3. Case Study #1: Simply Supported Beam
Two key questions regarding the use of smartphones for SHM can be posed as follows: is the (poor) sensitivity of their accelerometers enough to endorse the correct calculation of the frequencies needed for damage detection? And how large should the damage be to be recognizable by the application? This section firstly assesses the quality of the smartphone’s measurements in terms of accelerations, power spectral density, and natural frequencies. Subsequently, the damage detection capabilities of App4SHM are evaluated.
A simply supported steel beam was used to compare both the time series and the power spectral densities obtained using App4SHM and a conventional data acquisition system.
3.1. Structural Description
The steel beam, illustrated in Figure 5, was composed of two equal leg angles (L 20 × 20 × 3 mm) connected at given intervals by 100 × 3 mm plates to ensure that a global flexural behavior is achieved. The steel beam was simply supported with a free span of 4480 mm.
3.2. Vibration Tests
The steel beam was subjected to non-destructive forced vibration tests using an impact hammer, a conventional sensing system composed of two piezoelectric uniaxial accelerometers from PCB Piezotronics, Inc. (Depew, NY, USA), and a smartphone. The two piezoelectric accelerometers, with a sensitivity of 10 and 0.5 V/g and a measurement range of ±0.5 and ±10 g, were fixed on the underside of the steel plate by means of an auxiliary base, which was glued to the steel structure using a quick bonding gel. Figure 6 depicts the two piezoelectric accelerometers used in the laboratory tests. Additionally, the piezoelectric accelerometers required the use of a portable PC, running LabViewTM rel 11 in Windows, a digital analogue card, and a signal conditioning module. Alternatively, a smartphone was simply placed on the topside of the same steel plate and required no additional equipment. A total of 30 non-destructive impact tests were performed, resulting in the corresponding 20 s time series from the conventional sensing system and smartphone. The (extremely short) impact time is removed from the time series that are analyzed for natural frequencies.
3.3. Comparative Study
The comparative study was first carried out in terms of the first three natural frequencies, which were computed resorting to a peak-picking method on a power spectral density determined using the Welch algorithm. The accuracy of the natural frequencies is instrumental for the functioning of the damage detection algorithms. Next, in order to conduct the remainder of the comparative study, two of the 30 non-destructive tests were chosen. These tests correspond to those that deviate the most and the least from the average in terms of natural frequencies. The time series and power spectral densities, determined using the two accelerometers and the smartphone, were compared both qualitatively and by means of time-lagged cross-correlation. The latter allows one to automatically synchronize the conventional sensing system and smartphone’s time series and determine the maximum cross-correlation between them. It is noted that, unlike the correct estimation of the natural frequencies, the good correlations between the power spectral densities and (even more so) between the acceleration time series are not instrumental for the damage detection algorithms. However, the analysis of the correlation between the measurements obtained using the conventional sensing system and the smartphone supports a more complete perspective of the limitations of the latter as compared to the former.
3.3.1. Natural Frequencies
The first three natural frequencies for the 30 time series are estimated and presented in Table 1. Additionally, the mean, standard deviation, and confidence interval (95% confidence level) were determined for all tests. Comparing the average frequencies obtained resorting to the smartphone with those determined using the conventional sensing system, a maximum difference of 0.54% is obtained. Consequently, one may conclude that the smartphone was able to consistently determine the natural frequencies when compared with a conventional sensing system. Moreover, when analyzing the observations’ standard deviation, the smartphone presented a smaller standard deviation regardless of the considered natural frequency, which indicates excellent consistency.
Additionally, a correlation test was made to verify if the natural frequencies computed resorting to the conventional accelerometers and smartphone are linearly related to each other. Figure 7 depicts a matrix with the correlation heatmap, where the size and color intensity of the square are proportional to the correlation’s magnitude, while the color itself separates the negative and positive correlations. A strong positive correlation is attained between the conventional accelerometers, with correlation values between 0.84 and 1.00 obtained when comparing the same frequencies. The correlation values between the smartphone and the accelerometers are smaller (between 0.63 and 0.76), but still strongly positive. Lastly, if one compares different frequencies, regardless of the considered accelerometer, only small correlations were observed as suggested by the off-diagonal entry values.
Due to the strong positive correlation observed for the same frequencies, one may conclude that an excellent agreement between the natural frequencies computed using the smartphone and those determined using conventional sensing systems was attained.
The ANOVA test is commonly used to analyze the differences between the means of several populations based on their variances. It was used in the present study to verify the similarity between the average frequencies attained with different accelerometers. Through the comparison of the values attained for the first, second, and third natural frequencies, the ANOVA test determined p-values of 0.77, 0.99 and 0.68, respectively. Therefore, since a p-value larger than 0.05 indicates a valid null hypothesis (population means are equal), one may conclude that the frequencies computed using the reported sensing systems are statistically similar.
Although the main output of App4SHM is the indication of damage based on the first three natural frequencies of a given structure, a more in-depth analysis is made in terms of the recorded time series and power spectral densities. The following two subsections present the results.
3.3.2. Time Series
Due to the impracticality of testing all the available time series, and due to their similarity in terms of natural frequencies, only two sets of time series were chosen for further analysis. These sets, which are highlighted in Table 1, are the closest and furthest away from the average values of the frequencies. Firstly, the time series obtained using the conventional sensing system were compared in order to verify their similarity, both qualitatively and by means of cross-correlation. Next, if the readings of the accelerometers are highly similar, only one of their time series is compared with the time series recorded using the smartphone.
Figure 8 illustrates the comparison between the time series obtained resorting to the two accelerometers of the conventional sensing system. As expected, an excellent agreement is observed between the two accelerometers for all tests. Alternatively, the cross-correlation method, which determines the similarity between a fixed time series and a time-lagged one as a function of the lag, was also used to compare the time series. As expected, a null time lag was achieved for both tests, while correlation values of 0.97 and 0.96 were attained, indicating a strong positive correlation between the acceleration time series.
Therefore, the time series recorded by the smartphone is only compared with one of the conventional sensing systems, namely the 333B40 accelerometer, due to its larger measurement range.
Figure 9 depicts the time series obtained using the smartphone and the 333B40 accelerometer. The signals have a good general agreement, presenting only a limited number of small differences. Resorting to the cross-correlation method in order to synchronize the time series, correlation values of 0.71 and 0.55 were achieved for tests number 1 and 8, respectively. These values are considerably lower than those obtained for the time series measured with conventional accelerometers. This is caused by the relatively poor sensitivity of the smartphone’s accelerometer as compared with professional equipment. However, despite the difference between the time series, their frequency content is remarkably similar, as shown in the next section.
3.3.3. Power Spectral Densities
The entire set of time series obtained during the experimental study was used to compute their respective power spectra, which were determined using the Welch algorithm. Figure 10 illustrates all power spectral densities in grey lines, along with their averages, in blue lines, for the smartphone and conventional accelerometers. Similar to that reported in Table 1, the analysis of Figure 10 reveals a low variability in terms of natural frequencies, i.e., the local maxima, regardless of the considered accelerometer.
Additionally, Figure 11 depicts the average power spectral densities (using a logarithmic scale) obtained with the three devices, which exhibit an excellent agreement concerning the structure’s natural frequencies. Moreover, in order to compare the entire frequency range of the power spectral density, a cross correlation analysis was performed. This analysis yielded correlation values of 1.00 and 0.99 between each accelerometer and the smartphone, which reveals an excellent agreement between the considered power spectral densities.
3.4. Damage Detection Capabilites
An additional experimental campaign was performed to evaluate the damage detection capabilities of App4SHM. One hundred dynamic impulse tests were conducted on the simply supported steel beam presented in Figure 5 to extract its first three natural frequencies under normal (‘undamaged’) conditions. These observations yield the standardized natural frequencies cluster illustrated with blue circles in Figure 12. The standardization was performed by subtracting the mean of the training data from each frequency and dividing the result by the standard deviation.
Subsequently, three different masses (25, 50 and 100 g, see Figure 13) were placed at two different locations, namely at mid-span (position A in Figure 5) and at one quarter of the span (position B in Figure 5), to simulate several levels of damage and, consequently, identify the damage required by App4SHM to yield a positive reading. It is noted that the total mass of the structure is 9.7 kg. The addition of mass to simulate damage is a relatively standard procedure when the structure cannot be damaged, e.g., [28].
Five dynamic tests were executed for each mass/location combination and the first three natural frequencies were extracted for each case, yielding the squares, diamonds and upper triangles illustrated in Figure 12. Note that the marker indicates the amount of mass added to the steel beam, whilst the marker’s color specifies if damage was identified by App4SHM (red) or not (green). Figure 12a, which illustrates the results obtained when the mass is placed at mid-span, shows that the first and (especially) third frequencies are the most affected, while the second frequency suffered no significant change. This is physically consistent, as the second natural mode of vibration has a node at mid-span, so should be less affected by an additional mass placed at that position. Figure 12 shows that App4SHM correctly identifies damage when at least one natural frequency deviates by more than two standard deviations from the training cluster.
The clusters obtained when the mass is located at approximately one quarter of the span are presented in Figure 12b. In this case, the second frequency was by far the most affected, with limited degradation of the first and third frequencies. Again, the insensitivity observed for the third frequency is to be expected as the added masses were located close to a vibration node of the third modal shape. Conversely, the location corresponds to an anti-node of the second mode of vibration, which is therefore the most affected. Similarly to the observations made when the mass was placed at mid-span, App4SHM starts detecting damage when the differences between the normal and measured frequencies exceed two standard deviations in at least one natural frequency.
The evolution of the damage index (DI), defined by Equation (1), as a function of the added mass and its location is presented in Figure 14. Each cyan circular marker corresponds to a damage detection observation, the circle with a dot represents the mean DI, and the vertical lines indicate the ± standard deviation range. The DI threshold that separates undamaged from damaged data is shown as a horizontal line. The lowest added mass tends to yield higher DIs than the training data, but they always remain below the threshold. The 50 g mass is recognized as damage in most cases, especially when the mass is located at mid-span. However, the threshold is located at less than one standard deviation from the mean, so it is expected that the Mahalanobis-based algorithm would yield a considerable number of false negatives. Finally, the highest mass has a marked impact on the DI, with the measurements always located more than two standard deviations higher than the threshold. This means that false negatives should very seldomly occur for this level of added mass. Lastly, as expected, placing the mass at mid-span leads to larger DIs as compared to those attained with the mass at one quarter of the span.
For comparison purposes, the damage detection analysis was also performed using an alternative algorithm, based on Gaussian Mixture Models with an adaptive choice of the number of clusters to define the baseline condition [24]. The algorithms yield the same classification for the readings, as listed in Table 2.
CC BY 4.0. "Smartphone Application for Structural Health Monitoring of Bridges" by Figueiredo E, Moldovan I, Alves P et al., Sensors (Basel, Switzerland) (2022). https://doi.org/10.3390/s22218483. Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Peer-reviewed and open access; full text from PubMed Central (PMC9655388). Headings and layout adapted for reading; figures and tables are in the original.