Sediment has been identified as the major problem in the hydropower plants of the Hindu Kush Himalayan region due to the young and fragile geological formations, which contribute to the erosion of hydraulic turbines and their components within the region (Xu et al., 2019; Thapa et al., 2015; Chitrakar et al., 2018; Sapkota et al., 2022). Due to sediment erosion, most power plants experience forced failures, unnecessary downtime, reduced turbine efficiency, and financial losses. In the case of the Francis turbine runner, erosion is more prominent around the trailing edge of the runner, as shown in Figure 1, which belongs to power plants in Nepal and India (Thapa et al., 2015; Sapkota et al., 2022; Sharma, 2010). While the problem of sediment erosion is also observed in Pelton turbines, Francis runners are found to offer higher operating efficiency and hence, preferred typically in medium head and flow applications (Breeze, 2019).

Another type of fault found in the runner is breakage due to fatigue loading and high-pressure fluctuations produced by rotor-stator interaction. Past studies suggest that cracks initially grow at a slow rate but increase significantly only after the onset of severe cycle fatigue (Avellan, 2000; Gagnon et al., 2012). Therefore, detecting cracks before the onset of high-cycle fatigue is both crucial and challenging due to the nature of their growth.

The flow inside the turbine is incorporated with instabilities and transients, which is more pronounced during deep part-load conditions. Large vibrations are produced in the machines due to these fluctuations (Zhao et al., 2020). Erosive cavitation in the runner can be detected by utilizing high-frequency acceleration signals from the turbine bearing (Valentín et al., 2018). Vibrations caused by abnormal operation and early-stage damage might remain undetected until they lead to significant damage in the turbine (Gagnon et al., 2012). Several common damages in hydraulic turbines are caused by vibration amplitudes at the rotational frequency, such as runner obstruction and damaged components (Egusquiza et al., 2011). The amplitude of the rotational frequency increases as associated damage appears (Vashishtha and Kumar, 2022).

Vibration monitoring is typically conducted on machinery such as turbines, pumps, and fans, which consist of several rotating parts like bearings and gearboxes (Tung and Yang, 2009). In one study, measurements were made in the axial direction as well as 90° apart in the vertical and horizontal directions (perpendicular to the shaft axis) for horizontal shaft machines (Egusquiza et al., 2018). Peak frequency shifts were less pronounced than peak amplitude shifts, with the trend of increasing amplitude shifts noticeable as the crack or fault progressed away from the hub (Gillich et al., 2015; Awadallah and El-Sinawi, 2020). The physical, geometrical, and boundary conditions also affect the vibration signature of the structure (Awadallah and El-Sinawi, 2020). A higher amplitude of vibration at harmonics of the fundamental frequency (1X) indicates faulty conditions such as unbalance, misalignment, and a bent shaft. However, misalignment issues and bent shafts typically cause strong vibrations in the 2X order synchronous frequency as well (Betta et al., 2002; Patel and Darpe, 2009; Adams, 2009).

In condition monitoring, signal processing techniques are crucial as they examine and establish the relationship between input data such as vibration, sound, and current signals and the damage in tools and machines (Mohanraj et al., 2020). Frequency spectrum analysis is advantageous because it shows how much of a signal exists within a given frequency band, whereas time-domain analysis only reveals changes in a signal over time. This makes frequency spectrum analysis ideal for signals with more non-stationary characteristics (Srinivasan and Eswaran, 2005). Vibration signature analysis is an effective technique for identifying and tracking machine faults, with levels of damage in tools and machines clearly indicated by shifts in the acceleration spectrum and changes in magnitude (Rmili et al., 2006). Sensors play critical role in condition monitoring, providing real-time data for predictive maintenance. However, sensor faults such as bias errors, gain faults and drift faults can lead to inaccurate diagnostics. Recent advancements in sensor-based fault diagnosis integrate AI-driven models that enhance accuracy of fault detection and isolation (Chauhan et al., 2024). Despite the advancements in ML and Senosr-based fault detection, challenges remain in data availability, model interpretability and adaptability to real-world environments. High-dimensional sensor data require efficient feature selection techniques, and deep learning models often lack transparency, making them difficult to interpret for industrial applications (Vashishtha et al., 2025).

This study aims to predict faults in a locally manufactured Francis runner by analyzing the frequency spectrum of the runner in both healthy and faulty conditions. A uniaxial accelerometer is placed at the turbine bearing to capture the shaft amplitude and its harmonics. Fast Fourier Transform (FFT) and Wavelet Transform along with EMD algorithm are applied to analyze the frequency components of the vibrational signal.

The experimental setup consists of a Francis Test Rig, as shown in Figure 2. The rig features a main motor that pumps hydraulic fluid into the turbine and has a power capacity of 11 kW. It can produce up to a 36-m head with a discharge of 23.33 L/s. The pump is controlled using a Variable Frequency Drive (VFD) to achieve the required head and flow rate.

Similarly, the rig includes a sump tank to store the hydraulic fluid, which circulates from the main pump through a pipe into the turbine and exits from the draft tube back into the sump tank. The major components of the test rig include the Francis turbine, spiral casing, draft tube, turbine shaft, and induction generator. The speed of the induction generator is also controlled using another VFD, and it has a power capacity of 3 kW. To maintain the position of the draft tube, the turbine components and generator are located above the sump tank.

For measuring the pressure, two absolute pressure sensors are used: one at the inlet of the turbine and the other at the outlet of the runner. The flow through the pipe is measured using an ultrasonic flowmeter, and the rotational speed of the turbine shaft is determined using an optical sensor. A uniaxial accelerometer is installed at the top of the turbine bearing to capture the radial vibrational signal. The data acquisition devices are from National Instruments, and the data acquisition program is developed in LabVIEW. The data acquisition rate is 12,000 samples per second for the vibration sensor and 2,000 samples per second for the pressure and rotational speed sensors. The data were acquired for each operating point for a duration of 12 s. Repeatability tests were conducted to ensure the repeating nature of the result. The specifications of the sensors and devices used during the experiments are listed in Table 1.

A locally manufactured Francis turbine (Figure 3, with a design head of 12 m and a design flow of 32 L per second, was selected as the test specimen for the experiment. The detailed design specifications of the Francis turbine are listed in Table 2. The turbine has ten blades and is optimally designed to operate at 2250 RPM.

The Fast Fourier Transform (FFT) algorithm was used to analyze the vibrational signal around the turbine bearing. FFT is one of the most efficient algorithms used to obtain the frequency components of a time-domain signal. (Cooley et al., 1969). Figure 6 shows the frequency spectrum of the vibrational data acquired at a rotational speed of 600 RPM under the Case One Wet condition. The amplitudes of the fundamental frequency (10 Hz) of the shaft and its harmonics (20, 30, 40, 50, 60) are the frequencies of interest for this specific speed and can be visualized in the graph. It also includes the vibrational frequency around 24 Hz, which corresponds to the frequency of main pump in operation. While comparing this plot with the FFT plot at dry condition, the presence of sub harmonic components can be seen in wet condition while it is not visible at the dry condition. This might be the vibrations due to the flowing fluid, resulting in various fluctuations.

Similarly, for other test cases and operating speeds, the amplitude of the fundamental frequency and its harmonics were determined using FFT, referenced to the speed of the shaft. The speed of the shaft ranges from 7 to 16 Hz (420–960 RPM) for both dry and wet conditions. After determining the amplitudes of the shaft frequency, graphs were plotted for both dry and wet conditions with all test cases, as seen in Figures 7, 9.

While analyzing the result for dry condition (Figure 7), overlapping amplitudes can be observed between 7 and 12 Hz, followed by a peak in amplitude around the 12–14 Hz frequency, and a sharp decrease in amplitude after 14 Hz for all test cases. The peak amplitudes of 1.484, 1.449, and 1.371 m/s² can be seen in decreasing order as the faults increase, at frequencies of 13.36, 13.45, and 13.72 Hz for Case One Dry, Case Two Dry, and Case Three Dry, respectively. These frequencies can be referred to as the resonant frequencies, as they experience peaks due to the application of external forces. According to the literature, resonance occurs in a system when the excitation frequency matches its natural frequency, which depends on the system’s mass, stiffness, and geometry (Kraige, 2015; Singiresu, 2010).

Figure 8 shows the comparative plot of vibrational amplitude for the fundamental frequency in the dry condition. The dry condition is defined as the state when no hydraulic fluid is flowing through the turbine components, with experiments conducted under normal turbine conditions (emptying and filling the sump tank), abbreviated as Case Empty and Case Filled, respectively. The result of Case Filled shows almost a 1000% reduction in the vibrational resonant amplitude peak compared to Case Empty. This indicates that vibration is absorbed when the sump tank is filled with hydraulic fluid.

Similarly, while performing experiments in the wet condition (Figure 9), the results are comparable to the dry condition. The resonating peak has shifted to 10 Hz, while it was about 13.5 Hz in the dry condition. The peak amplitudes for Case One Wet, Case Two Wet, and Case Three Wet at the resonating frequency of 10 Hz were obtained as 0.429149, 0.362469, and 0.521942 m/s², respectively. The amplitude of vibration is lower in the wet condition (0.429149 for Case One Wet) compared to the dry condition (1.484 for Case One Dry) for all the cases. When comparing Case One Wet (normal) and Case Two Wet (faulty), the resonating peak amplitude decreases in Case Two Wet. In contrast, in Case Three Wet (faulty), the resonating peak amplitude is the highest.

The scalogram plot is useful to visualize the amplitude variations of a signal in the both time and frequency domain. Figures 10, 11 represent the scalogram plot for wavelet transform result of faulty cases, i.e., dry and wet. Referring to Figure 7, wavelet transform algorithm for dry condition has been applied to visualize the transient behavior of signal. Figure 10 is the scalogram plot for the Case Three Dry (faulty) in which the turbine is operated at 13 Hz, and it shows the dominant frequency components which is concentrated at 13 Hz and aligns with the FFT result. The plot shows clear and consistent energy distributions at the operating frequency over time.

Similarly, with reference to Figure 9, scalogram plots for wet condition has been generated to visualize the signal in time frequency domain. Figure 11 is the plot for Case Three Wet (faulty) in which turbine’s operational frequency is 10 Hz and is the dominating frequency component in the scalogram plot as well. Overall concentration of the energy is at 10 Hz and is consistent along the time domain. Comparing the plot with dry case, the amplitudes of the dominating frequency in wet case is low and it aligns with the FFT result in Figure 9. In both of the plots, there is a presence of weaker but visible components at the multiple of turbine’s frequency.

The study aims to predict turbine faults by applying the frequency transform algorithm and signal de-noising followed by wavelet transform algorithm to visualize vibrational time series data. The turbine faults are self-induced by adding masses to the top of the turbine blades. The experiment revealed that the turbine fault (adding masses) primarily affects the amplitude of the peak resonating frequency, with minimal effect on the amplitudes of other frequencies. The induced fault in the turbine can either increase or decrease the amplitude of the peak resonating frequency. In the dry condition, as the severity of the fault increased, the amplitude of the peak resonating frequency decreased from 1.484 to 1.371 m/s². In the wet condition, as the severity of the fault increased, the peak resonating frequency initially decreased from 0.429 to 0.3762 m/s², and then increased to 0.521 m/s² as the severity further increased. It can be concluded that turbine faults can be predicted by studying the amplitude of the resonating frequency.

While analyzing the results with the wavelet transform algorithm in scalogram plot, it aligns with the result of Fast Fourier Transform algorithm. In both dry and wet condition, there is clear and consistent energy distributions at the turbines operating frequency in Figures 10, 11. Also, the stronger amplitudes of turbine’s frequency can be seen in scalogram plot in dry condition compared to wet condition and it is the same with FFT result.

As suggested by the literature (Kraige, 2015; Singiresu, 2010), any changes in the mass of the system can alter the natural frequency of the system. In this experiment, when transitioning from dry to wet conditions, hydraulic fluid was added to the sump tank, causing the natural frequency of the system to shift from about 13.5 to 10 Hz, along with a decrease in the amplitude of the vibrational signal. Thus, any changes in the mass of the system can shift the resonating frequency.