A Scalable Fog Computing Solution for Industrial Predictive Maintenance and Customization

1. Introduction

The rapid proliferation of devices and data-intensive applications in the Internet of Things (IoT) presents significant challenges for industries in managing and processing large amounts of data [1]. As the volume of data generated at the edge increases, traditional cloud computing solutions face limitations, particularly latency. Cloud computing typically involves sending data to centralized servers far from the edge, which introduces delays due to the distance between devices and cloud data centers [2,3,4].

Figure 1. Fog computing architecture [5].

Figure 1. Fog computing architecture [5].

Edge computing aims to address this by processing data directly at or near the source of data generation, either on the devices themselves or at local edge nodes, minimizing latency and reducing the load on centralized cloud systems. However, while edge computing improves responsiveness by processing data locally, it can be limited in its ability to manage large-scale data and complex tasks due to constraints in processing power, storage, and connectivity [6].

Fog computing builds on edge computing by creating an intermediate layer between edge devices and the cloud [7]. Fog nodes, placed closer to the data sources, process and store data locally while maintaining the flexibility and scalability of the cloud [8]. This hybrid architecture improves the performance of real-time data processing and reduces latency more effectively than cloud-based solutions alone [9]. Moreover, by managing data processing closer to the edge and at multiple points in between, fog computing reduces the need to transmit large volumes of data to the cloud, thus improving privacy and minimizing network congestion [10]. The disadvantages of fog computing were carefully evaluated compared to edge and cloud computing [11]. It was concluded that, for this specific application, fog computing offered the most suitable solution. This choice was driven by the need to manage a high number of sensors, which would generate an excessive data flow for cloud computing, and edge computing was not feasible due to the limited space available within the capping machine.

Although fog computing is particularly beneficial for industrial environments, where real-time data processing and low latency are crucial, its applications extend to nonindustrial sectors such as smart cities [12], transports [13], and environmental or health monitoring [14]. These systems benefit from reduced reliance on centralized cloud services, which improves efficiency and responsiveness [15]. However, a key limitation of current fog solutions is that they often need to be customized to the specific needs of industrial applications, where high levels of customization are required to integrate a wide range of devices, sensors, and predictive models [16]. The need for more flexibility in existing systems is one of the central issues that this paper aims to address.

The flexibility of these systems, while offering benefits such as improved resource optimization and adaptability [17], also increases vulnerability to cyber threats [18]. Distributed Denial of Service (DDoS) attacks [19] and data breaches [20] could be some examples. DDoS attacks can overwhelm network resources, causing service disruptions or delays in real-time data collection and anomaly detection [21]. In contrast, data breaches can lead to unauthorized access to sensitive information, compromising operational integrity and regulatory compliance [22].

To address these evolving risks, IIoT systems must be designed with adaptable security protocols. The customization capabilities of these systems provide an opportunity to integrate tailored defense mechanisms that help mitigate specific threats. For example, encryption techniques (such as AES-256 [23]) and multifactor authentication (MFA) [24] can be employed to secure data both at rest and in transit, while traffic monitoring and anomaly detection algorithms can identify and mitigate DDoS attempts in real-time [25].

By aligning security measures with operational needs, IIoT systems can provide reliable predictive maintenance and remain resilient to potential interruptions. This ability to integrate robust security at every system layer - especially within sandbox environments that allow for external customization - ensures the platform can withstand known and emerging threats, ultimately protecting critical industrial operations.

This work introduces Concentrator, a fog computing platform for industrial environments with numerous interconnected nodes. Through Docker-based containerization, Concentrator offers extensive customization, enabling the integration of various sensors, algorithms, and real-time data processing capabilities [26]. Docker allows isolated containers to operate independently on the same kernel, enabling efficient management of resources without interference unless specifically allowed [27]. This modular approach is advantageous in predictive maintenance scenarios, where timely data analysis and interventions are critical for preventing equipment failures [28].

AROL Closure Systems, a manufacturer of capping machines, collaborated to demonstrate the system, using Long Short-Term Memory (LSTM) models for predictive maintenance. By predicting temperature anomalies and scheduling maintenance before equipment failures occur, the system helps optimize machine uptime [29]. LSTM models excel in forecasting time-series data by capturing long-term dependencies, offering advantages over traditional predictive models [30,31,32].

The main contributions of this paper are:

• The development and deployment of Concentrator in a real-world industrial setting.
• Customization capabilities for integrating specific sensors and algorithms.
• The application of LSTM models to enhance predictive maintenance and reduce downtime.

The paper is structured as follows: Section 2 reviews related work on fog computing and predictive maintenance. Section 3 outlines the design and implementation of the Concentrator system, with an emphasis on its customization and machine learning features. Finally, Section 4 discusses the real-world deployment of the system and its performance in a production environment.

2. State of Art

This work primarily revolves around two key aspects: device customization and predictive maintenance functionalities, all developed through a fog computing system.

Authors often pursue device customization by laying the groundwork for diverse application scenarios. For instance, Isaac Lera et al. [33] developed YAFS, a dynamic fog computing system enabling edge components to select the nearest and most secure link with various nodes, ensuring stable connections. Despite its utility for highly dynamic systems, YAFS does not add anything about the user’s application, even though it can be useful for highly dynamic systems.

Cecil Wöbker et al. [34] present Fogernetes, a framework aimed at streamlining the deployment and management of fog computing applications. Fogernetes offers tailored features for fog environments, including edge node discovery, resource-aware scheduling, and dynamic adaptation to network conditions. While the authors successfully demonstrate the system’s effectiveness, its adaptability is contingent upon the network; customers have limited scope for additional customization.

D. D’Alessandro et al. describe a similar project in [35], exploring mass customization in IoT wireless sensor systems. This study focuses on the need for flexible IoT solutions that adapt to different application requirements and environments. The authors present a modular approach to designing and developing IoT sensor systems, which allows for flexibility and scalability during deployment. Integrating modular components like sensors, communication modules, and processing units allows the system to be customized to meet specific user needs without requiring major redesigns or reconfigurations.

Unlike traditional methods that rely on hubs and separate radio modules, the Concentrator emphasizes embedding custom firmware directly into embedded systems used in production lines. This streamlined approach improves efficiency and reduces the need for additional hardware components unless required by the client. Additionally, by using Docker and containerization, as previously mentioned, it is possible to create adaptable environments tailored to each customer’s needs. This method provides exceptional flexibility and customization options, making integrating into various environments and applications accessible. This solution also adapts for use in microservices and IoT scenarios on embedded systems, such as a Raspberry Pi; [36] presents an example of its usage, making the solution easily deployable in production lines.

Shifting the focus to predictive maintenance, [38] illustrates the critical significance of these methods for industries aiming to minimize downtime, cut costs, prolong equipment lifespan, enhance operational efficiency, and support safety measures. The authors employ charts to assess the Return on Investment (ROI) based on commonly used variables across various industries. However, they highlight a significant challenge: the indispensable requirement for high-quality data to ensure dependable predictions. They emphasize the crucial role of meticulous data preparation and cleaning in guaranteeing the reliability of predictive maintenance systems. These results underscore the importance of data quality in deriving actionable insights and maximizing the effectiveness of predictive maintenance strategies.

Both [37] and [39] explore various machine learning methods and demonstrate their effectiveness in performing machine predictive maintenance. However, their focus primarily centers on classification techniques, including Support Vector Machine, Logistic Regression, K Nearest Neighbors, Random Forest, Gaussian Naive Bayes, Decision Tree, and Neural Network. Based on this study, researchers have considered the LSTM the optimal solution for predicting values from temporal sequences.

Another example of the usage of AI alongside an embedded system is [40], in which F. Laganà presents an integrated system for surface electromyography (sEMG) signal acquisition, processing, and analysis developed using artificial intelligence (AI) techniques. This study addresses the challenge of analyzing noisy and complex sEMG signals by leveraging the computational capabilities of the Raspberry Pi. The author demonstrates how this affordable platform can process real-time data while running machine-learning algorithms for classifying muscle activity. The system’s modular design allows easy integration with various applications, such as myoelectric prostheses and robotic control systems, while maintaining flexibility and scalability.

As demonstrated in the following studies, researchers have developed machine learning and predictive maintenance on a fog computing system.

As discussed in [41], the authors present a comprehensive survey on deploying machine learning (ML) on edge devices and cloud networks. The paper explores various IoT devices, including Raspberry Pi, NVIDIA Jetson, and Arduino Nano 33 BLE Sense, that integrate edge intelligence for object detection and gesture recognition tasks. The study also analyzes 1000 recent publications on "ML in IoT" to identify emerging topics and application domains, highlighting challenges such as security, privacy, and energy limitations in edge device implementation. The survey provides an in-depth look at traditional ML and deep learning methods in IoT contexts, showcasing their potential in diverse application domains, including healthcare, agriculture, and smart cities.

In [42], the authors explore the use of edge computing (EC) for anomaly detection in Internet of Things (IoT) environments, focusing on the integration of machine learning (ML) and deep learning (DL) models with microcontroller units (MCUs). The paper provides a systematic literature mapping of 18 studies published between 2021 and 2023, addressing the benefits and challenges of anomaly detection using TinyML on MCUs. The survey investigates different ML/DL algorithms, validation metrics, data estimation techniques, and hardware/software configurations while proposing a taxonomy of algorithms used in TinyML applications. The findings demonstrate how researchers can apply ML effectively in EC environments to ensure accurate data and improve the reliability of IoT systems despite the unique challenges posed by IoT devices’ resource constraints.

The paper [43] presents a detailed analysis of using reinforcement learning (RL) and deep reinforcement learning (DRL) for offloading tasks in fog computing environments. The paper highlights how RL and DRL can be applied to manage the offloading of computationally intensive tasks from resource-constrained IoT devices to fog or cloud layers to reduce latency and improve system performance. The authors discuss various offloading mechanisms, categorizing them into value-based, policy-based, and hybrid-based algorithms, and compare them based on offloading modes, performance metrics, and evaluation tools. The study also identifies open research areas and future directions, providing insights into how DRL can help overcome the challenges of large-scale networks and complex state-action spaces.

An example of actual applications is [44], in which the authors propose a novel fog data prediction and recovery (FDPR) algorithm. This algorithm leverages deep learning techniques, specifically a deep concatenated multilayer perceptron (DC-MLP), to predict and recover missing sensor data in a fog computing layer of an IoT network. By efficiently addressing data prediction and recovery, the proposed algorithm achieves high accuracy (99.89%) and significantly improves IoT device lifetime (121%) while maintaining low overhead. The evaluation, conducted through simulations and experimental work with nine-edge devices, demonstrates its effectiveness in real-world scenarios.

Another example is [45], where a Fog-enabled Air Quality Monitoring and Prediction (FAQMP) system is proposed to address challenges in air quality monitoring and forecasting in smart cities. The system integrates IoT, Fog Computing (FC), Low-Power Wide-Area Networks (LPWANs), and Deep Learning (DL) to improve the accuracy and efficiency of air quality predictions. A lightweight DL-based Seq2Seq Gated Recurrent Unit (GRU) attention model, optimized through post-training quantization, is employed for multi-step forecasting. The model demonstrated superior performance with reduced computational load, making it suitable for deployment on resource-constrained fog nodes. The FAQMP system provides real-time air quality updates, forecasts, and alerts through the EnviroWeb application, assisting governments in maintaining air quality standards and fostering a healthier environment.

A novel decentralized air quality prediction and event detection system, DeepFogAQ, is proposed in [46] to tackle the issue of air pollution in future cities. The system predicts pollutant concentrations and detects environmental events by integrating deep learning models, fog computing, and complex event processing. The architecture utilizes various models, such as Transformers, CNN-LSTM, and GRU, to ensure accurate predictions. Experimental results highlight the superior performance of the Transformer model in air quality forecasting. DeepFogAQ’s decentralized, scalable, and fault-tolerant structure offers a promising solution for managing air pollution and supporting decision-making processes.

While these studies effectively utilize LSTM networks and machine learning in similar contexts, the system proposed in this work prioritizes adaptability. The platform facilitates easy integration of custom machine learning models through its sandbox environment, offering enhanced versatility for diverse industrial applications.

Researchers can employ several techniques to optimize LSTM models for these environments, for example:

• Model Pruning and Quantization: Reducing the size of the LSTM model through pruning unnecessary parameters and applying quantization techniques helps decrease the memory footprint and computational load, making the model more suitable for embedded systems [47].
• Knowledge Distillation: This involves training a smaller, less complex model (student model) to mimic the behavior of a larger, more complex LSTM model (teacher model). The smaller model can then be deployed on resource-constrained devices without significant loss of accuracy [48].
• Hardware Acceleration: Utilizing specialized hardware, such as Tensor Processing Units (TPUs) or Field-Programmable Gate Arrays (FPGAs), can significantly accelerate the computation of LSTM models on embedded devices, enabling real-time processing [49].
• Edge AI Frameworks: Using edge AI frameworks like TensorFlow Lite or ONNX Runtime helps convert and optimize LSTM models for execution on embedded systems, ensuring efficient performance while maintaining high accuracy [50].

By implementing these optimization strategies, industries can effectively deploy LSTM-based predictive maintenance solutions on embedded systems, ensuring robust and reliable performance even in resource-constrained environments [51]. In this case, a combination of model pruning, quantization, and hyperparameter tuning has been applied to strike an optimal balance between computational efficiency and prediction accuracy. This strategy enables the deployment of highly efficient models without compromising the reliability of predictive maintenance, making them ideal for resource-constrained environments. Additionally, these optimizations help manage memory usage, ensuring the model retains only the essential features for accurate predictions while avoiding unnecessary information overload.

Compared to these articles, the LSTM model of the Concentrator serves a dual purpose within the system. First, it is utilized for predictive maintenance, ensuring the machine operates correctly and alerting the user to potential issues if specific parameters deviate from expected values. Second, the model monitors the sensors’ performance. Since the system is homogeneous, if predictions from one sensor diverge significantly from those of its neighboring sensors, it may indicate a problem with the measurement system itself. Therefore, the LSTM model aids in verifying both the machine’s proper operation and the accuracy of the measurement system.

Numerous studies have attested to the superior performance of LSTM models compared to their counterparts, showcasing their ability to achieve remarkable accuracy and generalization on diverse datasets [52]. LSTM’s adaptability to varying data complexities and ability to learn intricate patterns make it a favored choice for addressing real-world industry challenges. Its versatility and proven efficacy solidify LSTM’s position as a key technology in deep learning [53].

While the LSTM model excels in predictive maintenance due to its ability to capture long-term dependencies, deploying such models on embedded systems poses unique challenges. Embedded systems, often characterized by limited computational power, memory, and energy resources, require efficient deployment of LSTM models to ensure real-time performance without exhausting system resources [54].

3. Methodology and Architecture

3.1. The Native System

This section will describe the entire system, showing the architecture of the Concentrator, as seen in Figure 2 and its functionalities, as seen in Figure 3.

The Concentrator system is a comprehensive framework for real-time data management and analysis within industrial environments. Essential functionalities or Use Cases (UCs) include:

-

UC.1: wireless data retrieval from sensors and devices

-

UC.2: versatile handling of process data from AROL and external systems

-

UC.3: rigorous technical data processing for real-time monitoring and optimization

-

UC.4: prompt anomaly detection with effective user notification

-

UC.5: multi-device connectivity for holistic monitoring

These functionalities collectively empower the Concentrator to serve as a central data aggregation, processing, and dissemination hub, facilitating informed decision-making, predictive maintenance, and operational optimization. Moreover, with provisions for seamless integration of custom sensors and adapters, the Concentrator ensures interoperability across diverse industrial ecosystems.

• Protocols requirements
  The Concentrator system employs robust communication protocols tailored for seamless interaction with IoT and operational technology (OT) systems in industrial settings. For IoT device communication, it utilizes wireless connectivity, supporting protocols like Bluetooth Low Energy (BLE) version 4.0 and Wi-Fi. The Wi-Fi capability enables transmission using various protocols such as Message Queuing Telemetry Transport (MQTT), Hypertext Transfer Protocol (HTTP), Transmission Control Protocol/Internet Protocol (TCP/IP), and Constrained Application Protocol (CoAP) [55]. Notably, the platform handles multiple connections from IoT devices, handling connection establishment, maintenance, and termination while automatically reconnecting during temporary disruptions. While communication with OT systems is pending development, the Concentrator will interface with Modbus Transmission Control Protocol (Modbus/TCP) and OPC Unified Architecture (OPCUA) protocols, facilitating data exchange with industrial automation devices such as Programmable Logic Controllers (PLCs) and Supervisory Control and Data Acquisition (SCADA) systems. Additionally, it features synchronization inputs for the temporal alignment of devices or actions, ensuring precise coordination within industrial processes. Moreover, potential integration of Ethernet-based protocols may be necessary to accommodate client-specific devices, highlighting the Concentrator’s adaptability to diverse industrial environments.
• Functional requirements
  The Concentrator system has various functionalities essential to operate within industrial contexts. Firstly, the system addresses the platform power management requirement, including startup, shutdown, and configuration procedures. The Concentrator then establishes robust connections with various external systems regarding system communication. Notably, the Concentrator can communicate simultaneously with a maximum of 64 sensor node devices (usually mounted on a capping machine) while maintaining connectivity to additional devices as required. Moreover, to ensure future scalability and adaptability, compatibility with diverse IoT devices can be possible as long as the hardware limitations permit it.
• Data Processing
  The Concentrator system plays a crucial role in enhancing the quality and utility of data received from IoT devices. Initially, it performs preprocessing on data from IoT devices to ensure compatibility and efficiency before transmitting it to external systems. Depending on the user’s configuration, it also filters data based on specific criteria, potentially removing invalid, out-of-range, or non-compliant data to maintain data integrity and relevance. Additionally, the system aggregates data from various IoT devices to offer a comprehensive overview of collected information, aiding informed decision-making. It can also apply calibrations or normalizations to ensure data aligns with external system standards, improving interoperability.
• Error Management
  The Concentrator is responsible for robust error detection and mitigation to maintain operational reliability and data integrity. It identifies errors and anomalies like connection failures, transmission errors, out-of-range data, or invalid values. Once detected, the system communicates these errors via messages, system logs, or other notifications for timely intervention and resolution. Additionally, it can address critical errors through emergency procedures such as system shutdown or activating safety modes to prevent potential hazards. Moreover, the Concentrator supports temporary data storage during system disconnection, ensuring uninterrupted data transmission upon reconnection.
• Cybersecurity and Updates and Maintenance
  The primary responsibility of the Concentrator is to authenticate IoT devices, ensuring that only authorized devices can communicate with the network. This authentication process is pivotal in safeguarding sensitive data from unauthorized access, thereby preserving the confidentiality and integrity of the system. Moreover, the platform manages data access dynamically, preventing potential data breaches and ensuring that only permitted users or devices can interact with sensitive information.
  In addition to authentication and access control, the system supports remote administration and monitoring, enabling the continuous execution of system updates and maintenance. This capability ensures the platform remains up-to-date with the latest security patches, protecting it from emerging vulnerabilities. The system can proactively identify issues and optimize performance through remote monitoring, maintaining robust security and functionality. This approach to remote control also facilitates timely intervention for system enhancements and problem resolution, ensuring and introducing continuous security measures to address evolving cyber threats.

3.2. The Host System

The Host system is designed as a sandbox environment, enabling users to customize its functionality by integrating external machine learning models or algorithms. This adaptability enhances the system’s usability across various industrial applications. As an example, this paper demonstrates the integration of an LSTM model for predictive maintenance. The model processes sensor temperature data to forecast future values in real-time, enabling proactive operational decisions.

Time series forecasting models for resource-constrained embedded systems have been extensively studied, each offering distinct trade-offs. Traditional methods such as ARIMA [56] and Exponential Smoothing (ETS) [57] are computationally efficient and suitable for more straightforward, stationary datasets. However, their inherent limitations in capturing non-linear relationships and long-term dependencies reduce their applicability to more complex datasets. Gradient Boosting algorithms, including XGBoost and LightGBM [58], have demonstrated strong performance in modeling non-linear patterns, achieving high predictive accuracy when supported by robust feature engineering. Nevertheless, these methods lack native mechanisms to model sequential dependencies or long-term temporal patterns, often necessitating labor-intensive preprocessing, such as creating explicit lag features. TinyML [59], a framework designed for deploying machine learning models on highly resource-constrained devices, was initially considered due to its efficiency. However, its limitations in handling complex temporal dependencies and non-linear dynamics made it less suitable for the application. Instead, we selected Long Short-Term Memory (LSTM) networks, leveraging the computational capabilities of the Raspberry Pi to deploy these models using TensorFlow. LSTMs are designed explicitly for sequential data, excelling in their ability to model long-term dependencies and non-linear temporal relationships. By utilizing Raspberry Pi’s relatively advanced processing power, we avoided the need for additional hardware while achieving robust forecasting performance. This alignment of the LSTM’s advanced modeling capabilities with the Raspberry Pi’s computational potential underscores the suitability of this approach for addressing complex time series forecasting challenges in embedded systems.

• Model Development
  The development of the LSTM-based temperature prediction model followed a systematic process involving data preprocessing, model architecture design, training, evaluation, and deployment. The primary goal was to predict future temperatures based on historical sensor data from multiple nodes while ensuring the model’s compatibility with edge devices such as the Raspberry Pi. The dataset consisted of over one million temperature readings from 50 sensor nodes stored in a CSV file. The dataset was then split into training, validation, and test sets using an 80-10-10 ratio. The data was normalized to a range of [0, 1] using MinMax scaling to enhance training efficiency.
  Sequences of 20-time steps were generated using a sliding window approach for time series modeling. Each sequence provided the input for the model to predict the temperature for the subsequent hour. This preparation enabled the LSTM model to capture temporal dependencies and patterns within the data.
  The LSTM model architecture was designed to balance predictive accuracy and computational efficiency. The model comprised two LSTM layers with 50 units capable of processing sequential input and capturing long-term dependencies inherent in the temperature data. Two Dense output layers, one with 50 neurons and one with a single neuron, were used to predict the temperature for the next hour. The input shape for the model was defined as $(n\_steps,n\_features)$, where $n\_steps$ represents the number of historical time steps (20) and $n\_features$ corresponds to the number of sensor nodes. The model was trained using the Adam optimizer with a root mean squared error (RMSE) loss function.
  The training process was conducted over 20 epochs with a batch size of 30. Early stopping was implemented with a patience of 5 epochs to mitigate overfitting by halting training when the validation loss ceased to improve. Model performance was evaluated on the test set, with root mean squared error (RMSE) as the primary metric for assessing prediction accuracy, obtaining a result of 0.1838, and a prediction time of approximately 200 ms.
  Following training, the model was prepared for deployment on edge devices. The trained LSTM model was saved in TensorFlow’s SavedModel format and converted to TensorFlow Lite format using the TFLiteConverter. The TensorFlow Lite interpreter validated the converted model’s performance, ensuring compatibility and consistent predictions.
  This structured approach enabled the LSTM model to leverage the computational capabilities of the Raspberry Pi effectively, providing robust and scalable temperature forecasts while adhering to the constraints of resource-constrained edge environments.
• The inference code
  The software for the inference is based on Python 3.11, leveraging libraries such as Numpy 1.26.3, Pandas 2.1.4, Ujson 5.4.0, TensorFlow 2.15.0, and Sci-kit learn 1.2.2.
  Data acquisition begins with readings from a TCP server, which are preprocessed to eliminate erroneous values or measurements outside the acceptable range. Values exceeding this range trigger a warning, which the notification system records. Similarly, if the model predicts a value that is too high or too low, a warning is also triggered to signal potential anomalies. For error values (e.g., NaN), the algorithm ignores them to prevent distortion in the data. In the event of a disconnection from the TCP server, the program automatically attempts to reconnect after one second. After collecting enough valid measurements, the pre-trained model is loaded, and predictions are generated using a sliding window approach with a span of 20 values. This method ensures that each new measurement dynamically updates the forecast, providing precise and timely insights into sensor behavior while maintaining accuracy and detecting potential anomalies in real-time. To tailor predictions to the unique characteristics of each node, a personalized MinMax Scaler is trained for every sensor. This personalized scaling not only enhances the accuracy of the predictions but also facilitates the monitoring and maintenance of individual nodes, ensuring consistent performance across the network.
  

  Figure 4. Flowchart of the LSTM algorithm.

  Figure 4. Flowchart of the LSTM algorithm.

  
• Requirements of the Host system
  The Host system must:

  −

  Provide a sandbox environment for effortless integration of custom algorithms, empowering users to adapt the platform to diverse industrial needs. In our specific case:

  ∗

  Receive data from the Concentrator via TCP connection and process the raw data files generated by the container

  ∗

  Ensure real-time data processing, perform data cleaning, and apply integrated models or algorithms

  ∗

  Manage a sliding window of 20 values for accurate predictions of future data points

  ∗

  Provide updated predictions with each new data input to enhance monitoring and process optimization

  −

  Integrate easily with external systems for data analysis and industrial management

4. Experiment and Discussion

The experiment consisted of two phases: the first was in a controlled laboratory setting, where 50 sensor nodes were connected to monitor and record the room’s temperature. In the second phase, we used an actual machine and tested it in a thermal chamber instead of a real-world environment due to ongoing testing constraints.

The Nicla Sense ME has been used as a sensor node, and it exemplifies a trio of strengths: compact design, cost-effectiveness, and energy efficiency, while seamlessly integrating four state-of-the-art sensors engineered by Bosch Sensortec. The acronym "ME" stands for "Motion" and "Environment," highlighting its capability to accurately detect rotation, acceleration, temperature, humidity, pressure, air quality, and CO2 levels with industrial-grade precision. This latest version sets a new benchmark with its ultra-compact size, enabling effortless data transmission via Bluetooth Low Energy (BLE) 4.2, powered by the ANNA-B112 module. In the experiment, 50 Niclas were used, both in the laboratory and on the machine. Their primary role was to transmit temperature, humidity, and pressure data to the Concentrator using the wireless protocol.

The Raspberry Pi 4 Model B, with Quad-core Cortex-A72 (ARM v8), has been used as a Concentrator, the core of our embedded hardware system, operating on Raspberry Pi OS 64-bit, built on Debian 12 (Bookworm). This device provides ample memory (8 GB of RAM) to handle the demands of running resource-intensive applications, including machine learning models and real-time sensor data processing, ensuring smooth performance and efficient multitasking within the system. Regarding power consumption, the system uses embedded devices, such as the Raspberry Pi, which inherently consumes very little energy (typically around 2-5 watts), making it highly suitable for long-term operation in resource-constrained environments. The choice of this hardware for the Concentrator and the low-power sensors, such as the Nicla Sense ME, ensures that the system can operate for extended periods without significant energy demands. This efficient power usage is crucial for deployment in remote or industrial environments, where resources must be carefully managed.

To support virtualization and sandboxing for external users, Docker version 25.0.4 (build 1a576c5) has been implemented, with Docker Compose version 2.24.7 enabling efficient management of multiple containers concurrently. Given our reliance on Bluetooth for wireless communication, the BlueZ library version 5.66 was integrated into the system. The overall experiment is represented in Figure 5.

The Concentrator’s primary function is to aggregate data from multiple peripherals. Utilizing a container built on Node.js version 16 and the open-source library node-ble, this system intelligently differentiates between Host and Native configurations based on peripheral addresses and settings. It establishes smooth connections with the targeted peripherals and carefully records the data streams from the Nicla Sense ME sensors. The acquisition frequency is calibrated to capture measurements at a rate of one per second. Once connections are established and the necessary data is acquired, the container publishes the raw data locally using a TCP server. The Native container employs the node-ble library and Bluetooth connections to broadcast native measurements. Each line of the data file includes the peripheral’s MAC address, the measured characteristic’s value, and the sensor it uses (internal or external, in case an external sensor is attached).

The results are evaluated using the following performance metrics:

• RMSE (Root Mean Squared Error): This metric quantifies the square root of the average squared differences between the observed and predicted values. It is calculated as:

  $$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{(y_i-{\widehat{y}}_i)}^2}$$

  where $y_i$ are the observed values, ${\widehat{y}}_i$ are the predicted values, and n is the number of data points.
• MAE (Mean Absolute Error): MAE represents the average absolute difference between predicted values and actual values. It is given by:

  $$\mathrm{MAE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n|y_i-{\widehat{y}}_i|$$

  where $y_i$ are the observed values, ${\widehat{y}}_i$ are the predicted values, and n is the number of data points.
• MAPE (Mean Absolute Percentage Error): This metric expresses the average absolute percentage difference between predicted and actual values, offering a relative measure of prediction accuracy. It is calculated as:

  $$\mathrm{MAPE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|{\displaystyle \frac{y_i-{\widehat{y}}_i}{y_i}}\right|\times 100$$

  where $y_i$ are the observed values, ${\widehat{y}}_i$ are the predicted values, and n is the number of data points.
• Accuracy: Accuracy is defined as the percentage of correct predictions made by the model out of all predictions. It can be calculated as:

  $$\mathrm{Accuracy}={\displaystyle \frac{\mathrm{Number}\mathrm{of}\mathrm{Correct}\mathrm{Predictions}}{n}}\times 100$$

  where the number of correct predictions refers to the instances where ${\widehat{y}}_i=y_i$ and n is the total number of predictions.
• R2 (R-squared): R-squared measures the proportion of variance in the dependent variable that is explained by the independent variables in the model. It is computed as:

  $$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2}{{\sum }_{i=1}^n{(y_i-\overline{y})}^2}}$$

  where $y_i$ are the observed values, ${\widehat{y}}_i$ are the predicted values, $\overline{y}$ is the mean of the observed values, and n is the number of data points. A higher $R^2$ value indicates a better fit of the model to the data.

4.1. Controlled Laboratory Results

The first phase aimed to assess the system’s performance and evaluate anomaly detection between the sensors. One was intentionally adjusted within its software to generate more noticeable temperature fluctuations. This sensor was modified to heat up or cool down, creating more pronounced temperature variations than the others and maintaining a uniform reading. These modifications resulted in a change in its prediction without knowing which node had been altered. The table below presents the resource consumption data for the architecture, showcasing the system’s efficiency across various configurations:

Table 1. LSTM experiment data.

Container	Number of nodes	Memory	CPU
None	0	570MB	0%
Only Native	1	671MB	10%
Only Native	50	700MB	17%
Only Host	1	800MB	10%
Only Host	50	1GB	20%
Docker Compose	1	1,2GB	35%
Docker Compose	50	1,4GB	55%

Table 1. LSTM experiment data.

Container	Number of nodes	Memory	CPU
None	0	570MB	0%
Only Native	1	671MB	10%
Only Native	50	700MB	17%
Only Host	1	800MB	10%
Only Host	50	1GB	20%
Docker Compose	1	1,2GB	35%
Docker Compose	50	1,4GB	55%

The results indicate that the system manages resource consumption effectively as the number of nodes increases. Even with 50 nodes, memory and CPU usage remain within manageable limits, demonstrating the system’s scalability without significant performance degradation.

Moreover, the system’s transmission efficiency was evaluated as the number of nodes increased. When scaling from a single node (with a detection rate of 60 times per minute) to 50 nodes, the same sensor was detected 44 times per minute, resulting in a 26.53 % decrease in transmission efficiency. Despite this, the overall impact on system performance remains acceptable, suggesting the system is well-suited for environments with multiple nodes, where transmission efficiency is still maintained at a reasonable level.

These findings highlight that the system balances performance, resource usage, and scalability, making it a strong candidate for deployment in industrial IoT environments that require efficient resource management and the ability to scale as needed.

Figure 6. Mean Square Error of laboratory results (logarithmic scale).

Figure 6. Mean Square Error of laboratory results (logarithmic scale).

The analysis in Figure 7 demonstrates that minimal error metrics mark the system’s predictive performance. The highest values for MAE and RMSE are noted for NODE_36 at 0.0126 and 0.0156, respectively, and an accuracy of 99.81 %. These values represent the most significant deviations among the nodes, yet they still indicate that even the node with the most critical errors remains within a relatively narrow accuracy range. In contrast, NODE_25 is identified as an anomaly within the dataset due to its intentional modification as a test case designed to assess the system’s capability of checking the node’s behavior. This deliberate alteration of NODE_25 provides insights into the system’s behavior under abnormal conditions. The generally low error rates across most nodes highlight the effectiveness of the predictive maintenance system, underscoring its ability to produce accurate and reliable forecasts. The discrepancies observed, especially with NODE_25, illustrate how the Concentrator monitors and detects anomalies.

4.2. Real Machine Results

In alignment with laboratory results, the performance metrics obtained from the real machine data further affirm the system’s efficiency, as depicted in Figure 8. Individual sensors were not subject to modification or experimental alteration in the scenario. The consistency is reflected in the results, where all sensors exhibited comparable errors in predictive accuracy and recorded similar temperature measurements. Maximum error values were observed for NODE_08, which recorded a maximum MAE of 0.1821, an RMSE of 0.288, R² (R-squared) of 0.8898, a MAPE of 0.7227%, and an accuracy of 99.2% (with an error of $\le 1$ °C), as shown in Figure 9. These values show only minimal deviation, highlighting the system’s precision in predicting machine behavior. While the results are less optimal than those from the laboratory tests, this is attributed to the presence of metal and other interferences typical of a real industrial setting. The lack of anomalies or sensor discrepancies in this test led to consistently low error rates, indicating that the system performs optimally when all components operate as expected. This outcome effectively demonstrates the system’s ability to maintain high accuracy in the absence of external disruptions, further confirming its reliability for practical use in real-world environments.

Table 2 compares the results obtained in the experiments conducted for these experiments.

5. Comparison with Existing Approaches

5.1. Comparison with FDPR [44]

Both studies focus on similar use cases involving sensor data prediction and recovery in IoT environments. However, notable differences exist in the models employed, the experimental setups, and the results achieved. Table 3 presents a comparative analysis of key performance metrics between the LSTM-TFLite model (The Concentrator) and the model from the previous study (FDPR, DC-MLP). The table highlights significant differences across several performance indicators, including RMSE, MAPE, prediction accuracy, and inference time.

The LSTM-TFLite model demonstrates a higher MAE (0.1821 compared to 0.06) and equal RMSE (0.28 compared to 0.28), suggesting that the DC-MLP model exhibits lower error rates, indicating slightly better performance in error metrics. Additionally, the MAPE for the LSTM-TFLite model (0.7) is significantly larger than the value for the DC-MLP model (0.002), further suggesting that the DC-MLP model offers more accurate predictions relative to the actual values in percentage terms. The DC-MLP model also achieves a higher prediction accuracy (99.86%) compared to the LSTM-TFLite model (99.2%).

The most notable discrepancy is inference time: our model takes 200 ms to process, significantly longer than the previous study’s 0.05 ms. This large gap in processing speed could be attributed to differences in the complexity of the models, hardware configurations, or the specific optimization techniques used in each approach. Despite the increased inference time, the overall accuracy of our model and its ability to handle sensor data predictions and recovery are still competitive with the prior work.

5.2. Comparison with the FAQMP [45]

Both studies focus on similar use cases involving sensor data prediction and recovery in IoT environments. Table 4 presents a comparative analysis of key performance metrics between the LSTM-TFLite model (The Concentrator) and the model from the previous study (FDPR, Seq2Seq GRU Attention-TFLite model).

The table highlights significant differences across several performance indicators, including MAE, RMSE, Mean Absolute Percentage Error (MAPE), R-squared (R2), prediction accuracy, and inference time.

The LSTM-TFLite model demonstrates a significantly lower MAE (0.1821 compared to 4.385) and RMSE (0.28 compared to 7.9016), indicating superior performance regarding error metrics. This result suggests that the LSTM-TFLite model provides more accurate predictions with less deviation from the actual values. Similarly, the MAPE for the LSTM-TFLite model (0.7) is much smaller than that of the Seq2Seq GRU Attention-TFLite model (22.0133), indicating that the LSTM-TFLite model has a lower relative prediction error.

Regarding R-squared (R2), the LSTM-TFLite model also performs better, with an R2 value of 0.8898 compared to the previous study’s value of 0.6566. This result indicates that the LSTM-TFLite model explains a higher proportion of the variance in the data, thus providing a better fit and more reliable predictions.

However, the most striking difference is observed in inference time: the LSTM-TFLite model requires 200 ms for inference, while the previous study’s model takes an extraordinarily long 52,6 s. This significant discrepancy in processing time may be attributed to the complexity of the Seq2Seq GRU Attention model and suggests that the LSTM-TFLite model is much more efficient for real-time applications.

5.3. Comparison with the DeepFogAQ [45]

Both studies focus on similar use cases involving sensor data prediction and recovery in IoT environments. Table 5 presents a comparative analysis of key performance metrics between the LSTM-TFLite model (The Concentrator) and the model from the previous study (DeepFogAQ, Transformers).

The table highlights significant differences across several performance indicators, including MAE and RMSE.

The LSTM-TFLite model exhibits a lower MAE (0.1821 compared to 3.15) and RMSE (0.28 compared to 3.79), suggesting that it outperforms the DeepFogAQ Transformer model regarding error metrics. This result indicates that the LSTM-TFLite model provides more accurate predictions with less deviation from the actual values.

Despite the differences in error metrics, both models aim to optimize performance for sensor data prediction and recovery. However, the LSTM-TFLite model demonstrates superior prediction accuracy and error reduction, giving it an edge in sensor data prediction tasks.

6. Conclusions and Future Works

This study demonstrated the practicality of implementing a predictive maintenance system within a resource-constrained IoT framework. Utilizing Nicla Sense ME sensors and a Raspberry Pi-based Concentrator, the system successfully monitored, processed, and predicted real-time temperature measurements. Additionally, including a sandbox environment allows external users to customize the system by adding different codes or sensors without altering the core software, enhancing its adaptability for diverse applications.

Using an LSTM-based machine learning model on an embedded system showed that advanced AI techniques can be effectively applied even in environments with limited computational resources. The model’s accurate predictions, with an RMSE of 0.1838, demonstrate the potential for these systems to be utilized in industrial settings to support proactive maintenance and reduce downtime.

Moreover, the architecture’s scalability was confirmed, as the system maintained reliable performance across various connected devices. This scalability, combined with the successful integration of machine learning for predictive maintenance, underscores the system’s potential for broader application in smart industrial environments.

In future work, the system will undergo significant enhancements in cybersecurity to address emerging threats more effectively. Additional studies will incorporate data from other critical machine parameters, such as accelerations and vibrations. By integrating these additional data streams, the system aims to provide a more comprehensive analysis of machine behavior, enhancing its predictive maintenance capabilities. Moreover, the system will be subjected to rigorous stress testing, including scenarios with high data loads and network interruptions. These tests will evaluate its resilience and performance under extreme conditions, ensuring reliability in real-world applications. These improvements will strengthen the system’s security, improve its robustness, and enable a deeper understanding of machine dynamics, paving the way for more accurate maintenance strategies.