This paper presents the comparison of three methodologies to detect if some fans in a matrix are not working properly. These methodologies are based on detecting fan failures by analysing acoustic images of the fan matrix, obtained using a planar array of MEMS microphones. Geometrical parameters of these acoustic images for different frequencies are then used to train a support vector machine (SVM) classifier, in order to detect the fan failures. One of the methodologies is based on the detection of the faulty fan in the matrix, under the hypothesis that only one fan can fail. Other methodology is based on the detection of the specific working situation of the matrix. And finally, the third methodology that is compared is based on determining individually if each of the fans of the matrix is working properly or not. The comparison shows that this third methodology is the most reliable.
In recent years, techniques for obtaining acoustic images have been developed greatly and rapidly. At present, acoustic images are associated with a wide variety of applications [
Fault detection lies in determining failures in machine structural components or abnormal behaviours of a system [
One of the classic approaches for machinery condition monitoring is based on making periodic vibration measurements of the equipment and then comparing them to known healthy/damaged data to assess the health status of the machine [
An array is an arranged set of identical sensors, fed in a specific manner. The array beampattern can be controlled by modifying the geometry of the array (linear and planar), the sensor spacing, the response, the amplitude, and phase excitation of each sensor [
The acronym MEMS (microelectromechanical system) refers to mechanical systems with a dimension smaller than 1 mm, manufactured with tools and technology arising from the integrated circuits (ICs) field, and used for the miniaturization of mechanical sensors [
Due to the high diversity of applications of arrays of MEMS microphones, the authors of this paper are working on widening these uses to other fields. The authors already had experience in the design and development of acoustic arrays to be used in surveillance systems [
A fan matrix, fan array, or fan wall is a system formed by several fans located on a surface, working together in order to improve the performance of one alone large fan with lower power consumption. Any type of application that requires specific temperature conditions is a candidate for a fan matrix.
An analysis of the systems which uses fans matrices reveals that they do not have a subsystem to control if any of the fans that compose the matrix is down or is not working properly. It would be very useful to detect these kinds of situations. The authors have developed a novel fault diagnosis methodology to detect faulty behaviours on the fans included in a matrix. This method is based on the analysis and classification of the acoustic images, obtained from the fan matrix, by means of using machine learning techniques.
On the first step of this research [
On the basis of these results, this work analyses the behaviour of different detection methodologies based on detection if any of the fans of a matrix is not working properly. These methodologies are based on geometrical parameters of the acoustic images of the fan matrix and in SVM classifiers.
An acquisition and processing system, previously developed by the authors [
(a) Array module with myRIO and MEMS array board. (b) Complete acoustic acquisition system.
After the acquisition of the acoustic signals by the MEMS microphones of the array, they are processed using deinterlacing, decimation, and filtering techniques, in order to generate the acoustic images using wideband beamforming. A set of
Software algorithms diagram.
This work is focused on obtaining acoustic images of a 3 × 3 fan matrix, shown in Figure
Test fan matrix.
For the tests, the fan matrix is placed 50 cm opposite the 16 × 16 MEMS array, inside a 5m × 3m × 2.5 m anechoic chamber, as shown in Figure
Experimental setup block diagram.
The first step of this work was an analysis of the acoustic signals received by the microphones of the array, in order to characterize the noise generated by the fans of the matrix and to select the frequencies used to obtain the acoustic images of the fan matrix. As each fan has 7 blades and it rotates at 3500 rpm, its noise has harmonics at 400 Hz and its multiples. In this previous study [
Preliminary tests [ Only the fans on the corners of the matrix are working (left column) The fans of the middle column of the matrix are faulty (middle column) The fans of the middle row of the matrix are faulty (right column)
Analysing these images, it can be observed that if several fans of the matrix are not running, i.e., are faulty, the acoustic image reveals this effect in some way.
In Figure
Acoustic images of faulty fan matrices: (a) only the fans on the corners are working, (b) the fans of the middle column are faulty, and (c) the fans on the middle row are faulty.
This section describes the 3 fault detection methodologies that are going to be compared. All these methodologies are found on a machine learning algorithm, based on a linear support vector machine (SVM). Geometrical parameters of the acoustic images were used in this machine learning algorithm. These geometrical parameters were the value and the position (azimuth and elevation) of the maxima of the acoustic images for each of the 10 selected working frequencies, defined in Section
Figure
Block diagram of the compared methodologies.
A set of tests [
In this methodology, the SVM algorithm was used to detect the faulty fan position. The employed SVM worked with 10 different classes: 1 class represented a healthy matrix (all working fans), and the other 9 classes represented the 9 possible one faulty fan situations. As it was pointed previously, the SVM algorithm used 30 geometrical parameters of the acoustic images. In these tests, it was noticed that if one fan failed, the maximum position and value of the acoustic image changed. One of these effects is shown in Figure
Maximum positions of the acoustic images of fan matrix with one faulty fan (the position of this faulty fan is represented with a cross). (a) Fan 1. (b) Fan 2. (c) Fan 3. (d) Fan 4. (e) Fan 5. (f) Fan 6. (g) Fan 7. (h) Fan 8. (i) Fan 9.
The obtained accuracy rate by the SVM algorithm, trying to detect the faulty fan, was 95.6%. This result showed that the purposed methodology was reliable when one fan of the matrix failed because it was accurate for detecting the position of the faulty fan.
Although it is really unusual that more than one fan fails at the same time, a set of tests was carried out in order to study if this defined methodology was robust enough in the presence of unexpected situations, i.e., if the number of faulty fans in the matrix increased [
The objective of these tests was to analyse if the SVM algorithm, trained to detect only one faulty fan, was robust enough to detect any of the two faulty fans of the working situations. With a robust methodology, if one of the two faulty fans was detected, it would be repaired or replaced. After that, when the matrix began to run again, the other faulty fan would be detected and replaced. In these tests, if the algorithm did not detect any of the two faulty fans, it was considered that the algorithm failed. The results obtained in these tests showed that if the distance between the two faulty fans increased, the accuracy of the algorithm to detect that one of them failed or decreased. Table
SVM accuracy rates of the robustness tests for the one faulty fan detection methodology based on training the SVM with one faulty fan parameters.
One faulty fan tests | Accuracy rate (%) | |
---|---|---|
(1) Faultyfaulty fan validation | 95.6 | |
(2) Faultyfaulty fans validation | ||
Case A: horizontal/vertical step distance (no central fan) | 48.3 | |
Case B: horizontal/vertical step distance (with central fan) | 44.1 | |
Case C: diagonal step distance | 23.5 | |
Case D: knight movement distance | 14.7 | |
Case E: row/column distance | 14.2 | |
Case F: diagonal distance | 10.9 |
As the one faulty fan detection methodology is not considered to be robust enough under unexpected situations (more than one faulty fan), the following implemented tests have been aimed to train the SVM algorithm to be able to detect any of the 512 different working situations of the fan matrix (512 = 29, since the matrix is composed of 9 fans, each one with 29 different working states: working/healthy or faulty). These 512 working situations are considered from the perfectly working situation (all 9 working fans) to the impossible working situation (all 9 faulty fans), including all the remaining ones (one faulty fan, two faulty fans, and three faulty fans). So, this SVM algorithm works with 512 classes and the same 30 geometrical parameters than in the previous methodology used.
The obtained accuracy rate by the SVM algorithm in these tests has been 45.3%. This result shows that this purposed methodology is not reliable. Training the SVM algorithm in order to be able to discriminate among 512 different classes is not a reliable option. The employed data (30 parameters for each test) is not enough to discriminate among so many classes (512 different working situations).
So, the next idea has been the definition of a methodology based on 9 independent SVM classifiers, each one related with a specific fan of the matrix. Each classifier has been trained to detect if the corresponding fan is working or not, that is, to detect 2 different classes. Again the methodology uses 30 geometrical parameters to detect if the matrix is not working properly. Table
SVM accuracy rates of each classifier related to one of the fans of the matrix.
SVM classifier | Accuracy rate (%) |
---|---|
Fan 1 | 99.4 |
Fan 2 | 91.0 |
Fan 3 | 98.0 |
Fan 4 | 91.4 |
Fan 5 | 95.4 |
Fan 6 | 98.9 |
Fan 7 | 90.7 |
Fan 8 | 91.6 |
Fan 9 | 98.5 |
The problem with this methodology is that despite the individual accuracies of each of the 9 SVM classifiers being high, the global accuracy of this methodology is not so high. The global accuracy rate of this methodology, with the 9 SVM classifiers working properly at the same time (all of them detecting if the corresponding fan is working properly or not), falls to a 65.7% value. It seems that detection failures of the SVM classifiers do not match between them. This global value takes into account all the possible matrix failures, the 512 possibilities.
It has been previously noticed that a failure situation in the matrix with more than two faulty fans is not probable. So, a group of tests is performed to assess the accuracy of this individual faulty fan detection methodology when there are only two faulty fans that have been carried out. The obtained global accuracy of the methodology for all these working situations is 87.9%, and it is shown in Table
SVM accuracy rates of the individual faulty fan detection methodology with two faulty fans working situation of the matrix.
Individual faulty fan detection methodology: two faulty fans tests | Accuracy rate (%) | |
---|---|---|
Global | 87.9 | |
Two faulty fans separation | ||
Case A: horizontal/vertical step distance (no central fan) | 89.5 | |
Case B: horizontal/vertical step distance (with central fan) | 89.3 | |
Case C: diagonal step distance | 84.3 | |
Case D: knight movement distance | 89.6 | |
Case E: row/column distance | 84.7 | |
Case F: diagonal distance | 92.5 |
In order to be able to compare the shown methodologies, the same two faulty fans working situations as those shown in Table Case A: two fans located on both ends of the largest diagonal of the matrix Case B: two fans located on both ends of one row or one column of the matrix Case C: two fans located on both ends of the “L-shaped” (Knight) movement in chess Case D: two fans separated one diagonal step Case E: one fan in the centre of the matrix, and the other one separated one vertical/horizontal step Case F: two fans separated one vertical/horizontal step, and none is in the centre of the matrix
The obtained accuracy rates for the different tested cases are shown in Table
Three different methodologies to detect faulty fans in a fan matrix have been defined. These methodologies are based on the maxima positions and values of the acoustic images obtained for the fan matrix at different working frequencies. A summary of these methodologies is shown in Table
SVM accuracy rates of the detection methodologies.
Methodology | SVM | Accuracy rate (%) | |
---|---|---|---|
# classifiers | # classes | ||
One faulty fan detection | |||
One faulty fan validation | 1 | 10 | 95.6 |
Two faulty fan validation | (10.6–48.3) | ||
Working situation detection | |||
512 working situations | 1 | 512 | 45.3 |
Individual faulty fan detection | |||
Individual SVM classifier | 9 | 2 | (90.7–99.4) |
512 working situations | 65.7 | ||
Two faulty fans situations | 87.9 |
The first defined methodology is based on detecting one faulty fan. This methodology is highly accurate if only one fan is not working properly, but it is not robust enough under unexpected situations. If one more fan fails, the accuracy rate of the methodology in detecting one of the two faulty fans fails to values between 10.6% and 48.3%.
The second methodology is based on detecting the specific working situation of the fan matrix. As the matrix has 9 fans, 512 different working situations can be defined. The accuracy rate of this methodology is 45.3%. This value shows that, as for the previous case, this methodology is not robust enough either.
The last implemented methodology is based on detecting if each of the fans is or not working properly. This methodology uses 9 SVM classifiers, instead of only one, but each classifier is trained to distinguish between only two classes (a faulty fan or a healthy fan). The individual accuracy rates of the SVM classifiers show high values between 90.7% and 99.4%. This methodology is not accurate enough to distinguish a matrix failure among its 512 possible faulty situations. In this case, this methodology shows a 65.7% accuracy rate value. Although it is not a high accuracy value, it is higher than the one obtained for the methodology based on detecting the working situation of the fan matrix. But it is not a problem because the probability of having a fan matrix with more than two faulty fans is not high. Under this consideration, this methodology is a proper option, given that it shows high accuracy rate values in detecting faulty fans under two faulty fan working situations. In these cases, the methodology shows accuracy rate values between 84.3% and 92.5%.
This paper shows the comparison of three fault detection methodologies developed to identify if the fans in a matrix are not working properly. These methodologies are based on geometrical parameters of the acoustic images of the fan matrix and in support vector machine algorithms.
The most promising methodology is based on 9 SVM classifiers, each one related to a specific fan of the matrix. Each individual classifier detects if the corresponding fan is or not working. These high accuracy rate values shown by this methodology balance out the complexity increment related to using a SVM classifier for each fan of the matrix, instead of only one, as in the other defined methodologies.
It could be pointed that the tests carried out in this study must be widen with other tests including background noise or even objects near the assembly, in order to create surfaces where the sound generated by the fans could be reflected. These tests would simulate a more real operation situation because an operative fan matrix is not isolated of the surroundings nor even placed inside an anechoic chamber.
As near future work, a new methodology which combines the information of the maxima and other features or geometrical parameters of the acoustic images, such as the centroids or the energy, could be defined. In this case, if the SVM algorithm gets more information to discriminate between the different working situations, it is expected to be able to improve its accuracy rate. Also as a midterm future work, the extension of the methodology is to be able to detect/classify/identify specific fan failures such as not only to detect if the fans are working or not but also to detect the reason why the fan is not working properly, that is, if it is not rotating at the right speed, if some of the blades are broken, and if the shaft is misaligned. So, the next step in this research is the fan future work which could be the comparison of this extended methodology with other methodologies based on rotating machines that classifies specific failures, such as MSAF methodologies.
The numeric data used to support the findings of this study are available from the corresponding author upon request.
The authors declare that there are no conflicts of interest regarding the publication of this paper.
This research was funded by MINECO/FEDER (EU) (SAM TEC 2015-68170-R).