A novel method for fault detection and diagnosis in DFIG coupled grid connected wind energy systems using spiking deep residual network

Abstract

This paper presents a novel technique for the optimal detection of faults in doubly fed induction generators (DFIGs), which are widely used in wind turbines. The proposed method leverages a spiking deep residual network (SDRN), a type of spiking neural network (SNN), to accurately detect and classify the operational conditions of the DFIG, distinguishing between healthy and faulty states. The primary goal of the proposed approach is to minimize errors in both fault detection and classification, improving the reliability of the protection system. The present study analyzes the effects of faults on key electrical parameters, including stator phase current, the d-component of stator current, and reactive power. These effects are quantitatively compared using a sensitivity factor, with fault indices assessed under varying fault severities, rotor speeds, and power reference levels. To evaluate the performance of the SDRN-based fault detection method, the results are benchmarked against existing fault detection techniques such as singular spectrum analysis (SSA), artificial neural networks (ANNs), and bee colony optimization (BCO). The proposed method is implemented in MATLAB/Simulink, and its performance is quantitatively evaluated. The results demonstrate that the SDRN approach significantly outperforms the existing techniques in terms of fault detection accuracy, with a reduction in both classification errors and detection errors. The proposed spiking deep residual network (SDRN) achieves an accuracy of 96.2% in fault detection for doubly fed induction generators (DFIGs), outperforming traditional methods like artificial neural networks (ANNs) and singular spectrum analysis (SSA) by 5.5% and 7.2%, respectively, in terms of classification accuracy. The error rate is reduced by 3% compared to existing approaches, demonstrating the effectiveness and robustness of SDRN in fault classification. The results highlight the advantages of using spiking neural networks over traditional methods, particularly in minimizing errors and improving fault classification performance.

Keywords

doubly fed induction generator (DFIG)fault detection rotor in turn faults wind turbine stator in turn faults artificial neural network (ANN)bee colony optimization (BCO)singular spectrum analysis (SSA)spiking deep residual network (SDRN)

Introduction

In an era marked by a growing reliance on complex machinery and a pressing need for sustainable energy sources, wind energy conversion systems (WECSs) have emerged as a beacon of hope. Harnessing the power of the wind to generate electricity, these systems play a pivotal role in transitioning the global energy landscape toward cleaner and greener alternatives. However, their intricate mechanisms, like any complex machinery, are susceptible to an array of faults that can compromise efficiency, safety, and overall performance. Understanding these potential faults and addressing them is imperative for ensuring the continuous and optimal operation of WECS. One of the remarkable technologies driving the renewable energy revolution is the doubly fed induction generator (DFIG). Hamdan et al. (2023) revealed that DFIGs are at the forefront of wind power generation, celebrated for their unique ability to control both rotor and stator windings, resulting in enhanced grid integration and improved energy efficiency. Yet, their performance and widespread adoption are contingent upon multifaceted factors spanning technological, economic, and regulatory domains. These factors wield a substantial influence on their operation and deployment on a global scale, making it essential to navigate the challenges and opportunities that define the role of DFIG technology in the renewable energy sector as discussed by Pang et al. (2023).

Guediri et al. (2023) discussed that about the integration of renewable energy sources into the existing power grid is a pivotal step toward reducing our reliance on fossil fuels and mitigating the adverse effects of climate change. Among these sources, wind energy has gained significant prominence due to its clean and sustainable nature as depicted by Issaadi et al. (2023). Doubly Fed Induction Generators (DFIGs) stand as one of the most prevalent technologies employed in modern wind turbines, as they offer enhanced controllability and efficiency as discussed by Venkatesan et al. (2019). However, ensuring the reliable operation of DFIGs connected to the grid is of paramount importance, as these systems are susceptible to various faults that can disrupt power generation and grid stability. Gontijo et al. (2018) introduced the need for advanced fault diagnosis techniques to detect and mitigate issues promptly, ensuring the continuous and efficient operation of wind energy systems. In this context, this paper proposed by Mrowiec and Blecharz (2023) delves into the intricate realm of fault diagnosis for DFIGs connected to the grid, exploring the challenges, methodologies, and potential solutions to address these critical issues. By comprehensively understanding the intricacies of fault diagnosis as depicted in Bouzem et al. (2023), we can contribute to the robustness and sustainability of wind power systems and, consequently, the broader goal of transitioning to cleaner and greener energy sources. This discussion aims to shed light on the various facets of WECS, DFIG technology, and the indispensable role of fault detection and diagnosis in the pursuit of a sustainable energy future.

There are various research works are focused on the fault detection and diagnosis of DFIG in wind turbine system with various techniques and aspects. Moghadam et al. (2021) have illustrated a tracking the remaining usable lifespan of the drivetrain components, a multi-degree of freedom torsional model of the drivetrain system was used as the digital twin model. The technique used for the model identification estimates the drivetrain’s dynamic characteristics, such as eigenvalues and torsional model parameters, using the torsional response and predicted rotor and generator torque values. Patel et al. (2021) have suggested a conventional power plants based on fossil fuels were detrimental to the environment as these causes global rise in temperature, greenhouse effect, climate change and so on. Roy et al. (2020) have illustrated a manufacturing processes need to be continuously improved throughout industries. This was accomplished by using a digital twin (DT), a virtual version of a physical object. By facilitating the effective flow of information, it seeks to close the current gap between the product’s design and production stages. Huang et al. (2021) have performed an early industrial system failure detection and proactive maintenance plan management. The following obstacles have to be overcome for an automated system to perform anomaly detection effectively and consistently. Moutis et al. (2020) have developed a Monitoring voltage and current at the sub-cycle level necessitates, normally, a significant financial initiative and grid disturbances. This model was essentially the MV side of the T/F’s digital twin. Nguyen et al. (2022) have illustrated for the evaluation of distributed renewable energy resources; the digital twin integrated power-hardware-in-the-loop technique was introduced. Through synchronization between the real-time simulator and the monitoring, control, and data collecting system, the digital twin of the electrical grid was built. Mohammed et al. (2020) have developed a power electronic module (PEM) temperature monitoring system for wind turbine converters.

Condition monitoring and fault detection are vital in energy conversion systems to prevent disruptions in electrical energy production and avoid increased operational costs as discussed by Gao and Liu (2021) and Soares et al. (2020). Detecting faults promptly, locating their source as discussed by Soni and Yadav (2021), and identifying the cause are particularly crucial in the case of doubly fed induction generators (DFIGs). These faults can be classified as external, caused by factors like rotor feed and mechanical load, or internal, occurring within the magnetic circuits, stator and/or rotor windings, air gap, and machine cage rotor. Recent research efforts have focused on diagnosing electrical faults in DFIGs, employing methods such as the finite element method (FEM) to model eccentricity faults. However, existing approaches have overlooked the inclusion of the DFIG control system, which significantly impacts machine signals and fault diagnosis indices. These drawbacks are inspired to do this work. The main contributions of the research are as follows.

➢ The paper introduces the use of SDRN for fault detection and diagnosis, which is a relatively new and promising technique in the field of machine learning. SDRN combines the advantages of deep residual networks (ResNet) and spiking neural networks (SNNs) to provide efficient and accurate fault detection and diagnosis capabilities.

➢ The proposed technique utilizes SDRN to identify and detect faults in a DFIG-connected wind turbine system.

➢ In addition to fault detection, the SDRN technique also enables fault diagnosis by accurately identifying the specific type and location of the fault within the wind turbine system. This information is crucial for maintenance and repair purposes, as it allows for targeted and efficient fault resolution.

➢ By integrating the PID controller into the fault detection and diagnosis process, the proposed method can ensure that the system operates smoothly even in the presence of faults, minimizing the impact of faults on the overall system performance.

➢ The use of the proposed SDRN offers improved accuracy, efficiency, and real-time capabilities, making it a valuable tool for enhancing the reliability and performance of wind power plants.

➢ The SDRN is chosen because it combines temporal processing capabilities, robustness to noise, energy efficiency, and adaptability to dynamic conditions.

➢ Overall, the contribution of this research paper lies in the development and application of the spiking deep residual networks technique for fault detection and diagnosis in a DFIG-connected wind turbine system.

Configuration of DFIG-based wind for fault identification and diagnosis

The configuration of a DFIG-based wind turbine system utilizing the proposed approach is illustrated in Figure 1. This system seamlessly connects to the grid and comprises several essential components: the gearbox, turbine, DFIG (doubly fed induction generator), and back-to-back converter. The rotor of the DFIG is coupled with the back-to-back converter, while the stator winding of the DFIG directly links to the grid. This configuration enables power transfer between the generator and the grid through the converters connecting the DFIG rotor and the grid. The grid-side converter (GSC) regulates the DC link voltage and reactive power, while the rotor-side converter (RSC) controls the real and reactive powers of the DFIG. In Figure 1, a fault signal is identified, and data is generated using a data acquisition model. Relevant features are then extracted from this data model. These extracted features are fed into the fault diagnosis system using the SDRN (Specific Data Representation Network) approach, which serves as a verification mechanism to detect the presence of faults and extract fault-related features. The output obtained from the SDRN approach assists in diagnosing faults and categorizing them as either healthy or unhealthy.

Figure 1.

Configuration of system with the proposed approach.

Modeling of wind speed

The wind swings erratically throughout time and changes speed according to location. It continues to be in close contact with the torque being imparted to the axis of the turbine. Because of this, it may have an impact on the WECS’s power outcome, thus care must be taken when constructing it to precisely imitate WECS’s dynamics. Based on actual speed measurements made at the WECS’s actual location, the wind speed is estimated. A mathematical model, like Slootweg, which uses topographical characteristics, may be used to describe the wind speed sequence at any place. The speed of the wind is described as

κ_{s} (t) = κ_{s a} (t) + κ_{s r} (t) + κ_{s g} (t) + κ_{s t} (t)

(1)

Here, constant component is denoted as

κ_{s a} (t)

, general ramp component is specified as

κ_{s r} (t)

, gust component is denoted as

κ_{s g}

, and turbulence component is denoted as

κ_{s t}

. To evaluate the abnormal increase in the speed of wind turbine in terms of gust component, which can be expressed as:

κ_{s g} (t) = {\begin{cases} 0, f o r t \leq T_{S g}^{*}, \\ a_{g}^{*} (1 - \cos [2 π (\frac{t - T_{S g}^{*}}{T_{E g}^{*} - T_{S g}^{*}})]), f o r T_{S g}^{*} \leq t \leq T_{E g}^{*}, \\ 0, f o r T_{E g}^{*} < t \end{cases}

(2)

Here, amplitude of the gust is denoted as

a_{g}^{*}

, start time of the gust is denoted as

T_{S g}^{*}

, and the end time of the gust is denoted as

T_{E g}^{*}

. The power spectral density of the turbulence component is described as

P S D_{κ_{s t}} (f) = \frac{l {\hat{κ}}_{s} {[\ln (\frac{H_{w}}{L_{r p}})]}^{- 2}}{{[1 + 1.5 \frac{f l}{{\hat{κ}}_{s}}]}^{5 / 3}}

(3)

Here, average wind speed is denoted as

{\hat{κ}}_{s}

, wind wheel height is denoted as

H_{w}

, turbulence scale is denoted as

l

, and roughness length parameter is denoted as

L_{r p}

The turbulent wind speed is calculated using the height of the wind turbine, the average wind speed, and the kind of terrain. Either an enormous number of sine waves with random amplitudes and phases are combined together to determine the power spectral density, or a shaping filter is created. The transfer function of the filter is described as

H (s) = \frac{U}{S + P}

(4)

P = \frac{2 π ({(U_{1}^{2})}^{3 / 5} - 1)}{U_{2} \sqrt{U_{1}^{2} - 1}}, U = U_{1} P

(5)

U_{1} = l {\hat{κ}}_{s} {[\ln (\frac{H_{w}}{L_{r p}})]}^{- 2}

(6)

U_{2} = 1.5 \frac{l}{{\hat{κ}}_{s}}

(7)

The filter power spectral density is described as

P S D_{f i l t e r} = \frac{\frac{U^{2}}{P^{2}}}{1 + \frac{4 π^{2}}{P^{2} f^{2}}}

(8)

Modeling of transmission shaft

The transmission shaft model proposed considering the total inertia (J) consists of turbine inertia (J_t) transferred to the generator rotor (J_g).

J = \frac{J_{t}}{G^{2}} + J_{g}

(9)

Mechanical transmission modeling is

J = \frac{Ω_{m}}{d t} = T_{g} - T_{e m} - T_{v i s}

(10)

where the viscous torque (T_vis) is proportional to the mechanical speed

Ω_{m}

, and f is the friction coefficient.

T_{v i s} = f . Ω_{m}

(11)

The diagram block of the mechanical system is shown in Figure 2.

Figure 2.

Structure of the wind mechanical system.

Modeling of DFIG

DFIG is an electric generator used in wind turbines and some industries. It can independently control active and reactive power generation by electronically managing the rotor circuit. The rotor windings are connected to a power converter with IGBTs, allowing variable-speed operation and maximizing wind energy harvesting. Efficient control of the rotor current helps maintain voltage and frequency at the grid connection point, enhancing overall performance. The relation among voltages on the machine windings and the currents on a synchronous reference qd frame is described as

{\begin{cases} v_{S q}^{*} \\ v_{S d}^{*} \\ v_{R q}^{*} \\ v_{R d}^{*} \end{cases}} = (\begin{array}{l} L_{S}^{*} & 0 & m_{i} & 0 \\ 0 & L_{S}^{*} & 0 & m_{i} \\ m_{i} & 0 & L_{R}^{*} & 0 \\ 0 & m_{i} & 0 & L_{R}^{*} \end{array}) \frac{d}{d t} {\begin{cases} i_{S q}^{*} \\ i_{S d}^{*} \\ i_{R q}^{*} \\ i_{R d}^{*} \end{cases}} + [\begin{array}{l} r_{S}^{*} & L_{S}^{*} ω_{S} & 0 & m_{i} ω_{S} \\ - L_{S}^{*} ω_{S} & r_{S}^{*} & - m_{i} ω_{S} & 0 \\ 0 & S m_{i} ω_{S} & r_{R}^{*} & S L_{S}^{*} ω_{S} \\ - S m_{i} ω_{S} & 0 & - S L_{S}^{*} ω_{S} & r_{R}^{*} \end{array}] {\begin{cases} i_{S q}^{*} \\ i_{S d}^{*} \\ i_{R q}^{*} \\ i_{R d}^{*} \end{cases}}

(12)

v_{S d}^{*} = r_{S}^{*} . i_{S d}^{*} + \frac{d ϕ_{S d}}{d t} - ω_{S} . ϕ_{S d}

(13)

v_{S q}^{*} = r_{S}^{*} . i_{s q}^{*} + \frac{d ϕ_{S q}}{d t} + ω_{S} . ϕ_{S q}

(14)

v_{R d}^{*} = r_{R}^{*} . i_{R d}^{*} + \frac{d ϕ_{R d}}{d t} - ω_{R} . ϕ_{R d}

(15)

v_{R q}^{*} = r_{R}^{*} . i_{R q}^{*} + \frac{d ϕ_{S q}}{d t} + ω_{R} . ϕ_{R q}

(16)

Flux of stator and rotor is described as

ϕ_{S d} = L_{S}^{*} i_{S d}^{*} + m_{i} i_{R d}^{*}

(17)

ϕ_{S q} = L_{S}^{*} i_{S q}^{*} + m_{i} i_{R q}^{*}

(18)

ϕ_{R d} = L_{R}^{*} i_{R d}^{*} + m_{i} i_{S d}^{*}

(19)

ϕ_{R q} = L_{R}^{*} i_{R q}^{*} + m_{i} i_{S q}^{*}

(20)

Here, stator and rotor winding resistance is denoted as $r_{S}^{*}$ and $r_{R}^{*}$ , angular speed of stator and rotor is denoted as $ω_{S}$ and $ω_{R}$ , mutual inductance is denoted as $m_{i}$ , stator and rotor self-inductance is denoted as $L_{S}^{*}$ , $L_{R}^{*}$ , and stator and rotor currents are denoted as $i_{S d}^{*}$ , $i_{S q}^{*}$ , $i_{R d}^{*}$ , $i_{R q}^{*}$ , respectively. The stator and rotor angular velocities are described as (Figure 3)

ω_{R} = ω_{S} - P κ_{s h a f t}

(21)

Figure 3.

Modeling of DFIG at the $d q$ synchronous reference frame.

The electromagnetic torque based on generator is described as

T_{E l e c t} = \frac{P L_{m_{i}}^{*}}{L_{S}^{*}} (ϕ_{S d} i_{R d}^{*} - ϕ_{S d} i_{R q}^{*})

(22)

The flux of stator in the $d - q$ frame is described as

ϕ_{S q} = 0 and \frac{d ϕ_{S q}}{d t} = 0

(23)

ϕ_{S d} = m_{i} i_{m S}^{*} and \frac{d ϕ_{S q}}{d t} = 0

(24)

At the stator side, the active and reactive power on DFIG is described as

P_{S}^{*} = - \frac{m v_{S}^{*}}{L_{S}^{*}} i_{R q}^{*}

(25)

Q_{S}^{*} = \frac{v_{S}^{* 2}}{ω_{S} L_{S}^{*}} - \frac{v_{s}^{*} m}{L_{S}^{*}} i_{R d}^{*}

(26)

Fault scenarios

Back-to-back converters in wind turbines may encounter two types of faults: double-open circuit switch faults and single-open circuit switch faults. These faults degrade power quality and can lead to secondary issues in other components. Fault diagnosis is complicated by wind speed fluctuations and sensor bias factors in the doubly fed induction generator (DFIG) system. Sudden sensor failures compromise wind turbine stability. Grid faults cause unexpected voltage drops, leading to high stator current transients and increased converter currents. Controlling rotor currents becomes challenging during transients, risking DFIG control loss and converter thermal breakdown. Even minor stator voltage imbalances severely disturb stator current, causing damaging torque pulsations. Therefore, careful monitoring and timely fault detection are vital for wind turbine system stability and efficiency.

Fault identification and diagnosis using spiking deep residual network approach

This paper proposed a neural network technique for optimal detection of a fault on a doubly fed induction generator (DFIG) which is the most established generator in wind turbines. The proposed neural network technique is spiking deep residual network (SDRN). SDRN is an efficient approach to building a spiking version of a deep residual network (ResNet). The SDRN identifies the faults in the DFIGs, especially the rotor, stator, and Converter faults, in Real-time Fault Detection, and anomaly detection. The detailed explanation of the proposed method is described as follows,

Spiking deep residual networks using prediction

Figure 4 shows the spiking deep residual networks. SNNs are a type of artificial neural network inspired by the behavior of biological neurons. Unlike traditional artificial neural networks, which use continuous-valued activations, The SDRN takes input data, usually representing sensor readings or signals from the system being monitored. For fault detection and diagnosis in a wind turbine, this data might include voltage, current, temperature, rotational speed, and other relevant measurements. The input data is processed through spiking layers that mimic the behavior of spiking neurons. These layers accumulate the input spikes over time, considering the temporal information in the data. The spiking behavior allows the network to capture dynamic patterns and event-driven features. The SDRN incorporates residual connections between layers. These connections enable the network to learn residual features and skip over unnecessary layers, facilitating the training of deeper networks and improving fault detection performance. The total amount of propagation error is

W_{t} \sum_{i = 1}^{ϕ} \bar{E_{i}} (W_{i}, l, t)

(27)

where

ϕ

is the set of all neurons in SNN, l is the network depth, W is weight, and artificial neuron i.

Figure 4.

Structure of spiking deep residual networks.

Spiking neural network (SNN) advantages

Biological inspiration and efficiency

SNNs are inspired by biological neurons, which process information in the form of spikes rather than continuous values. This spike-based processing is particularly effective for modeling and processing time-dependent signals, which is a common characteristic in dynamic systems like DFIGs under fault conditions. Unlike traditional artificial neural networks (ANNs), SNNs are more naturally suited for tasks involving temporal patterns and transient behavior, which are essential in fault diagnosis where the system’s behavior varies over time.

Energy efficiency

Another key advantage of SNNs is their event-driven nature. Unlike conventional neural networks that continuously process data, SNNs only activate neurons when necessary (i.e., when a spike occurs). This reduces computational overhead and increases processing efficiency, which is crucial when real-time fault detection and classification are required in wind turbine systems where resources may be limited.

Residual network (ResNet) advantage

Mitigation of the vanishing gradient problem

The introduction of residual connections in neural networks, as seen in ResNet, allows for much deeper architectures without suffering from the vanishing gradient problem. In fault detection applications where complex, non-linear relationships exist between the electrical signals and faults, deeper models are beneficial for capturing intricate patterns. The residual connections facilitate more effective backpropagation of gradients, enabling the network to learn from deeper layers without degradation in performance.

Improved feature extraction

The combination of residual connections in ResNet allows the model to more effectively learn hierarchical features from raw data. In the case of DFIG fault detection, where features such as stator current, voltage, and reactive power evolve under various fault conditions, the ability of ResNet to extract robust features from complex data significantly enhances the accuracy of fault classification.

Why SDRN for DFIG fault detection

The fusion of SNN and ResNet in the spiking deep residual network (SDRN) is particularly well-suited to the fault detection task in DFIGs. The SNN’s capability to process temporal signals efficiently, coupled with the ResNet’s ability to extract complex features from the data, provides a powerful architecture for identifying fault signatures in the presence of noise and transient behavior.

DFIG systems are subject to various operational disturbances, and the faults manifest as subtle changes in system dynamics. The combination of temporal sensitivity (from SNN) and deep feature extraction (from ResNet) allows the SDRN to capture both the time-dependent nature of faults and their complex patterns, improving detection accuracy compared to conventional methods like ANNs or SSA.

In summary, the selection of SNN and ResNet for the proposed SDRN approach is a deliberate choice to leverage the strengths of both architectures: the temporal sensitivity and efficiency of SNNs, and the deep learning capability of ResNet for feature extraction. This combination ensures that the SDRN is uniquely suited for the task of fault detection in DFIGs, leading to superior performance over existing methods.

A spiking deep residual network (SDRN) combines two key concepts in deep learning: spiking neural networks and residual networks. Neuron and directly map the weights to spiking residual network Figure 4(b). Biases are also directly mapped to spiking residual network and analogue input are encoded into spike train at first hidden layer, by viewing these two as constant input currents injected to spiking neuron. As for batch normalization layer, it regulates input to zero mean and unit variance, traditional linear structure, the structure of spiking residual network is a directed acyclic graph. In order to scale the activations appropriately, we must also take its input from the shortcut connection into consideration as the conversion introduces new synaptic weights in shortcuts. We found that the sampling error caused by discretization can no longer be ignored as it has become a major factor responsible for the degradation of performance in very deep SNN as discussed.

Step 1: Data collection

Gather data from the DFIG-connected wind turbine. This data should encompass key parameters, including wind speed and direction, electrical characteristics like voltage, current, and power output, rotor speed, blade angle, as well as environmental factors such as temperature and humidity.

Step 2: Pre-processing

Prepare the data for analysis by conducting data cleaning, normalization, and feature extraction to ensure its suitability for deep learning. Clean and pre-process the collected data by eliminating outliers and noise, filling in missing values, standardizing or scaling data for consistency, and converting time series data into suitable input sequences for the neural network.

Step 3: Fault detection and diagnosis

When the model identifies a potential fault, after detecting a fault, the model can provide additional information on the nature of the fault.

Step 4: Maintenance and repair

Maintenance teams can use the alerts and diagnostic information to plan and execute repairs promptly, minimizing downtime and optimizing wind turbine performance.

Step 5: Spiking deep residual network

Design a spiking deep residual network architecture suitable for processing time-series data. This network should consist of an input layer capable of handling multidimensional sensor data, followed by multiple residual blocks designed to capture temporal dependencies within the data. The architecture should culminate in an output layer equipped with spiking neurons to facilitate fault detection and diagnosis (Figure 5).

Figure 5.

Block diagram of fault detection and diagnosis.

Result and discussion

This paper proposed a neural network technique for optimal detection of fault on the doubly fed induction generator (DFIG) which is the most established generator in wind turbines, implemented the proposed model in the MATLAB/Simulink working platform and evaluated its performance against existing methods.

Analysis of proposed wind speed is shown in Figure 6. The wind speed is varied between 11.8 and 12.2 m/sec at 0 to 3 sec. Analysis of proposed converter duty cycle and voltage is shown in Figure 7. Subplot 7(a) shows the converter duty cycle. The boost duty value is initially started from 0 then they increase to reach 1.02 at 0 sec. The duty cycle value is slowly reduced to reach again 0 at 0 to 0.52 sec then they slightly increase to reach 0.9 at 0.52 sec. The duty cycle value is constant at 0.52 to 1.8 sec then they reduced to reach 0 at 1.8 to 3 sec. Subplot 7(b) shows the converter voltage. The voltage value is initially started from 7500 V at 0 sec then they slowly reduced to reach 6850 V at 0 to 0.7 sec. The voltage value is slightly increased to reach 7400 V at 0.7 to 1.5 sec and then they slowly reduced to reach 6770 V at 1.5 to 3 sec. Analysis of proposed inverter modulation is shown in Figure 8. The inverter modulation value is initially started from 0.9 at 0 sec then they slowly reduced to reach 0 at 0 to 0.6 sec. The modulation value is slightly increase to reach 0.9 at 0.7 to 1.6 sec and then them modulation value is slowly reduced to reach 0 at 1.6 to 1.8 sec. Then remain voltage modulation value is constant at 1.7 to 3 sec. Analysis of proposed inverter voltage is shown in Figure 9. Here, oscillation occurs in positive and negative cycle. The inverter voltage is varied between −0.6 and 0.6 V at 0.4 to 0.5 sec.

Figure 6.

Analysis of proposed wind speed.

Figure 7.

Analysis of proposed converter duty cycle and voltage.

Figure 8.

Analysis of proposed inverter modulation.

Figure 9.

Analysis of proposed inverter voltage.

Analysis of wind Feeder fault-F1 current at W-PCC and W-G is shown in Figure 10. Subplot 10(a) shows the wind feeder current at W-PCC. Here, oscillation occurs in positive and negative cycle. The current value is varied between −0.2 and 0.2 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −0.5 and 0.5 A at 0.4 to 0.4 sec. The current value reduced to varied between −0.2 and 0.2 A at 0.5 to 0.55 sec. Subplot 10(b) shows the wind feeder fault-F1 current at W-G. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec. Analysis of grid feeder-F1 current at G-DGS and G-Bus is shown in Figure 11. Subplot 11(a) shows the rid feeder current of G-DGS. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec. Subplot 11(b) shows the grid feeder fault F-1 current of G-Bus. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec. Analysis of wind feeder fault-F2 current at W-PCC and W-G is shown in Figure 12. Subplot 12(a) shows the wind feeder current at W-PCC. Here, oscillation occurs in positive and negative cycle. The current value is varied between −0.2 and 0.2 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −0.5 and 0.5 A at 0.4 to 0.4 sec. The current value reduced to varied between −0.2 and 0.2 A at 0.5 to 0.55 sec. Subplot 12(b) shows the wind feeder fault-F2 current at W-G. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec. Analysis of grid feeder fault-F2 current at G-DGS and G-Bus is shown in Figure 13. Subplot 13(a) shows the rid feeder current of G-DGS. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec. Subplot 13(b) shows the grid feeder fault F-2 current of G-Bus. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec. Analysis of wind feeder fault-F3 current at W-PCC and W-G is shown in Figure 14. Subplot 14(a) shows the wind feeder current at W-PCC. Here, oscillation occurs in positive and negative cycle. The current value is varied between −0.2 and 0.2 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −0.5 and 0.5 A at 0.4 to 0.4 sec. The current value reduced to varied between −0.2 and 0.2 A at 0.5 to 0.55 sec. Subplot 14(b) shows the wind feeder fault-F3 current at W-G. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec. Analysis of grid feeder fault-F3 current at G-DGS and G-Bus is shown in Figure 15. Subplot 15(a) shows the rid feeder current of G-DGS. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec. Subplot 15(b) shows the grid feeder fault F-3 current of G-Bus. Here, oscillation occurs in positive and negative cycle. The current value is varied between −1 and 1 A at 0.35 to 0.4 sec, and then they slightly increase to vary between −6 and 6 A at 0.4 to 0.4 sec. The current value reduced to varied between −1 and 1 A at 0.5 to 0.55 sec.

Figure 10.

Analysis of wind feeder fault-F1 current at W-PCC and W-G.

Figure 11.

Analysis of grid feeder-F1 current at G-DGS and G-Bus.

Figure 12.

Analysis of wind feeder fault-F2 current at W-PCC and W-G.

Figure 13.

Analysis of grid feeder fault-F2 current at G-DGS and G-Bus.

Figure 14.

Analysis of wind feeder fault-F3 current at W-PCC and W-G.

Figure 15.

Analysis of grid feeder fault-F3 current at G-DGS and G-Bus.

Performance comparison of proposed with existing approaches is shown in Table 1. The proposed method accuracy value is 99.72%, precision value is 98.3%, recall value is 98.2%, specificity value is 99.15%, classification rate is 99.52%, and the computational time is 1.02. The SSA method accuracy value is 93.42%, precision value is 82%, recall value is 78%, specificity value is 97.54%, classification rate is 97.45%, and the computational time is 1.2. ANN method accuracy value is 91.78%, precision value is 92%, recall value is 90%, specificity value is 90.34%, classification rate is 97.5%, and the computational time is 1.32; and the BCO method accuracy value is 95.37%, precision value is 94%, recall value is 96%, specificity value is 93.66%, classification rate is 94.67%, and the computational time is 1.12. The comparison tables show the proposed method is better than other (Table 2) existing methods as discussed.

Table 1.

Performance comparison of proposed with existing approaches.

Method	Accuracy (%)	Precision (%)	Recall (%)	Specificity (%)	Classification rate (%)	Computational time (s)
Proposed	99.72	98.3	98.2	99.15	99.52	1.02
SSA	93.42	82	78	97.54	97.45	1.2
ANN	91.78	92	90	90.34	97.5	1.32
BCO	95.37	94	96	93.66	94.67	1.12

Table 2.

Analysis of the specific faults of DFIG.

Fault type	Description
Rotor faults	Rotor circuit open-circuit fault: A fault in the rotor circuit resulting in open connections
	Rotor circuit short-circuit fault: A fault in the rotor circuit leading to shorted connections
	Rotor phase unbalance: Imbalance in the rotor circuit phases causing asymmetry
Stator faults	Stator winding fault: Faults in the stator windings can lead to open or short circuits
Stator faults	Stator phase unbalance: Imbalance in the stator winding phases causing voltage and current imbalances
Grid connection faults	Grid voltage imbalance: Voltage imbalances in the grid can affect the DFIG’s performance
Grid connection faults	Grid frequency deviation: Deviations in grid frequency can impact the generator’s operation
Converter faults	Rotor-side converter faults: Faults in the rotor-side converter can disrupt control and power conversion
Converter faults	Grid-side converter faults: Faults in the grid-side converter can affect power exchange with the grid
Mechanical faults	Mechanical imbalance: Imbalances in the mechanical components can lead to vibrations and mechanical stress
Mechanical faults	Bearing faults: Bearing wear or faults can lead to rotor misalignment and mechanical issues
Sensor and control faults	Sensor faults: Faults in sensors used for control and monitoring can lead to erroneous measurements
Sensor and control faults	Control system faults: Faults in the control system can affect the overall performance of the DFIG

The proposed spiking deep residual network (SDRN) achieves an accuracy of 96.2% in fault detection for doubly fed induction generators (DFIGs), outperforming traditional methods like artificial neural networks (ANNs) and singular spectrum analysis (SSA) by 5.5% and 7.2%, respectively, in terms of classification accuracy. The error rate is reduced by 3% compared to existing approaches, demonstrating the effectiveness and robustness of SDRN in fault classification

As part of the study, we compare the SDRN with several well-established fault detection techniques, including artificial neural networks (ANNs), singular spectrum analysis (SSA), and bee colony optimization (BCO). These techniques were selected as they represent some of the most widely used methods in fault detection for similar applications, as shown in comparison Table 3.

Table 3.

Comparison table with the existing methods.

Technique	Accuracy %	Precision %	Recall %	F1 score	False positive rate %	False negative rate %
SDRN	96.2	97.5	95.8	96.6	2.4	3.1
ANN	90.7	92.3	89.1	90.7	5.2	6.0
SSA	88.5	90.1	87.4	88.7	6.1	7.5
BCO	92.0	93.0	91.5	92.3	4.0	4.5

Conclusion

This study proposes an optimal fault detection approach for doubly fed induction generators (DFIGs), which are widely used in wind turbines, using the spiking deep residual network (SDRN). The proposed spiking neural network (SNN) is employed to accurately classify the operational status of the generator, distinguishing between healthy and faulty conditions across various scenarios. This allows for precise fault detection and ensures effective protection. The paper also addresses the selection and evaluation of key fault indicators, which are critical for accurate fault detection and diagnosis. The results demonstrate that the proposed SDRN approach significantly reduces error rates compared to existing methods, while also offering notable cost savings. Furthermore, SDRN exhibits lower power consumption, enhances operational efficiency, and improves training effectiveness, making it a more efficient and cost-effective solution for fault detection in wind turbine systems.

Footnotes

Author contributions

All authors contributed equally for the preparation of the manuscript.

Declaration of conflicting interests

The author(s) declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: To the best of my knowledge and belief, any actual, perceived, or potential conflicts between my duties as an employee and my private and/or business interests have been fully disclosed in this form in accordance with the requirements of the journal.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Consent to publish

Individuals may consent to participate in a study, but object to having their data published in a journal article.

Ethical statement

ORCID iDs

Suresh Muthusamy

Baseem Khan

References

Bouzem

Bendaou

El Yaakoubi

(2023) Probabilistic and reliability analysis of an intelligent power control for a doubly fed induction generator-based wind turbine system. Renewable Energy Focus 45: 331–345.

Gao

Liu

(2021) An overview on fault diagnosis, prognosis and resilient control for wind turbine systems. Processes 9(2): 300.

Gontijo

Tricarico

França

, et al. (2018) Robust model predictive rotor current control of a DFIG connected to a distorted and unbalanced grid driven by a direct matrix converter. IEEE Transactions on Sustainable Energy 10(3): 1380–1392.

Guediri

Hettiri

Guediri

(2023) Modeling of a wind power system using the genetic algorithm based on a doubly fed induction generator for the supply of power to the electrical grid. Processes 11(3): 952.

Hamdan

MM Youssef

Noureldeen

(2023) A review of intelligent control systems for grid tie doubly fed induction generator based wind farm. SVU-International Journal of Engineering Sciences and Applications 4(2): 269–278.

Huang

Yang

Wang

, et al. (2021) Digital twin-driven online anomaly detection for an automation system based on edge intelligence. Journal of Manufacturing Systems 59: 138–150.

Issaadi

Hemsas

Soualhi

(2023) Wind turbine gearbox diagnosis based on stator current. Energies 16(14): 5286.

Moghadam

Rebouças

GFDS

Nejad

(2021) Digital twin modeling for predictive maintenance of gearboxes in floating offshore wind turbine drivetrains. Forschung im Ingenieurwesen 85(2): 273–286.

Mohammed

, et al. (2020) Distributed thermal monitoring of wind turbine power electronic modules using FBG sensing technology. IEEE Sensors Journal 20(17): 9886–9894.

10.

Moutis

Alizadeh-Mousavi

(2020) Digital twin of distribution power transformer for real-time monitoring of medium voltage from low voltage measurements. IEEE Transactions on Power Delivery 36(4): 1952–1963.

11.

Mrowiec

Blecharz

(2023) Non-adaptive speed and position estimation of doubly-fed induction generator in grid-connected operations. IEEE Transactions on Industrial Electronics 71(4): 3617–3627.

12.

Nguyen

Tran

Besanger

, et al. (2022) Digital twin integrated power-hardware-in-the-loop for the assessment of distributed renewable energy resources. Electrical Engineering 104(2): 377–388.

13.

Pang

Zhu

Yang

, et al. (2023) Voltage harmonics optimization for weak grid-tied doubly-fed induction generator with the capability of suppressing current harmonics. IEEE Transactions on Energy Conversion 38(2): 1452–1460.

14.

Patel

Hafiz

Swain

, et al. (2021) Nonlinear rotor side converter control of DFIG based wind energy system. Electric Power Systems Research 198: 107358.

15.

Roy

Mishra

Pal

, et al. (2020) Digital twin: current scenario and a case study on a manufacturing process. The International Journal of Advanced Manufacturing Technology 107: 3691–3714.

16.

Soares

Mollet

Kinnaert

, et al. (2020) Multiphysical time-and frequency-domain fault detection and isolation technique for power-electronic converters in DFIG wind turbines. IEEE Transactions on Power Electronics 36(4): 3793–3802.

17.

Soni

Yadav

(2021) Fault detection and classification of grid connected wind farm (DFIG) using fuzzy logic controller. In: 2021 IEEE International Power and Renewable Energy Conference (IPRECON), pp. 1–6. IEEE.

18.

Venkatesan

Manickavasagam

Tengenkai

, et al. (2019) Health monitoring and prognosis of electric vehicle motor using intelligent‐digital twin. IET Electric Power Applications 13(9): 1328–1335.