Integrating Industrial Sensors for Predictive Maintenance of Squirrel Cage Induction Motors and Three-Phase Synchronous Motors Using ESP8266 and Message Queuing Telemetry Transport Protocol

Predictive maintenance has become an essential strategy for improving the reliability and operational efficiency of industrial motors. Traditional condition monitoring methods relied mainly on signal processing techniques applied to electrical and vibration signals. Daviu presented a transient-based monitoring approach that analyzes motor start-up signals to assess machine health under dynamic operating conditions. This method demonstrated improved accuracy in fault detection compared to conventional steady-state monitoring techniques.

Recent studies have also explored advanced predictive maintenance frameworks for industrial motors. Singh et al. [15] proposed a digital twin–based predictive maintenance system for squirrel cage induction motors, enabling early fault detection and estimation of remaining useful life. By integrating data modeling with sensor measurements, the digital twin architecture provided improved monitoring capability and enhanced maintenance decision-making.

Machine learning and deep learning techniques have further improved motor fault diagnosis performance. Ponce et al. [16] introduced an IoT-enabled deep learning approach for robust fault recognition in induction motors using sensor data. Similarly, Martinez-Roman et al. [17] developed a deep neural network architecture to detect stator winding faults under varying load conditions. These approaches demonstrated improved fault detection accuracy; however, many of these models rely on single-sensor data or require complex feature engineering [18]. Industrial Internet of Things (IIoT) technologies have also enabled real-time monitoring of industrial equipment through distributed sensor networks. Kazmi et al. [19] developed an IIoT-based smart sensor node capable of collecting vibration signals from induction motors and transmitting the data to cloud platforms for predictive maintenance analysis. Such IIoT architectures provide efficient data acquisition and communication mechanisms for large-scale industrial monitoring systems. Although previous studies have explored predictive maintenance using signal processing, machine learning, and IIoT-based monitoring frameworks, several limitations remain. Many studies focus only on single-sensor analysis, which may not capture the complete operating condition of industrial motors. Other works rely on traditional machine learning techniques requiring manual feature extraction, which limits their scalability. Furthermore, existing IIoT monitoring systems often lack advanced deep learning models capable of capturing both spatial and temporal dependencies in sensor signals. To address these limitations, this research proposes a multi-sensor IIoT-enabled predictive maintenance framework using a CNN-GRU hybrid deep learning architecture, enabling automatic feature learning and improved fault classification performance.

Table 1. Comparison of existing predictive maintenance methods

Table 1 presents a comparison of commonly used machine learning and deep learning approaches for motor fault diagnosis. The comparison highlights the advantages and limitations of existing techniques and demonstrates the motivation for adopting the proposed CNN-GRU architecture. The comparison in Table 1 shows that traditional machine learning techniques, such as Support Vector Machines rely heavily on manual feature extraction and domain expertise. Deep learning approaches such as Convolutional Neural Networks can automatically extract spatial features from sensor data, while recurrent architectures such as Long Short-Term Memory networks capture temporal dependencies in sequential signals. However, these methods often focus on either spatial or temporal characteristics individually. The proposed CNN-GRU architecture integrates convolutional layers for spatial feature extraction with GRU units for temporal sequence modeling. This hybrid structure improves fault classification accuracy while maintaining computational efficiency for real-time predictive maintenance applications.

In the methodology for integrating industrial sensors for predictive maintenance of squirrel cage induction motors and three-phase synchronous motors using ESP8266 and MQTT protocol, the process initiates with the detection and analysis of faults in the motors. Fault occurrences are monitored through a set of industrial sensors. The acquired data is then wirelessly transmitted to a central processing unit, powered by an alternative current supply in a three-phase configuration. A pivotal component of the methodology involves leveraging deep learning techniques for more advanced fault classification. Convolutional Neural Network-Gated Recurrent Unit architecture is utilized for feature extraction and classification. This approach allows for the automatic learning of hierarchical features directly from the raw sensor data, eliminating the need for manual feature selection. The classification process is executed using the CNN-GRU model, enhancing the system's ability to accurately identify and categorize motor faults. This streamlined approach optimizes the system's efficiency and performance in fault detection. To facilitate user interaction and visualization of results, a user-friendly graphical user interface has been developed. This interface provides a comprehensive platform for real-time monitoring and maintenance decision-making. The ESP8266 module and MQTT protocol continue to serve as the backbone for interconnecting the entire system, enabling seamless communication and data exchange between sensors, processing units, and the user interface. Figure 1 illustrates the overall predictive maintenance framework.

4.1 Air Conditioning drive systems failures

The construction of an induction motor operated by a voltage source inverter is depicted in Figure 1. It is made up of a three-phase Alternating Current (AC) supply, a Direct Current (DC) connection, a regulated or uncontrolled rectifier, and an inverter that may use Pulse-Width Modulation (PWM) pulses to provide variable voltage and frequency. The modulation coefficient and various PWM approaches will affect the stator currents and inverter line output voltages. To generate a DC voltage at the front end, a diode bridge rectifier is linked to the three-phase supply. The rectified voltage is received over a DC connection by the power inverter system, which is made up of sophisticated power electronic switching components like Metal-Oxide-Semiconductor Field-Effect Transistors (MOSFETs) or IGBTs (Insulated-Gate Bipolar Transistors) which are driven by gate drive circuits to supply changeable voltage and frequency to the changing loads for control of speed. Many different kinds of electrical failures, including those in the rectifier, inverter, motor terminals, and mechanical load, might affect the power electronic circuitry used in AC drives.

Figure 1. Proposed Industrial Internet of Things (IIoT)-based predictive maintenance framework for industrial motors

The three-phase motor input terminal receives the three-phase variable frequency and variable voltage output from the VSI. Problems at the motor terminal can occur from faults such as SLG, L-L, and lines open at the AC power line. If there is a line short, the induction generator will shut down, and ery high power will flow across the shorted terminals, damaging the diodes and IGBTs in the process. Induction motor malfunctions can also be caused by broken bars in the squirrel cage rotor, insulating failure, damage to the stator windings, and bearings. Unlike short-circuit faults, incipient errors do not result in the motor stopping. Even when there are minor defects, the induction motors continue to function. Over time, the machines' dependability tends to decrease and become damaged. They are therefore risky, as they might go unnoticed for long time and damage the engine without being noticed. In industrial procedures that entail failure identification and categorization based on position and severity, fault identification is essential. In order to prevent failure and machine damage, the operator must also be notified of the defect via an interface and be able to take prompt corrective action or precautions. We refer to the complete procedure as reliability-centered management. Of all the methods used for reliability-centered regular consumption, machine learning seems to be a promising one.

4.2 Data collection

The Resistance Temperature Detector, piezoelectric accelerometer, ultrasonic sensors, Hall effect current sensor, Voltage transducer were used to collect temperature, vibration, acoustic, current and voltage data for the research. The dataset contains sensor measurements collected under different motor operating conditions, including normal operation and fault conditions.

The dataset includes 10,000 samples with five sensor features:

• Temperature
• Vibration
• Sound
• Current
• Voltage

Each sample is labeled as either normal or faulty motor operation. The dataset was divided into:

• 70% training data
• 20% validation data
• 10% testing data

This split ensures that the model is trained effectively while maintaining independent data for evaluating generalization performance. The dataset used for this study was obtained from the publicly available Induction Motor Fault Dataset available on Kaggle [24].

The induction machine specifications include 4 poles, 1 HP mechanical power, a delta configuration, 220V supply voltage, and 3A rated current.

Table 2. Data description

In Table 2, the comprehensive dataset captures various levels of short-circuit emulations and operational parameters such as vibration, temperature, sound, current, and voltage, crucial for the proposed predictive maintenance system utilizing ESP8266 and MQTT protocol [24].

4.3 Data preprocessing

Before training the machine learning model, the raw sensor data undergoes several preprocessing steps to improve data quality and model performance.

The preprocessing process includes the following stages:

Noise filtering: Sensor signals may contain noise due to mechanical disturbances and environmental interference. A filtering process is applied to remove high-frequency noise components.

Outlier detection: Extreme values caused by sensor errors or abnormal measurements are identified and removed using statistical thresholds.

Data normalization: The sensor values are normalized using min-max normalization, which scales all features to the range [0,1]. This ensures that all features contribute equally during model training.

The preprocessing of data acquired from IoT devices, such as the ESP8266, for various applications, including predictive maintenance in industrial settings. When dealing with sensor data from devices like ESP8266, it's essential to preprocess the raw data before feeding it into machine learning models. This normalization method is applied to ensure that all features contribute equally to the analysis and prevent features with larger magnitudes from dominating the model. The formulation for min-max normalization is as described in Eq. (1):

$Y_{\text {norm }}=\frac{Y-\min (Y)}{\max (Y)-\min (Y)}$            (1)

Here, $Y$ represents the original feature values, $\min (Y)$ is the minimum value in the dataset for that feature, and $\max (Y)$ is the maximum value.

4.4 Classification of fault occurrence using Convolutional Neural Network-Gated Recurrent Unit

The CNN component consists of a one-dimensional convolution layer with 64 filters and a kernel size of 3 followed by batch normalization. The extracted spatial features are then passed to a GRU layer with 50 hidden units to capture temporal dependencies in the sequential sensor data. The output is flattened and processed by a dense layer with 128 neurons before the final sigmoid classification layer.This approach uses Scaled Exponential Linear Units (SELU) as the activation function to fire the neurons within our CNN. This selection offers significant benefits over traditional activation functions such as Rectified Linear Units (ReLU). Eq. (2), which represents the SELU activation function, has better convergence qualities, which improves the model's training process. It successfully addresses the problem of gradient vanishing, which is an important factor to take into account for predicting squirrel cage faults, since precise predictions depend on the capture and analysis of small spatiotemporal patterns [25].

$S E L U=\lambda\left\{T\right.$ if $t>0, \alpha e^{\wedge} t-\alpha$ Otherwise$\}$               (2)

$\operatorname{ReLU}=\max (0, t)$                 (3)

$B N(t)=\gamma * \frac{(X-\mu)}{\sigma+\beta}$                  (4)

where, $B N(t)$ represents batch-normalized output, '$\gamma$' Scaling factor, '$t$' Input data, '$\mu$' Mean of the batch, '$\sigma$' Standard deviation of the batch, '$\beta$' Shifting factor.

Figure 2. Proposed Convolutional Neural Network-Gated Recurrent Unit (CNN-GRU) architecture

The predictive maintenance for squirrel cage induction motors and three-phase synchronous motors, the classification of fault occurrences is achieved through an innovative approach leveraging Convolutional Neural Network-Gated Recurrent Unit architecture is presented in Figure 2. The methodology integrates industrial sensors for monitoring crucial parameters such as vibration, temperature, and current. These sensors capture real-time data on motor conditions, providing a comprehensive dataset for analysis. The acquired data is wirelessly transmitted to a central processing unit powered by an alternating current supply in a three-phase configuration. Unlike traditional methods, the CNN-GRU model is employed for classification, combining the strengths of Convolutional Neural Networks for spatial feature extraction and Gated Recurrent Units for capturing temporal dependencies in the sequential sensor data. The CNN component operates by applying convolutional layers to extract spatial patterns and pooling layers for spatial dimension reduction. This feature-rich output is then fed into the GRU component, which effectively captures the temporal dynamics of the motor data. The GRU utilizes update and reset gates along with a candidate hidden state to process sequential information, enhancing the model's ability to discern patterns indicative of motor faults. Eqs. (5)-(8) represent the derivation of the two fundamental gates that comprise the GRU: the gate for the update (A) and the gate for the reset (r).

$A t=\sigma\left(W_a \cdot\left[h_{s-1}, x_t\right]+b_a\right)$               (5)

$r t=\sigma\left(W_r \cdot\left[h_{t-1}, x_t\right]+b_r\right)$                (6)

$\tilde{h} t=\tanh \left(W_a \cdot\left[r t * h_{t-1}, x_s\right]+b_a\right)$            (7)

$h t=(1-A t) * h_{n-1}+A t * \tilde{h} t$               (8)

Figure 3. Flowchart for proposed predictive maintenance of squirrel cage induction motors and three-phase synchronous motors

Figure 3 illustrates the comprehensive flowchart of the proposed predictive maintenance system for Squirrel Cage Induction Motors and Three-Phase Synchronous Motors. The flowchart delineates the systematic steps involved in the process, beginning with the deployment of industrial sensors to monitor critical parameters such as vibration, temperature, and current in real-time. The acquired data is then wirelessly transmitted to a central processing unit via ESP8266 and MQTT protocol [26]. Notably, the flowchart incorporates the innovative integration of Convolutional Neural Network-Gated Recurrent Unit architecture for fault classification. The CNN-GRU model optimally captures spatial and temporal features from the raw sensor data, enhancing the system's accuracy in identifying and categorizing motor faults. The flowchart further illustrates the user interface for visualization, facilitating real-time monitoring and maintenance decision-making. The depicted flowchart provides a clear roadmap of the proposed methodology, emphasizing the synergy of advanced sensor technology, wireless communication, and deep learning for effective predictive maintenance in industrial motor systems. The CNN layers extract spatial features from sensor signals, while the GRU layer captures temporal dependencies in the sequential data as mentioned in Table 3.

Table 3. Proposed model configuration

4.5 Experimental setup

The proposed CNN-GRU model was implemented using Python programming language with TensorFlow and Keras libraries. The training experiments were performed on a workstation with the following hardware configuration:

• Ubuntu operating system
• Intel Core i7 processor
• 16 GB RAM
• NVIDIA GPU acceleration

The model training process used the Adam optimizer, which is widely used for deep learning optimization due to its fast convergence and adaptive learning capability [27]. The batch size was set to 32, the learning rate of 0.002, and the model was trained for 50 epochs. To prevent overfitting and improve training stability, early stopping was implemented by monitoring validation loss with a patience of five epochs. Additionally, a learning-rate reduction strategy (ReduceLROnPlateau) was used to dynamically decrease the learning rate when the validation loss plateaued.

This study presented an IIoT-enabled predictive maintenance framework for squirrel cage induction motors and three-phase synchronous motors. The system integrates industrial sensors with an ESP8266-MQTT communication architecture and a CNN-GRU deep learning model for fault classification. Experimental results demonstrate that the proposed model achieves 98.94% accuracy, outperforming conventional machine learning methods. The framework enables reliable real-time monitoring and early fault detection in industrial environments. Future work will focus on validating the system in real industrial settings and exploring advanced deep learning architectures for improved predictive accuracy.

Declaration

• The authors have no financial or proprietary interests in any material discussed in this article.
• The authors have no relevant financial or non-financial interests to disclose.
• The authors have no competing interests to declare that are relevant to the content of this article.