Bearing monitoring plays an important role in the condition monitoring of wind turbines. Failures in this critical component can lead to unexpected shutdowns, significantly reducing the efficiency of the energy system. Therefore, the implementation of advanced monitoring tools is essential to ensure reliability, availability, and cost-effective operation. Digital Twin (DT) technology enables the achievement of these objectives and helps address the issue of limited data. This paper presents a monitoring methodology based on Digital Twin technology for bearing fault diagnosis. It relies on a four degrees of freedom (4-DOF) bearing model and the simulation of faults that can affect this component. A machine learning-based adaptation is proposed, starting with dimensionality reduction through an encoder network, followed by a hybrid network combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM). This network extracts physical information from the model and adapts it to the noise in the experimental fault signatures. The results were obtained through experimental validation using a dataset from the Machinery Failure Prevention Technology (MFPT) demonstrate the effectiveness of this methodology for bearing fault diagnosis under various operating conditions.

The increasing adoption of renewable energy has accelerated the deployment of photovoltaic infrastructure, introducing new challenges for fault diagnosis systems. Among these, classification-based diagnosis holds considerable practical significance. However, relying solely on classification accuracy to evaluate model effectiveness proves insufficient, particularly in the context of fault data scarcity commonly observed in commercial applications. While models may exhibit accuracy degradation under real-world constraints. This discrepancy reveals that accuracy alone fails to reflect performance deterioration under boundary scenarios, whereas the proposed indicators can expose potential reliability risks. To address this limitation, this work introduces two complementary evaluation metrics: inference-weighted accuracy and coverage-weighted accuracy. Inference-weighted accuracy is defined as the product of classification accuracy and the normalized inference scope, penalizing models that perform well but operate on only a limited subset of data. Coverage-weighted accuracy is defined as the product of classification accuracy and the overall scenario coverage, highlighting models capable of maintaining consistent performance across diverse and complex operating conditions. These metrics offer a more comprehensive and practically meaningful evaluation framework for photovoltaic fault diagnosis systems.

Accurate prediction of Remaining Useful Life (RUL) is critical for ensuring the reliability and safety of Lithium-ion Batteries (LiBs) in applications ranging from portable electronics to electric vehicles. However, capacity fading and complex aging mechanisms make RUL estimation challenging. This work presents a hybrid model based on Particle Swarm Optimization (PSO) and Variational Mode Decomposition (VMD) to evaluate the RUL of LiBs. The optimal hyperparameters of the VMD are determined with PSO, before the decomposition of the capacity data into Intrinsic Mode Functions (IMFs). Then, each IMF is input to a Transformer model for time-series forecasting and RUL estimation. The performance of the methodology is assessed with the NASA B0018 and CALCE C37 datasets with a 50%-50% training–testing split. The results show that this proposed hybrid model achieves superior accuracy and robustness compared to other state-of-the-art models.

Prognostic learning systems (PLS) primarily rely on and delve into understanding the behavior of degradation lifecycles. Linear degradation often serves as the prevailing trend for labeling performance indicators (PI) in Remaining Useful Life (RUL) or Health Index (HI) estimations. Nevertheless, some studies explore deriving HI directly from PI, capturing state of health (SoH) variations more effectively. Disregarding the impact of operating condition disturbances on PI measurements reveals that systems do not strictly follow a linear degradation path during deterioration. At certain Health Stages (HS), a system maintains a specific SoH before continuing to deteriorate, sometimes even briefly returning to a better SoH. These multivariate conditions (MVC) arise from factors such as integrated control systems and system resilience. In this less-explored perspective, following procedural data preprocessing, this study aims to enhance data quality in terms of RUL ground truth labeling by considering aspects of temporary stability and performance backup. To assess data quality improvements, a comparative study investigates various adaptive learning systems across multiple evaluation stages and metrics, focusing on generalization as the main criterion for accuracy assessment. Recognizing the inherent non-linearity in PI degradation patterns has proven to be a pivotal element in cleansing feature spaces for PLS, consequently bolstering its generalization capability. This enhancement ultimately contributes to the development of more effective maintenance schedules.

It has been shown that an analytical model including amplitude, phase shift, and the mean value of the phase currents in seven-phase electrical machines can be used as relevant information for fault detection, isolation, and estimation. However, this model fails to estimate the fault severity accurately under noisy conditions, especially for faults affecting the mean value. Therefore, the model is extended with the noise as the fourth parameter. Two approaches are considered before the estimation performance is evaluated: the noise level is estimated, or the currents are first denoised. The simulation results with different combinations of noise level and fault severity show that both approaches are efficient and enhance the fault severity estimation, even under high-noise conditions (SNR as low as 5dB). Besides, the analytical model, including the noise, agrees well with the numerical model.

Switched Reluctance Generator (SRG) is a sustainable and low-cost solution for future small-scale wind-energy-conversion system (WECS). However, the SRG has intrinsic spatial and magnetic nonlinearities and the study of WECS is multidisciplinary and complex. To simplify the modelling and controller design process, this paper proposes a comprehensive step-by-step design method for an SRG based DC micro grid. Thanks to the Energetic Macroscopic Representation, a natural decomposition of the studied WECS with respect to physical laws is obtained. Moreover, a control scheme is easily deduced from this description using inversion-based rules. Three operating modes are evaluated: maximum power tracking, direct power control, and output voltage control. The effectiveness of the methodology is validated under variable-speed wind conditions by simulation in MATLAB/Simulink environment.

Among the data-driven techniques for abnormalities detection in complex systems, Principal Component Analysis (PCA) is popular because of its simplicity and it does not require prior knowledge. However, the used of PCA is limited to stationary data. This work proposes a methodology to address this limitation. It consists of applying Nuisance Attribute Projection (NAP) in the preprocessing stage before the fault features are transformed with PCA to remove the nonstationarity effects due to the variable operating conditions. The proposal is evaluated to detect inter-turn short-circuits in the stator of a Permanent Magnet Assisted Synchronous Reluctance Motor (PMaSynRM) used in the powertrain of electric vehicles. The variances of the phase currents, computed in moving windows, are used as fault features. The results, obtained with seven fault severities and three load conditions, show that monitoring the Hotelling$T^{2}$in the principal subspace leads to good performance, with probabilities of missed detection and false alarms lower than 0.02 and 0.05, respectively. To provide a safety metric, an estimate of the fault level is obtained with an analytical model of the evolution of the slope of the CUmulative SUM decision function with an accuracy greater than 97%.

This paper presents a fast and efficient fault isolation method in 7-phase electrical machines based on the phase currents projections in the stationary reference frames. The study considers both non incipient faults with 15% to 30% fault severities, and incipient ones whose severities vary 1% to 6%. The noise level effect on the fault isolation is also considered. The fault features are extracted from the transformed currents in the frequency domain. The features are processed with a multi-step classification methodology based on usual techniques (principal component analysis, linear discriminant analysis and support vector machine). The simulation results show that the fault classification under low noise level conditions is effective with an accuracy higher than 98%. However, when the noise level increases, the proposal fails to classify incipient faults.

International audience

International audience

International audience

It is always challenging to choose a suitable control strategy to regulate the output voltage of a dc-dc boost converter in dc microgrids (MGs). Indeed, besides the standard requirements on dynamic performance and robustness to parameter variations and environmental nuisances, the controller should handle the power converter’s intrinsic non-minimum phase (NMP) behaviour and the stochastic load variations. This paper compares two typical control strategies: a linear controller-based averaged current-mode control (LC-ACMC) and a finite control set model predictive controller-based voltage mode control (FCS-MPC-VMC). The converter model and control strategies are developed in Matlab/Simulink environment. Qualitative and quantitative comparisons are made in circuit modelling methods, controller design complexity, and dynamic performance. The simulation results show that LC-ACMC is a good trade-off between dynamic performance and controller design complexity although it cannot operate on a wide operating range. However, FCS-MPC-VMC has higher dynamic performance and can operate in continuous conduction mode (CCM) and discontinuous conduction (DCM).

The core technical challenges faced by the battery industry is lifespan, cost, and toxic lead recycling. Increase in battery’s health and use life is a research hotspot to reduce cost and delay recycling. In the first part, a battery-supercapacitor (SC) hybrid is designed and simulated. In this two-stage design, the SC protects the battery from fatal damage and deep discharge. A simple decision-based control strategy is employed to proportionally share the power delivery capabilities of hybrid battery-SC storage units. Later, the rate-limiting strategy is simulated using the fresh and weak batteries for a scaled 125W load with a peak of 1000W in a typical rural home. The simulated results show effective coordination of such a low-cost fresh and weak battery-battery combination for handling light load variation. The proposed hybrid configuration demonstrates practical and economic advantages and best suits community electrification with an increased social acceptance.

Demand for power quality analysis in industrial networks has increased in recent years. Voltage sags are the most frequent and impactful disturbances, with especially high financial implications for industrial clients. Understanding the underlying causes behind voltage sags is a first step towards a mitigation solution. In this paper, we propose a voltage sag cause identification algorithm based on instantaneous symmetrical components and dynamic time warping applied to voltage and current measurements. Short-Time Fourier Transform and Fortescue transform are implemented in the pre-processing stage, obtaining a four-dimension time series signature. Then, a simple but effective multivariate time series classification approach is used. It is based on 1-Nearest Neighbor classifier and dependent Dynamic Time Warping as distance measure (1NN-DTWD). The main advantages of the proposed method are the electrical interpretability of the signatures and the reduced size of the training data. A classification accuracy of 100% is reached with synthetic data, representing seven voltage sag sources. The method reaches a classification accuracy ratio higher than 98% with a reduced real dataset representing five event classes.

The studies dedicated to the battery open-circuit-voltage (OCV) online estimation are not as much as the research efforts on the state-of-charge (SOC) determination and the parameter identification such as capacity and resistance. However, as an important term that represents the distinct characteristic of different Lithium-Ion batteries, OCV should also be estimated. Four estimation algorithms, namely, Luenberger observer, Kalman filter, Recursive least-square with forgetting factor and Recursive least-square with variable forgetting factor are selected and compared in terms of estimation accuracy, computational cost, parameter tuning and robustness to model parameters variations. Simulation results have shown that observer-based methods exhibit better estimation performances than regression-based ones.

The diagnosis (Detection, Estimation and Isolation) of incipient faults, i.e. faults with severity variation2statistics. The fault estimation performances validates the theoretical modeling for incipient faults in noisy environments. An estimation error lower than 3% is obtained even for a Signal to Noise Ratio (SNR) as low as 25dB.

Battery short circuit (SC), including both internal short circuit (ISC) and external short circuit (ESC), is an important stage before thermal runaway (TR). Therefore, on-line incipient battery SC detection is of vital importance to guarantee a safe and reliable operation of Lithium-Ion Batteries (LIBs). In this paper, based on a slightly modified battery equivalent circuit model (ECM), the purpose of battery incipient SC detection is achieved from the perspective of fault estimation. The proposed soft SC diagnosis method is independent of battery intrinsic properties as much as possible. A robust fault estimator, under the form of proportional-integral observer, is designed by solving two linear-matrix-inequality (LMI) constraints. Simulation studies based on an A123-M1 cell have verified the effectiveness of the proposed SC detection method.

Bearing early fault detection and diagnosis (classification, estimation, ...) is a key issue in Condition Monitoring (CM) of rotating machinery. In this context, we propose in this paper a multi-fault classification comparison between traditional Support Vector Machine (SVM) solutions and wavelet SVM (WSVM). For this work several kernel and wavelet functions were considered in the Kullback Leibler Divergence (KLD) framework. First, Empirical Mode Decomposition (EMD) is employed to preprocess vibration signals acquired from the rolling bearings elements. Second, a specific statistical analysis study is performed to select the most salient components from the different obtained Intrinsic Mode Functions (IMFs). Then, the KLD of the retained IMFs is calculated to carry out the classification of three bearing ball fault severities for several operating conditions. Thanks to four criteria, namely the classification accuracy rate average (ARA), the support vector average (SVA), the training time (Trt) and testing time (Tst), our results are derived to highlight the technique allowing to obtain the better results.

Most of signals in the electrical machines and drives are non-Gaussian and highly nonlinear in nature. A useful set of techniques for examining these kinds of signals relies on the spectral representations of higher-order statistics (HOS), well-known as polyspectra. They describe statistical dependences of frequency components that are neglected by traditional spectral measures. The bispectrum is the most used HOS, and studying higher-order correlations provides more information about the electromechanical system's behavior. It helps in building more accurate diagnostic models. Based on this proper relationship the overall aim of the current work is the interpretation of the stator current tidal turbine bispectrum under imbalanced rotor blades condition. Based on this proper relationship, the overall aim of the current work is the interpretation of the permanent magnet synchronous generator (PMSG)-based tidal current turbine (TCT) stator current bispectrum for the diagnosis of biofouling. The proposed bispectrum-based diagnosis method has been tested using experimental data issued from a TCT experiencing biofouling emulated by an attachment on the turbine rotor. The achieved results clearly indicate the feasibility and efficacy of the proposed method.

The cascaded H-bridge multilevel inverter for grid-connected photovoltaic(PV) system has the advantages of high power quality and easy modularization, but as the levels of the inverter increase, the failure probability of the power switching devices will also increase. In the open-circuit faults of the power switching devices, there are two groups of similar faults that are difficult to distinguish. To solve this problem, a secondary classification fault diagnosis strategy based on PCA-SVM is proposed. The first classification is used to make a preliminary fault diagnosis between all types of faults, the second classification is to make a further diagnosis of the two groups of similar faults. Finally, compared with other fault diagnosis strategies, the proposed strategy improves the accuracy of fault diagnosis.

The nondestructive crack detection is an important issue in industrial engineering. However, the detection of incipient cracks that can cause non obvious changes in the conductive material impedance map is difficult. In our paper, we propose a new method based on wavelet and Jensen-Shannon divergence in the framework of Noisy Independent Component Analysis (NICA) to address this problem. The source signals with fault features are obtained by the application of the Independent Component Analysis regarding the noise. Then, the wavelet decomposition is considered as the denoising method to partially reduce the noise influence. The Jensen-Shannon divergence(JSD) which has been proved as an efficient incipient fault detection algorithm in previous works is used here for incipient crack detection. The detection performances of the proposed method is compared with the ones obtained with the Kullback-Leibler divergence often proposed in the literature.

The high penetration of Wind Turbine (WT) in the grid is a promising solution to increase the electricity production with renewable energies. In this work, we propose a data-driven methodology for dip voltage fault detection and diagnosis. From experimental measurements the current vector trajectory deformation in the (αβ) reference frame is derived and a statistical-based analysis (first four statistical moments) of two relevant features are extracted (the ratio between the two axis and the rotation angle) is conducted. Thanks to this ratio, the method is robust to load variations. The fault isolation is done accurately with the analysis of the shift angle. The fault detection performances are evaluated with the ROC curves that reveal a probability of detection equal to 1 and a null probability of false alarm.

In this paper we develop a fault detection and isolation method based on data-driven approach. Data-driven methods are effective for feature extraction and feature analysis using statistical techniques. In the proposal, the Cumulated Sum (CUSUM) efficiency is explored for incipient fault detection. The fault is assumed to be a Gain variation, an Offset evolution, a Phase shifting or one of the multiple possible combination of such faults. A first preprocessing stage is proposed for this study and using some statistical other features we proceed to the operations of a Fault Detection and Diagnosis process: Detection, and Isolation. For the detection, the CUSUM efficiency is proved. For the isolation, we propose a specific algorithm based on the combination of several multivariate statistical techniques such as Support Vector Machines (SVM) and Linear Discriminant Analysis (LDA). The classification of each fault or one of their combination is accurately obtained. The results show that for incipient faults (<;10%), the fault detection and isolation is accurate with a relative classification error lower than 3%.

In this paper, we propose line current sensor fault detection for AC drives. The method is based on the measured currents and the features are extracted either in the natural reference frame or in the transformed Park synchronous rotating frame. The features are the first four statistical moments or the Kullback Leibler Divergence (KLD) of the Probability Density Functions (PDF). For offset fault, we show that if the offset is higher than 3% of the current amplitude, the mean value is the most relevant value among the first four statistical moments that leads to good detection performances (low probability of false alarm and low probability of miss detection). But for incipient faults (offset ranging from 1 to 2%), even the projection in the transformed Park reference frame cannot improve the fault detection. For these cases, we show that the fault information can be retrieved using the PDF and the KLD. This is confirmed by the results showing that the fault is detected with 100% probability of detection.

This paper addresses the estimation of a permanent magnet synchronous machine mechanical position and speed estimation at low speed and standstill. The method is based on the injection of an additive voltage at High Frequency (HF), which exploits the position dependency to the magnetic saliency. Unlike the usual HFI method, this estimator has a simple structure with only one filter. The simulation results prove the efficacy of the estimation under no load and with a load torque in the low speed region and at standstill. The mechanical position estimation errors are lower than 0.035rad (2°).

This paper proposes a data driven approach for dip voltage fault detection and identification using the grid current vector trajectory in the stationary reference frame. Three features are extracted for the different operating conditions to build the database and analysed using Linear Discriminant Analysis to identify the fault type and subtype. In the subspaces spanned by the factorial components the four faults and eight out of nine faults subtype are successfully identified and isolated with an error rate less than 5%. Simulation results prove the efficiency of the proposed algorithm.

Due to the increasing requirements of safety and reliability in more electrified applications (transportation for example), fault detection and diagnosis (FDD) of electrical systems has become a hot research topic. In the process of FDD, there are three steps: fault detection, fault isolation and fault estimation. The first two items are the commonly addressed while the last one is less tackled as it requires the development of a fault model which parameters are relevant of the fault. In this paper, we propose an analytical model of multiple fault effects on a three phases electrical system. The model is based on the expressions of the currents in the (d,q) or Park synchronous rotating reference frame. We prove through simulation results the efficiency of the model for fault combination of gain, offset and phase shift. The results show also the accuracy of the model that could be used for fault estimation purpose.

In this paper, a current sensor fault estimation using the transformed currents in the Park synchronous rotating frame is proposed. We show that from an analytical model, the fault characteristics can be retrieved. From experimental raw data collected from a Permanent Magnet Synchronous Machine drive, the gain or offset fault characteristics (amplitude and frequency) have been estimated with an average error of 10%. For incipient fault, despite the analytical model, the estimation might be tedious because the fault is concealed by the noise. In this case the Kullback-Leibler Divergence between two Probability Density Functions can be computed and the fault estimated from the value of the divergence. If the distributions are Gaussian, the closed form of the divergence allows the fault estimation. However even if the data are not perfectly Gaussian-distributed, the closed form can still be used despite an overestimation of the fault characteristics that is preferable as it is a safety margin.

The Switched Reluctance Machine is one of the most promising electrical machine in variable speed applications because of its intrinsic robustness and its fault tolerant capability. However, the performances of the machine are deteriorated when a defect affects the position sensor. This paper describes a sensorless control for the switched reluctance machine: two methods based on the characteristics of the machine are proposed to eliminate the requirement of the position sensor. The first one utilizes the phase inductance characteristic to determine the rotor position by injecting a test signal while the phase winding is non-energized. The second one utilizes the flux characteristic to estimate the position by measuring the phase voltage and current. Both methods are evaluated through intensive simulation and the results show their good performance respectively in low speed and high speed region. The flux-based method, implemented in a FPGA, has been evaluated on a test bed. The experimental results confirm the simulation ones with an average position estimation error of around 2° mechanical in the range between 1/3 and the rated speed.

Marine current turbine (MCT) have been widely used nowadays, it is important to monitor their health state. Unnecessary marine biological growth or marine pollutants attached to the moving parts will affect the operation of the system by introducing imbalance. The imbalance, regarded as faults, would change the performance of turbine and lead to progressively increasing damages. In this paper, a marine current turbine prototype with permanent magnet synchronous generator (PMSG) has been studied. An innovative imbalance fault detection method is proposed for marine current turbines under the condition of wave and turbulence. In the proposed method, the average frequency of current is calculated through synchronous sampling. Meanwhile, current fluctuation influence is reduced during one revolution. The empirical mode decomposition (EMD) and spectrum analysis are used to achieve fault characteristics. Theoretical analysis, simulation and experimental results under different conditions validate the proposed method. Moreover the proposed method could be used for long-term marine current turbine monitoring in respect to its simplicity and low time cost.

This paper is a contribution to the detection and characterisation of small cracks using Eddy Current Testing in the Non Destructive Evaluation framework. Small cracks are considered as incipient faults defined as gradual faults whose signature is weak and concealed by the noise. They are characterized by high signal to noise ratio and low fault to noise ratio. The detection and diagnosis of such faults is still an open challenge. For complex systems, model-based incipient fault detection and diagnosis (FDD) methods usually fail because of the inaccuracy of the model to describe all the phenomena and their interactions. Data-driven methods using statistical features are very promising as long as historical data are available. However in the case of incipient faults, there is not a significant variation of a single feature. The fault signature lies in the global variation of the signal properties. The proposed method relies on the Kullback-Leibler Divergence (KLD) as a nonparametric fault indicator. It measures the slight dissimilarities between the probability density functions of the current signal compared to the faultless or healthy one. Through experimental results, the KLD exhibits a higher sensitivity than the usual statistical features for the detection of small cracks (with dimensions in the order of 0.1 mm) realized in a nickel-based superalloy plate. Moreover, the detection is done with zero missed detection probability. Furthermore, the fault severity is assessed through the characteristics of the crack (surface, length, and depth). In the principal component analysis framework, the analysis of four statistical features (KLD, mean, variance, and maximum) dependency to the excitation frequency allows to discriminating among the cracks.

In this paper, we present a comparative study of four voting algorithms for two Induction Motor drive fault tolerant control to speed sensor failure schemes. In both Output and Input Fault Tolerant controller, the voting algorithm chooses the most appropriate output signal to ensure the best behavior in degraded mode. The performances are evaluated through simulations of a 7.5kW Induction Motor drive with robustness testing against parametric variations, as well as under load testing. The results show that Euler, Newton-Raphson and Maximum Likelihood voting algorithms are more efficient than the Weighted Average in both Fault Tolerant Control schemes.

The paper describes a Fault Detection and Diagnosis structure of mechanical and current sensors faults of a Permanent Magnet Synchronous Machine (PMSM) drive. The method is based on two interconnected observers: an Extended Kalman Filter (EKF) and a Model Reference Adaptive System (MRAS) observer. The EKF, thanks to its optimality is designed to estimate the position in case of mechanical sensor fault and despite sensor current fault. The MRAS estimates the phase currents using the actual position and speed. The computation and sort of the residuals (difference between measured and estimated values) allows the fault isolation. The structure is evaluated on a 1.1 kW test bed with mechanical and phase current sensor faults. The experimental results are so far promising with the capability of detection and diagnosis of the proposed structure.

Usually, bearing faults are diagnosed by the search of bearing characteristic frequencies in the spectrum of current or vibration signals. This local approach, even efficient, has the drawback of requiring the a prior knowledge of these frequencies. Moreover, characteristic bearing frequencies are only a part of the global spectral signature induced by natural bearing damages. In real situations, a fault on a particular bearing element may not produce the corresponding characteristic frequency. Several multiple harmonics of this frequency and sidebands related to their modulations by rotational frequencies can be quite dominant. An effective diagnosis should rather consider the global fault signature. Based on the fact that the global information encoded in the frequency domain is usually descriptive enough to diagnose and classify bearing faults, the present work proposes a classification scheme for bearing conditions which does not require the characteristic frequencies to be known or estimated. The method combines the envelope analysis, the sliding Fast Fourier Transform (FFT) technique and Principal Component Analysis (PCA). The application on experimental data shows that bearing faults can be diagnosed and classified accurately and without overlapping, irrespective of the system operating point. The extracted spectral features are informative enough to discriminate between different conditions of bearing.

Process-history based methods are very commonly used for fault diagnosis and detection. However their efficiency is closely related to the quality of the measured data. In noisy environments, they usually fail particularly for incipient faults. This paper is an attempt to determine an analytical model allowing to estimate a theoretical threshold for fault detection based on the Fault to Noise Ratio (FNR). This model is developed using the Kullback-Leibler Divergence (KLD). For feature extraction, the used data are previously processed through Principal Component Analysis (PCA). The model is validated with simulated data and the results are so far very encouraging.

In fault diagnosis studies two main approaches are mostly used. The first one consists in designing the full physical or empirical model of the system in healthy and faulty conditions. The major drawback of this approach is the difficulty to obtain an accurate model reflecting all the operating conditions and phenomena. The second approach, used in this work, consists in using signal processing techniques for the characterization of the healthy and faulty behaviors. This paper deals with the study of a fault detection and isolation procedure on a three phase inverter feeding an induction machine drive using pattern recognition techniques. The diagnosis procedure relies on the use of classifiers after the collection of the output currents of the inverter flowing in the machine windings. The proposed classifiers are based on Support Vector Machines (SVM). We show in this paper how it is possible to tune the SVM and also the influence of the data normalisation to perform an effective diagnosis with experimental data.

Fault Detection and Isolation (FDI) based on Principal Component Analysis (PCA) is achieved through the construction of control charts. Control charts differ, primarily, by the subspace into which they were defined, namely, the principle and the residual subspaces. Abnormalities are detected in the plotted monitoring chart if the confidence limit is violated. Often, the Hotelling's T2control chart, defined in the principal subspace, is applied for process monitoring. But to detect a fault with the T2chart, it must cause significant changes in the principal subspace, because little disturbances may be hidden by the large amount of variabilities present in the principal subspace. In this paper, we propose to use the Kullback-Leibler divergence, a probabilistic measure taken from information theory, as a diagnosis criterion. We show the efficiency of this criterion for which we find that small faults which might not be detected by the Hostelling test, become detectable without ambiguity. The simulation results show a significant improvement in the fault detection.