Multi-Layer Perceptron (MLP)
Fault Detection in Electric Vehicles	Vinothini et al. [56]	2020	Naive Bayes Classifier (NBC)
Electric Motor Fault Diagnosis	Akcan et al. [57]	2021	Extreme Learning Machine (ELM)
Engine Vibration Monitoring	Matthaiou et al. [58]	2020	Gaussian Process (GP)
Vehicle Powertrain Fault Detection	Ahmed et al. [59]	2021	Artificial Neural Networks (ANNs)
Hybrid Electric Vehicle Fault Diagnosis	Xiao et al. [60]	2022	Transfer Learning (TL)
Automotive Sensor Fault Detection	Liu et al. [61]	2020	Recurrent Neural Networks (RNNs)
Vehicle Vibration Diagnosis	Shen et al. [62]	2021	Deep Belief Networks (DBNs)
Autonomous Vehicle Fault Detection	Chen et al. [41]	2021	Multimodal Feature Fusion
Driver Behavior Prediction	Sun et al. [63]	2020	Reinforcement Learning (RL)
Vehicle Monitoring in Smart Cities	Band et al. [64]	2020	Hybrid Machine Learning Models
Fault Detection in Autonomous Cars	Chen et al. [65]	2022	Multi-Task Learning (MTL)
Battery Management System Faults	Lee et al. [66]	2020	Deep Convolutional Networks (DCNs)

).

4.2. Applications in Renewable Energy Systems

Renewable energy systems, such as wind turbines and photovoltaic (PV) solar panels, rely heavily on sensors to monitor performance and detect potential faults in real time. While these sensors continuously collect vast amounts of data, the fusion of data from different modalities greatly improves fault detection, particularly under the dynamic and unpredictable conditions in which these systems operate. Recent advancements in machine learning (ML) and deep learning (DL) have significantly accelerated the development of more accurate and responsive diagnostic systems. For instance, wind turbines can benefit from predictive maintenance models that analyze data from sensors measuring vibration, temperature, and wind speed to identify potential issues before failures occur.

Similar approaches are applied to PV systems, where environmental factors such as temperature and solar irradiance heavily influence performance. Data-driven models now enable the detection of issues like panel degradation, electrical faults, and the impact of environmental conditions on PV efficiency [67].

4.2.1. Wind Turbine Fault Diagnosis

Wind turbines are complex machines with numerous mechanical components, each of which can degrade or fail over time. Early fault detection is crucial for avoiding costly repairs and reducing downtime. Typically, wind turbines are equipped with sensors that monitor key parameters such as vibration, temperature, and acoustic emissions [68]. Analyzing these data sources enables early identification of mechanical issues before significant damage occurs.

Vibration sensors are especially effective at diagnosing misalignments or gear wear, whereas acoustic sensors can detect early signs of cracks or other structural faults [69]. By integrating data from multiple sensors, diagnostic systems obtain a more comprehensive view of turbine health, thus enhancing the accuracy and reliability of fault detection.

ML techniques are essential for processing sensor data and identifying fault patterns. Algorithms such as decision trees (DTs), support vector machines (SVMs), and deep neural networks (DNNs) classify fault types. For instance, DNNs excel at identifying patterns from vibration data, while long short-term memory (LSTM) networks track temporal changes that indicate gradual deterioration.

Recently, hybrid models that combine convolutional neural networks (CNNs) and LSTMs have demonstrated great promise. CNNs are well suited to extracting spatial features, like the frequency-domain characteristics of vibration signals, while LSTMs excel at modeling time-dependent patterns [41]. Together, these models provide a robust framework for high-accuracy fault diagnosis in wind turbines.

4.2.2. Photovoltaic (PV) System Fault Detection

Photovoltaic (PV) systems convert sunlight into electricity, but they are prone to faults such as shading, panel degradation, and inverter failures. Prompt fault detection ensures that the system operates at peak efficiency [70].

Sensors are positioned at key locations such as the solar panels, inverters, and grid connections to gather data on voltage, current, and temperature. These data are analyzed to identify anomalies like underperforming panels or malfunctioning inverters that could lead to energy losses.

Before feeding data into fault detection models, dimensionality reduction techniques such as principal component analysis (PCA) and kernel PCA (KPCA) are commonly applied. These methods minimize complexity while retaining crucial information that may indicate faults, particularly amid environmental noise like shading or temperature fluctuations [71].

Machine learning methods play a crucial role in photovoltaic (PV) fault detection. Convolutional neural networks (CNNs) are effective at identifying spatial anomalies in sensor data, such as irregular current distributions that indicate partial shading. Random forests (RFs) are also widely used for fault classification, providing robust and interpretable models for identifying issues based on sensor readings.

Hybrid models are becoming increasingly common as they combine the strengths of various algorithms. For instance, CNNs can extract features, while SVMs manage fault classification. This integration leads to more accurate, real-world-ready diagnostic systems.

4.2.3. Integrating Multiple Data Sources for Better Fault Detection

A key advantage of modern diagnostic systems is their ability to integrate data from multiple sensor types. In wind turbines, combining vibration, temperature, and acoustic data provides a comprehensive assessment of mechanical health. Similarly, in PV systems, integrating current, voltage, and temperature readings enables the more reliable detection of underperformance and hardware faults.

However, multimodal integration presents challenges. Sensors may collect data at different frequencies or formats, and some data streams may be missing or noisy. To address these challenges, advanced preprocessing techniques such as smoothing, filtering, or interpolation are employed to create a consistent, clean dataset suitable for reliable fault detection.

4.3. Applications in Automotive Systems

Automotive systems are becoming increasingly sophisticated with the rise of technologies such as autonomous driving and electric vehicles (EVs). Real-time monitoring and fault detection are essential to ensure these systems’ safety, reliability, and efficiency. This section examines how sensor data from modern vehicles can be integrated and analyzed to identify faults, enhance performance, and improve driver and passenger safety.

4.3.1. Autonomous Vehicle Fault Detection

Autonomous vehicles (AVs) depend on a range of sensors including cameras, radar, lidar, and ultrasonic sensors to navigate and understand their environment. These sensors produce large quantities of real-time data that must be processed accurately to ensure the safe operation of the vehicle. In this context, fault detection is crucial for identifying sensor failures or perception issues that could result in accidents.

One of the main challenges in AVs is the integration of multimodal sensors. For example, lidar offers 3D obstacle detection, while cameras provide visual information. Discrepancies between these data sources may indicate a fault in sensor calibration or perception logic [72].

Machine learning plays a central role in identifying such faults. Algorithms like decision trees (DTs), support vector machines (SVMs), and deep neural networks (DNNs) are trained on historical sensor data to detect anomalies [73]. For example, machine learning models can identify miscalibrated cameras or malfunctioning lidar sensors that inaccurately represent object positions.

Deep learning models, such as CNNs, excel at processing spatial information from camera feeds and can detect visual anomalies, including lane departure errors and object detection failures [41]. RNNs, especially LSTM networks, analyze temporal patterns to identify issues that develop gradually, such as system degradation or sensor drift [73].

4.3.2. Electric Vehicle (EV) Battery Health Monitoring

Battery management is critical in EVs, as battery performance directly affects range, reliability, and safety. Monitoring systems track parameters like voltage, current, and temperature to assess battery health. Detecting faults early such as overheating, overcharging, or imbalanced cells is essential for avoiding serious failures [74].

Sensor data from EV batteries allows systems to detect anomalies, such as rapid temperature increases or voltage drops, which may indicate thermal runaway or other critical issues [37]. By combining multiple sensor readings, fault detection becomes more precise and actionable. Machine learning models trained on historical data can predict when a battery is likely to fail, enabling predictive maintenance. These models analyze trends in temperature, charge cycles, and voltage patterns to forecast issues and suggest service interventions before failure occurs [75].

4.3.3. Vehicle Engine Fault Detection

Engine fault detection represents another crucial application area. Modern vehicles are outfitted with a range of sensors such as oxygen, pressure, and temperature sensors to monitor engine performance. These devices produce data regarding combustion, fuel efficiency, and emissions.

Through data analysis, machine learning models can identify issues such as misfires, clogged injectors, or emission control failures [76]. Algorithms such as DT, SVM, RF, and DNNs are used to classify fault types and aid maintenance decisions [77].

These models facilitate predictive maintenance scheduling, reducing breakdowns and ensuring vehicle reliability.

4.3.4. Integrating Multiple Data Sources for Automotive Fault Detection

Combining data from various vehicle subsystems such as the engine, brakes, battery, and ADAS sensors creates a more comprehensive and accurate fault detection framework. This multimodal approach improves the vehicle’s ability to detect faults early and ensures safe operation (See Table 3).

However, integrating diverse data streams presents challenges as well. Differences in data format, frequency, and completeness can introduce inconsistencies. Preprocessing techniques such as normalization, interpolation, and outlier removal are essential to guarantee high-quality inputs for diagnostic models [78]. The use of integrated sensor data and advanced machine learning techniques in automotive systems especially for autonomous vehicles (AVs) and electric vehicles (EVs) enables accurate and proactive fault detection. These systems enhance vehicle safety, reduce maintenance costs, and contribute to a more reliable and enjoyable driving experience. Continued development in hybrid and multimodal learning techniques promises even greater capabilities in the future.

Table 3. Presents multimodal learning applications in automotive contexts.

Application	Paper	Year	Technique Used
Autonomous Vehicle Fault Detection	Javed et al.	2020	MSALSTM-CNN: Hybrid Learning (CNN + LSTM) [79]
Electric Vehicle Battery Health Monitoring	Deng et al.	2022	Support Vector Machine (SVM) [80]
Engine Fault Diagnosis	Auran et al.	2024	Decision Trees (DT) [77]
Battery Monitoring for EVs	Sulaiman et al.	2024	Random Forest (RF) [81]
Autonomous Vehicle Sensor Fusion	Wang et al.	2022	Multimodal Learning (Fusion of Radar + Lidar) [82]
Predictive Maintenance for EV Batteries	Naresh et al.	2024	Deep Neural Networks (DNNs) [83]
Autonomous Vehicle Fault Detection	Safavi et al.	2021	Recurrent Neural Networks (RNNs) [84]
Engine Performance Monitoring	Okumucs et al.	2023	Gradient Boosting Machines (GBMs) [85]
Battery Fault Detection in EVs	Trivedi et al.	2022	Convolutional Neural Networks (CNNs) [86]
Vehicle System Health Monitoring	Rahman et al.	2022	Multilayer Perceptron (MLP) [87]
Autonomous Vehicle Navigation Faults	Jeong et al.	2023	Long Short-Term Memory (LSTM) [88]
EV Battery Fault Diagnosis	Shah et al.	2024	K-Nearest Neighbor (KNN) [89]
Vehicle Diagnostics in Autonomous Driving	Chen et al.	2020	Transfer Learning [90]
Electric Vehicle Fault Detection	Shen et al.	2024	Naive Bayes Classifier (NBC) [91]
Autonomous Vehicle Fault Classification	Kuutti et al.	2020	Deep Learning (DL) [92]
Engine Fault Classification	Zhao et al.	2022	Extreme Learning Machine (ELM) [93]
Battery Performance Degradation Detection	Valladares et al.	2022	Gaussian Process (GP) [94]
Vehicle Engine Health Monitoring	Fotias et al.	2021	Multiscale Learning (MSL) [95]
Vehicle Fault Detection using Sensor Data	Cui et al.	2022	Multi-Task Learning (MTL) [96]
Autonomous Vehicle Sensor Reliability	Anyanwu et al.	2023	Random Forest (RF) [97]

Table 3. Presents multimodal learning applications in automotive contexts.

Application	Paper	Year	Technique Used
Autonomous Vehicle Fault Detection	Javed et al.	2020	MSALSTM-CNN: Hybrid Learning (CNN + LSTM) [79]
Electric Vehicle Battery Health Monitoring	Deng et al.	2022	Support Vector Machine (SVM) [80]
Engine Fault Diagnosis	Auran et al.	2024	Decision Trees (DT) [77]
Battery Monitoring for EVs	Sulaiman et al.	2024	Random Forest (RF) [81]
Autonomous Vehicle Sensor Fusion	Wang et al.	2022	Multimodal Learning (Fusion of Radar + Lidar) [82]
Predictive Maintenance for EV Batteries	Naresh et al.	2024	Deep Neural Networks (DNNs) [83]
Autonomous Vehicle Fault Detection	Safavi et al.	2021	Recurrent Neural Networks (RNNs) [84]
Engine Performance Monitoring	Okumucs et al.	2023	Gradient Boosting Machines (GBMs) [85]
Battery Fault Detection in EVs	Trivedi et al.	2022	Convolutional Neural Networks (CNNs) [86]
Vehicle System Health Monitoring	Rahman et al.	2022	Multilayer Perceptron (MLP) [87]
Autonomous Vehicle Navigation Faults	Jeong et al.	2023	Long Short-Term Memory (LSTM) [88]
EV Battery Fault Diagnosis	Shah et al.	2024	K-Nearest Neighbor (KNN) [89]
Vehicle Diagnostics in Autonomous Driving	Chen et al.	2020	Transfer Learning [90]
Electric Vehicle Fault Detection	Shen et al.	2024	Naive Bayes Classifier (NBC) [91]
Autonomous Vehicle Fault Classification	Kuutti et al.	2020	Deep Learning (DL) [92]
Engine Fault Classification	Zhao et al.	2022	Extreme Learning Machine (ELM) [93]
Battery Performance Degradation Detection	Valladares et al.	2022	Gaussian Process (GP) [94]
Vehicle Engine Health Monitoring	Fotias et al.	2021	Multiscale Learning (MSL) [95]
Vehicle Fault Detection using Sensor Data	Cui et al.	2022	Multi-Task Learning (MTL) [96]
Autonomous Vehicle Sensor Reliability	Anyanwu et al.	2023	Random Forest (RF) [97]

4.4. Applications in Aerospace Systems

Aerospace systems including aircraft and spacecraft are highly complex and require continuous monitoring to ensure safe and efficient operation. Fault detection and diagnosis (FDD) are essential for maintaining system reliability, minimizing downtime, and preventing catastrophic failures. The adoption of advanced machine learning (ML) and deep learning (DL) techniques has significantly improved the accuracy and speed of fault detection in aerospace applications. These technologies facilitate the integration of data from multiple sensors embedded in engines, structural components, and subsystems, thereby enhancing the overall diagnostic process. One key area of fault detection in aerospace is monitoring aircraft engines. Since engines are critical to aircraft safety and performance, the early detection of faults is essential. Chen et al. (2021) utilized convolutional neural networks (CNNs) to detect faults in aircraft engines. Their model, trained on large-scale sensor datasets, demonstrated the ability to assess engine health in real time and identify anomalies before they escalate [41].

Another significant area is the monitoring of aircraft power systems, which supply electrical energy for essential flight operations. Faults in these systems can disrupt mission-critical functions. Das et al. (2020) used support vector machines (SVMs) to monitor key parameters such as voltage, current, and frequency, identifying patterns that indicate potential faults. Their model achieved high diagnostic accuracy and provided actionable alerts for maintenance teams [98].

Monitoring the structural integrity of aircraft is equally important. Components such as wings and the fuselage endure stress and fatigue over time. Dhanagopal et al. (2020) applied recurrent neural networks (RNNs) to structural health monitoring data, including strain and vibration signals. The model effectively detected small cracks and signs of corrosion, enabling timely maintenance [69].

Hydraulic systems are critical because they control landing gear, brakes, and flight surfaces. Khan et al. (2020) applied decision tree algorithms to detect hydraulic system faults, such as pressure loss and leakage. Their method provided interpretable results and facilitated in-flight decision-making [99].

Moreover, integrating sensor data enhances understanding of aerospace system health. Wang et al. (2022) explored transfer learning to improve fault detection across various aircraft platforms. By combining data from temperature, pressure, and vibration sensors, their model effectively recognized complex failure patterns and generalized across differing conditions [100].

In prognostics, machine learning is used to estimate the remaining useful life (RUL) of aerospace components. Liu et al. (2021) employed long short-term memory (LSTM) networks to predict RUL from sequential sensor data, allowing airlines to plan maintenance more effectively and reduce operational costs [101].

Fault detection and diagnosis (FDD) is particularly challenging for spacecraft due to extreme environmental conditions and limited maintenance opportunities. Zhang et al. (2022) developed a hybrid model combining convolutional neural networks (CNNs) and RNNs to diagnose faults in spacecraft electrical systems. Their model effectively identified common issues such as wiring faults and power irregularities [48].

Predictive maintenance is another essential application. Jamil et al. (2021) employed deep neural networks (DNNs) to analyze engine performance data, vibration patterns, and temperature variations. Their model successfully predicted maintenance needs, helping airlines minimize unexpected downtime [3].

Given the harsh and dynamic environments in which aerospace vehicles operate, real-time performance monitoring is crucial. Fentaye et al. (2020) applied multilayer perceptron (MLP) networks to monitor flight parameters such as speed, altitude, and attitude. Their model effectively identified anomalies during complex flight scenarios [102]. Additionally, multi-task learning (MTL) is gaining traction in this field due to its ability to handle multiple FDD objectives simultaneously.

4.5. Trends in Multimodal Learning for Fault Diagnosis

The emergence of sensor-rich environments and advanced computing platforms has established multimodal learning as a fundamental aspect of contemporary FDD systems. Several trends are set to influence future developments in this domain.

4.5.1. Integration of Advanced Deep Learning Models

Advanced architectures such as transformer networks, attention mechanisms, and graph neural networks allow models to learn more complex relationships among various data types.

$$\widehat{y}=\mathrm{Transformer}(x_1,x_2,\dots x_n)$$

(14)

where $x_1,x_2,\dots x_n$ are different modalities (e.g., vibration, temperature, and acoustic), and $\widehat{y}$ is the fault prediction.

4.5.2. End-to-End Learning Systems

A shift is underway toward systems that process raw sensor data directly, eliminating the need for handcrafted features and manual preprocessing.

$$f_{\theta }\left(x\right)=\mathrm{DeepModel}\left(x\right)$$

(15)

where $f_{\theta }\left(x\right)$ is the model’s learned output from raw input x, with $\theta$ as the set of parameters.

4.5.3. Multiscale and Multi-Resolution Approaches

Analyzing data at multiple temporal and spatial scales helps capture faults that evolve over time or occur in different parts of a system.

$$\mathcal{L}=\sum _{i=1}^n{\mathcal{L}}_i\left(f_{{\theta }_i}\left(x_i\right)\right)$$

(16)

where each model $f_{{\theta }_i}$ works at a different resolution, and ${\mathcal{L}}_i$ is the corresponding loss.

4.6. Challenges in Scaling Up to Large, Real-Time Systems

Although multimodal systems perform well on smaller scales, real-time industrial applications pose significant challenges.

4.6.1. Data Volume and Storage

High-resolution sensor data across modalities can overwhelm storage and compu ting resources.

$$D=\{d_1,d_2,\cdots ,d_n\}$$

(17)

where $d_1,d_2,\cdots ,d_n$ are the raw sensor data streams, and D is the integrated dataset.

4.6.2. Real-Time Data Processing

To ensure timely fault diagnosis, low-latency models are essential especially in safety-critical domains like aerospace.

$$\widehat{y}=\mathrm{FastModel}\left(x\right)$$

(18)

where $\widehat{y}$ is the fast-inference result.

4.6.3. Sensor Fusion and Synchronization

Sensors often collect data asynchronously or with noise. Synchronization is crucial for valid data fusion.

$$D\left(t\right)=\mathrm{Fusion}(d_1\left(t\right),d_2\left(t\right),\cdots ,d_n\left(t\right))$$

(19)

where $D\left(t\right)$ is the synchronized data snapshot at time t.

4.7. Promising Research Directions

4.7.1. Uncertainty Handling in Fault Diagnosis

Robust fault detection requires accounting for uncertainties caused by sensor noise or incomplete data.

$$\widehat{y}=arg\underset{y}{max}P\left(y\right|x)$$

(20)

Bayesian neural networks and ensemble methods help quantify this uncertainty:

$$\widehat{y}={\displaystyle \frac{1}{M}}\sum _{m=1}^M{\widehat{y}}_m$$

(21)

4.7.2. Edge Computing for Fault Diagnosis

Edge AI enables localized, fast diagnosis near the data source, minimizing latency.

$${\widehat{y}}_{\mathrm{edge}}=\mathrm{EdgeModel}\left(x\right)$$

(22)

4.7.3. Real-Time Data Fusion

As real-time demands grow, new models must perform data fusion incrementally with minimal delay.

$$\widehat{y}\left(t\right)=\mathrm{FusionModel}(d_1\left(t\right),d_2\left(t\right),\cdots ,d_n\left(t\right))$$

(23)

4.8. Enhancing Data Preprocessing in Multimodal Systems

• Improving Data Preprocessing with Multiscale Representation
  Raw data are often noisy and autocorrelated. Multiscale representation using high-pass and low-pass filters can isolate noise and highlight relevant features, improving fault detection accuracy.
• Addressing Uncertainty with Interval-Valued Representation and Dimensionality Reduction
  Interval-valued techniques account for environmental uncertainty. Combined with dimensionality reduction (e.g., Euclidean distance-based filtering), they enhance robustness and reduce computational load.
• Simplifying Multimodel Design
  Multimodel systems benefit from domain knowledge and data-driven optimization. Methods like PSO and GA can optimize parameters such as hidden layers and activation functions to reduce complexity and improve adaptability.
• Improving Decision-Making in Multimodel Systems
  Combining techniques like PCA, KPCA, and Fourier analysis in a unified framework enhances interpretability and decision-making in noisy or rapidly changing environments.
• Building Enhanced Multimodel Systems with Multiple Learners
  Integrated learning systems that combine multiple algorithms can handle temporal dependencies more effectively. Dynamic kernel PCA and ensemble learners improve classification and prediction speed while ensuring robustness.

5. Conclusions

The integration of multimodal learning techniques into fault detection and diagnosis (FDD) has emerged as a transformative solution for addressing the increasing complexity of modern industrial systems. As industries become more interconnected and generate vast quantities of data from diverse sources, effective data fusion and predictive maintenance strategies are critical for ensuring operational reliability and extending system lifespan.

This paper has underscored the vital role of multimodal learning encompassing deep learning, transfer learning, and hybrid models in enhancing diagnostic accuracy. By integrating data across multiple domains, these methods significantly improve the precision of fault detection and support the development of more reliable, proactive maintenance solutions. The review examined a wide array of methodologies, assessing their respective strengths and identifying potential limitations when applied in real-world scenarios.

Looking ahead, the future of multimodal FDD research is promising. Key areas for advancement include enhancing real-time diagnostic capabilities, streamlining data fusion techniques, and broadening the adoption of these systems across various industrial sectors, such as power grids, smart manufacturing, and automated operations.

The findings and insights presented in this review serve as a valuable resource for researchers and practitioners. They offer strategic guidance for ongoing and future efforts aimed at strengthening system resilience through advanced diagnostic frameworks. As industrial systems evolve in scale and complexity, the role of multimodal learning will become increasingly essential, solidifying its place as a cornerstone technology for the future of fault detection and predictive maintenance.