Low-Cost Wireless Structural Health Monitoring of Bridges · Komarizadehasl S, Lozano F, Lozano-Galant JA et al.

3. Laboratory Tests and Results

This section aims to validate the performance of LARA. Firstly, the resolution, sensitivity, and ND of a single MPU9250, CHEAP, and LARA are presented. For that, an extended duration test was performed in an office. Secondly, a test has been carried out on the hydraulic jack of the laboratory. In that test, LARA was validated and compared with CHEAP. Next, the performed tests for validating LARA’s accuracy in terms of various frequencies are investigated. Finally, the acceleration amplitude evaluation is shown.

It should be noted that the performance of CHEAP was validated using two high-precision accelerometers (PCB 393A03 and 356B18) [23] in several laboratory experiments for a range of frequencies from 0.5 to 10 Hz. These tests were performed on the same shaking platform used in this work. This platform is an INSTRON 8803 model located at the Structural Laboratory Lluís Agulló of the Technical University of Catalonia (Barcelona, Spain). In addition, it is essential to mention that LARA was calibrated in the laboratory of the Applus company. In that certification, the acceleration amplitude accuracy of LARA was studied in several experiments with a fixed RMS acceleration amplitude of 0.5 g within the range of 5 to 160 Hz.

3.1. Sensitivity, ND, and Resolution Evaluation

This section compares the sensitivity, ND, and resolution of a single MPU9250 accelerometer with CHEAP and LARA.

The accelerometers’ sensitivity can be defined as the ratio of the input (induced vibrations of the shaking table) to the output (the information reported by the accelerometer). This concept is measured and calculated differently for analog and MEMS sensors. Although an analog accelerometer usually produces electronic pulses related to the input vibrations, MEMS accelerometers convert these electronic pulses into digital signals. While the sensitivity of analog sensors is reported as V/g (Voltage per gravitational acceleration), MEMS accelerometers report their sensitivity in the Least Significant Bit per gravitational acceleration (LSB/g) [56]. Since this value is a characteristic of a sensor, it should be the same for MPU9250, CHEAP, and LARA. The datasheet of MPU9250 reports a sensor sensitivity for the acceleration amplitude range of ±2.0 g of 16,384 LSB/g. MPU9250 accelerometers work with an operating voltage of 2.5 volts and their scale for converting data from analog to digital (Analog to Digital Converter ADC) is 16 bits [47]. The ADC formula for calculating the LSB is shown in Equation (1) [56].

In Equation (1), with an input voltage of 2.5 and an ADC scale of 16, each unit of LSB is equivalent to 0.03814 mv. In this way, with 16,384 LSB/g for acceleration amplitude range of ±2.0 g, the comparable sensitivity of MPU9250, CHEAP, and LARA is calculated as 625 mV/g.

To calculate ND and resolution of the devices, a long-term test was performed. The ND results of the different accelerometers are calculated using Equation (1) [57]. It should be noted that CHEAP was located to read the signals from the Z direction. For validating this test and the used formula, an MPU9250 accelerometer was tested. In addition, the reported results of its datasheet are compared with the results of the tests. It should be noted that the presented noise-density measurement of this work is a standard procedure to characterize the noise of the developed accelerometer [58]. This, typically, appears on the datasheet of commercial accelerometers [59]. However, it should be noted that those applications that aim to characterize the noises throughout time (such as Allan variance or Allan deviation) are out of the scope of this paper. Typically, these applications are used to investigate the noises of sensors (such as gyroscopes) which data drift throughout time in the time-domain series.

In Equation (2), the xi is the reported values of the accelerometers in the time-domain, μ is the average of all xi values, N is the number of used samples, and f is the sampling frequency of the accelerometer.

The calculated ND of MPU9250, CHEAP, and LARA for the z-directions are 390, 162, and 81 µg/√Hz, respectively. In addition, the ND of LARA for both x and y directions is 51 µg/√Hz.

The datasheet of MPU9250 reports its ND as 300 µg/√Hz. In fact, the illustrated information of datasheets is usually acquired under the best possible circumstances. It can be deducted from the calculated ND that using shorter wires and better connections made LARA 50% less noisy than CHEAP by reducing the ND value from 162 to 81 µg/√Hz in the z-direction. The calculated ND values also shows that LARA has almost 79% less ND on the z axes than a single MPU9250 accelerometer by reducing the ND value from 390 to 81 µg/√Hz. Further studying of these values shows a lower ND of LARA on the x and y axes. These two axes do not measure the gravitational acceleration of the earth; subsequently, the ND of LARA is 51 µg/√Hz.

Since the evaluation of the accelerometer’s resolution depends on the number of samples [60], for a fair comparison between devices with various sampling frequencies, the same number of the acquired samples were used in the test. For illustrating the resolution of the accelerometers, their reported data have been transformed from time-domain to frequency-domain by a Fast Fourier Transform (FFT). The results for different sensors are reported in Figure 2.

Analysis of Figure 2 shows that while LARA and CHEAP resolution are almost equal, the resolution of a single MPU9250 is almost twice as the CHEAP or LARA. By studying Figure 2, MPU 9250’s resolution, CHEAP and LARA for the z-axis are reported as 0.00016 m/s2, 0.00009, and 0.00009, respectively. It is to be considered that this test needed a long duration of data capture; consequently, it was done in an office at midnight. The FFT method reports more accurate outputs when it has a higher number of inputs [60]. For a fair comparison between MPU 9250, CHEAP, and LARA, the same number of data had to be evaluated by the FFT method. Although sampling two million sets of data took MPU9250 and CHEAP 6.5 h, LARA acquired two million sets of data in about 1.7 h. Figure 3 shows the frequency domain diagrams of acquired data of LARA for all the axes.

Analysis of Figure 3 shows that the resolution of LARA for x and y axes is 0.00005 m/s2. The resolution for x and y-directions is 44% better than the z-axis because of the absence of gravitational acceleration of the earth in those directions. As a result, the resolution of this accelerometer is evaluated as 0.00005 m/s2.

After calculating LARA’s sensitivity, ND, and resolution, LARA can be compared with instrument-graded accelerometers. In Table 2, various commercial triaxial MEMS applications are presented. The contents of this table are ordered by the noise density of the accelerometers. This table includes the following information, which is reported by the datasheets provided by the producers or measured in this paper, organized in columns: (1) sensor number, (2) sensor name, (3) acceleration range, (4) sampling frequency speed, (5) Noise Density (ND): the RMS resolution can be calculated by multiplying the ND by the square root of the sampling frequency, (6) sensitivity: for a better comparison, the analog converted sensitivities by the producer companies are reported for each product, (7) price of the sensor (VAT excluded): prices are based on the recent declaration of the producer. The reported price of LARA, CHEAP, and MPU9250 refers to research prototypes and includes the used inceptions (such as accelerometer, Arduino, wires, multiplexor, and Arduino) in them. It should be noted that the rest of the sensors are commercial solutions. The information in this column is to illustrate the price ranges and not to compare the price of the prototype sensors with the commercial ones, and (8) acquisition equipment.

Analysis of Table 2 shows that noise level of the commercial solutions (IAC-Hires [61], 3713F112G [62], Unquake [63], and Recovib Tiny [64]) ranges from 8 to 30 µg/√Hz. In addition, analyzing Table 2 indicates that LARA has a more comparable ND with the introduced commercial accelerometers in Table 2 than with any of the low-cost accelerometers (5–7) presented in Table 1. The noise level of the commercial solutions (IAC-Hires [61], 3713F112G [62], Unquake [63], and Recovib Tiny [64]) ranges from 8 to 30 µg/√Hz.

It is shown in Table 2 that 3713F112G and IAC-Hires need extra data acquisition equipment for acquiring the vibrations. In fact, 3713F112G needs a signal conditioner such as 482C27 with a price of 5070 € with four channels, and the IAC-Hires requires a data acquisition such as Recovib Monitor with a price of 3700 €, which can provide up to eight reading channels. Moreover, CHEAP and MPU9250 are dependent on an attached computer for their signal monitoring and saving.

However, Unquake, Recovib Tiny, and LARA do not need any acquisition system for their data acquisition. Accelerometer Unquake samples internally, and the measurements are timestamped with absolute time from GPS. The measurements are synchronized in the post-processing stage based on the timestamps by using the software that comes for free from the company. In addition, Recovib Tiny and LARA are both wireless sensors.

Further analysis of Table 2 shows a wide range of prices (varying between 1192 € up to 2500 €) of commercial solutions. In fact, as stated in [65], the cost of accelerometers is known to be one of the critical limitations of SHM analysis and long-term monitoring.

3.2. Frequency Validation

This subsection illustrates the experiments aiming at the frequency measurement accuracy of MPU9250, CHEAP, and LARA on a vibration platform.

An experiment to validate the frequency report of LARA is presented in this section. A jack that could induce displacements with a known frequency was used. Although the jack is very accurate in reproducing a specific frequency, its performance is limited in movement. Therefore, for every experiment, the jack presents a report. This report is the time-domain information of the activities of its lower jaw.

It should be noted that the vibration platform used in this paper (INSTRON 8803) was programmed using WaveMatrix2 Dynamic Software This jack can create various waveform types with the frequency range of 0.1 to 100 Hz and its resonant frequency is 134 Hz. The movement direction of this vibration platform (z-direction) is shown in Figure 4b.

In this work, the waveform is set to a sine wave. All the additional technical features of this hydraulic jack, its programming software, and its datasheets are presented in [23].

Figure 4 shows the placement of the CHEAP and LARA on the hydraulic jack.

It can be seen from Figure 4 that both CHEAP and LARA are attached to a rigid metal plate. This attachment was carried out by using an industrial adhesive (X60). X60 made by HBM company, is a two-component methyl-methacrylate adhesive that is widely used for accelerometer mounting [66]. Moreover, the metal plate containing CHEAP and LARA is bolted to the shaking table. For validating the accelerometers’ accuracy, 11 tests with various frequency ranges (from 0.1 Hz to 32 Hz) were performed. For finding the frequency report of the accelerometers, the accelerometers’ time-domain raw data were converted to the frequency domain by an FFT program written in the MATLAB software. To show the methodology of the calculation of the accelerometer’s frequency errors from the jack’s results, four frequency domain plots of LARA for 0.1 Hz (Figure 5a), 0.2 Hz (Figure 5b), 0.3 Hz (Figure 5c), and 0.5 Hz (Figure 5d) are presented in Figure 5.

The accelerometer’s frequency error is estimated by calculating the percentage error of the plot value from the reference value of the excitation device. Table 3 compares the frequency errors of LARA, CHEAP, and MPU9250 with the jack’s reference values. Table 3 is organized in columns: (1) Frequency (Hz): the reference value of the excitation device, (2) LARA’s error (%), (3) CHEAP’s error (%), and (4) MPU9250’s error (%).

Analysis of Table 3 shows that a single MPU9250 is reporting the same frequency error as the CHEAP from 2 Hz on. Since this particular MPU9250 is one of the CHEAP accelerometers, its frequency report is synchronized with CHEAP. It can also be deduced from this table that a single MPU9250 has not enough resolution for frequencies smaller than 2 Hz. Further analysis of Table 3 shows that CHEAP cannot visualize signals with a frequency smaller than 0.4 Hz either. In fact, LARA has a broader range of frequency than MPU9250 and CHEAP and can locate signals up to 0.1 Hz with 0.5% of error. In fact, the frequency range of LARA, based on the presented information, is between 0.1 Hz and its bandwidth (165 Hz).

Further investigation of Table 3 illustrates that MPU9250, CHEAP, and LARA have similar errors for signals with frequencies bigger than 2 Hz, but the error is still not 0.000%. In fact, the FFT evaluation can be influenced by irregularities, such as the number of sampled data and the sampling frequency speed [60]. However, controlling all the data for such accuracy is not the aim of this paper.

3.3. Acceleration Amplitude Validation

This subsection first compares the acceleration amplitude accuracy of MPU925, CHEAP, and LARA using a sine wave with a known frequency and acceleration amplitude. Then, to check the ultimate acceleration range of LARA, a laboratory experiment using a sine wave with the RMS value of 1 g was performed.

An experiment was carried out with the jack to compare the acceleration acquisition of MPU925, CHEAP, and LARA. The jack was calibrated to move with a frequency of 4 Hz and a displacement range of 0.1 mm from its null axis. During each test, the jack reported its time-domain information with a sampling frequency of 500 Hz. Figure 6 illustrates jack’s report for this experiment in the frequency-domain diagram.

Analyzing Figure 6 shows that the jack was moving with a frequency of 4 Hz and an averaged displacement of 0.10181 mm; consequently, the jack was working with a 1.81% error rate from 0.1 mm that the programmed displacements.

For validating the accelerometers’ accuracy, their acceleration amplitude report of the sensors was converted to displacements. Then, high-pass and low-pass filters removed signals with frequencies smaller than 0.1 Hz and bigger than 1/10 of the accelerometer’s bandwidth. These filters are MATLAB functions that can filter the signals out of the interest range.

Finally, the FFT analysis was carried out, and obtained results are presented in Figure 7. Figure 7a–c show the reported displacement of MPU9250, CHEAP, and LARA, respectively. It should be noted that the accuracy of the accelerometers for measuring the magnitude of the induced vibration is analyzed in the frequency domain representation. It is due to the fact that making a unique noiseless signal on a shaking table or an actuator is significantly challenging [67]. In addition, for the OMA of bridges, accelerometers are mounted to measure the structural response of the bridges under ambient vibrations, such as those induced by traffic, wind, and temperature variation. Therefore, this sum of ambient vibrations is usually evaluated in the frequency domain representation [34].

Analysis of Figure 7 shows that even though all accelerometers are reporting the same frequency, they measure the induced signal’s magnitude inconsistently. MPU9250, CHEAP, and LARA have measured the jack’s displacement as 0.098913, 0.10023, and 0.10068 mm. The MPU9250’s displacement measurement is off by 1.31% from the CHEAP’s report and 2.85% off from jack’s report (Figure 6). Further analysis of Figure 7 shows that CHEAP and LARA’s magnitude report has less than 0.5% error from each other. In fact, CHEAP and LARA have measured the jack’s displacements with 1.55% and 1.2% errors from the introduced displacement measurement of the jack in Figure 6.

It should be pointed out that the shown displacement magnitude of Figure 7c corresponds to a measured acceleration amplitude of 0.006 g. Furthermore, for measuring the top acceleration amplitude range of LARA, another test is carried out on a dynamic actuator. This time, the used dynamic actuator was programmed to induce vibrations with a Root Mean Square (RMS) value of one g. This test was performed for one minute, and, after that, the RMS value of the acquired vibrations was calculated. The time-domain representation of this test for a time duration of one second is presented in Figure 8. The calculated RMS value of the acquired data of LARA for one minute is 0.991 g. This value shows a 0.9% error from the reference RMS value of the test.

It should be noted that LARA, such as any other accelerometer, has an acceleration amplitude range for each of its axes. The sensor gets overloaded when a signal has an acceleration amplitude that exceed a certain range. It should be noted that when MEMS accelerometers reach their saturation level (get overloaded) [68], they are not able to measure the magnitude of the impact anymore [69]. In fact, LARA, unlike piezoelectric solutions, does not require a parametric analysis as it does not experience a drift after saturation [21]. More information about the saturation of MEMS accelerometers can be found in [68].

LARA is set to have an acceleration amplitude range of ±2 g. It should be noted that the LARA is calibrated in Applplus [70] under RMS acceleration amplitude of 0.5 g and frequency range of between 5 and 160 Hz. In addition, Figure 8 presented acceleration amplitude verification of LARA for 1 g where the maximum measured impact was 1.42 g. In order to investigate its theoretical saturation level (±2 g), an experiment with an acceleration amplitude higher than 2 g was performed. Figure 9 shows the time-domain representation of this experiment.

The analysis of Figure 9 shows that LARA could not measure the acceleration amplitude of the produced signal beyond the ±2 g magnitude as the top and bottom of the signal are cut. This means that LARA is overloaded. It is also noted in the literature that when substantial impact happens, the output voltage of MEMS accelerometers (such as LARA) reach a fixed value that does not vary (saturation) [71]. To decrease the chance of overloading of accelerometers (saturation), low-pass filters are traditionally recommended [72].

To summarize the collected information of the laboratory tests in Section 3.2 and Section 3.3, LARA accurately measured frequencies of sine waves within the range of 0.4 and 32 Hz. Furthermore, LARA accurately measured the acceleration amplitude of a sine wave with a magnitude of 0.006 g. Moreover, LARA’s maximum acceleration amplitude measurement was investigated using an induced sine wave with RMS acceleration amplitude of 1 g.