Deep Residual Neural Network for EMI Event Classification Using Bispectrum Representations

Mitiche, Imene; Jenkins, M.; Boreham, Philip; Nesbitt, A.; Stewart, Brian G.; Morison, Gordon

doi:10.23919/eusipco.2018.8553177

Cited by 11 publications

(7 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The second stage solved a multi-class problem of 8 categories which resulted in a light performance drop when compared to stage 1. However, this performance is considered high and sufficient for a real-world multi-class problem and is in line with the results obtained in [15] and [16]. Note that the method in [16] used a sub-set of data that is used in this paper, this sub-set consists of balanced data classes.…”

Section: Resultssupporting

confidence: 84%

See 1 more Smart Citation

1D‐CNN based real‐time fault detection system for power asset diagnostics

Mitiche

Nesbitt

Conner³

et al. 2020

IET Generation, Transmission & Distribution

Self Cite

View full text Add to dashboard Cite

Electromagnetic Interference (EMI) diagnostics aid in identifying insulation and mechanical faults arising in High Voltage (HV) electrical power assets. EMI frequency scans are analysed to detect the frequencies associated with these faults. Timeresolved signals at these key frequencies provide important information for fault type identification and trending. An end-to-end fault classification approach based on real-world EMI time-resolved signals was developed which consists of two classification stages each based on 1D-Convolutional Neural Networks (1D-CNN) trained using transfer learning techniques. The first stage filters the in-distribution signals relevant to faults from out-of-distribution signals that may be collected during the EMI measurement. The fault signals are then passed to the second stage for fault type classification. The proposed analysis exploits the raw measured timeresolved signals directly into the 1D-CNN which eliminates the need for engineered feature extraction and reduces computation time. These results are compared to previously proposed CNN-based classification of EMI data. The results demonstrate high classification performance for a computationally efficient inference model. Furthermore, the inference model is implemented in an industrial instrument for HV condition monitoring and its performance is successfully demonstrated in tested in both a HV laboratory and an operational power generating site.

show abstract

Section: Resultssupporting

confidence: 84%

“…Previously the authors of this paper proposed deep learning methods based on 2D‐CNN for EMI signal classification [15, 16]. This work provided highly accurate results at the expense of high computational complexity, due to signal transformation and feature extraction using the bispectrum and the 2D‐CNN for classification.…”

Section: Introductionmentioning

confidence: 99%

1D‐CNN based real‐time fault detection system for power asset diagnostics

Mitiche

Nesbitt

Conner³

et al. 2020

IET Generation, Transmission & Distribution

Self Cite

View full text Add to dashboard Cite

show abstract

“…For data-based classification, In [30], authors combined the method of ''voting by comparing the polarizability with the library data'' and the method of ''Bayesian statistics on features extracted from the polarizabilities'' to generates an ordered dig list. Mitiche et al mapped time-domain signal to a Bispectrum image and classified the high-pressure discharge signals by using a deep residual neural network to classify the Bispectrum images [31]. Ammari et al identified conductive objects by extracting geometric features from the induction data and matching the features data for known objects from a given dictionary [14].…”

Section: Related Workmentioning

confidence: 99%

Sensitivity Analysis and Classification Algorithms Comparison for Underground Target Detection

Duan

Wan

et al. 2019

IEEE Access

View full text Add to dashboard Cite

Underground target detection technology has been widely used in urban construction and resource exploration. With the development of industrial modernization, the demand for underground target detection is becoming more specific, such as the material and shape of underground targets. Therefore, it is necessary to classify the properties of underground targets. In this paper, sensitivity analysis was performed on the spheroid model and the approximate forward model at first, and the influence of the target properties on the model output is obtained. Secondly, we utilized the fitting algorithm to obtain the model parameters of the simulation data (model response of targets with varying shapes and materials), and analyzed the influence of the fitting algorithm on the classification results at different SNR. Finally, eight machine learning algorithms: support vector machine(SVM), neural network(NN), quadratic discriminant analysis (QDA), Gaussian process (GP), decision tree (DT), random forest (RF) and AdaBoost were used in this study to compare the obtained results. From the above analysis, we found that the shape (radius) have a greater influence on the model than the material (permeability) in the spheroid model. According to the approximate forward model, we found that it is not feasible to classify targets when the orientation is unknown. The influence of the fitting algorithm on the classification performances is related to the noise level. The obtained results using neural network demonstrated that the proposed method outperformed in material-based classification and shape-based classification. In the material-based classification, the classifier generally has a weaker ability to distinguish between permeable materials.

show abstract

“…The existing classification strategies for underground metal targets mainly focus on two aspects: model-based methods [17], [20], [21] and data-based methods [22]- [24]. The model-based method needs to establish a forward model [25] and obtains the properties of the target through inversion.…”

Section: Introductionmentioning

confidence: 99%

“…In terms of the mapping relationship, abundant classification models have been applied for data-based classification. These methods include machine learning (ML) models such as support vector machine (SVM), artificial neural networks (ANN), random forest, AdaBoost algorithm, and so on [21], [22], [26]- [29]. Also, pattern matching methods, such as comparing the extracted features of targets with a given dictionary [30] and voting scheme which compares field data polarizabilities against templates in the library [31], are frequently employed for classification.…”

Section: Introductionmentioning

confidence: 99%