In this section, numerical examples and TEP are used to explore the effectiveness of the proposed method in fault detection. In addition, the mutual protection phenomenon of outliers is explored and solved using the elbow method to improve the detection performance of FD-MkNN.
3.1. Numerical Simulation
The number of generated training samples is 300. The outliers follow the Gaussian with mean 2 and variance 2 [
25], the proportion of outliers compared to the training samples is set to 0%, 1%, 2%, 3%,
, and
, respectively. In addition, there are 100 testing samples, of which the first 50 samples are normal, and the rest are faulty.
where
is a latent variable with zero mean and unit variance, and
is a zero-mean noise with variance
.
FD-kNN is first applied to detect the faults in the data set. The number of nearest neighbors is 3. At the confidence level of 99%, the detection result is shown in
Figure 3. It can be seen that, as the proportion of outliers increases, the detection performance of the FD-kNN method degrades seriously. As shown in
Table 1, when the ratio of outliers is 5%, the fault detection rates (FDR) of the FD-kNN approach is only 20.00%. Due to outliers in the training samples, part of the neighbors of the samples found using kNN rule in the training phase are pseudo-neighbors. These pseudo-neighbors seriously affect the determination of the control threshold (that is, the control limit will be much greater than the average level) and result in poor fault detection performance.
For FD-MkNN, the parameters
and
are set to 3 and 5, respectively. At the same confidence level (that is,
), the detection result is shown in
Figure 4. As shown in
Table 1, when the proportion of outliers increases from 0 to 2%, the detection performance of the FD-MkNN method is not significantly affected, and the FDR always remains above 90%. When the proportion of outliers increases from 2% to 5%, the FDR of the FD-MkNN method is significantly reduced but the FDR is always better than that of FD-kNN.
The false alarm rates (FAR) of the two methods are shown in
Table 2 (Note that the FAR is obtained based on the normal training samples). Due to outliers, the control limit or threshold of the FD-kNN method seriously deviates from the average level. Therefore, the FAR of the FD-kNN method is all zero.
The reason why the fault detection superiority of the FD-MkNN is better than that of the FD-kNN is as follows:
Before the training phase, part of the outliers in the training samples are removed so that the outliers will not affect the determination of the control limit in the training phase;
In the fault detection phase, MkNN carries more valuable and reliable information than kNN. Furthermore, the effect of PNN is eliminated.
3.1.1. Experimental Results of FD-MkNN with Different Values of k
The values of
k in the outlier elimination and fault detection stages are different and can be denoted as
and
, respectively. The larger the value of
k, the higher the probability that the query sample finds its mutual neighbors. Therefore, MkNN can more easily identify outliers when the value of
is generally smaller than
. However, the value of
cannot be too small because the MkNN method will misidentify the normal training samples as outliers and eliminate them. For example, as shown in
Table 3, when the value of
is set to 1, the MkNN method will eliminate all 300 training samples (the actual proportion of outliers introduced is 5%), resulting in the failure of the MkNN fault detection stages. As the value of
increases, the number of outliers removed decreases, which makes the monitoring threshold deviate from the normal level, and the FDR decreases seriously.
3.1.2. Mutual Protection Phenomenon of Outliers
As shown in
Figure 5, when two outliers are relatively close, an interesting phenomenon will appear: they will become each other’s mutual nearest neighbors. Therefore, the MkNN rule cannot identify them as outliers. For example, in
Figure 5,
,
, and
are protected by 1, 2, and 3 outliers, respectively. When the outliers far from the normal training samples are kept in the training set due to mutual protection, it will cause the threshold or control limit calculated in the training phase to deviate seriously from the average level. We call this phenomenon the “Mutual Protection of Outliers (MPO)”, which is also the main reason why the detection performance of the FD-MkNN method decreases when the proportion of outliers increases from 2% to 5%.
It can be observed from
Figure 5 that, for outliers with mutual protection, the corresponding MkNN distance statistic is significantly larger than that of the normal training sample. Therefore, the elbow method [
26] is used to eliminate outliers with mutual protection: first, arrange the MkNN distance statistics of the training samples in descending order, then determine all samples before the elbow position as outliers with mutual protection, and finally eliminate these outliers from the training set.
As shown in
Figure 6, the outliers with mutual protection can be identified according to the elbow method, that is, all samples before the elbow point. After determining the outliers with mutual protection, these outliers need to be removed from the training set. Finally, the process monitoring method was repeated. The detection results are shown in
Figure 7. After eliminating outliers with mutual protection, the recalculated threshold (that is, the red dotted line in
Figure 7) is more reasonable, and the FDR has reached 100.00%, as shown in
Table 4.
3.2. The Tennessee Eastman Process
When comparing the performance or effectiveness of process monitoring methods, the TEP [
27] is a benchmark choice. In [
28,
29], Downs and Vogel proposed the simulation platform. There are five major operating units in the TEP, namely, a product stripper, a recycle compressor, a vapor–liquid separator, a product condenser, and a reactor. The process has four kinds of reactants (A, C, D, E), two products (G, H), contains catalyst (B), and byproducts (F). There are 11 manipulated variables (No.42–No.52), 22 process measurements (No.1–No.22), and 19 composition variables (No.23–No.41). For detailed information on the 52 monitoring variables and 21 fault patterns, see ref. [
27]. The flowchart of the process is given in
Figure 8.
The number of training samples and the number of validation samples are 960 and 480, respectively. In addition, there are 960 testing samples where the fault is introduced from the 161st sample. To simulate the situation that the training data contains outliers, outliers whose magnitude is twice the normal data are randomly added to the training data. The thresholds of different methods are all calculated at a confidence level of .
These three faults are chosen to demonstrate the effectiveness of the proposed method. The parameter
k of FD-kNN is 3. The parameters
and
of FD-MkNN are 42 and 45, respectively. For the FD-MkNN method, the outliers with mutual protection phenomenon are first eliminated by the elbow method, as shown in
Figure 9.
According to [
29,
30], fault 1 is a step fault with a significant amplitude change. When fault 1 occurs, eight process variables are affected.
Figure 10 and
Figure 11 are the monitoring results of fault 1 by FD-kNN and FD-MkNN, respectively. As the proportion of outliers increases, the detection results of kNN and MkNN for fault 1 are not significantly affected. For example, the FDR of MkNN for fault 1 remains at 99.00%, as shown in
Table 5. Because fault 1 is a step fault with a significant amplitude change, the outliers introduced in this experiment are insignificant in the face of this fault. Although these outliers also deviate the control limits from normal levels, they do not have much impact on the fault detection phase. The fault false alarm rate of FD-kNN and FD-MkNN is shown in
Table 6.
The fault 7 is also a step fault, but its magnitude changes are small, and only one process variable (i.e., variable 45) is affected.
Figure 12 and
Figure 13 are the monitoring results of fault 7 by FD-kNN and FD-MkNN, respectively. As shown in
Table 7, as the proportion of outliers increases, the FDR of FD-kNN drops from 100.00% to 18.75%, while the FDR of FD-MkNN does not decrease significantly and remains above 90.00%. The fault false alarm rate of FD-kNN and FD-MkNN is shown in
Table 8.
According to the detection results of fault 1 and fault 7, it can be seen that FD-MkNN is suitable for the processing of incipient faults. Because outliers will significantly increase the threshold, the detection statistic of incipient faults is lower than the threshold. The proposed method eliminates outliers by judging whether the samples have MkNNs, thereby improving the fault detection performance.
Fault 13 is a slow drift in the reaction kinetics.
Figure 14 and
Figure 15 are the monitoring results of fault 13 by FD-kNN and FD-MkNN, respectively. In
Table 9 and
Table 10, as the proportion of outliers increases, the FDR of the FD-MkNN is always better than that of FD-kNN, while the FAR is higher than that of kNN. Due to the appearance of outliers, the threshold of the kNN is increased so the FAR of FD-kNN is always 0.00%.