Search

This article presents a two-stage fault detection and classification scheme specifically designed for rotating electrical machines. The approach involves the use of new condition indicators that are specific to the frequency domain. The paper proposes two distinct features: one based on the extraction of peaks by using the prominence measure, a technique originating from the topology of mountains, and other based on the calculation of the occupied band power ratio for specific characteristic fault frequencies. A linear based feature reduction technique, the principal component analysis (PCA) has been employed to represent all the data. Afterwards, shallow neural networks have been used to detect and classify the three-phase current signals online. The effectiveness of the proposed scheme has been validated experimentally by using signals obtained with grid and inverter fed induction motors.

Detection of stator-based faults in Induction Machines (IMs) can be carried out in numerous ways. In particular, the shorted turns in stator windings of IM are among the most common faults in the industry. As a matter of fact, most IMs come with pre-installed current sensors for the purpose of control and protection. At this aim, using only the stator current for fault detection has become a recent trend nowadays as it is much cheaper than installing additional sensors. The three-phase stator current signatures have been used in this study to observe the effect of stator inter-turn fault with respect to the healthy condition of the IM. The pre-processing of the healthy and faulty current signatures has been done via the in-built DSP module of dSPACE after which, these current signatures are passed into the MATLAB® software for further analysis using AI techniques. The authors present a Growing Curvilinear Component Analysis (GCCA) neural network that is capable of detecting and follow the evolution of the stator fault using the stator current signature, making online fault detection possible. For this purpose, a topological manifold analysis is carried out to study the fault evolution, which is a fundamental step for calibrating the GCCA neural network. The effectiveness of the proposed method has been verified experimentally.

Hydrogen energy conversion using Fuel Cells is very promising for standalone power as well as transportation applications. Hydrogen gas production using renewable energy sources is possible through the use of electrolyzers in which DC-DC converters play an important role. This paper presents an accurate and robust method of fault diagnosis and condition monitoring applied to an interleaved DC/DC buck converter that supplies a proton exchange membrane (PEM) electrolyzer. This work mainly focuses on power switch open-circuit failures. The study gives excellent results in the early detection of faults to improve the reliability of PEM electrolyzers. A suitable experimental test bench has been realized to obtain data under healthy and faulty operating conditions. A preliminary exploratory analysis of the data has been carried out using a linear approach to understand the geometry of the data and suggest a suitable tool for classifying the system’s condition. The paper then proposes a shallow configuration Long Short Term Memory (LSTM) based neural network capable of detecting and localizing the fault at every time step without any preprocessing. The experimental results presented in this paper show that, after a detailed comparison with other 25 neural and non-neural based techniques for classification, the shallow LSTM neural network gives the best results.

Fault detection in electrical drives can be really challenging, especially when the input data is collected from an operational electrical machine. In order to prevent machine damages and downtimes, it is really important to detect pre-fault conditions. This paper presents the detection of stator inter-turn fault for Synchronous Reluctance Motor (SynRM) with a severity as low as 1.3%. After the transformation of the three-phase currents using Extended Park Vector (EPV) approach, the temporal features were calculated. Thereafter, the geometry of the features has been studied by using the Principal Component Analysis (PCA) and the Curvilinear Component Analysis (CCA) to estimate the best intrinsic dimensionality and extract the most significant features. Finally, a variety of classifiers have been trained with this feature-set (FS) and the shallow neural network has proved to give the best performance.

Topological learning is a wide research area aiming at uncovering the mutual spatial relationships between the elements of a set. Some of the most common and oldest approaches involve the use of unsupervised competitive neural networks. However, these methods are not based on gradient optimization which has been proven to provide striking results in feature extraction also in unsupervised learning. Unfortunately, by focusing mostly on algorithmic efficiency and accuracy, deep clustering techniques are composed of overly complex feature extractors, while using trivial algorithms in their top layer. The aim of this work is to present a novel comprehensive theory aspiring at bridging competitive learning with gradient-based learning, thus allowing the use of extremely powerful deep neural networks for feature extraction and projection combined with the remarkable flexibility and expressiveness of competitive learning. In this paper we fully demonstrate the theoretical equivalence of two novel gradient-based competitive layers. Preliminary experiments show how the dual approach, trained on the transpose of the input matrix i.e. $X^T$, lead to faster convergence rate and higher training accuracy both in low and high-dimensional scenarios., Comment: 12 pages, 20 figures

Continuous vital signal monitoring is becoming more relevant in preventing diseases that afflict a large part of the world’s population; for this reason, healthcare equipment should be easy to wear and simple to use. Non-intrusive and non-invasive detection methods are a basic requirement for wearable medical devices, especially when these are used in sports applications or by the elderly for self-monitoring. Arterial blood pressure (ABP) is an essential physiological parameter for health monitoring. Most blood pressure measurement devices determine the systolic and diastolic arterial blood pressure through the inflation and the deflation of a cuff. This technique is uncomfortable for the user and may result in anxiety, and consequently affect the blood pressure and its measurement. The purpose of this paper is the continuous measurement of the ABP through a cuffless, non-intrusive approach. The approach of this paper is based on deep learning techniques where several neural networks are used to infer ABP, starting from photoplethysmogram (PPG) and electrocardiogram (ECG) signals. The ABP was predicted first by utilizing only PPG and then by using both PPG and ECG. Convolutional neural networks (ResNet and WaveNet) and recurrent neural networks (LSTM) were compared and analyzed for the regression task. Results show that the use of the ECG has resulted in improved performance for every proposed configuration. The best performing configuration was obtained with a ResNet followed by three LSTM layers: this led to a mean absolute error (MAE) of 4.118 mmHg on and 2.228 mmHg on systolic and diastolic blood pressures, respectively. The results comply with the American National Standards of the Association for the Advancement of Medical Instrumentation. ECG, PPG, and ABP measurements were extracted from the MIMIC database, which contains clinical signal data reflecting real measurements. The results were validated on a custom dataset created at Neuronica Lab, Politecnico di Torino.

Facial Emotion Recognition (FER) is the automatic processing of human emotions by means of facial expression analysis [1]. The most common approach exploits 3D Face Descriptors (3D-FD) [2], which derive from depth maps [3] by using mathematical operators. In recent years, Convolutional Neural Networks (CNNs) have been successfully employed in a wide range of tasks including large-scale image classification systems and to overcome the hurdles in facial expression classification. Based on previous studies, the purpose of the present work is to analyze and compare the abstraction level of 3D face descriptors with abstraction in deep CNNs. Experimental results suggest that 3D face descriptors have an abstraction level comparable with the features extracted in the fourth layer of CNN, the layer of the network having the highest correlations with emotions.

Exploiting the availability of the largest collection of patient-derived xenografts from metastatic colorectal cancer annotated for a response to therapies, this manuscript aims to characterize the biological phenomenon from a mathematical point of view. In particular, we design an experiment in order to investigate how genes interact with each other. By using a shallow neural network model, we find reduced feature subspaces where the resistance phenomenon may be much easier to understand and analyze.

Parkinson’s is a disease of the central nervous system characterized by neuronal necrosis. Patients at the time of diagnosis have already lost up to 70% of the neurons. It is essential to define early detection techniques to promptly intervene with an appropriate therapy. Handwriting analysis has been proven as a reliable method for Parkinson’s disease diagnose and monitoring. This paper presents an analysis of a Parkinson’s disease handwriting dataset in which neural networks are used as a tool for analyzing the problem space. The goal is to check the validity of the selected features. For estimating the data intrinsic dimensionality, a preliminary analysis based on PCA is performed. Then, a comparative analysis about the classification performances of a multilayer perceptron (MLP) has been conducted in order to determine the discriminative capabilities of the input features. Finally, fifteen temporal features, capable of a more meaningful discrimination, have been extracted and the classification performances of the MLP trained on these new datasets have been compared with the previous ones for selecting the best features.

Automated electrocardiogram analysis and classification is nowadays a fundamental tool for monitoring patient heart activity and, consequently, his state of health. Indeed, the main interest is detecting the arise of cardiac pathologies such as arrhythmia. This paper presents a novel approach for automatic arrhythmia classification based on a 1D convolutional neural network. The input is given by the combination of several databases from Physionet and is composed of two leads, LEAD1 and LEAD2. Data are not preprocessed, and no feature extraction has been performed, except for the medical evaluation in order to label it. Several 1D network configurations are tested and compared in order to determine the best one w.r.t. heart-beat classification. The test accuracy of the proposed neural approach is very high (up to 95%). However, the goal of this work is also the interpretation not only of the results, but also of the behavior of the neural network, by means of confusion matrix analysis w.r.t. the different arrhythmia classes.

This paper presents a comparison between two recurrent neural networks (RNN) for arterial blood pressure (ABP) estimation. ABP is a parameter closely related to the cardiac activity, for this reason its monitoring implies decreasing the risk of heart disease. In order to predict the ABP values (both systolic and diastolic), electrocardiographic (ECG) and photoplethysmographic (PPG) signals are used, separately, as inputs of the networks. To train the artificial neural networks, the synchronized signals are extracted from the Physionet MIMIC database. The output-error Neural networks (NNOE) and the Long Short Term Memory (LSTM) architectures are compared in terms of RMSE and absolute error. NNOE neural network, with ECG signal as input, results the best configuration in terms of both the proposed metrics. The predicted ABP falls within the values of the normative ANSI/AAMI/ ISO 81060-2:2013 for sphygmomanometer certification.

Despite all the progress made in biomedical field, the Electrocardiogram (ECG) is still one of the most commonly used signal in medical examinations. Over the years, the problem of ECG classification has been approached in many different ways, most of which rely on the extraction of features from the signal in the form of temporal or morphological characteristics. Although feature engineering can led to adequately good results, it mostly relies on human ability and experience in selecting the correct feature set. In the last decade, a growing class of techniques based on Convolutional Neural Network (CNN) has been proposed in opposition to feature engineering. The efficiency and accuracy of CNN-based approaches is indisputable, however their ability in extracting and using temporal features from raw signal is poorly understood. The main objective of this work was to uncover the differences and the relationships between CNN feature maps and human-curated temporal features, towards a deeper understanding of neural-based approaches for ECG. In fact, the proposed study succeeded in finding a similarity between the output stage of the first layers of a deep 1D-CNN with several temporal features, demonstrating that not only that the engineered features effectively works in ECG classification tasks, but also that CNN can improve those features by elaborating them towards an higher level of abstraction.

In the field of artificial intelligence, agents learn how to take decisions by fitting their parameters on a set of samples called training set. Similarly, a core set is a subset of the training samples such that, if an agent exploits this set to fit its parameters instead of the whole training set, then the quality of the inferences does not change significantly. Relaxing the constraint that restricts the search for core sets to the available data, neural networks may be used to generate virtual samples, called archetype set, containing the same kind of information. This work illustrates the features of GH-ARCH, a recently proposed self-organizing hierarchical neural network for archetype discovery. Experiments show how the use of archetypes allows both ML agents to make fast and accurate predictions and human experts to make sense of such decisions by analyzing few important samples.

This article presents a data-driven approach for the classification of faults in induction machines. The designed scheme involves newly engineered features extracted from the line current signals, which provides an improved fault discrimination. For this purpose, a topological-based fast projection technique (curvilinear component analysis) is used as a tool to reduce the dimensionality of the data and interpret the feature behavior. Consequently, a shallow convolutional neural network has been designed to classify the three-phase stator current signals. Experimental tests at different operating conditions have assessed the procedure, confirming its effectiveness and suitability for online and real-time diagnostics.

In the last few years, neural networks are effectively applied in different fields. However, the application of empirical-like algorithms as feed-forward neural networks is not always justified from an epistemological point of view [1]. In this work, the assumptions for the appropriate application of machine learning empirical-like algorithms to dynamical system learning are investigated from a theoretical perspective. A very simple example shows how the suggested analyses are crucial in corroborating or discrediting machine learning outcomes.

Applications such as surveillance, banking and healthcare deal with sensitive data whose confidentiality and integrity depends on accurate human recognition. In this sense, the crucial mechanism for performing an effective access control is authentication, which unequivocally yields user identity. In 2018, just in North America, around 445K identity thefts have been denounced. The most adopted strategy for automatic identity recognition uses a secret for encrypting and decrypting the authentication information. This approach works very well until the secret is kept safe. Electrocardiograms (ECGs) can be exploited for biometric purposes because both the physiological and geometrical differences in each human heart correspond to uniqueness in the ECG morphology. Compared with classical biometric techniques, e.g. fingerprints, ECG-based methods can definitely be considered a more reliable and safer way for user authentication due to ECG inherent robustness to circumvention, obfuscation and replay attacks. In this paper, the ECG WATCH, a non-expensive wristwatch for recording ECGs anytime, anywhere, in just 10 s, is proposed for user authentication. The ECG WATCH acquisitions have been used to train a shallow neural network, which has reached a 99% classification accuracy and 100% intruder recognition rate.

Due to the huge amount of available omic data, classifying samples according to various omics is a complex process. One of the most common approaches consists of creating a classifier for each omic and subsequently making a consensus among the classifiers that assigns to each sample the most voted class among the outputs on the individual omics.

In the field of machine learning, coresets are defined as subsets of the training set that can be used to obtain a good approximation of the behavior that a given algorithm would have on the whole training set. Advantages of using coresets instead of the training set include improving training speed and allowing for a better human understanding of the dataset. Not surprisingly, coreset discovery is an active research line, with several notable contributions in literature. Nevertheless, restricting the search for representative samples to the available data points might impair the final result. In this work, neural networks are used to create sets of virtual data points, named archetypes, with the objective to represent the information contained in a training set, in the same way a coreset does. Starting from a given training set, a hierarchical clustering neural network is trained and the weight vectors of the leaves are used as archetypes on which the classifiers are trained. Experimental results on several benchmarks show that the proposed approach is competitive with traditional coreset discovery techniques, delivering results with higher accuracy, and showing a greater ability to generalize to unseen test data.

In the last few years, the reduction of energy consumption and pollution became mandatory. It became also a common goal of many countries. Only in Europe, the building sector is responsible for the total 40% of energy consumption and 36% of $CO_{2}$ pollution. Therefore, new control policies based on the forecast of buildings energy behaviors can be developed to reduce energy waste (i.e. policies for Demand Response and Demand Side Management). This paper discusses an innovative methodology for smart building indoor air-temperature forecasting. This methodology is based on a Non-linear Autoregressive neural network. This neural network has been trained and validated with a dataset consisting of six years indoor air-temperature values of a building demonstrator. In detail, we have studied three characterizing rooms and the whole building. Experimental results of energy prediction are presented and discussed.

In the world, energy demand continues to grow incessantly. At the same time, there is a growing need to reduce CO2 emissions, greenhouse effects and pollution in our cities. A viable solution consists in producing energy by exploiting renewable sources, such as solar energy. However, for the efficient use of this energy, accurate estimation methods are needed. Indeed, applications like Demand/Response require prediction tools to estimate the generation profiles of renewable energy sources. This paper presents an innovative methodology for short-term (e.g. 15 minutes) forecasting of Global Horizontal Solar Irradiance (GHI). The proposed methodology is based on a Non-linear Autoregressive neural network. This neural network has been trained and validated with a dataset consisting of solar radiation samples collected for four years by a real weather station. Then GHI forecast, the output of the neural network, is given as input to our Photovoltaic simulator to predict energy production in short-term time periods. Finally, experimental results for both GHI forecast and Photovoltaic energy prediction are presented and discussed.

Three dimensional face analysis is being widely investigated since it appears as a robust solution to overcome the limits of two dimensional technologies. 3D methods allow to relate the recognition process on features not depending on lightning, head poses, make up and occlusions. This paper proposes a new approach to the problem consisting of a novel image representation, where specific facial descriptors replace the RGB traditional channels and a convolutional neural network performs the classification. We chose to use MobileNetV2, a relatively new network, as it has a low amount of parameters to train. The method has been evaluated on the Bosphorus database, and even though it is still a preliminary study, the results obtained with our method are extremely encouraging; the recognition rate achieved is 97.560% and it is comparable to the state of the art. This result, reached despite the fact that the Bosphorus database has a great number of subjects (105) but a low number of scans (4666), shows the effectiveness of this representation combined with convolutional neural networks.

Automated ECG analysis and classification are nowadays a fundamental tool for monitoring patient heart activity properly. The most important features used in literature are the raw data of a time window, the temporal attributes and the frequency information from the eigenvector techniques. This paper compares these approaches from a topological point of view, by using linear and nonlinear projections and a neural network for assessing the corresponding classification quality. The nonlinearity of the feature data manifold carries most of the QRS-complex information. Indeed, it yields high rates of classification with the smallest number of features. This is most evident if temporal features are used: Nonlinear dimensionality reduction techniques allow a very large data compression at the expense of a slight loss of accuracy. It can be an advantage in applications where the computing time is a critical factor. If, instead, the classification is performed offline, the raw data technique is the best one.

The analysis of complex systems, such as cancer resistance to drugs, requires flexible algorithms but also simple models, as they will be used by biologists in order to get insights on the underlying phenomenon. Exploiting the availability of the largest collection of patient-derived xenografts from metastatic colorectal cancer annotated for response to therapies, this manuscript aims to (i) forecast the response to treatments on human tissues using murine information; (ii) providing a trade-off between model accuracy and interpretability, evolving a shallow neural network using a genetic algorithm.

Fault diagnostics for electrical machines is a very difficult task because of the non-stationarity of the input information. Also, it is mandatory to recognize the pre-fault condition in order not to damage the machine. Only techniques like the principal component analysis (PCA) and its neural variants are used at this purpose, because of their simplicity and speed. However, they are limited by the fact they are linear. The GCCA neural network addresses this problem; it is nonlinear, incremental, and performs simultaneously the data quantization and projection by using the curvilinear component analysis (CCA), a distance-preserving reduction technique. Using bridges and seeds, it is able to fast adapt and track changes in the data distribution. Analyzing bridge length and density, it is able to detect a pre-fault condition. This paper presents an application of GCCA to a real induction machine on which a time-evolving stator fault in one phase is simulated.

International audience; This paper proposes a rotor speed and position estimation scheme for the synchronous reluctance motor (SynRM) drive system based on its active flux model in the stator reference frame and the total least squares (TLS) EXIN neuron. Firstly, the active flux model of the SynRM in the stator reference frame is reconstructed to the overdetermined matrix equations. On the basis of that, the estimation of the rotor speed of the SynRM is transferred into solving a TLS problem. The TLS EXIN neuron, which is a recursive TLS algorithm, is used to solve this problem online to extract the rotor speed. The estimated rotor position is obtained from the estimated rotor speed based on the integrator. The feasibility and effectiveness of the proposed rotor speed estimation scheme have been verified by the simulation results.

Hierarchical clustering is an important tool for extracting information from data in a multi-resolution way. It is more meaningful if driven by data, as in the case of divisive algorithms, which split data until no more division is allowed. However, they have the drawback of the splitting threshold setting. The neural networks can address this problem, because they basically depend on data. The growing hierarchical GH-EXIN neural network builds a hierarchical tree in an incremental (data-driven architecture) and self-organized way. It is a top-down technique which defines the horizontal growth by means of an anisotropic region of influence, based on the novel idea of neighborhood convex hull. It also reallocates data and detects outliers by using a novel approach on all the leaves, simultaneously. Its complexity is estimated and an analysis of its user-dependent parameters is given. The advantages of the proposed approach, with regard to the best existing networks, are shown and analyzed, qualitatively and quantitatively, both in benchmark synthetic problems and in a real application (image recognition from video), in order to test the performance in building hierarchical trees. Furthermore, an important and very promising application of GH-EXIN in two-way hierarchical clustering, for the analysis of gene expression data in the study of the colorectal cancer is described.

Nowadays, green energy is considered as a viable solution to hinder \( CO_{2} \) emissions and greenhouse effects. Indeed, it is expected that Renewable Energy Sources (RES) will cover \( 40\% \) of the total energy request by 2040. This will move forward decentralized and cooperative power distribution systems also called smart grids. Among RES, solar energy will play a crucial role. However, reliable models and tools are needed to forecast and estimate with a good accuracy the renewable energy production in short-term time periods. These tools will unlock new services for smart grid management.

In pattern recognition, neural networks can be used not only for the classification task, but also for feature selection and other intermediate steps. This paper addresses the 3D face recognition problem in order to select the most meaningful geometric descriptors. At this aim, the classification results are directly integrated in a biclustering process in order to select the best leaves of a neural hierarchical tree. This tree is created by a novel neural network GH-EXIN. This approach results in a new criterion for the feature selection. This technique is applied to a database of face expressions where both traditional and novel geometric descriptors are used. The results state the importance of the curvedness novel descriptors and only of a few Euclidean distances.

International audience; Nowadays, the use of electrolyzers to cleanly and efficiently generate hydrogen from renewable energy sources is an attractive solution. Like fuel cell systems, DC/DC converters are needed to interface the DC bus with the electrolyzer. Classic buck converters are generally used for this purpose. However, these topologies must meet several requirements from output current ripple reduction and availability in case of electrical failures point of view. Hence, an interleaved DC/DC buck converter has been chosen as a viable solution in order to meet these requirements. With the aim of improving the availability and reliability of the DC/DC converter, the development of a power switch fault detection method is mandatory. This paper proposes a novel power switch fault detection method based on a non-linear dimensionality reduction technique, implemented by means of a neural network. Experimental tests are carried out for demonstrating the performance of the proposed strategy to accurately detect the faulty leg.

This paper presents a subspace-based approach to identify and extract harmonics of interest for the diagnosis of stator and rotor related faults in induction machines. The maj or goal of this paper is firstly to introduce and highlight the effectiveness of prominence measure upon preparing features for classification of faults. Secondly, a new approach is presented here to retrieve harmonics by using prominence measure of the peaks for each case of the fault. Finally, a hierarchical multi-layer perceptron neural network has been used as the classifier and compared with other existing classification algorithms to deduce the best model. The effectiveness of the developed scheme is verified experimentally.

This paper proposes a methodology to assess the discriminative capabilities of geometrical descriptors referring to the public Bosphorus 3D facial database as testing dataset. The investigated descriptors include histogram versions of Shape Index and Curvedness, Euclidean and geodesic distances between facial soft-tissue landmarks. The discriminability of these features is evaluated through the analysis of single block of features and their meanings with different techniques. Multilayer perceptron neural network methodology is adopted to evaluate the relevance of the features, examined in different test combinations. Principle component analysis (PCA) is applied for dimensionality reduction.

Non-stationary topological representation can be addressed in two ways, according to the application: life-long modeling or by forgetting the past. Life-long learning requires neural networks equipped with a tool for judging if a neuron has to be created for tracking the input distribution. It is always implemented as an isotropic criterion (a hypersphere centered at the winner weight vector represents the domain of the neuron). Instead, the G-EXIN neural network, presented here, uses an anisotropic convex polytope, which, models the shape of the neuron neighborhood. This idea allows to consider the boundaries of the Voronoi sets of data and controls the extent of the extrapolation. It also employs a novel kind of edge, called bridge, which carries information on the extent of the distribution time change. Indeed, the analysis of bridges, mainly their density, yields a deeper insight to the kind of non-stationarity. Both artificial and real examples are given of the advantages of this approach with regard to the ESOINN neural network, which is the best existing approach to life-long modeling.

This paper presents a fault diagnosis and classification scheme for induction machines by using motor current signature analysis together with neural networks. The adopted strategy utilizes three-phase stator current sensors and calculates appropriate features using non-parametric and a statistical approach. The feature-set is reduced by means of the principal component analysis which acts as a pre-processor for the multilayer perceptron neural network. This two stage classification is carried out for detection and classification of faults. The efficacy of the proposed scheme is validated experimentally by using grid and inverter fed induction motors.

This paper presents a data driven approach where at first the most significant features of the three phase current signal are analyzed and then a Curvilinear Component based analysis (CCA), which is a nonlinear manifold learning technique, is performed to compress and interpret the feature behaviour. Finally, a multi-layer perceptron network is used to develop a classifier. The effectiveness of the developed model is verified experimentally with data provided on-line and in real-time.

Dealing with time-varying high dimensional data is a big problem for real time pattern recognition. Only linear projections, like principal component analysis, are used in real time while nonlinear techniques need the whole database (offline). Their incremental variants do no work properly. The onCCA neural network addresses this problem; it is incremental and performs simultaneously the data quantization and projection by using the Curvilinear Component Analysis (CCA), a distance-preserving reduction technique. However, onCCA requires an initial architecture, provided by a small offline CCA. This paper presents a variant of onCCA, called growing CCA (GCCA), which has a self-organized incremental architecture adapting to the nonstationary data distribution. This is achieved by introducing the ideas of “seeds”, pairs of neurons which colonize the input domain, and “bridge”, a different kind of edge in the manifold graph, which signal the data nonstationarity. Some examples from artificial problems and a real application are given.

Real time pattern recognition applications often deal with high dimensional data, which require a data reduction step which is only performed offline. However, this loses the possibility of adaption to a changing environment. This is also true for other applications different from pattern recognition, like data visualization for input inspection. Only linear projections, like the principal component analysis, can work in real time by using iterative algorithms while all known nonlinear techniques cannot be implemented in such a way and actually always work on the whole database at each epoch. Among these nonlinear tools, the Curvilinear Component Analysis (CCA), which is a non-convex technique based on the preservation of the local distances into the lower dimensional space, plays an important role. This paper presents the online version of CCA. It inherits the same features of CCA, is adaptive in real time and tracks non-stationary high dimensional distributions. It is composed of neurons with two weights: one, pointing to the input space, quantizes the data distribution, and the other, pointing to the output space, represents the projection of the first weight. This on-line CCA has been conceived not only for the previously cited applications, but also as a basic tool for more complex supervised neural networks for modelling very complex high dimensional data. This algorithm is tested on 2-D and 3-D synthetic data and on an experimental database concerning the bearing faults of an electrical motor, with the goal of novelty (fault) detection.

The paper presents a neural approach for the parameter estimation of adaptive IIR filters for linear system identification. It is based on a novel neuron, the TLS EXIN neuron, capable of resolving the TLS problem present in this kind of estimation, where noisy errors affect not only the observation vector but also the data matrix. After a survey of other techniques for solving such parameters estimations, the TLS EXIN neuron is compared both theoretically and numerically with the former techniques, resulting in improved performance. Moreover, it is also proved that the TLS EXIN neuron permits some powerful acceleration techniques, unlike the other approaches. These results are also shown numerically.

Mostly the faults in electrical machines are related with the bearings. Thus, a reliable bearing condition monitoring scheme able to detect either local or distributed defects are mandatory to avoid a breakdown in the machine. So far, the research has been carried out mainly in the detection of local faults, such as balls and raceways faults, but surface roughness is not so reported. This paper deals with a novel and reliable scheme capable to detect any fault that may occur in a bearing, based on EXIN Curvilinear Component Analysis, CCA, and Neural Network. The EXIN CCA, which is an improvement of the Curvilinear Component Analysis, has been conceived for data visualization, interpretation and classification for real time industrial applications. The effectiveness of this condition monitoring scheme has been verified by experimental results obtained from different operation conditions.

EXIN CCA is an extension of the Curvilinear Component Analysis (CCA), which solves for the noninvariant CCA projection and allows representing data drawn under different operating conditions. It can be applied to data visualization, interpretation (as a kind of sensor of the underlying physical phenomenon) and classification for real time industrial applications. Here an example is given for bearing fault diagnostics in an electromechanical device.

This chapter contains sections titled: High-Order MCA Neurons Robust MCA EXIN Nonlinear Neuron (NMCA EXIN) Extensions of the Neural MCA ]]>

This paper presents a new linear neuron, the GeTLS EXIN neuron for the on-line estimation of the electrical parameters of the induction motor. This estimation problem is a non generic one, particularly for the flatness of the surface error along the K 2 direction, and current OLS neurons are not capable of facing it. Only the TLS EXIN neuron and the GeTLS EXIN neuron can cope with this issue for on-line applications. However, unlike the TLS EXIN neuron, the GeTLS EXIN neuron with its variation of the parameter ς, varies the landscape of the error surface, which then varies at each batch of data. This causes the surface to crumple up and reduce the flatness of the critical directions (K 2 in this case) with resulting improved accuracy.

Systems with high nonlinearities are, in general, very difficult to model. This is particularly true in geostatistics, where the problem of the estimation of a regionalized variable (RV) given only a small amount of measurement stations and a complex terrain surface is very challenging. This paper introduces a novel strategy, which couples the Curvilinear Component Analysis (CCA) and the Generalized Mapping Regressor (GMR). CCA, which is a nonlinear projector of a data manifold, is here used in order to find the intrinsic dimension of the data manifold, just giving an insight on the nonlinearities of the problem. This analysis drives the pre-processing of the data set used for the training phase of GMR. GMR is an incremental, self-organizing neural network which is able to model nonlinear functions by transforming the approximation into a pattern recognition problem. The presented approach is tested on the spatial estimation of the wind speed over the complex terrain of the isle of Sicily, in Italy. Wind speed maps resulting from this technique are presented and compared with a deterministic interpolator, the Inverse Distance Weighting (IDW) method.

A very important problem in computer vision is the matching of features extracted from pairs of images. At this proposal, a new neural network, the Double Asynchronous Competitor (DAC) is presented. It exploits the self-organization for solving the matching as a pattern recognition problem. As a consequence, a set of attributes is required for each image feature. The network is able to find the variety of the input space. DAC exploits two intercoupled neural networks and outputs the matches together with the occlusion maps of the pair of frames taken in consideration. DAC can also solve other matching problems.

The motor current signature analysis has been considered as a standard for both electrical and mechanical fault detection in three-phase induction machine. However, even if the spectrum analysis is well known and has been published intensively in the literature, the problem of automatic classification for fault detection is still open. The aim of this paper is to present a new neural network approach for fault and speed detection using the current spectrum analysis. After the presentation of an original database creation, the classification using a particular neural network is developed. This network is conceived in order to output the posterior probabilities of class membership of the input, which allow the estimation of the level of confidence of the classification. Then, the classical motor slip detection algorithm is verified and the classification is experimentally performed on a squirrel-cage three-phase induction machine. The efficiency of the iteration process is shown together with the confusion matrix for the current spectrum analysis in the proposed method.

In this paper a neural network based architecture, which combines supervised and unsupervised learning for the static security assessment of power systems, is presented. The proposed method allows the on-line security evaluation of a possible outage simply by considering the position of the neuron activated by the pre-fault state vector in an output map, allowing an easy and immediate view of the contingency risks. The mapping capabilities of two unsupervised neural networks, SOM (self-organising map) and CCA (curvilinear component analysis), are compared. Numerical tests, carried out on a study system, are presented and discussed.

This paper deals with the application of the General Mapping Regressor (GMR) neural network to the direct torque control DTC of an induction motor. In particular it shows that the GMR neural network is able to correctly learn the classical DTC, as well as any other more involved control strategy. A suitable test bench has been set up in order to verify the performance of the neural controller.