Post on 09-Mar-2016
description
Cascade Fault Detection and Diagnosis for Aluminium Smelting Process using Multivariate Statistical Techniques
Nazatul Aini Abd Majid nabd040@aucklanduni.ac.nz
Improving the fault detection and diagnosis of the aluminium smelting process by using multivariate statistical techniques.
Introduction
The developed framework captures the dynamic trends in process variables by using two multivariate statistical techniques: Multi-way Principal Component Analysis (MPCA) and Multi-way Partial Least Square (MPLS).
To improve the fault detection and diagnosis
This new framework of multivariate analysis was developed because the existing multivariate approaches ( e.g. Tessier et al, 2008) were not capable of capturing the dynamic trends in process variables.
Literature
This new framework, called Cascade fault detection and diagnosis, was divided into two main parts:
•Part I for fault detection •Part II for fault diagnosis
For each part, there were three main phases: data training, model development and process monitoring.
Method
Cascade fault detection frameworkPhase I: Data training
I
JxK
X
J
K
W I
unfolded
Phase II: Multiple Models development
M1M2
M3M4
….. M10
PT Sm s
Scalar limits for the multivariate charts
Phase III: On-line process monitoring
Xnew Anode faults
Calculate t, T2 and SPE values
Abnormal pattern within a cycle
T2 SPE
I: number of alumina feeding cycles
In order to capture the dynamic trend in process variables:1. Process data were organised in a 3-dimensional data array W(IxJxK) in
order to treat an alumina feeding cycle as a batch operation2. In model development, different models were developed for different
phases in an anode changing cycle.
The dynamic patterns related to abnormal events were detected by using two monitoring charts, the Hotelling’s T2 and SPE charts.
Cascade fault detection framework(cont’d)
Phase II: Multiple Models developmentM1
M2M3
M4….. M10
W mQ s
Reference set for
diagnosing faults
B
Variability patterns
discovered using Knowledge
Discovery from Databases (KDD)
Phase III: On-line fault diagnosis
Calculate Y and T
Abnormal pattern within a cycle
Diagnosis results
Part I: Cascade Fault Detection
A) C)
Ci) Cii) Ciii)
B)
Hierarchal diagnosis
Phase I: data training
3-D data unfolded to 2-D data matrix
Cascade fault diagnosis framework
• The fault diagnosis framework also captures the dynamical trend by organising the data in a 3-dimensional data array W(IxJxK) and developing multiple models.
• The fault diagnosis framework was developed based on a hierarchal diagnosis approach. This approach makes the fault diagnosis more effective.
Cascade fault diagnosis framework(cont’d)
ResultsThe results of this thesis show that this multivariate framework is capable
of capturing the dynamic trends in process variables.
Contributions• The major contribution of this thesis is the development and
evaluation of a new multivariate framework, Cascade fault detection and diagnosis, that was based on 3 key factors• The treatment of the overfeed/underfeed cycle as a batch using
MPCA and MPLS• The incorporation of the behaviour of the cell during anode
changing• The development of a hierarchal diagnosis.
• The application of this thesis will help aluminium smelting plants to operate at full capacity.
• The algorithm developed in this thesis is limited to the aluminium smelting process with a point-feed strategy for point-feed cells.
• This algorithm would have limitations for bar-breaker cells that use demand feed. • This algorithm is also limited to alumina smelting processes that have different patterns
during anode changing.• In this thesis, the pattern was a step increase in voltage, followed by a more
gradual downward trend in cell voltage. The model developed for the phase when the downward trend ended can still be used for cells with different operating conditions for anode changing.
Questions -1a
• We can develop a statistical indicator by measuring the false alarm rate. • This is to examine the model accuracy and determine the appropriate time for
model updating. • The statistical indicator could be the rate of false alarms of the process. When the
rate of false alarms exceeds the pre-defined limit, this indicates that the model should be updated.
Examples of strategy for model updating: 1) Periodic offline rebuilding of models, 2) The development of automated model updating methods, and 3) Combination of these activities.
(Miletic et al. 2004)
Questions-1b
• By treating an alumina feeding cycle as a batch operation, W (I x J x K) • The W (I x J x K) were unfolded and rearranged into a two-dimensional data
matrix, Xold (I x JK) where every row contained all observations (J x K) within a feeding cycle.
• The data matrix X were mean centred by subtracting the mean of each column of this matrix.
• This way of subtracting the mean trajectory is actually subtracting the non-linear trajectory within an alumina feeding cycle.
• This is how the non-linearity in the process data is removed (in section 6.4 : addressing the 1st research question, 2nd paragraph).
Questions-2
Questions-2(cont’d)Multi-way Principal Components Analysis (MPCA)
A 3-dimensional data array
WNumber of feeding cycles (I)
Variables (J) Observations (K)
XI
J x K
unfolded
Apply PCA
MPCA models
This is supported by arguments in the thesis:
This unfolding is particularly meaningful because, by subtracting the mean of each column of this matrix X, we are in effect subtracting the mean trajectory of each variable, thereby removing the main nonlinear and dynamic components in data. A PCA performed on these mean-corrected data is therefore a study of the variation in the time trajectories of all the variables in all batches about their mean trajectories (Nomikos and MacGregor (1995)(p.43).
The advantage of modelling the deviations from the target trajectory was also pointed out by Kourti (2003) is as converting a non-linear problem to a linear one so that it is easy to tackle with linear latent variable methods such as MPCA and MPLS (thesis, section: 4.4.1 pp 73).
Questions-2(cont’d)
• Comments on robustness: • During the course of this research, after pre-processing, process data were recorded
as deviation variables from their set points before the development of the reference models.
• Therefore, the differences in age and temperature of the cells were removed by monitoring the deviation of the process from its target (current state).
• For example, when the set point of cell voltage for the new data is increased because of the low bath temperature, the reference model of the Cascade monitoring system is still valid for the monitoring of the new data because their set points will be subtracted.
• Thus, the reference model can be used for monitoring a number of cells and at different time periods of plant operation. To support this, and by way of validation, the recorded faults used in the offline data testing were from different cells and were also from different periods of plant operation.
• If the feeding pattern and anode setting cycles are changed, the models need to be updated.
Questions-3a
• Yes, we do expect a general trend of anode changing when anodes at different locations are changed. We expect the trend is based on the four different phases that occur during anode changing (page 23).
Questions-3b
• It is possible to develop a fault detection and diagnosis based on frequency domain data.
• This is based on the research done by Hesong Li, Chi Meia, Naijun Zhoua, Qian Tang, Yongbo Huang (2006).
• They developed a method for diagnosis of working conditions of an aluminium reduction cell using frequency domain process data. However, they used wavelet transform instead of Fourier transform. By using this technique, the frequency spectrum characteristics of the voltage vibration signal from aluminium reduction cells were extracted.
• In my research, we may use wavelet or Fourier transform to extract a signal for anode changing but we still need multiple models because we monitor the trend within an alumina feeding cycle.
Questions-4
Rationale of using PCA/PLS instead of FDA: • FDA is used to classify multiple fault classes by using a set of projection vectors that
maximize the scatter between the classes and minimize the scatter within each class (Chiang, 2000).
• However, although FDA is superior than PCA and PLS for classification with its slightly more complex algorithm, FDA has been less adopted for process monitoring (Chiang, 2000).
• In fact, the research efforts in PCA/PLS based process monitoring date back more than a decade.
Questions-5
•Since the use of PCA and PLS for multivariate process monitoring is more practical and widespread, PCA/PLS was used this thesis. •However, a future study investigating the use of FDA for monitoring the aluminium smelting process would be very interesting.
Thank you
Summary of Results
Fault diagnosis
Fault detection T2 SPE
> 95% &< 99%
>99% > 95% &< 99%
>99%
Average detection rate for anode
spikes 7.52 57.34 9.73 9.37
Average detection rate for anode
effect 25 75 33 25
Anode spikes Anode effect
Average diagnosis success rate
37 72