·Table of Contents
·Terrestrial and Marine Transportation
Unsupervised Classification of Acoustic Emission Sources from Aerial Man Lift DevicesAthanassios Anastassopoulos, Dimitrios Kouroussis, Apostolos Tsimogiannis.
ENVIROCOUSTICS S.A., El. Venizelou 7 & Delfon, 114 52 Athens GREECE
Various types of AE sources are expected during testing of Aerial Man Lift Devises, arising from the fibreglass components, the metal parts of the arm, the high strength pins, the welds, as well as the hydraulic systems and the lift mechanisms. The use of pattern recognition analysis, as applied in the present work, aims to identify noise sources from the mechanisms used to manipulate the arm movements and to discriminate signals from various failure mechanisms arising from the different materials.
Results from different unsupervised classification schemes, applied either on the AE feature set, or to its principal component projection are presented. Discussion is focused on the validity of the resulting partitions by using numerical optimisation criteria and common Acoustic Emission practices such as cumulative plots and emissions during load hold.
The proposed methodology proved efficient for the discrimination of AE sources recorded during proof testing of Aerial Man Lift Devices and can be used as a basis for automating the evaluation of Acoustic Emission data from future tests of similar devices.
AE testing of Insulated Aerial Man Lift devices has been applied successfully and became standard procedure for testing both new and in-service vehicles -. The existing standards -, for such type of testing, do not outline pass-fail criteria for each structural part and material, while only ASTM standards , present guidelines for determining when the test should be stopped immediately to avoid damage, i.e. emergency test termination. Conventional analysis - is based on the events ("zonal" location - first hit) recorded by channel, amplitude distribution for metal and for FRP channels and counts per channel. In some cases, major industrial users  established evaluation criteria for specific types of devices in co-operation with the device manufacturers. The aforementioned evaluation criteria are based on the total counts and the number of high threshold events (first hit) during second loading and second hold period. In any case, evaluation criteria refer to filtered data, where noise sources have been identified and filtered.
The main problem associated with data analysis is the discrimination between genuine emission and the various potential noise sources such as the mechanisms (usually chain and gears) used for arm movements and the hydraulic system which remains under pressure during testing. Furthermore, since both metal and composite parts are simultaneously tested, signature recognition becomes more complicated. In the present study, unsupervised pattern recognition techniques - are used for the analysis of AE data, acquired during testing of five Insulated Aerial Man Lift devices. The aim of the work is to enhance analyst efficiency in discriminating various AE sources and establish an automated procedure for noise rejection and AE signal classification in order to evaluate AE activity from future tests of similar devices.
Results from different unsupervised classification schemes, applied either on the AE feature set, or to its principal component projection are presented. Discussion is focused on the validity of the resulting partitions by using numerical optimisation criteria and common Acoustic Emission practices such as cumulative plots, emissions during load hold or during unloading as well as location of AE signals from the different classes identified by means of pattern recognition analysis. Further to the validation of the unsupervised classification, supervised algorithms were successfully trained and applied.
|Fig 1: Sensors position and overall assembly of the device|
Channels 1 to 6 were attached to the composite/insulated parts, while the remaining channels to the metal parts.
A PAC-SPARTAN-2000 AE system was used for real time data acquisition. Acquisition was performed at 40dB threshold and 23dB gain.
The aerial device was positioned and loaded according to the ASTM  guidelines for Non-Over-Centre Models (upper boom horizontal, lower boom vertical, 90o angle between upper-lower boom). Two loading cycles were performed with 10 minutes load hold at the maximum load (1.5 times the rated capacity), as shown in Figure 2 for the first of the five devices tested.
|Fig 2: Amplitude, Cumulative Events, Cumulative First Hit Counts Vs. Time for low level Threshold 40dB (left column) and high level Threshold 70dB (right column). Loading cycles in background plot.|
ASTM guidelines were used for immediate test stop in case of extreme damage. The associated graphs for the two emergency criteria, i.e. 150 total first hit events with amplitude greater than 70dB (high level threshold) and 150000 total first hit counts with amplitude greater than 40dB (low level threshold) are shown in Figure 2.
AE hits, are represented in the multidimensional space by a pattern vector, whose components are the recorded AE features (Amplitude, Counts, etc.). Euclidean distance is used as a measure of similarity between pattern vectors. Data close to the cluster centre is grouped together to form a class, while data further apart is assigned to a different class. Data clustering can be easily understood and the results visually confirmed in two dimensional feature spaces, where data in the same class are tight together and well separated from data of different classes. However the main difficulty in evaluating and validating the clustering results in multidimensional space, is human inability to visualise the geometrical properties of such space.
Furthermore, in the absence of any apriori information regarding the signal classes, the features composing the pattern vector, cannot be selected on the basis of discriminant analysis. To overcome this problem, the technique, known as Principal Components Analysis, is used. Six out of the eight recorded AE features, are selected and used as the default feature space in order to define a transformation and project the data on a set of orthogonal axes where maximum variance is achieved . The original pattern vector comprising Amplitude, Duration, Energy (MARSE), Average Frequency, Rise Time and Counts to Peak, is projected on their principal axes, defined by the eigenvalues-eigenvectors of the correlation matrix. The first three principal components, named PCA0, PCA1 and PCA2, resulted in a degree of fit 95.52% and are further used for the unsupervised pattern recognition analysis and classifier design.
Four different clustering algorithms were used for a parametric study in partitioning the AE data in classes ranging from two to twenty-five. The algorithm performance was evaluated by means of Rij and Tou criteria -. Both are heuristic criteria based on the ratio of average within-class distances to the distance between classes. The Rij criterion is an average measure of such a ratio, calculated using all of the different pair of classes, while the Tou criterion is defined by the ratio of the minimum distance between any pair of classes to the maximum of the average within-class distances. Therefore, the lower the value of Rij, (or the higher the value of Tou), the higher the discrimination efficiency of the resulting data partition. Furthermore, minimisation of Rij or maximisation of Tou criteria as a function of the resulting number of classes can be used to estimate the number of classes in the data.
Two out of the four clustering algorithms, named Max-Min Distance and Cluster Seeking, are heuristic procedures based on predefined distance threshold for the definition of classes. The remaining two algorithms, named K-Means and Forgy, are iterative procedures aiming to minimise the sum of squared error for a predefined number of classes.
The results of the parametric study performed with K-Means clustering algorithm are presented in figure 3. As can be seen from Figure 3, the Rij criterion is minimised at four classes (a second local minimum observed at eleven classes), while Tou criterion is maximised for three and four classes. Therefore the resulting data partition in four classes was considered to be the most representative, from the numerical point of view, for the available AE data.
|Fig 3: Clustering optimisation & estimation of number of clusters|
Experimentation with the remaining clustering algorithms produced comparable results from the criteria point of view. A better performance resulted by the Max-Min distance, clustering algorithm. The algorithm resulted in lower value of Rij for a seven classes partition. A closer investigation showed that the bulk volume of the four main classes resulted from the selected K-Means partition remained the same, while some of the extreme AE hits were split to form additional classes. Selection between different clustering results and the ultimate validity of the resulting data partitions can be performed only in relation to the physical phenomena investigated. For this purpose, standard AE practices are used.
The respective clustering results of the K-Means algorithm are presented in Figure 4. As can be realised from figure, the left column refers to the working copy of the data, i.e. the projected data presented to the clustering algorithms. The scatter plot of the 1st vs. the 2nd principal component (PCA0 vs. PCA1) is indicative of the geometrical shape of the partition, while the bar distribution plot of 1st principal component is indicative of the class population and contribution to the square error. Although such presentation is useful for the assessment of the algorithm performance, it does not allow correlation with the physical phenomena and source characterisation.
|Fig 4: K-Means Clustering results on the principal component space (left column) and the original AE feature space (right column).|
A better insight on the results is gained by simultaneous evaluation of the results by means of standard AE plots. Indicative graphs (Amplitude-Energy scatter and the corresponding Max-Min cumulative distribution of Amplitude) of the clustering results in the original feature space are presented in the right column of figure 4. The characteristics of each class for the resulting K-Means partition (data of Figure 4) are summarised in the following table:
|Class||Hits||AMP (dB)||Aver. Freq. (KHz)||Energy (MARSE)|
|0||438||40-71 (46.8)||3-333 (59.3)||1-300 (28.1)|
|1||103||50-77 (62.2)||39-120 (82.8)||31-1829 (371.5)|
|2||44||40-42 (40.2)||400-1000 (690.9)||1-4 (2.8)|
|3||50||69-97 (88.8)||69-105 (91.8)||764-11396 (4312.2)|
Hits of Class 0, were recorded by all channels. The majority of hits in Class 1 were recorded by channels 1, 4, 7 and 8 while very few hits of this class were recorded by channel 6 and 9 to 13. Hits of Class 2, were recorded by all channels except 2, 3 and 5. Finally hits of Class 3 were recorded only by channels 6 and 7 (only 1 hit in channel 8).
|Fig 5: Activity vs. time and linear location of class 1 data||Fig 6: Activity vs. time and linear location of class 0 data|
On the other hand, data of Class 0, for which the amplitude ranges between 40dB and 71dB, are characterised by lower energy and the lowest average frequency among the four signal classes. AE hits from this class locate between sensors 9 and 10 as can be seen from the bottom graph of Figure 6. In addition to that, Class 0 is active during all the different stages of the loading cycle (mainly during loading and unloading) as can be seen from the hits versus time graph of Figure 6. Most important is the fact that the located events shown in Figure 6 appear only during unloading.
AE data attributed to classes 2 and 3 are not located by the linear location group.
The zonal location results indicate that data of Class 3 are located only by channels 6 and 7, while first-hit data of Class 2 are located by the channels associated with the metal parts of the aerial device. Considering Figure 4 and Table 1, Class 2 has very low amplitude, very low energy and very high average frequency. On the contrary, data of Class 3 has amplitude greater than 75dB and is observed only in channels 6 and 7. As in the case of class 1, class 3 is active during the first loading, while class 2 has similar behaviour with that of class 0, i.e. it remains active during both loading, load hold and unloading at both cycles.
Summarising the results, Class 0 might be characterised as structurally insignificant or friction or noise related classes. Class 2, is also structurally insignificant with very few hits. Class 1 is considered as structurally significant while characterisation of Class 3, requires further investigation. As said, the first-hit data of Class 3, are seen only in sensors number 6 and 7, which are very close to the pins connecting the basket with upper boom and the one connecting the two booms (see Figure 1).
|Fig 7: Back Propagation Net Topology|
Convergence during training stage was achieved after 575 iterations. The network was applied for the classification of the test data (of known classification - 50% of the available data from unsupervised), resulting in 0.63% misclassification error (2 hits). The nearest neighbour classifier resulted in 1.57% misclassification error (5 hits).
The Back Propagation Neural Net was successfully used to classify the AE data recorded during testing of the remaining four aerial man lift devices. The number of hits classified in class 1 and class 3, can be used as evaluation criteria for the assessment of the device structural integrity. At this point it is worth noting that the device number 3 (indicated as AE aerial 3 test), during the first loading cycle was accidentally loaded to higher than 1.5 times the rated capacity. The overloading resulted in the highest number of hits classified in class 1 (and the highest percentage) among all the devices tested.
The network was then used to classify AE data recorded without load (weight) and while the device operator performed standard movements of the boom to position it to its normal position. Using the location set-up presented in figures 5 and 6, these data located between sensors 9 to 11. The classifier assigned 97.3% of the total hits to the noise classes (89.7% in class 0 and 8.2% in class 2). In addition to that, the classifier was used for the discrimination of AE signals recorded during device operation (movements of the boom) with a weight hanging from the basket. In this case 4.4% of the data were classified in classes 1 and 3.
|© AIPnD , created by NDT.net|||Home| |Top||