A Centralized Federated Learning Algorithm Based Multi Classification Predictive Maintenance in Industrial Internet of Things System

Predictive maintenance (PdM) and the Industrial Internet of Things (IIoT) together mark a significant change in equipment management, especially when deep learning (DL) is used to interpret time series data from IIoT devices [1]. IoT is a scalable network system of distributed devices and connectivity that is having a dramatic impact on a wide range of application areas. Particularly in the context of Industry 4.0, the increasing velocity and volume of data generated by Industrial IoT (IIoT) devices, combined with advanced analytics capabilities, enables intelligent industrial operations, which improve manufacturing processes and operational efficiency [2].

Four main types of maintenance in industry: Predictive Maintenance (PdM) improves equipment longevity and prevents unplanned machine downtime by identifying possible problems before they materialize. While Reactive Maintenance (RM) focuses on fixing equipment as soon as it breaks down, Proactive Maintenance (PRM) aims to prevent failures by addressing their root causes. This suggests that because failure duration is unpredictable, downtime may be significant. Consequently, regardless of the machine's real condition, Preventive Maintenance (PM) is carried out regularly. Unlike reactive, which is less costly but results in higher expenses if equipment breaks down and creates unexpected downtime. Predictive methods enable planning and minimize unexpected downtime, while preventative measures may result in needless maintenance. Because PdM entails condition monitoring, it only minimizes unexpected downtime, while preventative measures may result in needless maintenance. Because Predictive maintenance entails condition monitoring, it is only carried out when required. As a result, there should be less need for replacement components, which would save expenses. Additionally, it helps prevent excessive maintenance, which frequently leads to problems. While predictive maintenance primarily focuses on scheduling maintenance for equipment situation, preventive maintenance may entail reevaluating systems to prevent breakdown. Using contemporary technology like data analysis, machine learning, and Internet of Things sensors yields further benefits. As a result, forecasts became more precise. Equipment life can be increased by addressing issues early. PdM necessitates real-time data analysis and storage while accounting for the many facets and impacts of the signals gathered. Predictive maintenance use various techniques, such as artificial intelligence, time series analysis, data analytics, , and different failure modes ,etc. [3].

In an IIoT setting, federated learning enables several devices (or clients) to work together to train machine learning models without disclosing private information. In PdM, senssory data should de secured when uploaded to a cloud server. Federated learning resolves privacy issues while facilitating efficient model training by maintaining data decentralization [4]. Combining both PdM and FL will enhance prediction accuracy while protecting privacy through local data processing. This tactic promotes inter-organizational collaboration while reducing the cost of data transit. FL is therefore an effective tool in the context of Industry 4.0 [5].

Currently, deep learning is applied across various fields, including PdM. When ample historical data is provided, deep learning models outperform both statistical and conventional machine learning approaches. At the core of deep learning there are artificial neural networks (ANN), which mimic the brain's functioning. Unlike shallow networks with one or two hidden layers, deep learning utilizes more complex architectures [6]. Sensors directly attached to industrial equipment produce data that condition monitoring (CM) systems analyze to detect irregularities and forecast potential failures. This enhances reliability and optimizes maintenance costs. CM relies on diverse sensors, including those measuring temperature and vibration, to track system parameters, model behavior, and detect anomalies signaling wear or damage. Such monitoring is critical, as equipment failures can disrupt entire production processes, resulting in costly repairs and significant financial losses. As simple sensors alone cannot provide enough power, the data is usually transferred to an external server or cloud for analytics jobs such as machine learning models.

Industrial facilities are often located far away from the actual cloud data centers, which can limit network capabilities. For example, they can suffer from high latency and unstable network connections, which can cause network link congestion. Early fault diagnosis leads to faster repairs and reduced follow-up costs [7]. It is the reason why FL as a distributed machine learning paradigm has been more and more popular for the IIoT recently. FL involves training the models locally using real data sources, and then aggregating them into a global model without sharing any training data. This makes it possible to integrate the knowledge of other faculties while keeping training ML models inside the organizational limits of IIoT locations [8].

This work introduces and applies a PdM algorithm that combines FL with a DNN. The propsed system has been tested on a benchmark system of three AC motors equipped with four sensor types. A substantial offline dataset was gathered and trained the DNN model on a public server. During the online process, real-time data is gathered from different sensors and fed into the DNN model to predict potential failures. Additionally, the method integrates condition monitoring technologies with the FL framework, enabling local model training at individual IIoT sites while facilitating cloud-based model aggregation to enhance knowledge sharing across different locations.

The rest of this paper is organized as follows: Section 2 presents the current research in PdM systems. prerequisites necessary to design the proposed system is introduced in section 3. Section 4 shows how the custom dataset being gathered from real system. Section 5 describes the proposed PdM design and section 6 introduced the discussion and results of the proposed system. Finally, Section 7 introduce the conclusion of this work.

To enable preventive maintenance, predictive maintenance leverages IIoT and data analytics to check machines status and predict faults. In this work, a predictive maintenance system has been built based on federated deep learning.

3.1 Working sensors

Data has been gathered from the workbench using a variety of sensor types, including: As a 13-bit, 3-axis shaking sensor, the ADXL345 is an ideal device for monitoring motor vibrations. It has good qualities such as a tiny profile, a compact form factor, and extremely low power consumption. With the ACS712 current sensor module, the motor's current is tracked. It is an entirely integrated current sensor that does not experience magnetic hysteresis. With just one 5V power source, it provides linear current sensing for both AC and DC currents via the Hall Effect. It also uses a low-resistance current conductor and an analog-to-digital converter (ADS1115). Certain essential qualities must be possessed by the temperature sensor that was employed in this study. The MLX90614 contactless temperature sensor was chosen for its wide range, high accuracy and precision. Its affordability and compact size also make it a good fit for the requirements of this study. A SHT21 digital humidity and temperature sensor was also used to measure the outdoor temperature [19]. That operating efficiency is greatly increased by predictive maintenance with vibration sensors. Makes highly accurate predictions about production stoppages by using machine learning techniques. This proactive strategy reduces downtime and boosts overall productivity by enabling prompt maintenance interventions [20]. The specifications of the sensors used in this work are shown in Table 1.

Table 1. Specification of the sensors [20]

3.2 I2C protocol

The Raspberry Pi's mobility, parallel, affordability, and low power consumption make it a very practical, promising, and perfect technology for IoT applications. The Raspberry Pi is an affordable, energy-efficient computer capable of performing many tasks typically handled by a traditional desktop PC, making it ideal for IoT applications. One tool for accelerating processes and calculations is the Raspberry Pi 4, it is a quad core processor [21]. The different types of sensors is connected using the I2C protocol ( a two-line interface enables communication between different sensors at the same time). Sensors are connected to the RPi using two wires Serial Clock (SCL) and Serial Data (SDA).

3.3 Deep neural networks

The proposed DNN model is described in Table 2. The model consists of eight layers, with the input layer matching the feature dimensions and the output layer represents the five classes of faults [22]. In this work, ReLU is used as the activation function in hidden layers while Softmax is used in the output layer which enable multi classification required in this work [23]. The proposed DNN model layers used in this work is chosen such that a decision impacted by a number of input features. By striking a balance between model capacity and issue complexity, it enables the network to learn abstract patterns without going into too much detail. This setup maintains computational efficiency while reducing the chance of overfitting, particularly when data is few. It also conforms to the empirical results of experiments and the principles of architectural design. In general, eight layers each of 50 neurons work well for a variety of datasets and classification applications. The loss function used in this work is categorical cross-entropy which is best work with multiclass labels.

Table 2. Hyperparameter of DNN

3.4 Federated learning

In the conventional federated learning (FL) framework as shown in Figure 1, multiple edge devices function as independent clients connected to a central global server. Typically, a predetermined number of FL clients (D) are randomly chosen from a group of edge devices that have expressed willingness to participate in the training process. During a training iteration at time t , each selected client (d) retrieves the global model parameters $\left(\theta_t\right)$ from the server and trains the model using its local dataset. If $\mathrm{m}_{\mathrm{d}}$ denotes the number of data samples for a client, then $\sum_{d=1}^D m_d=\mathrm{M}$, where M is the total size of data samples from D clients. The FL attempts to optimize $f(\theta)$ [24].

1.png

Figure 1. Traditional fedrated learning [24]

$\begin{gathered}\min f(\theta) \\ f(\theta)=\sum_{d=1}^D \frac{m_d}{M} F_{d(\theta)} \\ F_d(\theta)=\frac{1}{m_d} \sum_{i \in m_d} f^i(\theta)\end{gathered}$     (1)

In the local training phase, each client d optimizes the global model by minimizing its loss function $f^i(\theta)$, which corresponds to the i-th sample in its datasetOptimization techniques like stochastic gradient descent (SGD) or Adam are employed for this purpose. Once local training is complete, the client transmits its updated version of the global model, denoted as $\theta_{t+1}^d$, back to the global server.

$\theta_{i+1}^d=\theta_i-\alpha_d \lambda_d$     (2)

Here, $\alpha_d$ represents the learning rate, while $\lambda_d$ denoting the gradient calculated by client d using its local dataset and the parameters $\theta_t$.

The global server combines the collected local models and calculates the enhanced global model $\left(\theta_{t+1}\right)$ in the following manner.

$\theta_{i+1}=\sum_{d=1}^D \frac{m_d}{M} \theta_{i+1}^d$     (3)

In FL, clients download the improved global model for subsequent training rounds, continuing this iterative process until the model converges. Federated Averaging (FedAvg) is a decentralized learning method where multiple participants collaboratively train a model while keeping their raw data private. Each client performs local training using its dataset and sends only the adjusted model parameters—not the raw data—to a central server. The server aggregates these updates by taking a weighted average of the model weights, with the weights determined by the size of each client's dataset. This approach enhances privacy, reduces bandwidth usage, and scales well across numerous clients. However, it faces challenges like data heterogeneity and communication efficiency. Overall, FedAvg is crucial for effective federated learning. The FedAvg formula can be describe as:

$w_{t+1}=\sum_{k=1}^K \frac{n_k}{n} w_{t+1}^k$     (4)

where, $n_k$ represents the aggregate number of data samples from all devices; $w_{t+1}^k$ is the weight vector of the local model modified by device k in round $t+1$. Given a loss function $\eta$ and a gradient $\nabla \mathrm{L}\left(w_t^k ; \mathrm{b}\right)$ computed regarding the model weights $w_t^k$ and input data $b$, the local and global model updates can be presented as illustrated in Algorithm 1.

The proposed benchmark testbed consists of three motors each equipped with four sensors: 3-axis vibration sensor, two temperature sensors a contactless infrared sensor and ambient sensor, and current sensor. All these sensors have been interfaced with three Raspberry Pi 4 for each motor via I2C communication protocol where each sensor has its unique address to avoid packet collision. Each Raspberry Pi then communicates with the cloud server via the MQTT protocol.

The Raspberry pi has a local DNN model (i.e., three DNN model or clients) while a global model is presented in the cloud server. Figure 4 shows the proposed system architecture.

4.png

Figure 4. Proposed system architecture

The proposed system has two mode of operation, local and global models training and predictive mode. Before train the models, a realistic dataset has been collected from the motor in different failed types which are deliberately done.

Within the training mode, a FL approach has been used where in each round the local models are trained on their part of the dataset where each part consists of all features (sensor) but with two failure classes only. After the local model is trained to its desired accuracy and loss values, the final weights of the model are uploaded to the cloud server. In this stage, the three local models upload their weights to the cloud and an aggregation algorithm based on FedAvg is used to compute the new global model weight. Then the global model will train again for the new weights. After that, the global model will download its weight to all local models which will retain their models again. This process is the first round, the proposed system consists of 20 rounds. By doing so, the local model which trains only on its part of the dataset and fewer classes of failure will be trained for all types of classes and this is the main advantage of FL in addition to the security and privacy of the sensory data. Algorithm 1, mentioned earlier, illustrates the machine learning training process. In the second mode of operation, the local model cloud predicts the failure types, and also the global model does. Also, the proposed system included data visualization to show the different types of failures.