FAULT DIAGNOSIS AND CONDITION MONITORING

Fault diagnosis is an established field of control systems. However, due to the intrinsic characteristics of the processes being monitored, fault detection and health monitoring methodologies need to be adapted for the particular problem at hand. The aim of this research area is to develop fault detection and condition assessment algorithms for industrial applications, drawing from fields such as statistics, machine learning and system identification. This field has nowadays received increased attention under various synonyms such as predictive maintenance, and it is expected to be even more important with the widespread of sensor technologies. The research activity in this area is specifically focused on:

  1. Fault diagnosis and condition monitoring of electro-mechanical actuators in aerospace environment
  2. Fault diagnosis and condition monitoring of industrial machines (packaging, workcenters, etc…)
  3. Fault diagnosis of smart meters networks

 

Fault diagnosis and condition monitoring of electro-mechanical actuators in aerospace environment

Fault detection and isolation via hybrid particle filtering

This approach consists in  modification of the standard particle filter algorithm, and it is applied to face the fault detection issue on an electro-mechanical actuator. The variant, based on a hybrid system interpretation of the health monitoring problem, is known as OTPF (Observation and Transition Particle Filter). By modeling each fault condition as a hybrid system mode, the method is able to assess the most likely regime for each time stamp. Simulation results show the superiority of the method with respect to the EKF (Extended Kalman Filter), especially because the distribution of the disturbances which affect the system is usually not gaussian. Funded by CleanSky 1 JTI under the HOLMES FP7 Project.

Figure 1: OTPF single-step procedure.
Figure 2. Fault detection with gaussian disturbances. Top: OTPF estimated system mode (blue dots) vs. real system mode (black dashed line). Bottom: EKF estimation.

 

Fault detection and isolation via machine learning models

With this method, we propose the use of various features, extracted from an electro-mechanical acquator measurements, as input for machine learning models used for fault detection and isolation. These indexes are based largely on the motor driving currents, in order
to avoid the installation of new sensors. Finally, a Gradient Tree Boosting algorithm has been chosen to detect the system status: the choice has been dictated by a comparison with other known classification algorithms. Furthermore, the most promising features for a classification
point of view are reported. Funded by CleanSky 1 JTI under the HOLMES FP7 Project.

Figure 3. Model-free methodology flowchart.
Figure 4. Classifiers comparison summary.

 

 

 

 

 

 

 

Condition monitoring using batch change detection algorithms

This method consists of a change detection algorithm to monitor the degradation of mechanical components of Electro-Mechanical Actuators (EMA) employed the aerospace industry. Contrary to the standard on-line application of change detection methods, the presented approach can be applied in a batch mode, leveraging on the knowledge of when the data were collected. Results show how the method is able to assess the degradation of the actuator over time, constituting a first step towards a condition monitoring solution for the more-electric-aircraft of the future. Funded by CleanSky 2 JTI under the REPRISE H2020 Project.

Figure 5. Condition assessment via the batch RuLSIF method, with a one-dimensional time series.
Figure 6. First results of the Health Monitoring algorithm. Starting from the top of the figure: 1) time behaviour of the first feature (Crest Factor); 2) time behaviour of the second feature (Root Mean Square); 3) Index computed by the Change Detection algorithm using the “Always Health” policy; 4) Index computed by the Change Detection algorithm using the “Always Previous” policy; 5) Index provided by the Change Detection algorithm using the “Last Change” policy. In this last figure it is also depicted the “Health Status” index (red diamonds)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Condition monitoring using statistical process control methods

We developed a Condition Monitoring (CM) algorithm for Electro-Mechanical Actuators (EMAs) that relies on the Statistical Process Monitoring (SPM) literature. SPM  approaches give a dichotomous answer, i.e. the presence/absence of a fault. In this work, we propose different monitoring indicators to continuously evaluate the health state of the system, using results obtained from SPM methods. We test the approach using a dataset collected from a large experimental campaign on a 1:1 scale EMA for primary flight controls of small aircrafts, that led to EMA failure. Results show the effectiveness of the method. Funded by CleanSky 2 JTI under the REPRISE H2020 Project.

 

Figure 1: Indicator 1 and Indicator 2, as function of test dates (bottom axis) and total number of screw revolutions (top axis).
Figue 2: Indicator 3 and Indicator 4, as function of test dates (bottom axis) and total number of screw revolutions (top axis).

 

 

 

 

 

 

 

 

 

Figure 3: Estimated Weibull distributions (green line) and empirical distribution from data (dashed black line), for different experimental tests.

 

 

Fault diagnosis and condition monitoring of industrial machines

We are currently working on these research projects. News will come soon!

 

Fault diagnosis of smart meters networks

We are currently working on these research projects. News will come soon!