Table 1. Summary of methods used for detecting cavitation in centrifugal pumps. The scope of this study was to assess and compare four standard classification meth- ods using raw signals from a low-cost data acquisition platform, ultimately identifying a simple yet highly reliable algorithm for detecting cavitation. By using signal and vibration data with minimum processing (no feature extraction) and leveraging machine learning and deep learning techniques, the research explored their effectiveness in fault detection and condition monitoring of centrifugal pumps, providing invaluable insights for future advancements in IoT-integrated real-time fault detection and operation monitoring ap- plications. What sets this research apart is the emphasis given to simulating a real-life irrigation system and utilizing a low-cost device that simultaneously acquires sound and vibration data, commonly owned by farmers and operators. Combining those data with a well-trusted, robust algorithm, real-time early detection can be achieved, and preventive maintenance and proactive control of centrifugal pumps are possible. Figure 1. Experimental Setup Piping and Instrumentation Diagram. The experimental setup involved utilizing a complete water pumping station configu- ration. To replicate a range of operating conditions, throttling valves were incorporated into both the suction and discharge lines. Additionally, a flow sensor was installed in the discharge line to measure the water flow rate. To monitor the pressure levels, sensors were placed in both the suction and discharge lines. Figure 1 provides a visual representation of the experimental setup in the form of a simple piping and instrumentation diagram (P&ID), while Table 2 presents a detailed list of the components comprising the setup. These components collectively form an integral part of the experimental apparatus. Table 2. Main equipment characteristics. The pump that was used for the experimental setup is shown in Figure 2. The data acquisition platform employed for the experiment was a Samsung Galaxy A52s 5G smartphone, securely attached to the pump’s casing using a rigid metal plate, as depicted in Figure 3. To capture acceleration data, the smartphone’s embedded acceleration sensor (ICM42605M Accelerometer -TDK InvenSense, resolution: 0.002394 m/ s?, maximum range: 78.453201 m/s2). The sampling frequency was set to the maximum available range of 500-510 Hz. The recorded acceleration data was stored in a .txt file format using the smartphone’s onboard SD card. Figure 2. Pump and corresponding 3D axis system (courtesy of SAER Elettropompe Spa). Figure 3. Actual mounting of the smart phone on the pump casing, using a rigid metal plate. Figure 4. SAER IR 40-160 NB/B Performance curves based on water’s kinematic viscosity of 1mm2/s, density of 1000 kg/ m°, and temperature of 15 °C. Tolerance and curves according to UNI EN ISO 9906. Figure 5. Indicative vibration data plots for State 2, as exported by MATLAB® 6.5 (R13), (a) time domain—amsec vs. amplitude; (b) frequency domain—mHz vs. amplitude; (c—e) spectrogram for x, y, and z axes (normalized frequency (x 7 radians/sample) and power/frequency (dB/rad/sample)). Figure 5. Indicative vibration data plots for State 2, as exported by MATLAB* 6.5 (R13), (a) time Table 3. Experimental setup characteristics, State 1 highlighted to indicate nominal state. Auxiliary data was recorded for each state to define the operating state of the pump compared to the manufacturer’s running curves and as opposed to the “Best Efficiency Operating State (State 1) and are presented in Table 4”. Recorded auxiliary data are: When using the Classification Tree method, as shown in Figure 7a, only the datasets for State 0 are 100.00% correctly classified, which was expected since it is the state that differs the most from the others. The highest success classification rate apart from State 0 is for States 2, 6, and 8, with 80.00% of the data correctly classified. The method is greatly unsuccessful in identifying State 1, which is wrongly classified as State 2 for 20.00% of the datasets, as State 4 for another 20% of the datasets, and as State 8 for 60.00% of the datasets. Additionally, although 60.00% of the data for State 9 is correctly classified, 20.00% of the data for State 9 is classified as State 0, and 20.00% of the data is classified as State 8. Figure 9. Cont. What is of greater value, rather than just observing the success rate of the method, is the confusion matrix of each axis. Figure 9 shows the actual and predicted states for each axis for the Classification Tree method. Classes 0-9 correspond to States 0-9 as described in Table 2. Different colors in Figure 9 subfigures are used for consistency with the colors in Figure 8. Figure 8. Success rate (%) of the classification tree method for all states and for all axes. The figure reveals that most states and axes achieve a 100.00% success rate. However, for State 0 on the z-axis, the success rate drops to 60.00%. States 1, 3, and 5 show an 80.00% success rate on the z-axis, while States 3 and 5 exhibit an 80.00% success rate on the y-axis. State 2 has a 60.00% success rate on the x-axis, and State 4 achieves a 100.00% success rate on all axes. States 6 through 8 attain 100.00% accuracy in classification. In contrast, State 9 shows deviations on all axes, with less than half of the data correctly classified for the x- and y-axes, resulting in a 40.00% success rate and a 60.00% success rate for the z-axis. For the x-axis, 40.00% of the data are classified in State 2 rather than in their correct State 3. Similarly, 60.00% of the data is wrongly classified in State 8 rather than in State 9. For the y-axis, where the overall success rate is lower, 60.00% of the data are classified in State 0 instead of State 1. 20.00% of the data are classified in State 2 instead of State 1, 20.00% of the data are classified in State 0 instead of State 5, and finally, 40.00% of the data are classified in State 2 instead of State 9. The z-axis confusion matrix shows similar behavior to the y-axis. 80.00% of the data for State 3 is classified correctly, with the balance being classified as State 4. 20% of the data for State 5 is classified as State 0. Finally, part of the State 9 data is classified as State 7 (20.00%) and State 8 (40.00%). Figure 10. Success rate (%) of the KNN method for all states and for all axes. Figure 12. Success rate (%) of the SVM method for all states and for all axes. Figure 13c shows the deviation in success rate for the z-axis discussed above. In more detail, 20.00% of the data in State 4 is classified in State 3, while 20.00% of the data in State 7 is classified in State 6. Different colors in Figure 13 subfigures are used for consistency with the colors in Figure 12. SVM x-axis Figure 13. Cont. Figure 13. (a) SVM confusion matrix for x-axis (b) SVM confusion matrix for y-axis (c) SVM confusion matrix for z-axis. Figure 14. Success rate of the CNN method for all states and all axes. CNN x-axis Figure 15. Cont. Figure 15. (a) CNN confusion matrix for x-axis (b) CNN confusion matrix for y-axis (c) CNN confusion matrix for z-axis. Predicted Class CNN y-axis Predicted Class Table 5. Accuracy of the different classification methods for all three directional axes. Table 6. Summarized key findings.
