Energy theft and defective meters not only lead to non-technical losses (NTLs) that are extremely detrimental to energy distributors and power infrastructure, but NTLs are also a major cause of damages to electricity and massive losses of revenue each year. Automatic meter reading (AMR) is a system used by the Provincial Electricity Authority (PEA) of Thailand for NTLs detection that works in conjunction with physical inspection. At present, using only AMR is unable to classify and identify the types of abnormalities that occur. In addition, another important issue that has been rarely studied and should not be neglected is the balancing of data. Because this issue has a negative impact on minorities, learners misclassify and leads to incorrect predictions. This paper proposes unbalanced data techniques for classifying energy theft, defective and normal customers based on AMR monitoring of PEA. In order to handle the multiclass imbalance problems, three methods are evaluated and compared: anomaly models (AM), adaptive synthetic sampling (ADASYN), and image data augmentation (IA). The data were extracted from time series into an image using a recurrence plot (RP) and classified as abnormal patterns of imaging time series using six deep learning models: LeNet5, AlexNet, VGGNet19, DenseNet121, ResNet50, and InceptionV3. The experimental results demonstrate that data generation techniques using anomaly models and DenseNet121 for classifying provided the best results compared to other techniques, and data extraction using images yields better results than time series. Moreover, compared to balanced and unbalanced data, classification evaluation using AUC-ROC and F1-score are the most appropriate evaluation methods, and importantly, balancing the data before classification improved the model performance.

Feature selection and ensemble classification increase system efficiency and accuracy in machine learning, data mining and biomedical informatics. This research presents an analysis of the effect of removing irrelevant and redundant features with ensemble classifiers using two datasets from UCI machine learning repository. Accuracy and computational time were evaluated by four base classifiers; NaiveBayes, multilayer perceptron, support vector machines and decision tree. Eliminating irrelevant features improves accuracy and reduces computational time while removing redundant features reduces computational time and reduces accuracy of the ensemble.

This paper presents a processing system of vehicle detection and counting for a camera on the city street using a heterogeneous ARM/FPGA processor and Xilinx SDSoC (Software-Defined System on Chip). An adaptive background method for reducing the impact of environment have been developed by analyzing luminance approximation changes over time and luminance approximation changes suddenly to improve background image of the object and to eliminate shadows so that the process of vehicles detecting and counting worked effectively.

Deep learning algorithms are employed in many applications, especially in medical fields such as gait analysis and human pose detection for rehabilitation. However, creating the desired model with deep learning algorithms requires high memory and computing costs, which is problematic because deep learning technologies must be run on low-power devices such as edge computing equipment. To deal with these problems, feature reduction methods reduce the memory and energy costs. This paper presents an empirical analysis of deep learning with feature reduction. The method classifies foot images for knee rehabilitation using convolutional and dense autoencoders. The obtained results are compared with those of conventional methods (histograms of oriented gradients and local binary pattern algorithms). The features were classified and compared using support vector machine, k-nearest neighbor, and multilayer perceptron methods. The experimental results demonstrate that the conventional method uses fewer features than the deep learning method with higher accuracy because its algorithm projects pixels onto the histogram. In addition, using fewer features in deep learning layers maintains high accuracy, which is beneficial for edge computing implementations.

Non-technical loss (NTL) is one of the problems that has been a major issue in lost revenue for many years. Electricity distributors have attempted to reduce NTL by detecting electricity theft using various methods. Some events are difficult to detect that conventional meters inspection is inadequate. Moreover, many anomaly patterns found are very complex, confusing in identifying or distinguishing what types of electricity customers are at abnormal risk or energy theft that affects NTL. This paper proposes five key feature extraction methods and six classifying electricity customers using supervised learning. The main problem was studied and collected information, including kilowatt meters, electronic meters, TOU meters, and AMR meters, which cover four customer types that were recorded in the Provincial Electricity Authority (PEA) of Thailand. An electrical profile to be extracted for in-depth analysis of the behavior of each type of electricity customer, combined with the information of physical data to help enhance and increase efficiency. All features examined the relationships in each feature using Pearson correlation and handled unbalanced data using random oversampling (ROS). Then, the extracted data has been trained, validated, and tested to classify three classes: normal, risk, and theft, where we evaluate the results with performance metrics. The results show that random forest (RF) outperforms the rest of the classifiers by achieving a precision-recall area under the curve of 90% and a receiver operating characteristic curve of 78%. Significantly, the results were compared to previous studies and benchmark datasets, which revealed that the proposed method gave better results than other techniques.