A voiceprint-based fault identification method for a spiral case of a hydropower station

Abstract

The application discloses a kind of acoustic fingerprints in the fault identification method of hydroelectric station volute, comprising the following steps: S1, multi-modal acoustic sensing and volute specificity array optimization, for the characteristics of volute solid-liquid coupling acoustic cavity, adopt double-band sensor collaborative layout strategy, establish volute sound propagation attenuation model, optimize sensor coordinates by multi-objective genetic algorithm, realize the acoustic coverage of volute full area;S2, self-adaptive signal enhancement under the hydraulic noise environment, establish cavitation noise feature model, use improved normalized least mean square algorithm to carry out cavitation noise reference cancellation, combined with wavelet packet-blind source separation joint denoising, separate out fault acoustic emission signal.The technical problem to be solved by the application is to provide a kind of acoustic fingerprints in the fault identification method of hydroelectric station volute, realize early warning, accurate diagnosis and intelligent positioning of fault.

Description

Technical Field

[0001] This invention relates to the field of hydropower station fault diagnosis technology, and in particular to a method for fault identification of acoustic signatures in the spiral casing of a hydropower station. Background Technology

[0002] The spiral casing of a hydroelectric power station is a helical flow-through component that encloses the turbine runner. Its function is to uniformly and symmetrically introduce water flow into the guide vanes and runner chamber. As a pressure-bearing boundary, the spiral casing is subjected to enormous hydrostatic pressure, hydrodynamic pressure, and water flow excitation forces over a long period of time. Its structural health directly affects the safe, stable, and economical operation of the entire power station. Traditional methods for spiral casing fault detection and health management mainly rely on the following technologies: 1. Regular overhaul and manual offline inspection: This is the most traditional and common method. It is usually combined with the unit's overhaul cycle (typically every few years) to evacuate and clean the volute, performed by experienced inspection engineers using the following methods: Visual inspection (VT): Check the inner wall for cracks, cavitation, and wear marks.

[0003] Tapping inspection: By listening to the sound of hammering, one can judge whether there is any void or internal defect based on experience. This method is highly subjective.

[0004] Non-destructive testing (NDT) includes (1) Ultrasonic Testing (UT): High-frequency ultrasonic waves are emitted by a piezoelectric probe and the internal cracks, inclusions and other defects are detected by receiving the echoes, and the depth of the defects is quantified. However, it has high requirements for surface roughness and coupling agent, and requires point-by-point scanning, which is extremely inefficient.

[0005] (2) Magnetic particle inspection (MT) / penetrant inspection (PT): only applicable to surface or near-surface defect detection, and cannot detect internal defects.

[0006] (3) Radiographic testing (RT): X-rays or gamma rays penetrate materials and internal defects are displayed through film or digital imaging. This method poses radiation safety risks, is costly, and is extremely difficult to implement on-site.

[0007] 2. Conventional Vibration Monitoring: Vibration acceleration sensors (typically in the frequency range of 0.1Hz~20kHz) are installed on the outer wall of the volute to monitor parameters such as vibration velocity, acceleration, and displacement. Abnormal changes in characteristic frequency components such as rotational speed and blade passage frequency are identified by analyzing the vibration spectrum (e.g., FFT analysis). This method is relatively effective for faults in rotating components (such as bearings and spindles).

[0008] 3. Pressure pulsation monitoring: A pressure sensor is installed inside the volute to monitor the dynamic changes in water pressure inside the volute during operation. This is used to assess hydraulic stability and prevent resonance caused by hydraulic vibrations such as vortex bands and Karman vortices.

[0009] However, the above technologies have the following defects and limitations: 1. It suffers from severe offline and lag issues. Offline detection requires the machine to be shut down, making it impossible to achieve real-time status awareness. The cycle from the emergence of a fault to its discovery during the next major overhaul is long, which can easily develop into a serious accident, violating the core principles of Condition-Based Maintenance (CBM).

[0010] 2. Insufficient sensitivity in early fault detection: Conventional vibration monitoring is only effective for low-frequency, high-energy overall vibrations, and is extremely insensitive to high-frequency, low-energy local micro-damage such as micron-level crack initiation and initial cavitation. Fault signals are easily drowned out by strong background vibrations, leading to missed detections.

[0011] 3. Poor positioning capability and low efficiency. Non-destructive testing such as ultrasonic testing requires knowing the approximate fault area before scanning point by point. The workload of checking the entire volute is huge, and it cannot achieve automatic and real-time positioning of the fault source.

[0012] 4. Over-reliance on expert experience: The accuracy of diagnostic results such as percussion tests and vibration spectrum analysis is highly dependent on the experience level and subjective judgment of the testing personnel, making it difficult to achieve a standardized and intelligent diagnostic process.

[0013] 5. It cannot capture the dynamic development process of the failure. Traditional methods can only provide a static snapshot at a certain point in time. They cannot continuously record the dynamic information of the entire process of crack initiation, expansion and failure. Therefore, it is impossible to carry out failure trend prediction and remaining life assessment. Summary of the Invention

[0014] The technical problem to be solved by the present invention is to provide a method for fault identification of acoustic patterns in the spiral casing of hydropower stations, so as to realize early warning, accurate diagnosis and intelligent location of faults.

[0015] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for fault identification of acoustic signatures in the spiral casing of a hydropower station, comprising the following steps: S1. Multimodal acoustic sensing and volute-specific array optimization: Based on the solid-liquid coupling acoustic cavity characteristics of the volute, a dual-band sensor collaborative layout strategy is adopted to establish a volute acoustic propagation attenuation model. The sensor coordinates are optimized through a multi-objective genetic algorithm to achieve acoustic coverage of the entire volute area. S2. Adaptive signal enhancement under hydraulic noise environment: A cavitation noise feature model is established, and an improved normalized minimum mean square algorithm is used for cavitation noise reference cancellation. Combined with wavelet packet-blind source separation for joint denoising, the fault acoustic emission signal is separated. S3. Extraction and enhancement of acoustic signature features of volute failure: Establish an acoustic energy attenuation compensation model to correct the measurement signal, construct a fault-sensitive feature set containing special features of cavitation, cracks and wear, and extract time-frequency image features by generating acoustic signature spectrum through continuous wavelet transform. S4. Multi-condition adaptive feature selection and fusion: Establish an acoustic feature-condition mapping model to normalize the condition parameters of the original features, select sensitive features for different fault types based on physical mechanisms, and use a one-dimensional convolutional neural network to automatically extract condition-invariant features. S5. Intelligent diagnosis and localization of fused propagation model: Establish a forward model of volute sound propagation, use Bayesian inversion to locate sound sources in layered media, fuse multi-source information of acoustic emission, vibration, and pressure pulsation based on DS evidence theory for fault diagnosis, and construct a digital twin model to achieve predictive maintenance.

[0016] Preferably, in step S1, the dual-band sensor collaborative layout strategy specifically includes: A 30-100kHz low-frequency sensor array was used to monitor continuous acoustic emission signals generated by cavitation and wear, and was deployed in cavitation-prone areas with flow velocities greater than 25m / s. A 100-300kHz high-frequency sensor array is used to capture burst-type acoustic emission signals generated by crack propagation and is placed in stress concentration areas such as welds and structural abrupt changes. Sensor density is set according to regional classification: sensor spacing ≤ 2m in Class A area, sensor spacing ≤ 3m in Class B area, and sensor spacing ≤ 4m in Class C area.

[0017] Preferably, in step S1, the sound propagation attenuation model of the volute is as follows: A(d) = A0 + 20log 10 (d)+αd+βd²; Where: A(d) is the sound pressure level (dB) at a distance d from the sound source, and A0 is the initial sound pressure level (dB) of the sound source. α is the material absorption coefficient (dB / m), in steel α≈0.01-0.1dB / m / MHz, β is the geometric diffusion coefficient (dB / m²), and d is the propagation distance (m); A multi-objective genetic algorithm is used to optimize the sensor coordinates. The objective function is: minimizeF(x)=[f1(x),f2(x),f3(x)]; Where f1(x)=Σ[GDOP(x,y,z)] is the sum of geometric precision factors; f2(x)=Cov_gap is the coverage blind area; f3(x)=ΣC i Total deployment cost.

[0018] Preferably, in step S2, the cavitation noise characteristic model is: N_cav(f)=A·f^(-n)·exp(-f / f_c)+B·f^(-m); Where: A is the cavitation intensity coefficient, which is related to the cavitation number σ; n is the spectral attenuation index (1.5-2.5); f_c is the cutoff frequency (50-100kHz); B is the turbulence noise figure; m is the turbulence spectral index (5 / 3). An improved Normalized Least Mean Square (NLMS) algorithm is employed: w(n+1)=w(n)+μ·e(n)·x(n) / (||x(n)||²+δ); e(n)=d(n)-w (n)x(n); Where: w(n) is the filter weight vector at step n; μ is the step size factor (0 < μ < 2); δ is the regularization parameter (to prevent division by zero); x(n) is the reference noise input vector; and d(n) is the main sensor signal.

[0019] Preferably, in step S2, the wavelet packet-blind source separation joint denoising specifically involves using independent component analysis to separate the fault signal, based on the characteristics that turbulent noise exhibits a Gaussian distribution and strong spatial correlation, while fault signals exhibit a non-Gaussian distribution and strong locality. The objective function is: X=A·S;J(W)=ΣE[G(w x)]-log|detW| Where: X is the observed signal matrix; A is the mixing matrix; S is the source signal matrix; G(·) is a non-quadratic function (commonly tanh or cubic function); J(W) is the objective function, which measures non-Gaussianity.

[0020] Preferably, in step S3, the sound energy attenuation compensation model is: E_corrected=E_measured·exp(η·d)·d η=η_material+η_geometry+η_coupling Where: η_material is the material absorption coefficient, which is related to the frequency f by η_material∝f; η_geometry is the geometric diffusion coefficient, which is related to the radius of curvature R of the volute η_geometry=k / R (when R>λ); η_coupling is the coupling loss coefficient, which is related to the surface roughness; d is the propagation distance.

[0021] Preferably, in step S3, the fault-sensitive feature set includes: Cavitation characteristics: Cavitation intensity index CII=Σ(f·PSD(f)) / ΣPSD(f), f∈[20,100]kHz; Bubble collapse density BD=N_events / T_interval; Cavitation development rate CDR=d(CII) / dt; Crack characteristics: Abrupt change index SI = ( A / t)_max / A_rms; Frequency centroid offset ΔFC=FC_current-FC_baseline; Acoustic emission b value b=log 10 (N) / log 10 (M), where N is the number of events and M is the amplitude; Wear characteristics: Energy flow trend slope ETS=Δ(RMS) / Δt; Spectral roughness SR=Σ|PSD(f)-PSD_smoothed(f)|; Wear index WI=∫(PSD(f)·W(f))df, where W(f) is the weighting function.

[0022] Preferably, in step S4, the acoustic feature-operating condition mapping model is: F_normalized=F_raw / Φ(Q,H,P,T); Where F_raw is the original feature, Φ(Q,H,P,T) is the influence function, Q is the flow rate, H is the head, P is the power, and T is the temperature; The influence function Φ was obtained by fitting experimental data: Φ(Q,H,P,T)=a·Q^α+b·H^β+c·P^γ+d·T^δ+ε; Where a, b, c, and d are coefficients, α, β, γ, and δ are exponents, and ε is the error term.

[0023] Preferably, in step S5, the forward model for volute acoustic propagation is: T_calculated=Σ(L_i / v_i)+Σ(Δt_interface)+Σ(Δt_curvature) Where: L_i is the propagation path length in each medium; v_i is the corresponding wave velocity (related to frequency v(f)=v0+k·f); Δt_interface is the interface transition time delay; Δt_curvature is the curvature effect time delay.

[0024] Preferably, in step S5, the digital twin model comprises three levels: Physical model: The natural frequencies and modes of the volute are calculated through finite element acoustic simulation; Data-driven model: Using long short-term memory networks to predict fault development trends; Mechanism model: A fatigue crack propagation model is established based on the Paris formula.

[0025] This invention provides a method for fault identification of acoustic signatures in the spiral casing of a hydroelectric power station, which has the following beneficial effects: 1. It abandons the reliance on shutdown and venting for detection during major unit overhauls. It can continuously and in real-time monitor the propagation of cracks, cavitation, and wear inside the volute during normal unit operation. This avoids safety hazards from the onset of a fault to the next major overhaul, achieving full lifecycle health management.

[0026] 2. For high-frequency, low-energy signals generated by micron-level crack initiation or early cavitation, this invention utilizes a sensor layout within a specific frequency band to effectively capture these needle-in-a-haystack weak acoustic emission signals. Through an adaptive signal enhancement process in a hydraulically noisy environment, background noise such as water turbulence and mechanical vibration is effectively filtered out, significantly improving the signal-to-noise ratio and solving the problem of conventional vibration monitoring failing to identify early localized damage.

[0027] 3. Unlike traditional ultrasonic testing, which relies on a piecemeal, point-by-point scanning approach, this invention utilizes a multi-sensor array and a layered medium sound source localization algorithm to quickly pinpoint the specific coordinates of a fault based on the propagation time difference of sound waves in a steel-water composite medium. By introducing expert systems and digital twin technology, and extracting specific acoustic signature features, combined with DS evidence theory and fusion of vibration and pressure pulsation data, it can automatically distinguish between different fault types such as cracks, cavitation, and wear, reducing reliance on human experience. Attached Figure Description

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0029] like Figure 1 As shown, a method for fault identification of acoustic signatures in the spiral casing of a hydropower station includes the following steps: S1. Multimodal acoustic sensing and volute-specific array optimization: Based on the solid-liquid coupling acoustic cavity characteristics of the volute, a dual-band sensor collaborative layout strategy is adopted to establish a volute acoustic propagation attenuation model. The sensor coordinates are optimized through a multi-objective genetic algorithm to achieve acoustic coverage of the entire volute area.

[0030] As a large pressure vessel, the spiral casing of a hydroelectric power station exhibits unique acoustic propagation characteristics. The casing is filled with high-speed water flow, while the exterior is composed of thick steel plates, forming a solid-liquid coupled acoustic cavity. Sound waves propagate at approximately 5000-5900 m / s (longitudinal waves) in the steel plates, but only about 1480 m / s in the water. This difference in medium results in a complex sound wave propagation path, involving multiple reflections and mode conversions.

[0031] The spiral structure of the volute imparts directional characteristics to sound wave propagation, resulting in significant differences in acoustic sensitivity at different locations. Furthermore, the uneven distribution of water velocity inside the volute, with a distinct velocity gradient from inlet to outlet, further complicates acoustic monitoring.

[0032] Firstly, for the design of the composite medium sensor array, a dual-band sensor collaborative layout strategy is adopted: A low-frequency sensor array (30-100kHz) primarily monitors continuous acoustic emission signals generated by cavitation and wear, and is placed in areas prone to cavitation (flow velocity > 25m / s); a high-frequency sensor array (100-300kHz) specifically captures sudden acoustic emission signals such as crack propagation, and is placed in stress concentration areas (such as welds and structural abrupt changes). The sensor density is graded according to the importance of the area: Grade A areas (nose, transition section): sensor spacing ≤ 2m; Grade B areas (main body): sensor spacing ≤ 3m; Grade C areas (normal areas): sensor spacing ≤ 4m.

[0033] Then, array optimization based on acoustic topology is carried out. A volute acoustic propagation attenuation model is established: A(d) = A0 + 20log 10 (d)+αd+βd² Where: A(d): sound pressure level at a distance d from the sound source (dB), A0: initial sound pressure level at the sound source (dB), α: Material absorption coefficient (dB / m), α≈0.01-0.1dB / m / MHz in steel, β: Geometric diffusion coefficient (dB / m²), d: Propagation distance (m).

[0034] A multi-objective genetic algorithm is used to optimize the sensor coordinates. The objective function is: minimizeF(x)=[f1(x),f2(x),f3(x)] Where, f1(x)=Σ[GDOP(x,y,z)]: sum of geometric precision factors; f2(x)=Cov_gap: coverage blind area; f3(x)=ΣC i Total deployment cost The model considers the unique characteristics of hydropower station spiral casing fault identification scenarios: Reflection characteristics of the medium interface: The reflection coefficient of the solid-liquid interface R = (Z2 - Z1) / (Z2 + Z1), where Z is the acoustic impedance (Z_water ≈ 1.5 × 10⁻⁶). 6 Pa·s / m, Z_steel≈47×10 6 Pa·s / m); Volute curvature guiding effect: the waveguide characteristic of sound waves propagating along the spiral of the volute. The waveguide effect is obvious when the relationship between the radius of curvature R and the wavelength λ satisfies R / λ>5. Hydraulic load acoustic coupling: The effect of water pressure on the sound wave propagation speed Δv / v=k·P, where k is the pressure coefficient (k≈0.02-0.05GPa in steel). - ¹).

[0035] S2. Adaptive signal enhancement under hydraulic noise environment: A cavitation noise characteristic model is established, and an improved normalized minimum mean square algorithm is used for cavitation noise reference cancellation. Combined with wavelet packet-blind source separation for joint denoising, the fault acoustic emission signal is separated.

[0036] The internal noise sources of the volute are complex and diverse, mainly including: turbulent boundary layer noise: broadband continuous spectrum, frequency range 20Hz-20kHz; cavitation bubble collapse noise: pulsed, with main energy concentrated in 20-100kHz; rotor blade frequency modulation noise: line spectrum characteristics, frequency f=n×Z×RPM / 60; and mechanical vibration transmission noise: low-frequency noise transmitted through the structure. These noises largely overlap with the fault acoustic emission signal in the frequency domain, especially in the 50-150kHz frequency band, where cavitation noise has similar spectral characteristics to early fault signals, making them difficult to separate.

[0037] First, based on the cavitation noise reference cancellation technique, a cavitation noise characteristic model is established: N_cav(f)=A·f^(-n)·exp(-f / f_c)+B·f^(-m) Where: A: cavitation intensity coefficient, which is related to the cavitation number σ; n: spectral attenuation index (1.5-2.5); f_c: cutoff frequency (50-100kHz); B: turbulence noise figure; m: turbulence spectral index (5 / 3).

[0038] An improved Normalized Least Mean Square (NLMS) algorithm is employed: w(n+1)=w(n)+μ·e(n)·x(n) / (||x(n)||²+δ) e(n)=d(n)-w (n)x(n) Where: w(n): filter weight vector at step n; μ: step size factor (0 < μ < 2); δ: regularization parameter (to prevent division by zero); x(n): reference noise input vector; d(n): main sensor signal.

[0039] Then, wavelet packet-blind source separation is used for joint denoising. Considering the spatiotemporal characteristics of the volute noise: turbulent noise has a Gaussian distribution and strong spatial correlation; fault signals have a non-Gaussian distribution and strong locality. Independent component analysis (ICA) is used to separate the fault signals. X=A·S;J(W)=ΣE[G(w x)]-log|detW| Where: X: Observation signal matrix; A: Mixing matrix; S: Source signal matrix; G(·): Non-quadratic function (commonly tanh or cubic function); J(W): Objective function, measuring non-Gaussianity.

[0040] The model considers the unique characteristics of hydropower station spiral casing fault identification scenarios: Hydraulic noise frequency band characteristics: low frequency dominance (<50kHz), energy flux density is proportional to flow velocity v³.5; cavitation pulse periodicity: time modulation characteristics related to operating conditions, pulse repetition frequency f_p=k·v·√(ΔP / ρ); fluid-structure coupled vibration mode: coupling characteristics of structural natural frequency and water flow excitation, satisfying f_n=(β_n / 2π)·√(EI / ρA).

[0041] S3. Extraction and enhancement of acoustic signature features for volute failures: Establish an acoustic energy attenuation compensation model to correct the measured signal, construct a fault-sensitive feature set containing special features of cavitation, cracks, and wear, and extract time-frequency image features by generating acoustic signature spectra through continuous wavelet transform.

[0042] Common fault types in hydropower station spiral casings include cracks, cavitation, wear, and loosening. Each fault produces a unique acoustic signal: Crack propagation: generates a high-frequency burst acoustic emission signal with a short rise time (microseconds) and concentrated energy; Cavitation damage: a mid-frequency continuous signal with obvious amplitude modulation, strongly correlated with operating conditions; Sediment wear: a low-frequency continuous signal that changes slowly and is positively correlated with sediment concentration and flow velocity; Component loosening: an impact signal with periodic repetition, related to rotational speed harmonics.

[0043] First, conduct propagation attenuation compensation characteristic correction and establish a sound energy attenuation compensation model: E_corrected=E_measured·exp(η·d)·d η=η_material+η_geometry+η_coupling Where: η_material: material absorption coefficient, which is related to the frequency f by η_material∝f; η_geometry: geometric diffusion coefficient, which is related to the radius of curvature R of the volute η_geometry=k / R (when R>λ); η_coupling: coupling loss coefficient, which is related to the surface roughness.

[0044] Based on this, a fault-sensitive feature set was constructed. Corresponding dedicated features were designed for different fault types: Cavitation characteristics: Cavitation intensity index CII=Σ(f·PSD(f)) / ΣPSD(f), f∈[20,100]kHz; Bubble collapse density BD=N_events / T_interval; Cavitation development rate CDR=d(CII) / dt.

[0045] Crack characteristics: Abrupt change index SI = ( A / t)_max / A_rms; Frequency centroid offset ΔFC=FC_current-FC_baseline; Acoustic emission b value b=log 10 (N) / log 10 (M), where N is the number of events and M is the amplitude; Wear characteristics: Energy flow trend slope ETS=Δ(RMS) / Δt; Spectral roughness SR=Σ|PSD(f)-PSD_smoothed(f)|; Wear index WI=∫(PSD(f)·W(f))df, where W(f) is the weighting function.

[0046] Then, time-frequency image feature extraction is performed, and continuous wavelet transform is used to generate a speaker graph: CWT(a,b)=(1 / √|a|)∫s(t)ψ*((tb) / a)dt Morlet wavelet is chosen as the basis function: ψ(t)=π^(-1 / 4)e^(iω0t)e^(-t² / 2) Its frequency resolution is suitable for volute acoustic emission analysis, and it has good time-frequency localization characteristics.

[0047] The model considers the unique characteristics of hydropower station spiral casing fault identification scenarios: Hydraulic load modulation characteristics: the correlation between acoustic emission amplitude and head H is A∝H^m (m≈1.5-2.0); material fatigue acoustic memory effect: the correlation between cumulative acoustic emission count and stress history is ΣN=C·(Δσ)^n; multi-fault coupling intermodulation characteristics: the nonlinear interaction of acoustic emission signals of different fault types, satisfying y(t)=∫h(τ)x(t-τ)dτ+∫∫h(τ1,τ2)x(t-τ1)x(t-τ2)dτ1dτ2+...

[0048] S4. Multi-condition adaptive feature selection and fusion: Establish an acoustic feature-condition mapping model to normalize the condition parameters of the original features, select sensitive features for different fault types based on physical mechanisms, and use a one-dimensional convolutional neural network to automatically extract condition-invariant features.

[0049] Hydropower stations operate under complex and variable conditions. Key operating parameters include: head variation (50-200m, affecting flow regime and pressure distribution); load regulation (25-100% rated load, affecting flow rate and rotational speed); operating modes (generation, phase regulation, pumping, etc.); and transient processes (start-up, shutdown, sudden load changes, etc.). These variations in operating conditions lead to significant changes in acoustic characteristics, necessitating the establishment of a mapping relationship between these characteristics and operating parameters to achieve adaptive selection and normalization of these characteristics.

[0050] First, normalize the operating parameters. Establish a sound feature-operating condition mapping model: F_normalized=F_raw / Φ(Q,H,P,T) The influence function Φ was obtained by fitting experimental data: Φ(Q,H,P,T)=a·Q^α+b·H^β+c·P^γ+d·T^δ+ε The parameters a, b, c, d and the exponents α, β, γ, δ are determined by multiple nonlinear regression.

[0051] Then, feature selection based on physical mechanisms is carried out, and sensitive features are selected according to the physical mechanisms of different faults: Cavitation-related characteristics: strongly correlated with cavitation number σ=(P-P_v) / (0.5ρV²), choose CII_normalized=CII / σ^k (k≈0.5-0.7).

[0052] Crack-related characteristics: related to stress intensity factor ΔK=YΔσ√(πa), choose SI_normalized=SI / ΔK^m (m≈1.0-1.2) Wear-related characteristics: correlated with the product of sediment concentration C and flow velocity V; choose WI_normalized=WI / (C·V)^n (n≈0.8-1.0).

[0053] Based on this, we conduct adaptive feature extraction using deep learning. We employ a 1D-CNN network structure to automatically learn condition-invariant features.

[0054] The model considers the unique characteristics of hydropower station spiral casing fault identification scenarios: condition-transition invariant features: a subset of features that maintain stability under different operating conditions, satisfying... F / P≈0, F / Q≈0; Transient characteristics of the transition process: unique acoustic emission mode during start-up / shutdown, time constant τ=J·ω / M (J is the moment of inertia, M is the torque); Load-related modulation characteristics: linear / nonlinear relationship between acoustic emission amplitude and unit load A∝L^k (k≈0.8-1.2).

[0055] S5. Intelligent diagnosis and localization of fused propagation model: Establish a forward model of volute sound propagation, use Bayesian inversion to locate sound sources in layered media, fuse multi-source information of acoustic emission, vibration, and pressure pulsation based on DS evidence theory for fault diagnosis, and construct a digital twin model to achieve predictive maintenance.

[0056] The complex structure of the volute results in multiple propagation modes for sound waves: longitudinal waves in the steel plate (velocity approximately 5900 m / s, with relatively low attenuation); transverse waves in the steel plate (velocity approximately 3200 m / s, with relatively high attenuation); compression waves in the water (velocity approximately 1480 m / s, with moderate attenuation); and solid-liquid interface transition waves (complex mode transitions and significant energy loss). This multi-mode propagation characteristic makes source localization difficult, necessitating the development of an accurate propagation model for time-difference correction.

[0057] First, based on the layered medium sound source localization algorithm, a forward model of volute sound propagation is established: T_calculated=Σ(L_i / v_i)+Σ(Δt_interface)+Σ(Δt_curvature) Where: L_i: propagation path length in each medium; v_i: corresponding wave velocity (related to frequency v(f)=v0+k·f); Δt_interface: interface transition time delay; Δt_curvature: curvature effect time delay.

[0058] Location is achieved using Bayesian inversion: p(θ|D)∝p(D|θ)p(θ); θ=(x,y,z,TOA,v) The prior probability p(θ) contains structural constraint information.

[0059] Then, multimodal information fusion diagnosis is conducted, integrating acoustic emission, vibration, and pressure pulsation information from multiple sources: Acoustic emission provides microscopic damage information with high sensitivity; vibration reflects the overall structural state with good reliability; pressure pulsation characterizes hydraulic excitation and is strongly correlated with operating conditions. This fusion is based on DS evidence theory. m(C)=Σ(m1(A)·m2(B)) / (1-K),A∩B=C K=Σ(m1(A)·m2(B)),A∩B=

[0060] Where K represents the degree of conflict of evidence.

[0061] Based on this, a digital twin-driven predictive maintenance is proposed, and a volute acoustic digital twin model is established, which includes three levels: physical model: finite element acoustic simulation, calculating natural frequencies and modes [K-ω²M]Φ=0; data-driven: deep learning fault prediction, LSTM network predicting trends y(t+Δt)=f(y(t),y(t-1),...,u(t)); mechanism model: fatigue crack propagation model, Paris formula da / dN=C(ΔK)^m.

[0062] The model considers the unique characteristics of hydropower station spiral casing fault identification scenarios: 3D surface localization correction: considering the localization algorithm correction for the hyperbolic surface of the spiral casing, curvature tensor C=[κ1,κ2]; multipath propagation identification: utilizing the time difference of the first wave and reflected wave to invert the propagation path Δt_{ij}=(L_i-L_j) / v; acoustic emission b-value analysis: through amplitude distribution b-value=log 10 (N) / log 10 (M) Assess the severity of damage; sound velocity temperature compensation: v_corrected=v_measured·[1-α(T-T_ref)], α≈0.0001 / ℃ in steel.

[0063] This invention systematically proposes a five-step method for diagnosing hydropower station spiral casing faults based on voiceprint recognition technology. Each step closely integrates the unique acoustic environment, structural characteristics, and operating conditions of the hydropower station spiral casing. Through theoretical analysis, algorithm design, and experimental verification, the effectiveness and practicality of the method have been demonstrated. This technical system can be widely applied to the health monitoring of spiral casings in various types of hydropower stations, and is particularly suitable for: large mixed-flow turbine generator units, high-head impulse turbine generator units, pump-turbine turbines in pumped storage power stations, and the retrofitting and life extension assessment of aging units. The invention provides effective technical support for the safe operation of hydropower station spiral casings, and is of great significance for promoting condition-based maintenance and intelligent operation and maintenance in the hydropower industry. It is expected to improve equipment availability by 3-5% and reduce maintenance costs by 20-30%, resulting in significant economic and social benefits.

[0064] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for fault identification using acoustic signatures in the spiral casing of a hydropower station, characterized in that, Includes the following steps: S1. Multimodal acoustic sensing and volute-specific array optimization: Based on the solid-liquid coupling acoustic cavity characteristics of the volute, a dual-band sensor collaborative layout strategy is adopted to establish a volute acoustic propagation attenuation model. The sensor coordinates are optimized through a multi-objective genetic algorithm to achieve acoustic coverage of the entire volute area. S2. Adaptive signal enhancement under hydraulic noise environment: A cavitation noise feature model is established, and an improved normalized minimum mean square algorithm is used for cavitation noise reference cancellation. Combined with wavelet packet-blind source separation for joint denoising, the fault acoustic emission signal is separated. S3. Extraction and enhancement of acoustic signature features of volute failure: Establish an acoustic energy attenuation compensation model to correct the measurement signal, construct a fault-sensitive feature set containing special features of cavitation, cracks and wear, and extract time-frequency image features by generating acoustic signature spectrum through continuous wavelet transform. S4. Multi-condition adaptive feature selection and fusion: Establish an acoustic feature-condition mapping model to normalize the condition parameters of the original features, select sensitive features for different fault types based on physical mechanisms, and use a one-dimensional convolutional neural network to automatically extract condition-invariant features. S5. Intelligent diagnosis and localization of fused propagation model: Establish a forward model of volute sound propagation, use Bayesian inversion to locate sound sources in layered media, fuse multi-source information of acoustic emission, vibration, and pressure pulsation based on DS evidence theory for fault diagnosis, and construct a digital twin model to achieve predictive maintenance.

2. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S1, the dual-band sensor collaborative layout strategy is specifically as follows: A 30-100kHz low-frequency sensor array was used to monitor continuous acoustic emission signals generated by cavitation and wear, and was deployed in cavitation-prone areas with flow velocities greater than 25m / s. A 100-300kHz high-frequency sensor array is used to capture burst-type acoustic emission signals generated by crack propagation and is placed in stress concentration areas such as welds and structural abrupt changes. Sensor density is set according to regional classification: sensor spacing ≤ 2m in Class A area, sensor spacing ≤ 3m in Class B area, and sensor spacing ≤ 4m in Class C area.

3. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S1, the sound propagation attenuation model of the volute is as follows: A(d)=A0+20log 10 (d)+αd+βd²; Where: A(d) is the sound pressure level at a distance d from the sound source, and A0 is the initial sound pressure level of the sound source. α is the material absorption coefficient, in steel α≈0.01-0.1dB / m / MHz, β is the geometric diffusion coefficient, and d is the propagation distance; A multi-objective genetic algorithm is used to optimize the sensor coordinates. The objective function is: minimizeF(x)=[f1(x),f2(x),f3(x)]; Where f1(x)=Σ[GDOP(x,y,z)] is the sum of geometric precision factors; f2(x)=Cov_gap is the coverage blind area; f3(x)=ΣC i Total deployment cost.

4. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S2, the cavitation noise characteristic model is as follows: N_cav(f)=A·f^(-n)·exp(-f / f_c)+B·f^(-m); Where: A is the cavitation intensity coefficient, which is related to the cavitation number σ; n is the spectral attenuation index; f_c is the cutoff frequency; B is the turbulence noise figure; m is the turbulence spectral index; An improved Normalized Least Mean Square (NLMS) algorithm is employed: w(n+1)=w(n)+μ·e(n)·x(n) / (||x(n)||²+δ); e(n)=d(n)-w (n)x(n); Where: w(n) is the filter weight vector at step n; μ is the step size factor; δ is the regularization parameter; x(n) is the reference noise input vector; and d(n) is the main sensor signal.

5. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S2, the wavelet packet-blind source separation joint denoising specifically involves using independent component analysis to separate the fault signal, based on the characteristics that turbulent noise exhibits a Gaussian distribution and strong spatial correlation, while fault signals exhibit a non-Gaussian distribution and strong locality. The objective function is: X=A·S;J(W)=ΣE[G(w x)]-log|detW|; Where: X is the observed signal matrix; A is the mixing matrix; S is the source signal matrix; G(·) is a non-quadratic function; J(W) is the objective function, which measures non-Gaussianity.

6. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S3, the sound energy attenuation compensation model is as follows: E_corrected=E_measured·exp(η·d)·d; η=η_material+η_geometry+η_coupling; Where: η_material is the material absorption coefficient, which is related to the frequency f by η_material∝f; η_geometry is the geometric diffusion coefficient, which is related to the radius of curvature R of the volute η_geometry=k / R; η_coupling is the coupling loss coefficient, which is related to the surface roughness; and d is the propagation distance.

7. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S3, the fault-sensitive feature set includes: Cavitation characteristics: Cavitation intensity index CII=Σ(f·PSD(f)) / ΣPSD(f), f∈[20,100]kHz; Bubble collapse density BD=N_events / T_interval; Cavitation development rate CDR=d(CII) / dt; Crack characteristics: Abrupt change index SI = ( A / t)_max / A_rms; Frequency centroid offset ΔFC=FC_current-FC_baseline; Acoustic emission b value b=log 10 (N) / log 10 (M), where N is the number of events and M is the amplitude; Wear characteristics: Energy flow trend slope ETS=Δ(RMS) / Δt; Spectral roughness SR=Σ|PSD(f)-PSD_smoothed(f)|; Wear index WI=∫(PSD(f)·W(f))df, where W(f) is the weighting function.

8. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S4, the acoustic feature-operating condition mapping model is as follows: F_normalized=F_raw / Φ(Q,H,P,T); Where F_raw is the original feature, Φ(Q,H,P,T) is the influence function, Q is the flow rate, H is the head, P is the power, and T is the temperature; The influence function Φ was obtained by fitting experimental data: Φ(Q,H,P,T)=a·Q^α+b·H^β+c·P^γ+d·T^δ+ε; Where a, b, c, and d are coefficients, α, β, γ, and δ are exponents, and ε is the error term.

9. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S5, the forward model for acoustic propagation in the volute is as follows: T_calculated=Σ(L_i / v_i)+Σ(Δt_interface)+Σ(Δt_curvature); Where: L_i is the propagation path length in each medium; v_i is the corresponding wave velocity; Δt_interface is the interface transition time delay; Δt_curvature is the curvature effect time delay.

10. The method for fault identification of acoustic signature in the spiral casing of a hydropower station according to claim 1, characterized in that, In step S5, the digital twin model comprises three levels: Physical model: The natural frequencies and modes of the volute are calculated through finite element acoustic simulation; Data-driven model: Using long short-term memory networks to predict fault development trends; Mechanism model: A fatigue crack propagation model is established based on the Paris formula.

Images