Fault feature extraction for centrifugal pump impellers via EMD and cyclic bispectral slicing Authors: Xiaopei Liang, Hanxin Chen, Lei Wang, Shaoyi Li, Zhigang Wang, Zhaohui Shi, Xiangying Sheng Published: 2025 ABSTRACT AbstractThe vibration signal of early centrifugal pump impeller faults is a nonlinear, non-Gaussian, non-steady-state signal with inherent periodicity. These characteristics complicate the accurate extraction of fault features. This study aims to explore a novel feature extraction method for centrifugal pump impeller fault vibration signals. This method leverages the adaptive characteristics of empirical mode decomposition (EMD) for multi-scale decomposition of the original vibration signal and separation of each intrinsic mode function (IMF). The cyclic bispectrum secondary slicing technique is introduced to perform high-order statistical purification of the noisy IMF, and the modulation frequency characterizing the fault is accurately isolated through optimized slicing parameters. In the analysis of actual centrifugal pump impeller vibration signals, this method effectively enhances the separability and anti-noise robustness of the modulation component. Furthermore, the extracted features are input into SVM, XGBoost, and 1D-CNN classification models, with test accuracies of 85.7%, 92.1%, and 95.7%, respectively, significantly outperforming the single-feature method.Keywords: Centrifugal pump, Fault diagnosis, Cyclic bispectrum secondary slicing, Empirical mode decompositionSubject terms: Energy science and technology, Engineering Introduction Centrifugal pumps, widely used fluid-conveying devices in industrial production, are primarily composed of a pump body, pump shaft, impeller, sealing device, and other components. As the core rotating component, the safe and stable operation of the impeller is crucial. During long-term operation, the impeller is exposed to high temperature, high pressure, and high-speed rotation, making it susceptible to wear, cracking, deformation, perforation, and other failure modes. Because centrifugal pumps require high sealing performance, timely fault diagnosis enables accurate maintenance without disrupting normal operation, thereby preventing casualties and secondary failures in healthy components. Moreover, it reduces equipment downtime and enhances production efficiency. Consequently, increasing attention has been directed to impeller fault diagnosis, with emphasis on extracting key fault information to ensure the efficiency and reliability of centrifugal pumps. In recent years, various effective methods have been applied to early fault diagnosis of centrifugal pumps, providing valuable approaches to enhance system reliability and performance. Hanxin et al.1 proposed the use of parallel factor analysis to diagnose centrifugal pump faults. Wasim et al.2 proposed the use of Sobel edge scale mapping for centrifugal pump diagnosis. Because of differences in feature extraction approaches, the diagnostic performance of the two methods varies considerably. Wathan and Husaini3 investigated the diagnosis of pump shaft fractures; however, few studies have focused on impeller faults, even though the impeller is a key component with a high failure rate and plays a central role in energy conversion. On the one hand, the complex internal flow field of the impeller makes accurate extraction and identification of fault characteristics challenging; on the other hand, since the impeller typically operates with other mechanical components (such as shafts and bearings), the collected information is highly complex, complicating analysis and diagnosis4. Consequently, relatively few studies have addressed impeller fault diagnosis, and progress in this field remains slow. In impeller fault diagnosis, the vibration signal of centrifugal pumps is typically non-stationary, nonlinear, and non-Gaussian, containing rich time-domain, frequency-domain, and time–frequency-domain information5. For such signals, traditional processing methods—including time-domain analysis (mean value, peak index, etc.), frequency-domain analysis (Fourier transform, power spectrum estimation, etc.), and time–frequency analysis (wavelet transform, short-time Fourier transform, etc.)—yield limited effectiveness. The Fourier transform (FT) is applicable only to stationary processes6. Ladd7. proposed using the Discrete Wavelet Transform (DWT) to detect mechanical fault noise, but the results were inconclusive. Empirical mode decomposition (EMD), introduced by Bessam et al.8, is an adaptive signal decomposition technique designed to overcome the limitations of wavelet transform. To provide a novel approach for solving non-stationary signal processing, empirical mode decomposition (EMD) has been effectively applied to centrifugal pump fault diagnosis. However, EMD has inherent limitations, such as modal aliasing and endpoint effects. To address this, a series of studies have focused on improving EMD at the algorithmic level. Lei9 proposed adaptive EEMD to resolve the modal aliasing problem and enhance adaptability. Tang et al. utilized morphological filtering to assist EMD in mitigating modal aliasing10. However, these methods generally introduce higher computational complexity and subjective parameters, thereby compromising the simplicity and adaptability of EMD. In response to the above limitations of algorithm optimization, subsequent research has increasingly focused on the diagnostic strategy itself, specifically the development of feature extraction and intelligent diagnostic models that do not rely on specific decomposition algorithms. Ahmad et al.11 proposed a fault diagnosis method for centrifugal pumps based on three-stage discriminant feature extraction. Through cross-correlation analysis, a comprehensive feature pool containing mixed time-, frequency-, and time–frequency-domain features was constructed. When combined with a multi-class support vector machine (MSVM) classifier, the detection accuracy of weak early faults and the classification performance for different damage levels were significantly improved, achieving considerable gains compared with traditional methods. Sunal et al.12 applied motor current characteristic analysis (MCSA) to directly detect induction motor faults. Garousi et al.13 employed vibration signal analysis combined with a multi-layer perceptron (MLP) algorithm to achieve real-time diagnosis and severity quantification of impeller defects and cavitation faults in centrifugal pumps. This demonstrates that, whether focused on optimizing the decomposition algorithm itself or on post-processing for feature extraction, the effective extraction of fault features from complex background noise remains a central challenge. For practical applications, when second-order statistical methods are used to extract fault characteristics from centrifugal pump signals, the classification accuracy is very low. Therefore, to obtain more effective and higher-resolution fault features, it is essential to employ higher-order statistical methods. Higher-order moments in these statistics can effectively suppress Gaussian noise in the signal, thereby isolating the fault signal. The cyclic autocorrelation function in higher-order cyclic statistics can demodulate both the carrier and modulation frequencies, and different demodulated frequencies can be obtained by slicing the main frequency of the cyclic autocorrelation function. Zhou et al.14 employed cyclic bispectrum secondary slicing and obtained relatively good fault features. However, Guo et al.15 found that after the cyclic bispectrum was first diagonally sliced, the second slice—whether peak-frequency or multiple peak-frequency—still contained modulation frequency components, carrier frequency components, and numerous difference, combination, and harmonic frequency components. Thus, the frequency components of the processed spectrum remained overly complex. Yang Xiaoqing et al.16 demonstrated that the carrier frequency components of faulty and non-faulty vibration signals are essentially identical, with the primary distinction being in the modulation frequency component. They clearly stated that the primary distinction between fault and non-fault vibration signals lies in the modulation frequency component, while the carrier frequency remains essentially the same. Consequently, the accurate and efficient extraction of modulation frequency components has become critical to enhancing the fault diagnosis performance of centrifugal pumps17. Based on this, this paper proposes a novel approach: accepting the modal aliasing of EMD and utilizing the cyclic bispectrum secondary slice to achieve feature extraction that overcomes decomposition defects in the IMF component. Through simulation experiments, the influence of cyclic frequency parameters and slicing methods is systematically investigated, and the optimal parameter combination is identified: anti-angle slicing with the cyclic frequency set to the triple carrier frequency. After applying this parameter configuration to experimental data, the fault features of the centrifugal pump impeller are extracted, and the EMD-derived features are input into the same classification model for testing. The proposed EMD–cyclic bispectrum secondary slicing fusion method offers a novel technical approach for fault diagnosis of rotating machinery and has significant engineering application value in early fault feature extraction and accurate identification of centrifugal pump impellers. The remainder of this paper is organized as follows: “Basic theory” section presents the basic theory, introducing the principles of cyclic bispectrum slicing and the EMD method, and providing theoretical support for subsequent research. “Simulation signal analysis and parameter optimization” section details the simulation experiment process and determines the optimal cyclic slicing frequency and method. “Experimental data feature extraction and diagnostic verification” section verifies the feature extraction capability of the proposed method using actual centrifugal pump vibration data, and compares the results with those obtained from original signal processing. “Diagnostic performance verification based on classification model” section employs SVM and XGBoost classification models to systematically evaluate the diagnostic accuracy of different feature extraction methods (cyclic bispectrum secondary slicing, EMD decomposition, and EMD–cyclic bispectrum secondary slicing), thereby verifying the superiority of the fusion method. “Complimentary close” section presents the conclusions and prospects, summarizing the main contributions, highlighting the limitations of the method, and offering suggestions for future research. Basic theory The purpose of this paper is to systematically elaborate on the theoretical basis, implementation steps, and core innovations of the proposed ‘EMD-cyclic bispectrum anti-corner slice’ fusion method, and to provide theoretical support for subsequent experimental verification. The principle of cyclic bispectrum dimensionality reduction and feature extraction mechanism Cyclic bispectrum is the third-order cyclic cumulant spectrum. For a cyclic stationary signal x(t), its k-th order cyclic moment is defined as where τ1, τ2, …, τk−1 are the lag times, and α is the cyclic frequency. When k equals 3, if x(t) is a zero-mean cyclic stationary signal, the third-order cyclic moment is equal to its third-order cyclic cumulant18. According to the definition of bispectrum, the third-order cyclic spectrum (cyclic bispectrum) can be obtained from the third-order cyclic cumulant through a two-dimensional Fourier transform. Formula (3) demonstrates that the cyclic bispectrum is a two-dimensional spectral representation of the third-order cyclic cumulant of the signal, which can reflect the nonlinear relationship between different frequency components. The cyclic bispectrum includes three unknown four-dimensional distributions, namely cyclic frequency α, frequency ω1, and frequency ω219. To facilitate feature extraction and visualization, either frequency ω1 or ω2 in the cyclic bispectrum is typically fixed to reduce its dimensionality.The cyclic frequency α is chosen as three times the peak frequency in the spectrum, and this value is used for slicing analysis, yielding a two-dimensional frequency spectrum that is subsequently subjected to spectral analysis. The detailed procedure is illustrated in Fig. 1. Schematic diagram of slicing analysis steps. Based on the above principle, the calculation of the cyclic bispectrum slice can be divided into three steps: signal transformation, third-order cyclic cumulant calculation, and bispectrum slice construction. The mathematical expressions are given by (4–6). Assuming the cyclic stationary signal , the specific steps of the cyclic bispectrum slicing algorithm are as follows: The three transformations of the signal x(t):4where , M represents the number of sampling points, N2 represents the lag time, and . The three transformations of the signal x(t): 4where , M represents the number of sampling points, N2 represents the lag time, and . The third-order cyclic cumulant of the signal x(t) at the cyclic frequency α: 5 The third-order cyclic cumulant of the signal x(t) at the cyclic frequency α: 5 The cyclic bispectrum slice of the signal x(t): 6 The cyclic bispectrum slice of the signal x(t): 6 Principle of positive diagonal slice and anti-diagonal slice However, the direct use of complete two-dimensional bispectrum features is computationally intensive in practical applications, so two-dimensional bispectrum is usually converted into one-dimensional feature functions and feature slices. Among them, diagonal slice (ω1 = ω2) is the most commonly used slice form, which can effectively extract fault feature information, remove high-order noise, improve signal-to-noise ratio, and reflect the high-order correlation of signals20. In the diagonal slice, the cyclic frequency α is fixed, and the frequency ω1 is equal to the frequency ω2. It can be obtained from Eq. (3). At this time, the slice extracts the frequency component along the main diagonal, which mainly reflects the carrier frequency and its harmonic components. When ω1 = − ω2 = ω, the anti-angle slice is obtained in the form : The anti-diagonal slice corresponds to the anti-diagonal direction of the frequency plane, which highlights modulation frequency information, suppresses the carrier component, and achieves better feature separation for non-stationary and modulated signals. Empirical mode decomposition EMD decomposition principle Empirical Mode Decomposition (EMD) is an adaptive signal decomposition technique designed to decompose complex signals into several Intrinsic Mode Functions (IMFs) and a residual trend term21. The primary function of EMD is to decompose a signal into a set of components with distinct frequencies that are localized in the time domain, thereby enabling multi-scale signal analysis. The necessary conditions for EMD are as follows: (1) for each IMF, the difference between the number of extrema and the number of zero-crossings must not exceed one; and (2) the upper and lower envelopes must be locally symmetric with respect to the time axis22. The decomposition formula is expressed as follows: where is the ith-order IMF component, is the residual term, and n is the maximum number of IMF components decomposed. EMD can effectively achieve signal denoising and feature extraction. This method decomposes a non-stationary signal into a series of Intrinsic Mode Functions (IMFs) through an iterative sifting process. These IMFs capture different frequency components of the signal and characterize its vibration behavior across multiple time scales. Based on this property, the IMFs obtained through EMD are employed as input signals for cyclic bispectrum secondary slicing to assess their effectiveness in fault feature extraction. By analyzing the higher-order statistical characteristics of each IMF, this method reveals the hidden nonlinear features and modulation information in the signal with greater depth. EMD-cyclic bispectrum anti-corner slice combination The core concept of combining EMD decomposition with the anti-corner slice is that the complex vibration signal is decomposed into several IMFs by EMD, and the IMF component with the highest correlation to the original signal is selected as the input for the anti-corner slice, thus highlighting the modulation frequency and suppressing the carrier interference. Subsequently, cyclic bispectrum secondary slice analysis is introduced, and the anti-corner slice is compared with the traditional diagonal slice. Since the anti-corner slice intercepts the spectrum along the direction in the frequency plane, which corresponds to the difference frequency relationship between the signal frequency components, it can enhance the modulation frequency information. When the input signal is preprocessed by EMD, the carrier and its frequency-doubled energy are separated in each IMF, and the anti-corner slice can accurately extract the modulation frequency characteristics from the lower-dimensional IMF signal. The combination mechanism can be summarized as the following three steps : decompose the original signal x(t) into several IMF components ci(t) ; decompose the original signal x(t) into several IMF components ci(t) ; Calculate the correlation coefficient between each IMF component and the original signal, and select the component with the highest correlation as the subsequent analysis object. Calculate the correlation coefficient between each IMF component and the original signal, and select the component with the highest correlation as the subsequent analysis object. Anti-corner slicing stage : The selected IMF component is subjected to cyclic bispectrum anti-corner slicing analysis to extract its modulation frequency component. Anti-corner slicing stage : The selected IMF component is subjected to cyclic bispectrum anti-corner slicing analysis to extract its modulation frequency component. In summary, based on the adaptive decomposition ability of EMD and the unique advantages of the cyclic bispectrum anti-corner slice in feature extraction, although envelope demodulation can extract a single modulation frequency, its spectral focusing and robustness are limited under multi-carrier, strong noise, and nonlinear conditions. The introduction of EMD in this paper is not to extract the feature itself but to provide a low-noise, modulated input signal for the cyclic bispectrum, which makes the high-order statistical features more concentrated. In this section, the ‘EMD-cyclic bispectrum anti-corner slice’ fusion method is systematically constructed, and experimental verification will be carried out below. Simulation signal analysis and parameter optimization The vibration signal of a centrifugal pump consists of the pump’s inherent vibration frequency (carrier frequency), the fault characteristic frequency (modulation frequency), and Gaussian white noise from the surrounding environment. To verify the effectiveness of cyclic bispectrum secondary slicing in extracting fault features from centrifugal pump vibration signals, an amplitude modulation (AM) simulation signal, consistent with actual operating conditions, is first employed to model the impeller fault characteristics. After evaluating its discriminative capability through quantitative performance metrics, the influence of slicing methods and cyclic frequency parameters in the cyclic bispectrum secondary slicing process on fault feature extraction is systematically analyzed to identify the optimal parameter combination. Simulation signal construction and noise injection To simulate the “carrier-modulation-noise” characteristics of centrifugal pump impeller fault vibration signals, an amplitude modulation (AM) simulation signal is constructed with modulation frequencies of 10 Hz and 20 Hz, a carrier frequency of 100 Hz, and Gaussian white noise. The simulation signal is represented by Eq. (10): where fm1 and fm2 are the modulation frequencies, fc is the carrier frequency, the sampling frequency is 1000 Hz, the total number of sampling points is 4048, the time delay coefficient is τ = 2s, N1 = 2000, N2 = 2048. To quantitatively control the simulated noise level, the signal-to-noise ratio (SNR) defined in decibels (dB) is adopted: where xsignal is the noise-free simulation signal, xnoise is the additive Gaussian white noise, and σn is the standard deviation of the noise.In this study, experiments are conducted under low SNR conditions to simulate the operation of a centrifugal pump in a strong noise environment, thereby verifying the robustness and stability of feature extraction of the proposed method under adverse noise conditions. Optimization of cyclic bispectrum slicing method and cyclic frequency The baseline model employs a deterministic simulation signal without additive Gaussian white noise as the analysis object. where fm1 and fm2 are the modulation frequency, fcis the carrier frequency, the sampling frequency is 1000 Hz, the number of sampling points is 4048, the delay coefficient is and the Fourier change of the simulation signal is shown in Fig. 2.The signal-to-noise ratio (SNR) is set to − 3 dB to simulate harsh industrial conditions where the fault characteristics of centrifugal pump vibration signals are significantly obscured by strong background noise. Signal spectrum of Fourier transform of simulation signal. When diagonal slicing with different cyclic frequencies is applied, the resulting cyclic bispectrum secondary slices are illustrated in Fig. 3. Original spectrum and cyclic bispectrum secondary slice spectrum after different cyclic frequencies. Analysis of the cyclic bispectrum secondary slice (CBSOS) under diagonal slicing with different cyclic frequencies α shows that when α is set to the modulation frequencies (10 Hz, 20 Hz) or to the second harmonic of the carrier frequency (200 Hz), the CBSOS exhibits periodicity, with peaks appearing at integer multiples of the carrier frequency (100 Hz, 200 Hz, 300 Hz). However, this is accompanied by significant broadband spectral interference, which prevents effective extraction of modulation features. When α is adjusted to 80 Hz, 90 Hz, 100 Hz, 120 Hz, and 300 Hz, the CBSOS spectrum produces distinct peaks at the modulation frequency harmonics (10 Hz, 20 Hz), the second harmonic of the carrier (200 Hz), and its sum and difference frequencies with the modulation frequencies (180 Hz, 190 Hz, 210 Hz, 220 Hz), while the amplitude of the background noise components is attenuated to a negligible level. Based on these results, the cyclic bispectrum secondary slicing method is capable of extracting characteristic frequencies. However, because centrifugal pump impeller faults are primarily characterized by modulation frequencies, the traditional diagonal slicing approach exhibits inherent limitations: its output not only includes the target modulation frequency but also contains the carrier frequency and its intermodulation components (sum and difference frequencies). Definition of bispectral slicing and parameter settings To evaluate the theoretical capability of different slicing methods for fault feature extraction, a comparative analysis is initially conducted using ideal simulation signals without additive Gaussian white noise. In noisy conditions, Gaussian white noise introduces additional frequency components into the signal, leading to random peaks in the amplitude spectrum of the cyclic bispectrum. This effect complicates distinguishing the intrinsic differences between the two slicing methods in terms of frequency extraction capability. Therefore, during the method validation stage, analyses under “pure signal” conditions are conducted to provide an intuitive comparison of diagonal slicing and anti-diagonal slicing in three key aspects: modulation frequency extraction, spectral peak resolution, and completeness of frequency components. For the noise-free version of the simulation signal, a deterministic simulation signal without Gaussian white noise is used as the analysis object. The simulation signal is represented by Eq. (13). The Fourier transform of the signal is shown in Fig. 4, while the diagonal and anti-diagonal slicing results under different cyclic frequencies are presented in Fig. 5. The signal spectrum of the Fourier transform of the simulated signal. Comparison of diagonal slices and anti-diagonal slices at different cyclic frequencies: (a) diagonal slices at different cyclic frequencies (b) spectrum of anti-diagonal slices at different cyclic frequencies. Through cyclic bispectrum secondary slicing analysis of the simulation signal under noise-free conditions, it is observed that anti-diagonal slicing reveals richer frequency component characteristics than diagonal slicing. For example, when the cyclic frequency is set to 80 Hz, diagonal slicing reveals spectral peaks only at the carrier frequency, modulation frequency, and their first- and second-order sum and difference frequencies. In contrast, anti-diagonal slicing additionally presents third- and fourth-order sum and difference frequency components, indicating its higher sensitivity and resolution in extracting nonlinear modulation characteristics. Notably, when the cyclic frequency reaches 300 Hz, both slicing methods effectively extract the primary modulation frequency component, with the spectrum primarily displaying modulation frequency information, suggesting that the cyclic bispectrum at this frequency can focus on the dominant modulation energy distribution. Comparative analysis further reveals that anti-diagonal slicing demonstrates significantly enhanced spectral peaks at twice the modulation frequency compared to diagonal slicing. Based on the above results, anti-diagonal slicing is confirmed to have superior feature extraction capability. Therefore, in subsequent experiments, the cyclic frequency is uniformly set to 300 Hz, and both simulated signals and centrifugal pump impeller vibration signals are processed using cyclic bispectrum secondary slicing to verify the consistency and effectiveness of the proposed method in both simulation and practical signal analysis. As shown in Fig. 6, after the addition of noise, diagonal slicing reveals multiple non-modulation-related spectral peaks, in addition to the main modulation frequency observed in the noise-free case. Significant energy enhancements occur at the second and fourth harmonics of the carrier frequency, as well as at the sum and difference frequencies between twice the carrier frequency and the modulation frequency. This indicates that diagonal slicing is more susceptible to multi-frequency coupling effects under noisy interference, leading to complex and unfocused feature distributions. Cyclic bispectrum secondary slicing of the noisy simulation signal using diagonal and anti-diagonal slicing methods. (a) Diagonal slicing, (b) Anti-diagonal slicing. In contrast, anti-diagonal slicing presents prominent peaks primarily at the modulation frequency and its second harmonic, while the peak amplitudes at other frequencies remain very low, with noticeably reduced noise interference. This demonstrates that anti-diagonal slicing more effectively suppresses carrier-related and higher-order harmonic interferences in noisy environments, thereby achieving high-SNR extraction of modulation frequency features. Therefore, the comparative analysis of noisy simulation signals confirms that anti-diagonal slicing exhibits superior capability in identifying modulation frequencies under complex noise conditions, further validating its robustness and applicability in nonlinear modulation feature extraction and complex mechanical system signal analysis. Experimental data feature extraction and diagnostic verification Common fault types of centrifugal pumps A centrifugal pump, widely used as a fluid transport device in industrial production, primarily consists of the pump casing, impeller, inlet and outlet pipelines, and bearing and sealing components. The fluid enters the pump through the inlet, is accelerated by the impeller, and flows through the pump casing toward the outlet, as shown in Fig. 7. Due to the high sealing requirements during operation, direct intervention in the pump structure is avoided during intelligent fault diagnosis. Instead, accelerometers are typically mounted on the external surface of the pump casing to collect vibration signals. Impeller fault-induced vibrations are transmitted through the shaft–bearing–housing system to the exterior of the pump. Schematic diagram of a single-stage single-suction centrifugal pump. Because the impeller is located inside the pump body and is difficult to access directly, vibration analysis methods are used for indirect (non-invasive) diagnosis. Given the structural complexity of the centrifugal pump and the strong vibration coupling among its components, single-channel vibration signals are often insufficient for accurate impeller fault detection. Therefore, in this study, a multi-sensor configuration is employed to collect multi-channel vibration signals, forming a centrifugal pump impeller fault diagnosis system. This system provides comprehensive experimental data to support subsequent cyclic bispectrum-based feature extraction, thereby enhancing the accuracy and reliability of fault diagnosis. The centrifugal pump impeller is a mechanical component that frequently comes into contact with the working fluid and mechanical agitator. It is prone to varying degrees of damage during pump operation. In this experiment, three faulty impellers—blade-damaged (F2), edge-worn (F3), and through-hole (F4)—were selected and compared with a normal impeller (F1). The condition diagrams of the four impellers are shown in panels (a–b) of Fig. 8. Data acquisition system: (a–d) four kinds of impeller state diagram (e) sensor arrangement of centrifugal pump operation monitoring points (f) data acquisition system. To collect characteristic signals under nonlinear multi-fault conditions, a mechanical multi-source dynamic signal monitoring system was established. The motor speed was set to 1200 rpm, and the impeller had five blades. The system measured vibration signals at three measuring points in the X, Y, and Z directions at each point. As shown in panel (e) of Fig. 8, one sensor was installed on top of the centrifugal pump, another at the pump outlet, and a third on top of the bearing. Since acceleration sensors were used, the amplitude unit in the subsequent spectral analysis is m/s². The complete vibration signal acquisition system is shown in panel (f) of Fig. 9, consisting primarily of a signal analyzer and a notebook computer for data storage. The system is capable of collecting vibration signals in multiple directions. Data acquisition followed the requirements below: Spectrum of vibration signals of four impeller modes at three times the main peak cycle frequency. (a) Signal F1, (b) Signal F2, (c) Signal F3, (d) Signal F4. Stable operation of the centrifugal pump must be ensured before data collection. Stable operation of the centrifugal pump must be ensured before data collection. The sampling frequency was determined according to the sampling theorem. The sampling frequency was determined according to the sampling theorem. Multiple datasets were collected for each impeller state to reduce contingency. Multiple datasets were collected for each impeller state to reduce contingency. Cyclic bispectrum analysis of impeller fault vibration signals The vibration signal of a centrifugal pump impeller arises from the coupled effects of rotational excitation, fluid-induced perturbations, and structural nonlinearities during operation.Impeller fault characteristic frequencies rarely exist as singular, analytically predictable values; rather, they emerge as modulation sidebands and energy redistribution near the rotational frequency, blade passing frequency (BPF), and their harmonics, caused by intricate nonlinear coupling effects.In the normal operating state, the main periodic components of the impeller vibration signal consist of the rotor rotation frequency and the blade passing frequency (BPF).The blade passing frequency is defined by the rotational speed n and the number of blades Z, which can be expressed as follows: In this experiment, the motor speed is , the impeller has five blades, and the theoretical blade passing frequency is 100 Hz, which coincides with the main peak position of the normal condition (F1) in Fig. 9. When structural defects such as blade cracking, leading-edge erosion, or perforation occur in the impeller, time-varying stiffness and damping induce amplitude or frequency modulation of the carrier frequency (BPF), resulting in sideband components at . The amplitude variations at the blade passing frequency and its harmonic components serve as effective indicators of impeller fault characteristics23. Previous studies have suggested that periodic fluid–structure interactions and variations in impeller stiffness can introduce modulation effects near the blade passing frequency, which may lead to sideband-like spectral features in the vibration response. Consequently, various impeller fault types in centrifugal pumps induce different modulation frequencies; thus, extracting the modulation frequency corresponding to each fault is the most efficient way to obtain fault-related feat.Building upon this theoretical principle, the present study applies cyclic bispectrum analysis combined with anti-diagonal slicing to uncover the nonlinear modulation features in impeller vibration signals, thereby facilitating the recognition and interpretation of various fault conditions.To investigate how various impeller defects affect the modulation behavior of vibration signals, the Y-direction vibration signal at the second measurement point was used as the analysis target in the centrifugal pump impeller fault diagnosis experiment. The spectra under four impeller conditions are shown in Fig. 9. The spectral analysis results in Fig. 9 reveal that different faulty impellers exhibit additional peaks in the low-frequency range (approximately 65 Hz and 125 Hz), indicating the presence of clear modulation effects in the signal. These low-frequency components typically correspond to the impeller’s rotational frequency and its harmonics, which periodically modulate the high-frequency carrier components (around 1150 Hz, 1750 Hz, and 2300 Hz). However, the conventional frequency spectrum only reflects the overall energy distribution and fails to reveal the nonlinear coupling relationship between the carrier and modulation frequencies. In particular, under noisy conditions or compound fault scenarios, the modulation sideband peaks are often obscured by the dominant peaks (e.g., around 1750 Hz), making it difficult to accurately identify the specific modulation frequencies and their higher-order components. Comparative analysis of diagonal and anti-diagonal slicing To further verify the effectiveness of different slicing methods in extracting modulation features, cyclic bispectrum secondary slicing was applied to the four impeller conditions (F1–F4). When the cyclic frequency is set to three times the dominant peak frequency, the cyclic bispectrum concentrates on the modulation coupling region, facilitating a more distinct enhancement and separation of the modulation structures in the spectrum. Since the dominant frequency in the spectrum approximately corresponds to the carrier frequency, the cyclic frequency was set to three times the dominant frequency, as illustrated in Fig. 10. Frequency spectra of vibration signals under four impeller conditions at three times the dominant cyclic frequency. (a) Signal F1, (b) Signal F2, (c) Signal F3, (d) Signal F4. As shown in Fig. 10, when the cyclic frequency is set to three times the main peak frequency, the cyclic bispectrum secondary slicing results for the different impeller states (F1–F4) reveal clearer spectral peak structures at 65 Hz, 125 Hz, 1150 Hz, 1750 Hz, and 2305 Hz, compared to the original spectrum (Fig. 10). The amplitude of the low-frequency modulation peaks is significantly enhanced, while the carrier peaks become more dispersed, leading to more concentrated spectral lines and a noticeably reduced noise background. Although cyclic bispectrum secondary slicing at three times the dominant cyclic frequency clearly exposes the modulation characteristics of impeller faults, the choice of slicing direction still exerts a considerable influence on the presentation of spectral information. The traditional diagonal slicing method has advantages in revealing the dominant frequency and its harmonics; however, it often introduces numerous cross-terms and redundant harmonic components, causing energy coupling between modulation and carrier peaks, thus reducing the distinctness of fault features. To further compare the impact of different slicing directions on modulation feature extraction, this study employs both diagonal and anti-diagonal slicing approaches for performing cyclic bispectrum secondary slicing on the vibration signals of the four impeller conditions (F1–F4). The corresponding slicing results are shown in Figs. 11, 12, 13 and 14. Diagonal and anti-diagonal slicing results of signal F1. (a) Signal F1 – Diagonal slicing, (b) Signal F1 – Anti-diagonal slicing. Diagonal and anti-diagonal slicing results of signal F2. (a) Signal F2 – Diagonal slicing, (b) Signal F2 – Anti-diagonal slicing. Diagonal and anti-diagonal slicing results of signal F3. (a) Signal F3 – Diagonal slicing, (b) Signal F3 – Anti-diagonal slicing. Diagonal and anti-diagonal slicing results of signal F4. (a) Signal F4 – Diagonal slicing, (b) Signal F4 – Anti-diagonal slicing. As shown in Fig. 11, the main peaks of the diagonal slice appear at 205 Hz (modulation) and 1650 Hz (carrier). The energy distribution is relatively broad, with partial overlap between the low- and high-frequency regions. The modulation peaks are not well focused, and the spectral lines are relatively complex.The anti-diagonal slice shows its main peaks at 5 Hz (modulation) and 3500 Hz (carrier), with clearly separated distributions and low background noise. As shown in Fig. 12, the modulation peaks of the diagonal slice appear within 190–250 Hz, the carrier energy is relatively weak, and the spectral lines remain complex.The anti-diagonal slice shows significantly enhanced modulation peaks (at 60 Hz and 190 Hz), with noticeably increased amplitudes, while the carrier energy is suppressed. As shown in Fig. 13, the diagonal slice exhibits a multi-peak structure at 250 Hz and 1280 Hz, with diffused energy and a complex spectral pattern.The anti-diagonal slice shows concentrated main peaks at 250 Hz and 500 Hz, with significantly enhanced low-frequency energy and weakened carrier-region energy. According to Fig. 14, the diagonal slice shows a low-frequency region around 470 Hz, with high-frequency carriers at 2430 Hz and 4395 Hz. The energy distributionIn the anti-diagonal slice, high-frequency energy is concentrated at 2180 Hz and 2427 Hz, with significantly enhanced peaks, while the low-frequency modulation region almost disappears. is scattered, and the spectral structure remains complex. The spectral analysis results indicate that the diagonal slicing method generates spectra with significantly higher complexity. As shown in the figures, both slicing methods can extract the characteristic frequencies of each impeller fault signal; however, the diagonal slicing results are generally more complex than those of anti-diagonal slicing. In impeller modes F2 and F3, the maximum amplitude peaks appear around 250 Hz, corresponding to twice the dominant frequency of the original signal, while in mode F1, the secondary peak occurs at approximately 205 Hz, representing the second harmonic of the fundamental frequency. This observation suggests that when the dominant frequencies of the original spectra are close to each other, diagonal slicing tends to produce similar spectral peaks. Consequently, using only the frequency corresponding to the maximum amplitude as a feature may reduce feature resolution, making it difficult to distinguish multiple fault types. In contrast, when applying anti-diagonal slicing for cyclic bispectrum secondary slicing, the results show clearer and more distinct modulation frequency information. Different impeller fault modes in the centrifugal pump yield different modulation frequencies, and thus the peak frequencies of maximum amplitude also differ. Moreover, the anti-diagonal slicing spectra contain fewer spurious frequency components, resulting in cleaner and more interpretable fault characteristics. Therefore, for cyclic bispectrum secondary slicing-based feature extraction in centrifugal pump impeller fault diagnosis, the anti-diagonal slicing method, combined with a cyclic frequency set to three times the dominant frequency, proves more advantageous. However, directly applying cyclic bispectrum secondary slicing to raw signals may still result in considerable spectral clutter. To mitigate the influence of excessive spurious components, Empirical Mode Decomposition (EMD) can be introduced as a preprocessing step. By decomposing the signal into multiple Intrinsic Mode Functions (IMFs) representing different frequency components, one can select the IMF containing richer modulation frequency information for subsequent cyclic bispectrum secondary slicing. The remaining challenge is to establish a quantitative criterion for determining which IMF component contains the most significant modulation information. This will be discussed in the following section. EMD preprocessing and cyclic bispectrum secondary slicing feature extraction To investigate which IMF component obtained through Empirical Mode Decomposition (EMD) contains the richest modulation frequency information while minimizing interference from other frequency components, the noise-free simulation signal is analyzed first. The analysis is conducted using the noise-free simulation signal defined in Eq. (13). The decomposition results from the EMD process are shown in Fig. 15. EMD decomposition diagram. The decomposed IMF signals are subjected to cyclic bispectrum secondary slicing. The slicing is performed using the anti-diagonal method with a cyclic frequency of 300 Hz. The results are shown in Fig. 16. Spectrum of IMF1-4 components and cyclic bispectrum secondary slice. Comparative analysis of the cyclic bispectrum secondary slicing results for each IMF component reveals that the map corresponding to IMF2 contains almost exclusively the modulation frequency component. This indicates that the feature extraction approach of IMF component screening, based on EMD decomposition combined with cyclic bispectrum secondary slicing, is feasible. To further investigate the optimal role of the IMF2 component, the correlation between each IMF and the original signal is analyzed. The corresponding correlation coefficients are presented in Table 1. Table 1Correlation coefficients of each IMF with the original signal.ComponentIMF1IMF2IMF3IMF4Correlation coefficient0.5370.7180.3480.140 Correlation coefficients of each IMF with the original signal. The analysis results reveal that the correlation coefficient between each IMF component and the original signal is positively correlated with the effectiveness of modulation frequency feature extraction after cyclic bispectrum secondary slicing. IMF components with higher correlation coefficients exhibit more prominent modulation frequency characteristics. To verify this correlation, a simulation signal containing Gaussian white noise is introduced for experimental validation, as defined in Eq. (10). The correlation coefficients between the first four IMF components and the original signal are calculated, and the results are summarized in Table 2. Table 2Correlation coefficients of each IMF with the original signal.ComponentIMF1IMF2IMF3IMF4Correlation coefficient0.5750.6920.3260.119 Correlation coefficients of each IMF with the original signal. The analysis indicates that the IMF2 component exhibits the highest correlation coefficient with the original signal. IMF2 is then subjected to cyclic bispectrum secondary slicing. The Fourier transform spectrum of IMF2 and the corresponding spectrum after slicing are presented in Fig. 17. IMF2 component. (a) Spectral diagram, (b) secondary slice diagram. As shown in Fig. 17, the Fourier transform spectrum of IMF2 in panel (a) exhibits significant peaks at 80 Hz, 100 Hz, and 120 Hz, corresponding to the carrier frequency and its difference and sum frequency components with the modulation frequency, indicating that this component retains the main modulation information. In contrast, the cyclic bispectrum secondary slicing result in panel (b) shows distinct peaks only at 20 Hz and 40 Hz, corresponding to the modulation frequency and its second harmonic, while the carrier frequency and its high-frequency components are completely suppressed. The spectral analysis results confirm that by selecting the IMF component with the highest correlation to the original signal as the input for cyclic bispectrum secondary slicing, efficient extraction of modulation frequency features can be achieved. Based on the above simulation verification results, to further evaluate the applicability of this method to actual complex signals, the EMD–cyclic bispectrum secondary slicing method is applied to the centrifugal pump impeller vibration signal analysis. Specifically, the Y-direction vibration signal at the second measurement point for each impeller fault is used as the analysis signal. The signal is first subjected to EMD decomposition, then the IMF component with the highest correlation is selected as the input signal; subsequently, cyclic bispectrum secondary slicing is performed using a cyclic frequency three times the dominant peak frequency and the anti-diagonal slicing method. The cyclic bispectrum secondary slicing results for the four centrifugal pump impeller modes (F1, F2, F3, F4) are shown in Fig. 17. As shown in Fig. 18, the results of the four impeller states (F1–F4) analyzed using the EMD–cyclic bispectrum secondary slicing method show that the main peak frequency positions are consistent with those obtained from the direct slicing analysis in Figs. 11, 12, 13 and 14, while the spectral structure is significantly simplified, and spurious frequency components are greatly reduced. Specifically, in the F1 condition, the main peak is concentrated around 1650 Hz, indicating that modulation is weak in the fault-free state and that the carrier frequency dominates.In the F2 mode, distinct peaks appear at 60 Hz and 190 Hz, with pronounced low-frequency double peaks, indicating that blade imbalance leads to enhanced modulation while the carrier region is almost suppressed.In the F3 mode, the main peak is located at 250 Hz, showing a clear single modulation peak and a concise spectral structure, indicating a stable modulation period due to wear and effective noise suppression after EMD filtering.In the F4 mode, a double-peak structure is formed at 2177 Hz and 2427 Hz, with significantly enhanced high-frequency resonance and the disappearance of modulation, indicating that structural defects induce localized high-frequency responses. The EMD-cyclic bispectrum secondary slice diagram under four impeller modes. (a) EMD-cyclic bispectral secondary slices -F1, (b) EMD-cyclic bispectral secondary slices -F2, (c) EMD-cyclic bispectral secondary slices -F3, (d) EMD-cyclic bispectral secondary slices -F4. These main frequencies are consistent with the characteristic frequencies identified in the previous cyclic bispectrum analysis, reflecting typical modulation features corresponding to different impeller fault conditions. Therefore, by applying the EMD–cyclic bispectrum secondary slicing method, the frequency with the maximum amplitude in the resulting spectrum can be extracted as a feature for centrifugal pump impeller fault diagnosis. To verify the stability of this method, multiple experiments were repeated for the same fault type. Figure 19 presents the 3D superimposed spectra of three sets of similar signals processed using the EMD–cyclic bispectrum secondary slicing method. 3D superimposed spectra of similar F2 fault signals after EMD–cyclic bispectrum analysis. The results show that the main peak frequencies of the three signals remain consistent (approximately 60 Hz and 190 Hz), consistent with the modulation frequency analysis presented earlier. Slight fluctuations in amplitude and in the non-peak regions reflect the influence of experimental random noise; however, the overall spectral shapes remain consistent, with similar energy distribution patterns. This indicates that the proposed method demonstrates good stability and repeatability across multiple experimental trials. Diagnostic performance verification based on classification model Feature extraction Since each impeller state produces nine-channel vibration signals, nine frequency values corresponding to the highest amplitudes in the EMD–cyclic bispectrum secondary slicing spectrograms can be extracted. Thus, the features of a fault mode can be represented by nine eigenvalues. Using this feature extraction method, 140 training sets and 35 test sets were selected for various impeller states. To evaluate whether the EMD–cyclic bispectrum secondary slicing method improves the extraction of fault characteristics from centrifugal pump impeller vibration signals, the nine vibration signals were processed using cyclic bispectrum secondary slicing. The frequency values corresponding to the highest amplitude points in the nine secondary slice spectrograms were extracted. For each impeller state, 140 training sets and 25 test sets were selected using this method. The EMD method was also applied to identify the IMF component with the highest correlation to the original vibration signal. The frequency corresponding to the maximum amplitude of this component was taken as a feature. Thus, one eigenvalue was obtained for each channel, yielding a total of nine eigenvalues corresponding to an impeller state; similarly, 140 training sets and 35 test sets were constructed. The training and test sets obtained by these three methods served as the datasets for subsequent classification model evaluation. Classification results and performance analysis To comprehensively verify the classification performance of the proposed EMD–cyclic bispectrum secondary slicing method, three classification models—Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and one-dimensional Convolutional Neural Network (1D-CNN)—were used to classify and compare datasets obtained from three feature extraction approaches: cyclic bispectrum secondary slicing, EMD decomposition, and EMD–cyclic bispectrum secondary slicing. First, to minimize randomness in the results of a single model, both the SVM and XGBoost classifiers were applied to the datasets extracted using the three feature extraction methods, and their classification performances were compared. The test set results of the three feature extraction methods using the SVM classifier are shown in Fig. 20. Test results of three feature extraction methods under SVM classification model. (a) Cyclic bispectral secondary slice, (b) EMD decomposition, (c) EMD-cyclic bispectral secondary slice. The results under the XGBoost classification model are shown in Fig. 21 below. Test results of three feature extraction methods under the XGBoost. (a) Cyclic bispectral secondary slice, (b) EMD decomposition. The classification results obtained using the XGBoost model are presented in Fig. 21. The results indicate that both the cyclic bispectrum secondary slicing and EMD decomposition features can identify impeller fault modes across different models; however, their recognition accuracy remains limited. Specifically, the EMD decomposition method performs better in the SVM model (with a test accuracy of 80.71%), while the cyclic bispectrum method achieves higher accuracy in the XGBoost model (78.57%). In contrast, the proposed EMD-cyclic bispectrum secondary slicing method shows a significant improvement in both models, with test accuracies of 85.71% (SVM) and 92.14% (XGBoost), demonstrating its stronger discriminative capability and stability in multi-fault feature recognition.The detailed comparison results for both models are summarized in Table 3. Table 3Recognition results of different feature extraction methods.Classification modelFeature extraction methodTraining set accuracy (%)Test set accuracy (%)SVMCyclic bispectral secondary slice78.5776.43EMD decomposition78.0480.71EMD- cyclic bispectral secondary slice91.9685.71XGBoostCyclic bispectral secondary slice78.0478.57EMD decomposition83.0485EMD- cyclic bispectral secondary slice93.9392.14 Recognition results of different feature extraction methods. To further validate the applicability of the proposed EMD–cyclic bispectrum features across different network architectures, a one-dimensional convolutional neural network (1D-CNN) was used for classification. The results show that the training and testing accuracies reached 99.11% and 94.29%, respectively, reflecting an improvement of approximately 2.15% compared to the XGBoost model. This demonstrates that the deep network can effectively capture the latent high-order structural information embedded in the cyclic bispectrum features. Figure 22 shows the t-SNE feature distribution visualization.Different colors represent distinct fault categories (F1 – Normal, F2 – Blade, F3 – Edge, F4 – Hole). As shown in the figure, the four types of samples form clearly separated clusters in the two-dimensional feature space, indicating that the proposed EMD–cyclic bispectrum features, after being mapped through the 1D-CNN, exhibit excellent separability and inter-class discriminative capability. t-SNE feature distribution visualization. As illustrated in Fig. 23, the overall classification accuracy of the 1D-CNN model based on EMD–cyclic bispectrum features is 95.71%, with recognition rates exceeding 94% for all categories. Minor misclassifications occur between F3 (edge wear) and F4 (hole defect), suggesting that the model maintains high stability and generalization performance even under conditions of high feature similarity. Confusion matrix of 1D-CNN classification results. Comprehensive performance evaluation of the model To comprehensively evaluate the overall performance of the proposed EMD–cyclic bispectrum secondary slicing method, a systematic comparison and analysis were performed from three perspectives: model performance, robustness verification, and computational efficiency. Comparison of model performance and training efficiency Comparison of model performance and training efficiency To further assess the applicability of the proposed method across different classification models, a comparative analysis was conducted between the 1D-CNN and traditional models such as SVM and XGBoost. The results are summarized in Table 4. Table 4Performance comparison of different methods.MethodFeature extractionTraining accuracyTesting accuracyImprovement relative to XGBoostXGBoostEMD- cyclic bispectral secondary slice93.93%92.14%BaselineSVMEMD- cyclic bispectral secondary slice91.25%89.29%-2.85%1D CNNEMD- cyclic bispectral secondary slice95.71%94.29%+ 2.15% Performance comparison of different methods. The results in Table 4 show that the EMD–XGBoost method demonstrates clear practical advantages for engineering applications. As a highly efficient ensemble learning model, XGBoost offers fast training speed, simple parameter tuning, and strong interpretability, making it well-suited for scenarios with limited computational resources or where rapid validation of feature effectiveness is required. In summary, benefiting from the adaptive decomposition capability of EMD and the high-order feature representation of the cyclic bispectrum, integration with the 1D-CNN deep feature learning mechanism enables the proposed cascaded model to more accurately capture modulation features and enhance feature discriminability, significantly improving fault recognition performance. (2)Model robustness and cross-validation analysis Model robustness and cross-validation analysis To reduce uncertainty and the risk of overfitting under small-sample conditions, a five-fold cross-validation approach was adopted to evaluate the statistical reliability of the model. All samples were randomly divided into five subsets; in each iteration, one subset was used as the validation set, while the remaining four were used for training. After five iterations, the average accuracy was calculated as the final performance metric. This method is widely used in rotating machinery fault diagnosis because it allows sufficient interchange between training and testing sets under limited data conditions, thereby enhancing the stability and statistical credibility of the results. Using the enhanced feature set (EMD–cyclic bispectrum secondary slicing) and XGBoost as the classifier, the average accuracy of the five-fold cross-validation was 92.11% ± 1.05%. The small standard deviation indicates that the model is insensitive to sample partitioning and has good generalization capability. Considering the number of samples, feature dimensionality, and information content, this validation approach is sufficient to support the reliability of the study’s conclusions. (3)Comparison of computational performance and recognition accuracy for different feature extraction methods Comparison of computational performance and recognition accuracy for different feature extraction methods To ensure fairness, all algorithms were executed under the same hardware conditions. All experiments were conducted on a Windows 11 Home Edition operating system with an Intel Core i7-8550U processor (1.80 GHz, 4 cores, 8 threads), 16 GB of RAM, and executed in MATLAB R2023b. Each algorithm was run 10 times on the same dataset (sampling points: 2048, sampling frequency: 1000 Hz), and the average value was calculated as the final result. As shown in Table 5, the proposed method exhibits slightly higher runtime and memory usage compared to the traditional cyclic bispectrum approach (approximately 26% longer runtime), but achieves a 4–6% improvement in accuracy. Considering the balance between recognition accuracy, noise robustness, and interpretability, the proposed method strikes an excellent balance between computational cost and performance. Runtime was measured using MATLAB’s built-in tic–toc function, and memory usage was recorded using the memory function. Each method was executed 30 times, and the average values were used to mitigate random fluctuations. Table 5Comparison of computational performance and recognition accuracy of different feature extraction methods.MethodAverage runtime (s)Peak memory (MB)Average accuracy (%)DescriptionEnvelope spectrum (bandpass demodulation)0.160.0689.4Requires manual band selection; fast but noise-sensitiveTraditional cyclic bispectrum (diagonal slicing)0.234.0091.2Carrier interference not fully suppressedEEMD–cyclic bispectrum0.417.8093.0Computationally intensive due to multiple noise realizationsVMD–bispectrum0.365.9592.4Strong parameter dependencyEMD–cyclic bispectrum0.294.0195.7Adaptive decomposition + optimized parameters; best focusing performance Comparison of computational performance and recognition accuracy of different feature extraction methods. Complimentary close This paper proposes a fault feature extraction method for centrifugal pump impellers based on EMD–cyclic bispectrum secondary slicing. Building on the traditional cyclic bispectrum secondary slicing, the proposed method identifies two slicing strategies—anti-angle slicing and tripler-based cyclic frequency slicing—that effectively extract modulation frequency information, as demonstrated through both simulation and experimental analyses. Compared with single signal processing methods, this fusion approach significantly improves the extraction of modulation frequency components and enhances fault diagnosis accuracy. This paper proposes a fault feature extraction method for centrifugal pump impellers based on EMD–cyclic bispectrum secondary slicing. Building on the traditional cyclic bispectrum secondary slicing, the proposed method identifies two slicing strategies—anti-angle slicing and tripler-based cyclic frequency slicing—that effectively extract modulation frequency information, as demonstrated through both simulation and experimental analyses. Compared with single signal processing methods, this fusion approach significantly improves the extraction of modulation frequency components and enhances fault diagnosis accuracy. After optimizing the parameters of cyclic bispectrum secondary slicing, the experimental dynamic signal of the centrifugal pump impeller is decomposed by EMD. The IMF component with the highest correlation is then selected, and cyclic bispectrum secondary slicing with optimized parameters is applied. This process enables more effective extraction of fault features. After optimizing the parameters of cyclic bispectrum secondary slicing, the experimental dynamic signal of the centrifugal pump impeller is decomposed by EMD. The IMF component with the highest correlation is then selected, and cyclic bispectrum secondary slicing with optimized parameters is applied. This process enables more effective extraction of fault features. In the classification performance evaluation, the features extracted by the proposed method demonstrate higher diagnostic accuracy than those derived from individual approaches. Within the proposed fault diagnosis framework, the EMD–cyclic bispectrum features achieved test accuracies of 85.71% and 92.14% under the SVM and XGBoost models, respectively, and further increased to 94.29% when integrated with 1D-CNN. These results indicate that the proposed features exhibit excellent separability and robustness in both shallow and deep learning models. Compared to traditional EEMD, VMD, or envelope spectrum methods, the proposed fusion strategy demonstrates superior recognition accuracy, noise robustness, and interpretability, achieving an effective balance between computational complexity and diagnostic performance. In the classification performance evaluation, the features extracted by the proposed method demonstrate higher diagnostic accuracy than those derived from individual approaches. Within the proposed fault diagnosis framework, the EMD–cyclic bispectrum features achieved test accuracies of 85.71% and 92.14% under the SVM and XGBoost models, respectively, and further increased to 94.29% when integrated with 1D-CNN. These results indicate that the proposed features exhibit excellent separability and robustness in both shallow and deep learning models. Compared to traditional EEMD, VMD, or envelope spectrum methods, the proposed fusion strategy demonstrates superior recognition accuracy, noise robustness, and interpretability, achieving an effective balance between computational complexity and diagnostic performance. It should be noted that although the proposed method can effectively identify modulation features under low SNR and multi-fault conditions, its computational complexity remains relatively high. Further improvements are necessary for handling multi-modulation sources and strongly coupled signals. It should be noted that although the proposed method can effectively identify modulation features under low SNR and multi-fault conditions, its computational complexity remains relatively high. Further improvements are necessary for handling multi-modulation sources and strongly coupled signals. Acknowledgements The authors thank the Reliability Research Lab in the Department of Mechanical Engineering at the University of Alberta, Canada, for providing the original experimental data. This work was supported by the National Natural Science Foundation of China (Grant No. 51775390). Author contributions The algorithms proposed in this paper were conceived by all authors. H.X. Chen designed and performed the experiments and analyzed the results. X.P. Liang performed the experiments, analyzed the simulation results, and wrote the manuscript. L. Wang assisted in data processing and result validation. S.Y. Li contributed to theoretical analysis and manuscript revision. Z.G. Wang and Z.H. Shi provided guidance on experimental design and data interpretation. X.Y. Sheng supervised the overall research and reviewed the final manuscript. All authors read and approved the final version of the paper. Funding The experimental data were obtained from the Laboratory of Reliability at the University of Alberta, Canada. This research was financially supported by the National Natural Science Foundation of China (Grant No. 51775390, Principal Investigator: Hanxin Chen). Data availability All data generated or analysed during this study are included in this published article. Declarations Competing interests The authors declare no competing interests. Footnotes The original online version of this Article was revised: In the original version of this Article Zhaohui Shi was incorrectly listed as a corresponding author. The correct corresponding author for this Article is Hanxin Chen. Correspondence and request for materials should be addressed to pg01074075@ntu.edu.sg. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Change history 1/28/2026 A Correction to this paper has been published: 10.1038/s41598-026-37684-8 References Associated Data Data Availability Statement All data generated or analysed during this study are included in this published article.