A Deep Hybrid 1D-CNN-LSTM Framework for Mechanical Fault Detection in Switchgear Using Time-Frequency Domain Analysis

The suggested research will follow a series of phases, each aimed at the determination of the mechanical defects within MV switchgear by virtue of both time-like and frequency-like analysis. The first need is to develop an extensive dataset that is measured in situ: the dataset would include mechanical and non-mechanical arcing, corona, and tracking in addition to samples of relatively normal operating conditions.

A hybrid 1D-CNN-LSTM model is employed to derive patterned insights, which is a DL model specially trained to handle ultrasonic recordings made during fault occurrences and improvisation in audio format. Figure 1 presents a schematic representation of the consequent classification system of a mechanical fault.

Figure 1. Diagram illustrating the overall research approach

The data-collection part of the work involved the collection of raw distribution data from the seven states of Peninsular Malaysia, namely Kedah, Kuala Lumpur, Melaka, Selangor, Perak, Negeri Sembilan, and Johor, through the power utility company (PUC). MATLAB was used to achieve the subsequent processing. The choice of the hybrid model was driven by the fact that it is one of the DL techniques that is highly applicable to the problem.

The data preprocessing and processing via MATLAB were performed to initiate the workflow that continued with the development process of a neural network using Google Colab. Airborne Ultrasonic Testing (AUT) equipment is utilized as an essential tool to detect surface Partial Discharges (PD) and as such is useful as a diagnostic indicator of ineffective insulation of the electric circuit. As such, the source data that was studied was in the form of audio files, usually in the format of MP3, MPEG, or WAV. The following paragraphs explain how each step of the methodology described in Figure 2 took place.

Figure 2. Airborne ultrasonic test equipment visualization [47]

2.1 Data collection

The conducted research uses a combined data set, which combines switchgear recordings acquired earlier with a new ultrasonic recording set. The previous papers [11, 12, 41], provided comprehensive accounts of the data collection process, preprocessing procedure, and instrument strategy. This generated corpus is the complete set of ultrasonic signal traces of various operating states that include arcing, corona discharge, tracking events, mechanical failures, and normal operation, recorded in both the time and frequency domains.

Table 1. Provides an overview of the datasets, encompassing both mechanical and non-mechanical faults

It is built so that real-life operating conditions are reflected, and thus there is a balance in the dataset in terms of faulty and normal switchgear operating modes. Because of this variety, diagnostic models are more able to be trained and evaluated, and they are also better able to discriminate between various sorts of errors. A large-scale data set that is both believable and practical was prioritized so that it may be utilized to validate trustworthy and computationally efficient fault-detecting strategies in MV switchgear configurations. Investigating the use of DL approaches, particularly hybrid tools like the 1D-CNN-LSTM model, is especially crucial. The goal of this design is to improve performance in the face of different types of faults. There was a significant manifestation of this into the real world since the dataset was configured in a way that was near to the actual task. The dataset has been summarized in Table 1, representing both cases of mechanical and non-mechanical faults in the time domain as well as in the frequency domain.

2.2 Data pre-processing

At this stage, a highly structured algorithm of data processing was developed to make the data accessible to further studies. The selection of features took place through the usage of the Mel spectrogram technique. Along with the inclusion of new classes—including arcing, corona, and tracking—and the synthesis of mechanical faults, regular events have also been introduced. Training, validation, and testing were the three stages that the dataset went through once this data synthesis was finished. In order to provide a seamless dataset integration into the suggested hybrid 1D-CNN-LSTM model, this data processing step was necessary.

The aspect of data preprocessing is an essential part of DL projects and includes a step of procedures that should be carried out to ensure the input data are of high quality and provides a relevant interpretation of the data. This involves the elimination of noise, variable normalization and treatment of anomalies and missing values. The sequencing technique is a network of steps, that are interdependent:

1. Data Cleaning: The data were thoroughly audited, and this involved correcting any data that was missing or erroneous in order to maintain data integrity.

2. Data Normalization: Using standardization of variables, biases caused by scale were reduced and balance was achieved during model training.

3. Feature Engineering: Extraction of meaningful characteristics out of the raw data was considered necessary and held significance when it comes to intelligible operational representations of switchgear and to improve the performance of the model.

4. Sequence Development: Data sequences with successive and equal length data sequences allowed the LSTM network to discover repetitive trends across days. The stages will give a strong basis to the training of the proposed hybrid 1D-CNN-LSTM DL model that exhibits a fine predictive power to detect mechanical failures in MV switchgear. The offered methodology, therefore, contributes greatly to the establishment of efficient fault-detection systems, and the direct way of application could be related to regular switchgear maintenance and general efficiency.

Moreover, the experimental data used in the research was obtained through ultrasonic sensing procedures that were aimed at measuring acoustic emissions caused by mechanical faults in MV switchgear [41]. The ultrasonic sensors were chosen by the research team because it is more sensitive to partial discharges, arcing phenomena and mechanical defects, which produce high-frequency acoustic signals. These were mounted on the switchgear enclosure in fixed, repeatable positions to ensure that they received a consistent, reliable signal under all operating conditions.

The records of ultrasonic signals were taken at a fixed sampling frequency in the ultrasonic frequency that was selected to ensure that the high-frequency components associated with mechanical faults were well represented, as well as to allow easy time-frequency domain analysis. Data collection was conducted in controlled operating conditions that modeled realistic operating conditions of switchgear, including constant voltage levels and reproducible mechanical conditions. Noise and external disturbances were reduced when recording so that the quality of the signal was not affected.

Different recordings were taken under each operating condition, including normal and faulty conditions, to increase the robustness of data and minimize randomness. Before model training the obtained ultrasonic signals were divided into constant-length samples and were normalized. These preprocessed signals were further converted into time-frequency representations and utilized as inputs into the proposed hybrid 1D - CNN -LSTM framework.

2.3 Extract features by using "Mel spectrogram"

The Mel spectrogram is an easy-to-understand evaluation tool for practitioners working in the field of time domain analysis; it may be used to ultrasonic recordings of switchgear components captured under different fault states in order to identify useful characteristics. Transposing a discrete time-domain signal to a two-dimensional matrix, the Mel spectrogram shows the visual representation where the spectrum content is vertically shown and the time progression is horizontally represented [48].

To achieve this transformation, a series of Mel filters is used, which in fact reproduces the human ear's nonuniform frequency sensitivity. This enables precise timing of mechanical failure patterns to be determined by using Mel filters to boost certain frequency components in the signal. Ultrasound waveforms that change over time may be represented using the Mel Spectrogram in a way that incorporates both the frequency content and a temporal variation. Such an approach allows a successful capture of the frequency components in the process of changing over the signal duration. With this resulting Mel spectrogram, thus, complex fault-related patterns would not have otherwise been identified in the original time-domain alone [49, 50].

Due to these reasons, it is best to transform the Mel spectrogram into enriched data providing input to the proposed hybrid 1D-CNN-LSTM model, which, in its turn, will increase its ability to differentiate between the various types of faults by relying on their distinctive frequency patterns. Besides, the fact that the Mel spectrogram is implemented in the architecture used in the proposed fault detection system makes the frequency-domain analysis component a serious addition to the time-domain analysis core, which cannot be omitted and would serve as the addition to the recent defects recognized [51, 52].

Through translating ultrasound signals in the time domain into spectral display in the frequency domain with the Mel Spectrogram, we are able to see the characteristic patterns of the frequencies of a fault. Such transformation allows extracting spectral features that assist in a model discerning between the different kinds of faults. Under the frequency space, the Mel spectrogram could be used as input to a convolutional neural network with LSTM to drive prediction and classification of mechanical faults into their characteristic spectral signatures.

Figure 3 shows the Mel spectrogram aspect of the mechanical faults. The current mechanism couples time and frequency directions and is therefore accurate and robust. As a method of extracting features, the Mel spectrogram has been proven to enhance the functioning of the hybrid model and enhance production of the mechanical faults in switchgear systems.

Figure 3. The Mel spectrogram for mechanical faults in both time and frequency domains

The Mel spectrogram is more discriminating on the type of faults by highlighting frequency bands that are more perceptually relevant and which fault-induced variation of energy is most apparent. This code gives inter-class separability by modeling specific time-frequency images of different mechanical faults and, at the same time, reducing high-frequency noise sensitivity. Therefore, the model proposed will have stronger and more discriminatory feature learning.

2.4 Concatenate the mechanical fault data with the non-mechanical fault data

The subsequent step involved the integration of data sources, merging mechanical fault data with instances with occurrences of non-mechanical faults such as arcing, corona, and tracking, as well as normal cases. This all-inclusive integration enabled wholehearted examination of a broad sector of the possible operating conditions and fault manifestations.

The need for a full and balanced training set drove this integration. We subjected the model to many real-world patterns and variations by including mechanical and non-mechanical faults. Through this experience, the model learned to distinguish between normal and unhealthy behavior, improving its diagnostic accuracy. Even with new parameters, the model's versatility in handling arcing, corona, tracking, and normal instances enabled it to identify even the tiniest departure from normal behavior. Thus, the Hybrid 1D-CNN-LSTM model has mastered the complexity of mechanical and non-mechanical fault patterns, advancing analysis.

2.5 Dataset classification

Normalization and scaling of the input variables are carried out as part and parcel of the data consolidation process to guarantee uniformity of the data and are done to improve the quality of the samples. Such a preprocessing step increases the generalization ability of the proposed hybrid model, allows making the learning process more efficient, and adds to the prediction accuracy. After preprocessing, the data is divided into three subsets, namely training, validation, and testing. Randomized stratified data splitting is used to ensure that all fault classes are proportionately represented in the subsets. Particularly, 70% of the data is used to train, and 15% to validate and test data, respectively.

The same partitioning protocol is used in all the experiments to eliminate the possibility of data leakage as well as have a fair and replicable performance evaluation. At the training stage, the model learns discriminative time-frequency characteristics of ultrasonic MV switchgear signals, and thus it can effectively detect the signature of mechanical faults. The exposure of the model to different operating conditions during training further enhances the ability of the model to distinguish mechanical and non-mechanical events during switchgear operation.

2.6 Hybrid 1D-CNN-LSTM model

In this paper, we suggest the introduction of a new framework to be applied to fault detection and differentiation in switchgear: the Hybrid 1D-CNN-LSTM new model, which will be able to handle features that are the result of such a method as the Mel Spectrogram. Improved fault type and identification accuracy and resilience are achieved by the use of convolutional and neural networks, which enhance the system's capacity to detect intricate spatial and temporal patterns of electrical signals.

Both mechanical and non-mechanical faults constitute the dataset, which was processed through several important stages of data processing that were aimed at isolating the two classes of faults. These steps were carried out at the preprocessing stage and allowed proper preparation of the datasets before feeding them to the fault detection model.

• Input Layer

The architecture will start with an input layer where the time-domain features of the electrical signal in the form of a Mel spectrogram representation have already been computed. Through the Short-Time Fourier Transform (STFT) of the raw time-domain data, each Mel Spectrogram provides a succinct overview of its original signal spectrum along time-axis boundaries and therefore imparts a representation to the model that condenses both the time and the spectrum dimensions.

• 1D-CNN Module

A 1D-CNN block plays a key role in the process of spatial feature extraction by the model [35]. Applied in the form of a cascade of 1D convolutional, each convolutional filter is an optimized set of filters designed to recognize local patterns of Mel Spectrogram inputs. Following each convolution, there is a max pooling layer that shrinks feature maps by selecting the highest value in small local regions so that the computational cost is reduced and the risk of overfitting is combated. The mathematical background of such an operation is expressed in the following equations:

$X_i=\operatorname{Conv1D}\left(H_{i-1}, W_i\right)+b_i$     (1)

where, $H_{i-1}$ is the input feature map of the ($i-1$) - th layer, $W_i$ is the weight tensor of the $i$ - th and $b_i$ is the bias vector. The output of the $(i-1)-t h$ convolutional layer after applying ReLU activation is given as:

$A_i=\operatorname{ReLU}\left(X_i\right)$    (2)

Subsequently, the Maxpooling layer downsamples the feature maps to obtain $Y_i$:

$Y_i=\operatorname{Maxpool}\left(A_i\right)$        (3)

Finally, the downsampled spatial features $Y_i$ are passed through the activation function to obtain $H_i$:

$H_i=\operatorname{ReLU}\left(Y_i\right)$   (4)

All convolutional layers in the proposed 1D-CNN model have an iterative process that is initiated by transforming the Mel spectrograms to a series of feature vectors. The 2D convolution of these vectors takes place with the spatial extent of the kernel being mostly determined by the layer it is contained in: a kernel size of 1 × 1 is used in the first layer, with a later, wider kernel size of 2 × 2 used in all the other layers. After every convolution, max-pooling is done, and the downsampling burst pressure is done across the time as well as spectral axes.

The result of this layer, which is known as H, is the input of the next convolutional layer. The same is done to all the convolutional layers, giving rise to the production of H-final, which is the final output of the network that carries codes of the significant spatial characteristics of the original Mel spectrograms. The downsampled spatial features are routed to the next LSTM module, where they form the basis of capturing the complexity of temporal dependencies that are inherent in the electrical signals. Figure 4 provides a full visual description of the architecture of the 1D-CNN, given the time and frequency domains.

Figure 4. The representation of the 1D-CNN model in both the time and frequency domains

• LSTM Module

In the proposed neural architecture, 1D-CNN outputs whose features have been downsampled will be fed to an LSTM module whose purpose is to train the model to learn temporal dependencies within the Mel spectrogram sequence. This is because each of the LSTM cells in the module is used to model the time-series nature of the data in order to recognize patterns that are associated with non-mechanical and mechanical faults over time [53]. As follows, the equations that describe the work of LSTM cells (5)-(9) are performed:

$f_t=\sigma\left(W_f \cdot\left[h_{t-1}, x_t\right]+b_f\right)$     (5)

$i_t=\sigma\left(W_i \cdot\left[h_{t-1}, x_t\right]+b_i\right)$   (6)

$o_t=\sigma\left(W_o \cdot\left[h_{t-1}, x_t\right]+b_o\right)$  (7)

$c_t=f_t \odot c_{t-1}+i_t \odot \tanh \left(W_c \cdot\left[h_{t-1}, x_t\right]+b_c\right.$ (8)

$h_t=o_t \odot \tanh \left(c_t\right)$        (9)

where, $x_t$ is the input at time step $t, h_t$ is the hidden state at time step $t, c_t$ is the cell state at time step $t, W_f, W_i, W_o, W_c$ are the weight matrices, $b_f, b_i, b_o, b_c$ are the bias vectors, and $\sigma$ is the sigmoid function. Figure 5 depicts the LSTM model representation in the time and frequency domains.

Figure 5. The representation of the LSTM model in both the time and frequency domains

• Fully Connected Layer

Upon capturing the temporal dependencies within the LSTM, the final hidden representation is obtained, which is termed H-final and passed through a fully connected layer. This layer carries out a weighted combination of the hidden state, which is instantaneously activated through the SoftMax operation. In so doing, the probabilities expected to be encountered with respect to the different fault types in each of the samples to be considered are derived.

The SoftMax function lowers the outcome to make the possibilities adjusted to the sum of unity and hence make an inclusive probability distribution encompassing the expanse of fault classes.

• Training and Optimization

The current study explores how well the hybrid 1D-CNN-LSTM model performs after training using an aggregate data set that contains both non- and mechanical fault cases. The categorical cross-entropy loss is the key training metric that will be computed during the training phase by taking the difference between predicted probability distributions and known fault labels. The process of parameter refinement is performed by means of backpropagation, where the computed gradients are treated as feedback to gradient-based optimization.

To optimize the process, the Adam stochastic optimization algorithm is employed, and a well-furnished learning rate of 0.0001 is used. Also, epochs of 60 are run on each pair of faults and normal cases, i.e., a batch size of 16 samples is performed per epoch. The choice of Adam is beneficial due to its dynamical scaling of each of the learning rates based on historical gradients. This kind of adversative behavior leads to increased efficiency and quicker convergence in comparison with classical stochastic gradient descent (SGD).

The model can be very appealing to a diverse rich set of mechanical and non-mechanical fault scenarios since the unified dataset was used along with the powerful optimization strategy, which boosts the ability of the Hybrid 1D-CNN-LSTM model. This enhanced capability further increases the model's ability to generalize, and consequently, this model assists in the classification of a wide range of switchgear-related faults in an effective and correct manner across the whole system. The use of coding was completed on Google Colab.

• Evaluation and Performance Metrics

A range of performance measures was utilized to perform an arduous comparison of the performance of the suggested model, specifically the accuracy, precision, recall, and F1-score. It was tested in a systematic manner to determine its suitability to detect and discriminate different fault types in switchgear systems. In particular, the addition of the Mel Spectrogram features to the Hybrid 1D-CNN-LSTM architecture granted the compound approach the ability to make use of the temporal and spectral advantages present within electrical signals.

This synergetic merger gave rise to a method capable of providing true and repeatable diagnostics in the switchgear space. The Hybrid 1D-CNN-LSTM model was the only one that perfectly combined the spatial capability feature extraction of the 1D-CNN with the temporal consideration of the LSTM to provide an overall better result and stability in classifying the mechanical faults and non-mechanical faults. These findings are demonstrated in Figure 6, which shows the representation of the Hybrid 1D-CNN-LSTM model in both time and frequency universes.

A flawless testing data set that fully covers a wide array of fault categories was also developed and meticulously curated to include arcing, corona, tracking, mechanical faults, and normal cases. The model was tested in full. Quantitative evaluation of its performance showed that it could discriminate among these various diagnostic classes at a rate higher than 90%. These findings prove the feasibility of the hybrid 1D-CNN-LSTM model in practice concerning switchgear fault detection and promote the use of the model in preference to the alternative approaches in this field, classifying it as an innovative one.

Figure 6. Illustrates the representation of the hybrid 1D-CNN-LSTM model in both the time and frequency domains

While the general accuracy is rather good, there are instances when the performance falls short of 90%. This indicates that the model is not yet fully developed and needs further work. In reality, thorough examination of the testing dataset leads to the conclusion that the proposed model is useful and adaptable, while simultaneously producing the essential insights for enhancing switchgear performance and stabilizing power systems. Consequently, the model's future as a crucial decision-making tool in the switchgear fault detection and management sector will be highlighted by the accomplishment of a high accuracy rate that is unusual.