Andres Bustillo

An SVM-Based Solution for Fault Detection in Wind Turbines

Research into fault diagnosis in machines with a wide range of variable loads and speeds, such as wind turbines, is of great industrial interest. Analysis of the power signals emitted by wind turbines for the diagnosis of mechanical faults in their mechanical transmission chain is insufficient. A successful diagnosis requires the inclusion of accelerometers to evaluate vibrations. This work presents a multi-sensory system for fault diagnosis in wind turbines, combined with a data-mining solution for the classification of the operational state of the turbine. The selected sensors are accelerometers, in which vibration signals are processed using angular resampling techniques and electrical, torque and speed measurements. Support vector machines (SVMs) are selected for the classification task, including two traditional and two promising new kernels. This multi-sensory system has been validated on a test-bed that simulates the real conditions of wind turbines with two fault typologies: misalignment and imbalance. Comparison of SVM performance with the results of artificial neural networks (ANNs) shows that linear kernel SVM outperforms other kernels and ANNs in terms of accuracy, training and tuning times. The suitability and superior performance of linear SVM is also experimentally analyzed, to conclude that this data acquisition technique generates linearly separable datasets.

The evolutionary development of roughness prediction models

The vigorous expansion of wind energy power generation over the last decade has also entailed innovative improvements to surface roughness prediction models applied to high-torque milling operations. Artificial neural networks are the most widely used soft computing technique for the development of these prediction models. In this paper, we concentrate on the initial data transformation and its effect on the prediction of surface roughness in high-torque face milling operations. An extensive data set is generated from experiments performed under industrial conditions. The data set includes a very broad set of different parameters that influence surface roughness: cutting tool properties, machining parameters and cutting phenomena. Some of these parameters may potentially be related to the others or may only have a minor influence on the prediction model. Moreover, depending on the number of available records, the machine learning models may or may not be capable of modelling some of the underlying dependencies. Hence, the need to select an appropriate number of input signals and their matching prediction model configuration. A hybrid algorithm that combines a genetic algorithm with neural networks is proposed in this paper, in order to address the selection of relevant parameters and their appropriate transformation. The algorithm has been tested in a number of experiments performed under workshop conditions with data sets of different sizes to investigate the impact of available data on the selection of corresponding data transformation. Data set size has a direct influence on the accuracy of the prediction models for roughness modelling, but also on the use of individual parameters and transformed features. The results of the tests show significant improvements in the quality of prediction models constructed in this way. These improvements are evident when these models are compared with standard multilayer perceptrons trained with all the parameters and with data reduced through standard Principal Component Analysis practice.

Modelling of process parameters in laser polishing of steel components using ensembles of regression trees

Laser polishing of steel components is an emergent process in the automation of finishing operations in the industry. The aim of this work is to develop a soft computing tool for surface roughness prediction of laser polished components. The laser polishing process depends primarily on three factors: surface material, initial topography and energy density. Although the first two factors can be reasonably estimated, the third one is often unknown under real industrial conditions. The modelling tool developed solves this limitation. The application is composed of four stages: a data-acquisition system, a data set generated from the inputs, a soft computing model trained and validated with the data set. Finally, the model obtained is used to generate different plots of industrial interest. Different prediction models are tested until the most accurate one is selected, in order to generate the soft computing model, and due to the highly complex phenomena that influence surface roughness generation in laser polishing. Ensembles of regression trees yield the best results for the methods under consideration (multilayer perceptrons, radial basis function networks and support vector machines). It has been proven that the results of an ensemble, which is a combination of several models, are better than single methods in many applications.

Prediction, monitoring and control of surface roughness in high-torque milling machine operations

The development and testing of an application that will predict, monitor and control surface roughness are described. It comprises three modules for off-line roughness prediction, surface roughness monitoring and surface roughness control, and is especially designed for high-torque, high-power milling operations, which are widely used nowadays in the manufacture of wind turbine components. The application is tested in a milling machine with a high working volume. Due to the highly complex phenomena that generate surface roughness and the large number of factors that interact during the cutting process, models to calculate the average surface roughness parameter (Ra) are based on artificial neural networks (ANN) as they are especially suitable for modelling complex relationships between inputs and outputs.

A Virtual Sensor for Online Fault Detection of Multitooth-Tools

The installation of suitable sensors close to the tool tip on milling centres is not possible in industrial environments. It is therefore necessary to design virtual sensors for these machines to perform online fault detection in many industrial tasks. This paper presents a virtual sensor for online fault detection of multitooth tools based on a Bayesian classifier. The device that performs this task applies mathematical models that function in conjunction with physical sensors. Only two experimental variables are collected from the milling centre that performs the machining operations: the electrical power consumption of the feed drive and the time required for machining each workpiece. The task of achieving reliable signals from a milling process is especially complex when multitooth tools are used, because each kind of cutting insert in the milling centre only works on each workpiece during a certain time window. Great effort has gone into designing a robust virtual sensor that can avoid re-calibration due to, e.g., maintenance operations. The virtual sensor developed as a result of this research is successfully validated under real conditions on a milling centre used for the mass production of automobile engine crankshafts. Recognition accuracy, calculated with a k-fold cross validation, had on average 0.957 of true positives and 0.986 of true negatives. Moreover, measured accuracy was 98%, which suggests that the virtual sensor correctly identifies new cases.

Online breakage detection of multitooth tools using classifier ensembles for imbalanced data

Cutting tool breakage detection is an important task, due to its economic impact on mass production lines in the automobile industry. This task presents a central limitation: real data-sets are extremely imbalanced because breakage occurs in very few cases compared with normal operation of the cutting process. In this paper, we present an analysis of different data-mining techniques applied to the detection of insert breakage in multitooth tools. The analysis applies only one experimental variable: the electrical power consumption of the tool drive. This restriction profiles real industrial conditions more accurately than other physical variables, such as acoustic or vibration signals, which are not so easily measured. Many efforts have been made to design a method that is able to identify breakages with a high degree of reliability within a short period of time. The solution is based on classifier ensembles for imbalanced data-sets. Classifier ensembles are combinations of classifiers, which in many situations are more accurate than individual classifiers. Six different base classifiers are tested: Decision Trees, Rules, Naïve Bayes, Nearest Neighbour, Multilayer Perceptrons and Logistic Regression. Three different balancing strategies are tested with each of the classifier ensembles and compared to their performance with the original data-set: Synthetic Minority Over-Sampling Technique (SMOTE), undersampling and a combination of SMOTE and undersampling. To identify the most suitable data-mining solution, Receiver Operating Characteristics (ROC) graph and Recall-precision graph are generated and discussed. The performance of logistic regression ensembles on the balanced data-set using the combination of SMOTE and undersampling turned out to be the most suitable technique. Finally a comparison using industrial performance measures is presented, which concludes that this technique is also more suited to this industrial problem than the other techniques presented in the bibliography.

Data-mining modeling for the prediction of wear on forming-taps in the threading of steel components

Abstract An experimental approach is presented for the measurement of wear that is common in the threading of cold-forged steel. In this work, the first objective is to measure wear on various types of roll taps manufactured to tapping holes in microalloyed HR45 steel. Different geometries and levels of wear are tested and measured. Taking their geometry as the critical factor, the types of forming tap with the least wear and the best performance are identified. Abrasive wear was observed on the forming lobes. A higher number of lobes in the chamber zone and around the nominal diameter meant a more uniform load distribution and a more gradual forming process. A second objective is to identify the most accurate data-mining technique for the prediction of form-tap wear. Different data-mining techniques are tested to select the most accurate one: from standard versions such as Multilayer Perceptrons, Support Vector Machines and Regression Trees to the most recent ones such as Rotation Forest ensembles and Iterated Bagging ensembles. The best results were obtained with ensembles of Rotation Forest with unpruned Regression Trees as base regressors that reduced the RMS error of the best-tested baseline technique for the lower length output by 33%, and Additive Regression with unpruned M5P as base regressors that reduced the RMS errors of the linear fit for the upper and total lengths by 25% and 39%, respectively. However, the lower length was statistically more difficult to model in Additive Regression than in Rotation Forest. Rotation Forest with unpruned Regression Trees as base regressors therefore appeared to be the most suitable regressor for the modeling of this industrial problem.

Wind turbines fault diagnosis using ensemble classifiers

Fault diagnosis in machines that work under a wide range of speeds and loads is currently an active area of research. Wind turbines are one of the most recent examples of these machines in industry. Conventional vibration analysis applied to machines throughout their operation is of limited utility when the speed variation is too high. This work proposes an alternative methodology for fault diagnosis in machines: the combination of angular resampling techniques for vibration signal processing and the use of data mining techniques for the classification of the operational state of wind turbines. The methodology has been validated over a test-bed with a large variation of speeds and loads which simulates, on a smaller scale, the real conditions of wind turbines. Over this test-bed two of the most common typologies of faults in wind turbines have been generated: imbalance and misalignment. Several data mining techniques have been used to analyze the dataset obtained by order analysis, having previously processed signals with angular resampling technique. Specifically, the methods used are ensemble classifiers built with Bagging, Adaboost, Geneneral Boosting Projection and Rotation Forest; the best results having been achieved with Adaboost using C4.5 decision trees as base classifiers.

Interpreting tree-based prediction models and their data in machining processes

Machine-learning techniques frequently predict the results of machining processes, based on pre-determined cutting tool settings. By doing so, key parameters of a machined product can be predicted before production begins. Nevertheless, a prediction model cannot capture all the features of interest under real-life industrial conditions. Moreover, careful assessment of prediction credibility is necessary for accurate calibration; aspects that should be addressed through appropriate modeling and visualization techniques. A machine process test problem is proposed to analyze data-visualization techniques, in which a real data set is analyzed that describes deep-drilling under different cutting and cooling conditions. The main objective is the efficient fusion of visualization techniques with the knowledge of industrial engineers. Common modeling and visualization techniques were first surveyed, to contrast standard practice with our novel approach. A hybrid technique combining conditional inference trees with dimensionality reduction was then examined. The results show that a process engineer will be able to estimate overall model accuracy and to verify the extent to which accuracy depends on industrial process settings and the statistical significance of model predictions. Moreover, evaluation of the data set in terms of its sufficiency for modeling purposes will help assess the credibility of these decisions.

New methodology for the design of ultra-light structural components for machine tools

Energy consumption is the key to the ecological impact of many machine tools, especially milling machines. One promising strategy for minimising the energy consumption of machine tools is to reduce the mass of their structural components. This solution, however, has a clear drawback: the mechanical stiffness of the machine is reduced, impairing its performance and, in the long run, its productivity. This study proposes a new methodology to overcome such limitations, which involves the design of machine tools with ultra-light structural components, and the development of strategies to counteract the loss of productivity as a consequence of lightweight machines. The new methodology includes the use of modular boxes built with carbon-fibre trusses, calculation of the dynamic stiffness of the new design, the identification of its weaknesses in terms of its cutting processes, and the design and integration of active damping systems in the machine to soften the expected vibrations under the most critical cutting conditions. This methodology has been tested in the new design of a ram of a bridge-type machining centre of medium size. The results show that a 60% reduction in mass can be achieved and that an active damper system can compensate a 60% reduction in mechanical stiffness, maintaining a level of performance that is comparable to heavier standard machines under high-cutting conditions.