Andrea Bernieri

A neural network approach for identification and fault diagnosis on dynamic systems

The possibilities offered by neural networks for overcoming both system identification and fault diagnosis problems in dynamic systems are investigated. In particular, an original neural fault diagnosis procedure is illustrated. Its sensitivity and response time enables it to be used to great advantage in online applications. Some applications are also reported which, although pertaining to a simple linear dynamic system, highlight the general applicability and advantages of a neural approach. >

Crack Shape Reconstruction in Eddy Current Testing Using Machine Learning Systems for Regression

Nondestructive testing techniques for the diagnosis of defects in solid materials can follow three steps, i.e., detection, location, and characterization. The solutions currently on the market allow for good detection and location of defects, but their characterization in terms of the exact determination of defect shape and dimensions is still an open question. This paper proposes a method for the reliable estimation of crack shape and dimensions in conductive materials using a suitable nondestructive instrument based on the eddy current principle and machine learning system postprocessing. After the design and tuning stages, a performance comparison between the two machine learning systems [artificial neural network (ANN) and support vector machine (SVM)] was carried out. An experimental validation carried out on a number of specimens with different known cracks confirmed the suitability of the proposed approach for defect characterization.

A neural network approach to instrument fault detection and isolation

The growing diffusion of Artificial Neural Network (ANN) applications suggests the authors a possible solution to Instrument Fault Detection and Isolation (IFDI) problems. It is based on the modelling of both the measurement station and the system under analysis by a suitable ANN, having the input layer fed by instrument outputs and the output layer which gives information for faulty instrument detection and isolation. The methodologies adopted are described in detail and tested on a complex automatic measurement station for induction motor testing. The performance of the proposed IFDI scheme is experimentally evaluated mainly in terms of correct diagnosis, incorrect fault isolation, missed fault detection, and false alarm. The proposed diagnostic scheme proves to have good performance also out of the domain on which it was trained.<<ETX>>

On-line fault detection and diagnosis obtained by implementing neural algorithms on a digital signal processor

A measurement instrument for on-line fault detection and diagnosis is proposed. It is based on the implementation of a neural network algorithm on a processor specialized in digital signal processing and provided with suitable data acquisition and generation units. Two specific implementations are detailed. The former uses the neural-network to simulate on-line the correct system behavior, thus allowing the fault detection to be achieved by comparing the neural network output with the measured one. The latter uses the neural network to classify on-line the system as correct or faulty, thus allowing the fault detection and diagnosis to be achieved simultaneously. These two implementations are applied to detect on-line and diagnose faults on a real system in order to point out different fields of application and to highlight the performance of the measurement apparatus.

Crack Depth Estimation by Using a Multi-Frequency ECT Method

In many industrial application fields as manufacturing, quality control, and so on, it is very important to highlight, to locate, and to characterize the presence of thin defects (cracks) in conductive materials. The characterization phase tries to determine the geometrical characteristics of the thin defect namely the length, the width, the height, and the depth. The analysis of these characteristics allows the user in accepting or discarding realized components and in tuning and improving the production chain. The authors have engaged this line of research with particular reference to non-destructive testing applied to the conductive material through the use of eddy currents. They realized methods and instruments able to detect, locate, and characterize thin defects. In this paper, a novel measurement method able to improve the characterization of the crack depth is proposed. It is based on the use of a suitable multi-frequency excitation signals and of digital signal processing algorithms. Tests carried out in an emulation environment have shown the applicability of the method and have allowed the tuning of the measurement algorithm. Tests carried out in a real environment confirm the goodness of the proposal.

A measurement system based on magnetic sensors for nondestructive testing

The paper deals with a measurement system based on a low-cost eddy current probe for nondestructive testing (NDT) on conducting materials aimed at reconstructing the shape and position of thin cracks. The magnetic probe is characterized, highlighting good repeatability, linearity, and overall accuracy. A number of different measurement approaches are investigated, in order to choose the most appropriate for NDT applications. A numerical method is then illustrated; it proves to be able to reconstruct cracks starting from noisy measurement data.

Characterization of an eddy-current-based system for nondestructive testing

A low-cost probe for nondestructive testing of conductive materials was set up and tested. It is based on the measurement of the magnetic field produced by eddy currents (ECs) and perturbed by the presence of cracks. The magnetic sensor used, characterized by the authors in a previous work, gives an output that, suitably processed, allow both the amplitude and phase of the magnetic field to be measured. The preliminary tests reported confirm the suitability of the proposed probe for detecting cracks and correctly identifying their direction and depths.

Error compensation of A/D converters using neural networks

This paper describes a new technique for compensating errors in analog-to-digital converters (ADCs). It can be considered an improvement of the phase plane compensation technique: the idea is to exploit the generalization capabilities of Artificial Neural Networks (ANNs) to reduce the large number of experiments required. The ANN is built and set up in a simulation environment using an ADC behavioral model, whose errors can be fixed to known values. It is thus possible to simulate a set of ADCs with very different performances, thereby enabling the usefulness of the proposed approach to be investigated in very different working conditions. The results were analyzed by comparing the behaviour of uncompensated and compensated ADC outputs.

A digital-signal-processor-based measurement system for on-line fault detection

This paper deals with the design, construction, and setting up of a measurement apparatus, based on an architecture using two parallel digital signal processors (DSPs), for on-line fault detection in electric and electronic devices. In the proposed architecture, the first DSP monitors a device output on-line in order to detect faults, whereas the second DSP estimates and updates the system-model parameters in real-time in order to track their eventual drifts. The problems which arose when the proposed apparatus was applied to a single-phase inverter are discussed, and some of the experimental results obtained in fault and nonfault conditions are reported.

ADC neural modeling

The paper deals with a neural network-based approach for A/D converter (ADC) modeling. This approach is initially developed with reference to an ADC mathematical model, allowing the neural modeling to be properly setup. Subsequently, a neural model of a real ADC is identified and validated, highlighting the performance of the proposed approach in terms of ease of model building and result accuracy.