Olivier Ondel

Fault Detection and Diagnosis in a Set “Inverter–Induction Machine” Through Multidimensional Membership Function and Pattern Recognition

Nowadays, electrical drives generally associate inverter and induction machine. Thus, these two elements must be taken into account in order to provide a relevant diagnosis of these electrical systems. In this context, the paper presents a diagnosis method based on a multidimensional function and pattern recognition (PR). Traditional formalism of the PR method has been extended with some improvements such as the automatic choice of the feature space dimension or a ldquononexclusiverdquo decision rule based on the k-nearest neighbors. Thus, we introduce a new membership function, which takes into account the number of nearest neighbors as well as the distance from these neighbors with the sample to be classified. This approach is illustrated on a 5.5 kW inverter-fed asynchronous motor, in order to detect supply and motor faults. In this application, diagnostic features are only extracted from electrical measurements. Experimental results prove the efficiency of our diagnosis method.

Adaptive diagnosis by pattern recognition: Application on an induction machine

In this paper, a pattern recognition method is used to provide the tracking and the diagnosis of a system. To illustrate it, we used as application, an asynchronous motor 5.5 kW with squirrel-cage, in particular for the detection of broken bars, under any level of load. From measurements carried out on the system, parameters are calculated. These parameters are used to build up a pattern vector which is considered as the system signature. To determine this pattern vector, two methods are applied. One, well-known, sequential backward selection (SBS) and the other, which we developed, based on a genetic approach, with the advantage to determine the optimal dimension of the representation space and to give better results (value of criterion) than SBS. The determination of the decision space is carried out using a method of automatic classification called clustering. The decision phase is based on the ldquok-nearest neighborsrdquo rule, associated with an evolution tracking of system using trajectory allowing a diagnosis not only of states defined in the training set, but also of the intermediate states. The appearance of a new operating mode is taken into account in order to enrich the initial knowledge base and thus to improve the diagnosis.

Beyond the diagnosis: the forecast of state system Application in an induction machine

This paper deals with the tracking and the prediction of the evolution of the system operation. The aim is to define a forecast of future operating state of the process by using the previous state. First of all, a signature is determined in order to monitor the evolution of different operating modes. For this purpose, on the example of an induction machine, diagnostic features are extracted from current and voltage measurements without any other sensors. Then, a feature selection method is applied in order to select the most relevant features which define the representation space. A polynomial approach of tracking evolution is presented. Next, a Kalman algorithm is developed to predict evolution and to allow pre-empting on the appearance of a fault and the accelerated ageing of system. Finally these two approaches are applied and compared with an induction machine of 5.5 kW with squirrel-cage.

Using linear interpolation and Kalman prediction in Pattern Recognition: Application to an induction machine

This paper deals with pattern recognition (PR) method associated with a tracking and a prediction of evolution for various operating modes of a process. The aim is to improve diagnosis of a process by enhancing its knowledge database. Indeed, PR needs an initial database named training set. It is composed of different operating modes and obtained during the first step of PR. It is commonly named training phase. It is a laborious step and moreover the whole of operating modes is never available (generally poor experimental feedback). Thatpsilas why, using knowledge in training set, it is interesting to predict evolution of operating modes in unknown fields of representation space. PR steps are first presented and followed by a polynomial approach of tracking evolution. Next, a Kalman algorithm is used to predict evolution and finally two different asynchronous machines (5.5 kW and 18.5 kW) are used to illustrate our purpose.