As employed in industries, anomaly detection systems can be unstable due to the lack of training examples and often narrow feature extraction methods, both of which burden common models as abnormalities are rare and exceptionally unique. Since equipment used within oil and gas companies often tend to incur on great financial losses due to inherent unrecognized malfunctions, diagnosing components’ vibration signals beforehand becomes essential. However, as obtaining examples of a plethora of possible fault patterns can be difficult, developing a system based on the equipment’s normal behavior allows for better identification of irregularities within the signals. Additionally, as erroneous feature extraction methods can mask or dismiss a signal’s abnormal demeanor, approaching this issue with a system that has generalized knowledge over the dataset’s nature may grant a more thorough signal analysis. This paper advances the fault detection in the oil industry by investigating well-known feature extraction neural networks (such as ResNet, VGG and VGGish) together with one-class classifiers (such as Isolation Forest, Elliptic Envelope and OneClass SVM). With this approach, vibration signals are converted (using transformations like spectrogram) to images in order to leverage image-pretrained networks. Results show that features extracted with pretrained networks have similar performance to those hand-crafted with prior knowledge about the faults and are significantly better than standard statistical features. Furthermore, the audio-originated pretrained CNN, VGGish, tends to perform slightly better than those trained with natural images like ImageNET. Therefore, the proposed system can be applied as a fault-agnostic feature extraction method alternative to the problem specific hand-crafted feature extraction in one-class classification problems.

Abstract One major challenge multi‐label classification faces, are the conditions for evaluating multi‐label algorithms. Simplistic experimental setups based on artificial data may not capture crucial situations for analyzing these algorithms. This article introduces an experimental framework for evaluating multi‐label algorithms by artificially generating the probabilistic label distributions. The proposed framework has the benefits of considering a wide variety of labels distributions, and enables users to simulate probability label distributions with a better control of the label dependence and the difficulty of the problem. An experimental study was conducted using the framework where new findings with respect to five methods, binary relevance, classifier chain, dependent binary relevance, calibrated label ranking by pairwise comparison and probabilistic classifier chain, were revealed. This framework will facilitate conducting new experimental studies for analysing the effects of changing label dependence and the difficulty of the problem on the performance of new multi‐label algorithms.

Most multi-label classification methods are evaluated on real datasets, which is a good practice for comparing the performance among methods on the average scenario. Due to the large amount of factors to consider, this empirical approach does not explain, nor does show the factors impacting the performance. A reasonable way to understand some of the performance’s factors of multi-label methods independently of the context is to find a mathematical proof about them. In this paper, mathematical proofs are given for the multi-label method ranking by pairwise comparison and its extension for classification named by calibrated label ranking, showing their performance on a worst case scenario for five multi-label metrics. The pairwise approach adopted by ranking by pairwise comparison enables the algorithm to achieve the optimal performance on Spearman rank correlation. However, the findings presented in this paper clearly show that the same pairwise approach adopted by the algorithm is also a crucial factor contributing to a very poor performance on other multi-label metrics.

Flow instability is a class of abnormal operation in subsea oil wells. Applying machine learning models and improving detection and classification performance is decisive to reduce operational costs and downtime. In this paper we evaluate a pipeline of methods in order to increase correct classification rates. Our strategy is defined to avoid the similarity bias and approaches the binary problem in two distinct ways: A) using normal labels as negative and B) using both normal labels and all other kind of defects as negative, leveraging all data in the available dataset. The workflow includes feature extraction, hyperparameter tuning, feature selection with sequential algorithms, hybrid ranking wrapper and also with genetic algorithm. We show that hyperparameter tuning produces minor improvements and due to problem complexity a robust feature selection algorithm is required to deliver higher results.

Clustering is a recurrent task in machine learning. The application of traditional heuristics techniques in large sets of data is not easy. They tend to have at least quadratic complexity with respect to the number of points, yielding prohibitive run times or low quality solutions. The most common approach to tackle this problem is to use weaker, more randomized algorithms with lower complexities to solve the clustering problem. This work proposes a novel approach for performing this task, allowing traditional, stronger algorithms to work on a sample of the data, chosen in such a way that the overall clustering is considered good. Preliminary experimental results indicate that the proposed approach is competitive to classical algorithms in large datasets with the advantage of automatically adapting to many different datasets.

Machine learning classification algorithms are highly dependent of a dataset composed of high-level features. In this paper, a deep learning approach is combined with traditional machine learning classifiers in order to circumvent the need of a specialist for extracting relevant features from one dimensional frequency-domain vibration signals. Our approach relies on a convolutional architecture trained with a triplet loss function for extracting relevant features directly from the raw data. A previously hand-crafted feature set, created by a specialist over the course of many years of research, is compared with the newly extracted feature set. Six conventional classifiers models (K-Nearest Neighbors, Support Vector Machine, Decision Tree, Random Forest, Quadratic Discriminant Analysis and Naive Bayes) are trained in both features set separately and compared in terms of macro F-measure. Results shows statistical evidence towards to the acceptance that the extracted feature set is as good as or better than the hand-crafted feature set, for classification purposes.

Electric power distribution companies in Brazil assess the energy consumption of most of their costumers by reading the meters in loco. A human reader has an itinerary with clients that should be read on each working day. The number of meters per route tends to increase over time, as the number of customers is constantly growing. At a certain point, the route must be restructured so that it still runs in a single day. This paper proposes the use of clustering techniques integrated with community detection algorithms and a heuristic routing algorithm to solve the problem of globally restructuring the routes used by the readers of the energy distribution companies. Experimental results showed that the proposed method significantly reduced the number of routes required to perform the meters readings.

Electrical submersible pumps are devices frequently used in off-shore oil exploration. Vibration signals analysis and expert systems technology are used for detecting faults on these motor pumps. Fault diagnosis classifiers may need to be updated or expanded. This paper proposes a domain specific language for enabling non-programmer engineers to create and adjust rule-based fault diagnosis classifiers of electrical submersible pumps.

Every year, energy consumption grows world widely. Therefore, power companies need to investigate models to better forecast and plan the energy use. One approach to address this problem is the estimation of energy consumption in the customer level. Energy consumption forecasting problem is a time series regression task. It consists of predicting the energy consumption for the next month given a finite history of a customer. Machine learning techniques have shown promising results in a variety of problems including time series and regression problems. Part of these promising results are attributed to deep neural networks. Although investigated in other domains, deep architectures have not been used to address the energy consumption prediction problem. In this work, we propose a system to predict monthly energy consumption using deep learning techniques. Three deep learning models were studied: Deep Fully Connected, Convolutional and Long Short-Term Memory Neural Networks. Due to the sensitivity of these models to the input range, normalization techniques were also investigated. The proposed system was validated with real data of almost a million customers (resulting in over 9 million samples). Results showed that our system can predict monthly energy consumption with an absolute error of 31.83 kWh and a relative error of 17.29%.

This paper presents an extension of a comparative study of classifier architectures for automatic fault diagnosis, with a special emphasis on the Extreme Learning Machine (ELM), with and without kernel mapping. Besides the explanation of the ELM model, an attempt is made to find theoretical hints of the excellent generalization capabilities of this model, based on the findings of Cover about dichotomies and the equivalence of Mean Squared Error minimization in the high-dimensional feature spaces induced by kernels, and spaces defined by a finite sample set. The field of application is a practical problem in the context of offshore petroleum exploration where sophisticated submersible motor pumps are extensively tested before being deployed. The work juxtaposes the performance of ELM to an existing statistically sound comparison of state of the art classifier methods for a hand-crafted feature model tailored specially to the spectra of the vibrational signals of the pump. The results suggest the remarkably good generalization capability of ELM, exhibiting the highest scores for the chosen F-measure performance criterion.

This work addresses the application of a population based evolutionary algorithm called shuffled complex evolution (SCE) in the core of multidimensional knapsack problem (MKP). The core of the MKP is a set of items which are hard to decide if they are or not selected in good solutions. This concept is used to reduce the original size of MKP instances. The performance of the SCE applied to the reduced MKP is verified through computational experiments using well-known instances from literature. The approach proved to be effective in finding near optimal solutions demanding a small amount of processing time.

In pattern recognition systems, ensemble techniques claim a potential performance improvement compared to single classifier approaches. Decision templates (DT) were proposed as a simple and effective method for combining continuous valued outputs of an ensemble of classifiers. In this paper, the concept of decision template single-label multi-class classifier combination is extended to the multi-label case. The different classifiers needed for a combination are obtained from the continuous re-estimation used in the Recursive Dependent Binary Relevance multi-label classifier. Each base classifier used in this work, delivers besides the class label, a continuous output for the class that can be used to assemble the DTs.

In this paper we present a sampling approach to run the k-means algorithm in large data sets. We propose a genetic algorithm to guide sampling based on evaluating the fitness of each individual of the population through the k-means clustering algorithm. Although we want a partition with the lowest SSE, our algorithm tries to find the sample with the highest SSE. After finding a good sample the remaining points of the entire data set are clustered using the nearest centroid and, after that, the SSE of the final solution is calculated. Our proposal is applied on a set of public domain data sets and the results are compared against two other methods: the k-means running in a uniform random sample of the data set, and the k-means in the complete data set. The results showed that our algorithm has a good trade off between quality and computational cost, especially for large data sets and higher number of clusters.

In this paper, an artificial intelligence solution to diagnose faults before acquisition of submersible petroleum motor pump systems is presented. Proper fault identification is time consuming and demands highly trained human experts. The diagnosis system is intended to facilitate the work of the human component of this important process by replicating the decision of highly trained experts through a classifier. To perform the automatic diagnosis, firstly intermediate features are extracted as the vibration spectra. Subsequently, high level features are extracted and fed into a classifier that outputs the final diagnose. To validate our proposal and to select the best classifier (among K-Nearest-Neighbour, Random Forest, Support Vector Machine and Decision Tree) for this problem, we performed a comparative study using real data acquired in tests accomplished before acquisition of submersible motor pumps. Our dataset comprises thousands of entries of accelerometer sensors (vertically distributed along the particular system components) data labelled by an human expert to one of the considered scenarios (normal pump, faulty sensor, faulty pump with rubbing, misalignment or unbalance). Results have showed that the evaluated classifiers have equivalent performance for the given problem, and that the standardization procedure can improve the performance of some classifiers. The performance of the classifiers is sufficient to facilitate the work performed by humans and consequently reduce the time spent in the pump fault diagnosis process.

A hardware and software system was devised for solving a problem related to energy losses in public lighting services. A computational intelligence framework based on heterogeneous feature models, feature extraction and selection is used for classifying lamp types and wattages. This paper shows how radiometric sensors and lamp image data are used for the classification task. It is also presented an experimental comparison on the public lighting problem domain between three feature selection algorithms.

As the first main contribution, this work proposes a feature selection algorithm to be used as base driver for comparisons in fast feature selection experiments. This heuristic algorithm tries to eliminate the redundant and irrelevant features of the datasets by creating a univariate ranking, in decreasing order with respect to their individual performance, followed by a sequential selection to establish the final set. Secondly, it presents examples where feature selection surpasses the predictive power of classifier ensembles based on feature selection. The proposed algorithm is compared to two ensemble methods, one fast feature selection algorithm, one pure ranking method and one classifier algorithm without feature selection, achieving a better performance in 17 of a total of 20 microarray gene datasets.

This work presents a novel hybrid approach for feature selection using a combination of ranking and wrapper methods. Its main goal is to select features quickly, without significant loss of classification performance. Experiments comparing this approach with Sequential Forward Feature (SFS) selection showed its viability using Support Vector Machine and K-Nearest Neighbor classifiers in specific scenarios. As a test bed, vibrational signals were employed which need a previous feature extraction stage to create a classification system. In two experiments, 74 and 130 features were extracted from these databases. The proposed approach performed at least ten times faster than SFS, with 0.32% loss of accuracy in the worst case, requiring 26% to 57.5% less features to achieve its highest accuracy.

The Extreme Learning Machine is an attractive artificial neural network architecture due to its low computational cost during the training process. In this work this classifier architecture is evaluated in the context of automatic fault diagnosis. As a benchmark, the data provided by the Tennessee Eastman simulator is used. The results are compared to the Support Vector Machine, K-Nearest Neighbor classifiers and methods based on feature extraction techniques, like e.g. Principal Component Analysis, Partial Least Squares, Independent Component Analysis. The test results suggest that the Extreme Learning Machine is an attractive alternative classification method of process conditions.

This work presents a performance analysis of the Extreme Learning Machine (ELM) compared to the Support Vector Machine (SVM) and K-Nearest Neighbor (K-NN) classifiers for automatic diagnosis of machine conditions. Tests were performed using 5,314 real examples extracted from electrical submersible pumps. The vibration signal extraction was executed in laboratory and the samples were labeled by experts. Two feature extraction models were employed, statistical features from the time and frequency domains and amplitude peaks of harmonics and subharmonics of the shaft rotation frequency. Sequential feature selection was applied to improve classifier performance and to reduce dataset dimensionality. Experimental results suggest that the ELM may be used as a classification algorithm in automatic diagnosis systems. In certain scenarios, the ELM can outperform SVM regarding the quality of results and training speed.

Non-technical energy losses mostly arise from illegal use of energy and force energy distribution companies to inspect large batches of clients in order to make decisions on actions for reducing these losses. Since an exhaustive inspection is impractical due to the high inspection cost and the very large number of clients, a carefully designed sampling procedure is needed. A useful strategy is offered by stratified sampling based on a division of the clients into homogeneous subgroups (strata). In this work we formulate the stratification task as a non-linear restricted optimization problem, in which the variance of overall energy loss due to the fraudulent activities is minimized. Solving this problem analytically is difficult and an exhaustive algorithm is intractable even for small problem instances. Therefore, we propose a Genetic Algorithm for finding practical solutions for the problem. Numerical experiments and a comparison with Simulated Annealing algorithm and a proportional allocation scheme are presented.

We present a system for the detection and diagnosis of faults of a high performance electric submersible pump used in deep water oil exploration. During the installation phase 36 accelerometers acquire vibrational patterns under various load conditions. The machine condition is labeled with the help of human experts. The training set is submitted to an automatic model-free learning system based on Bayesian belief networks and compared to a reference Support Vector Machine classifier. Experiments are presented for three different condition classes, using sophisticated statistical evaluation methodologies to measure the classifier performance.

We perform an empirical performance analysis of the Multilayer Perceptron applied to the fault diagnosis of motor pumps installed on oil rigs. The conventional Multilayer Perceptron architecture is compared to a recently developed enhancement of this general purpose regression/classification paradigm, using an intermediate opaque layer which maps the original patterns to a reproducing kernel Hilbert space prior to learning the usual functional mapping of the network. State of the art statistical tools are used to corroborate our hypotheses that the kernel enhanced version improves the classification performance.

This paper presents the results achieved by fault classifier ensembles based on a model-free supervised learning approach for diagnosing faults on oil rigs motor pumps. The main goal is to compare two feature-based ensemble construction methods, and present a third variation from one of them. The use of ensembles instead of single classifier systems has been widely applied in classification problems lately. The diversification of classifiers performed by the methods presented in this work is obtained by varying the feature set each classifier uses, and also at one point, alternating the intrinsic parameters for the training algorithm. We show results obtained with the established genetic algorithm GEFS and our recently developed approach called BSFS, which has a lower computational cost. We rely on a database of real data, with 2000 acquisitions of vibration signals extracted from operational motor pumps. Our results compare the outcomes from the two methods mentioned, and present a modification in one of them that improved the accuracy, reinforcing the motivation for the usage of that method.

The classification problem is recurrent in the context of supervised learning. A classification problem is a class of computational task in which labels must be assigned to object instances using information acquired from labeled instances of the same type of objects. When these objects contain time sensitive data, special classification methods could be used to take ad- vantage of the inherent extra information. As far as this paper is concerned, the time sensitive data are sequences of values that represent the measured energy consumption of residential clients in a given month. Traditional classifiers do not take temporal features into account, interpreting them as a series of unrelated static information. The proposed method is to develop methods of classification to be applied in a real time-series problem that somehow consider the time series as being the same value being repeatedly measured. Two new approaches are suggested to deal with this problem: the first is a Hybrid classifier that uses clustering, DTW (Dynamic Time Warp) and Euclidean distance to label a given instance. The second is a Weighted Curve Comparison Algorithm that creates consumption profiles and compares them with the unknown instance to classify it.