Hydropower station mixed-flow unit runaway critical fault voiceprint advanced recognition and early warning method

Abstract

The invention discloses a voiceprint advanced recognition and early warning method for runaway critical faults of a mixed-flow unit of a hydropower station, and aims to solve the problems of lag, low sensitivity, difficulty in positioning and poor universal voiceprint adaptation of the existing monitoring technology. According to the method, a high-frequency acoustic emission AE and broadband sound pressure composite sensing array is arranged, a key phase sensor rotating speed pulse signal is embedded, and an equal-angle increment sampling mode is switched in a self-adaptive mode according to the rotating speed change rate; a sound signal is converted from a time domain to an angular domain, and a fault signal is effectively enhanced in combination with an order domain blind source separation technology; sensitive features such as rotating speed normalized sound energy and transient impact repetition frequency entropy TIRFE are extracted; simulation data are generated by means of a digital twinning technology to pre-train a one-dimensional convolutional neural network, a model is finely adjusted and optimized through measured data, grading early warning is achieved by combining a dynamic threshold value, faults such as the blade early cracks are accurately positioned and early recognized through a TDOA technology, positioning is accurate, the anti-interference capacity is high, the small sample problem is effectively solved, and the method is suitable for large-scale popularization and application. Safe and stable operation of the unit is ensured.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and monitoring technology for hydropower station equipment, and in particular to a method for early identification and warning of runaway critical faults of mixed-flow turbine units in hydropower stations using acoustic signatures. Background Technology

[0002] In hydropower station operation, the turbine-generator unit, as the core equipment, directly affects the continuity and reliability of power supply due to its safety and stability. Especially under extreme conditions, such as runaway, the challenges faced by the unit are particularly severe. Runaway refers to the extreme situation where, at maximum head, the load is suddenly completely dropped, and the speed control system malfunctions, preventing the guide vanes from closing. This causes the unit's speed to rise sharply and stabilize at an extremely high speed. In this case, the runaway speed can reach 1.6 to 2.4 times the rated speed, posing a fatal threat to the unit's structural strength, bearing stability, and the dynamic load on the runner blades and volute.

[0003] Currently, the monitoring and diagnosis of runaway conditions in hydro-generator units mainly rely on speed sensors and vibration sensors. These sensors assess the unit's operating status by capturing changes in the unit's speed and vibration signals. Specifically, speed sensors monitor the unit's real-time speed, while vibration sensors capture vibration signals generated during operation. By analyzing parameters such as the frequency and amplitude of the vibration signals, it is determined whether there are any faults or abnormalities in the unit.

[0004] However, existing monitoring technologies have significant limitations, mainly in the following aspects: 1. Delay: By the time vibration signals propagate and manifest within the structure, the fault has often already progressed to a certain stage, making "advanced" early warning impossible. This means that under current technology, managers can only take countermeasures after a fault occurs or when it has significantly worsened, increasing the risk of unit damage and maintenance costs. For example, under runaway conditions, early faults such as the initiation and propagation of microcracks in the blades generate high-frequency, low-energy acoustic emission signals. However, existing sensors are not sensitive to these signals and often only issue an alarm after the fault has developed to a certain extent and caused macroscopic vibration changes. This prevents operators from taking timely measures in the early stages of a fault, potentially leading to further deterioration and increasing the risk of unit damage and maintenance costs.

[0005] 2. Insufficient Sensitivity: Conventional vibration sensors (frequency range <20kHz) are insensitive to acoustic emission events such as early crack initiation and cavitation intensification at high frequencies and low energy levels. Although these early fault signals have low energy, they are often precursors to failure. Due to the limitations of existing sensors, this important information is often overlooked, leading to the inability to prevent the further development of the fault in time. For example, during the flyaway transition, the coupling between the dynamic excitation frequency and the structure's natural frequency can cause problems such as strong vibration, crack initiation, and even structural damage. However, existing speed and vibration sensors are insensitive to acoustic emission events such as early crack initiation and cavitation intensification at high frequencies and low energy levels, making it difficult to accurately identify these fault characteristics, thus reducing the accuracy of fault identification.

[0006] 3. Inability to pinpoint the exact component where the fault occurred: Existing monitoring technologies struggle to accurately determine which blade developed the crack. In complex turbine structures, faults can occur in any component. However, current technologies often only provide overall turbine status information, failing to pinpoint the fault precisely, increasing maintenance difficulty and time costs. For instance, when a turbine blade cracks, existing sensors cannot accurately identify which blade is faulty. This makes subsequent maintenance and troubleshooting lack focus, increasing repair time and costs, and reducing turbine operating efficiency and safety.

[0007] 4. Small Sample Size Problem: In practical engineering, the complex operating environment of mixed-flow turbine units in hydropower stations leads to uncertainties and randomness in fault occurrence, resulting in a limited number of on-site fault samples. Existing deep learning-based fault diagnosis methods often require a large number of fault samples for model training, which greatly limits the application of existing technologies and makes it difficult to adapt to different models of mixed-flow turbine units. For example, for newly commissioned or technically upgraded units, the lack of sufficient fault samples makes it difficult to effectively train existing deep learning models, thus affecting the accuracy and reliability of fault diagnosis.

[0008] To address the aforementioned issues, although some improved monitoring and early warning methods have been proposed, such as the fault early warning system and method based on hydroelectric generator acoustic signature recognition disclosed in CN115539277A, the acoustic fiber optic fault identification method, system, equipment, and medium for hydroelectric generator sets disclosed in CN117932525A, and the acoustic signature monitoring and diagnosis method for hydroelectric generator sets disclosed in CN119860313A, these methods still have their own limitations. For example, although the fault early warning system and method based on hydroelectric generator acoustic signature recognition utilizes acoustic signature recognition technology, it may still be affected by strong noise background interference when facing extreme conditions such as runaway operation; although the acoustic fiber optic fault identification method has high sensitivity and anti-interference ability, its deployment and maintenance costs are high; and although the acoustic signature monitoring and diagnosis method for hydroelectric generator sets combines multiple signal processing technologies, it may still face challenges in feature extraction and classification when dealing with complex and ever-changing fault modes.

[0009] Therefore, this invention proposes a method for early identification and warning of runaway critical faults in mixed-flow hydropower units based on voiceprint recognition technology, aiming to overcome the shortcomings of existing technologies and achieve early identification and accurate location of early faults in the units. Summary of the Invention

[0010] The technical problem to be solved by this invention is to provide a method for early identification and warning of runaway critical faults of mixed-flow turbine units in hydropower stations using acoustic signatures. This method addresses the technical problems of lag, insufficient sensitivity, inability to locate faults, and poor adaptability of general acoustic signature recognition technologies in existing technologies. It enables accurate identification, location, and early warning of early faults such as blade cracks and severe cavitation, thereby ensuring the structural safety and stable operation of the unit.

[0011] To achieve the above technical objectives, the present invention adopts the following technical solution: The acoustic signature-based early warning method for runaway critical faults in mixed-flow turbine units of hydropower stations achieves early identification and early warning of runaway critical faults through four core steps: "acoustic sensor array construction - acoustic signature signal preprocessing and enhancement - fault feature extraction - intelligent diagnosis and early warning". The specific technical solutions for each step are as follows: Step 1: Deployment of acoustic sensor arrays and synchronous signal acquisition for the fly-out transition process The core of this step is to construct an acoustic sensing system adapted to the time-varying characteristics of rotational speed, achieving synchronous and distortion-free acquisition of wideband acoustic signature signals. Key technologies include: Step 1.1: Design and Deployment of Composite Acoustic Sensor Array To address the frequency characteristics of the fault signal during the runaway critical process (coexistence of high-frequency acoustic emission signal and mid-to-low frequency continuous noise), a composite array consisting of a high-frequency acoustic emission sensor and a broadband acoustic pressure sensor is employed. Sensor selection: High-frequency acoustic emission (AE) sensors have a main frequency range of 100kHz~300kHz and are used to capture high-frequency burst acoustic signals such as blade crack propagation and micro-damage; broadband acoustic pressure sensors have a frequency range of 20Hz~100kHz and are used to monitor mid-to-low frequency continuous noise such as cavitation vortex bands and water flow impact, providing data support for noise separation. Deployment location: Sensors are deployed in key acoustic propagation paths and high-fault areas of the unit, including the inlet of the volute, the 45° section of the volute, the nose of the volute, and the area of ​​the turbine top cover near the guide vanes and the runner, to ensure coverage of the entire circumference of the runner blades and fault-sensitive parts such as key sections of the volute.

[0012] Step 1.2: Speed ​​pulse signal embedding and synchronization To achieve rotational speed phase synchronization of all acoustic signals, rotational speed pulse signals from the key phase sensor are synchronously embedded in the acquisition channels of all sensors—each pulse corresponds to one revolution of the unit's main shaft. A rotational speed phase reference φ(t) is constructed based on this, and its calculation formula is as follows: (1); In the formula, for Instantaneous angular velocity at time t (rad / s). for The instantaneous rotational speed (r / min) at time t; this phase reference provides a benchmark for subsequent time-domain to angular-domain conversion and equal-angle sampling.

[0013] Step 1.3, Adaptive Sampling Strategy To address the "spectral ambiguity" problem caused by the time-varying rotational speed during flyaway, an adaptive sampling strategy based on the rate of change of rotational speed is designed: The system monitors the rate of change of rotational speed in real time. When the value exceeds the preset threshold (calibrated experimentally based on the unit's rated speed, maximum head, and speed control system characteristics), the unit is determined to have entered the runaway transition process and automatically switches to the equal angle incremental sampling mode. In this mode, the sampling interval is no longer fixed. The decision is not made by the fixed spindle rotation angle. (Preferred 0.1°) Determines that each key phase pulse triggers a data block acquisition once, ensuring that each data block corresponds to a fixed angle period of spindle rotation, completely avoiding the spectral aliasing problem of conventional equal-time sampling when the rotation speed changes, and laying the data foundation for subsequent order analysis.

[0014] Step 2: Preprocessing and Enhancement of Acoustic Text Signal Based on Rotational Speed ​​Synchronous Order Tracking During the flyaway critical process, fault acoustic signals (such as AE signals) are overwhelmed by strong noise, and conventional filtering methods easily lose fault information. This step innovatively combines rotating machinery order analysis with blind source separation technology to achieve noise separation and fault signal enhancement. Key technologies include: Step 2.1, Time-Opposite Domain Resampling Using the rotational speed phase reference obtained in step 1 All acoustic time-domain signals Resampling as angular domain signal The core function of this conversion is to make periodic noise that is synchronized with or multiplied by the rotational speed (such as the main shaft rotational frequency, the blade passing frequency and its harmonics) appear as a stable periodic signal in the angular domain (with a fixed relationship between phase and angle), while fault characteristics that are unrelated to the rotational speed (such as random crack propagation AE signal) appear as a non-periodic signal, thus providing "distinguishing features" for the subsequent separation of noise and fault signals.

[0015] Step 2.2: Constructing the order spectrum and identifying periodic noise Diagonal domain signal Performing a short-time angular domain Fourier transform (STFT) yields the order spectrum of the voiceprint. ,in It is the order ratio (i.e., the ratio of the signal frequency to the spindle rotation frequency). This is an index for the corner domain blocks (corresponding to different rotation angle intervals). In the order spectrum, periodic noise sources (such as frequency conversion) Blade passing frequency , The number of blades will be concentrated on a few clear order ratio lines, while the fault acoustic emission signal and random noise have a broad spectrum distribution, and the two form a significant difference in the order ratio spectrum.

[0016] Step 2.3: Order Domain-Blind Signal Separation (OD-BSS) and Fault Signal Enhancement To accurately separate periodic noise from fault signals, this invention proposes an order mask algorithm, the specific process of which is as follows: Periodic component identification: by calculating the order spectrum In each order ratio kurtosis To distinguish between periodic noise and random signals—periodic components have a concentrated amplitude distribution and a much higher kurtosis value than random components. The formula for calculating kurtosis is: (2); In the formula, For expectation operator, order ratio The average of all amplitudes, Ω represents the standard deviation of all magnitudes over the order ratio Ω; Binary order ratio mask generation: setting kurtosis threshold (Determined by statistically analyzing the kurtosis distribution of the order spectrum during the runaway transition of healthy units, ensuring coverage of over 99% of the periodic noise kurtosis values), generating a binary mask. : (3); In the formula, The region corresponds to periodic noise. The area corresponds to fault signals and random noise; Noise separation and signal enhancement: The periodic noise component is extracted by multiplying the order mask with the order spectrum. (4); right Perform an inverse short-time angular domain Fourier transform to obtain the angular domain noise signal. Finally, the noise signal is subtracted from the original angular domain signal to obtain the enhanced fault acoustic signature signal: (5).

[0017] This process can improve the signal-to-noise ratio of fault signals by 10-20 dB, effectively preserving the weak characteristics of early faults.

[0018] Step 3: Extraction and Fusion of Voiceprint Features Sensitive to Faults During Flight Even after preprocessing, the enhanced signal still needs to extract "sensitive features" that characterize runaway critical faults. This invention focuses on extracting two types of features with strong anti-interference capabilities and high fault correlation, as follows: Step 3.1, Normalized sound energy characteristics of rotational speed ( ) During flight, acoustic energy increases non-linearly with increasing rotational speed. Directly using acoustic energy data cannot distinguish between "energy increase due to increased rotational speed" and "energy increase due to a malfunction." Therefore, this invention designs a rotational speed-normalized acoustic energy characteristic: Sound energy calculation: Extracting enhanced signal Sound energy in the 150~250kHz frequency band (This frequency band is the main distribution range of blade crack propagation AE signals, with minimal interference); Rotational speed normalization: A rotational speed correction term is introduced to eliminate the influence of rotational speed variations on sound energy. The normalization formula is as follows: (6); In the formula, The correction index (range 1.5~2.0) calibrated for the experiment is used to compensate for the nonlinear increase of acoustic energy with rotational speed—for healthy units. The value stabilizes within a fixed range, while faults such as blade cracks and cavitation can lead to... A significant increase in the value can serve as a fundamental characteristic for fault diagnosis.

[0019] Step 3.2, Transient Impact Repetition Frequency Entropy (TIRFE) During flight, each rotation of a cracked blade, passing a guide vane or adjacent blade, generates a transient acoustic emission (TIRFE) impact due to structural collision or stress release. The repetitive pattern of this impact is directly related to the fault. This invention quantifies the regularity of the impact using "entropy value," defining TIRFE characteristics. The calculation process is as follows: Envelope demodulation: for enhancing signals Perform a Hilbert transform to obtain the envelope signal. This highlights the peak characteristics of transient impacts; Impact Peak Detection and Angular Interval Calculation: In the angular domain, an adaptive threshold method is used to detect the impact peak of the envelope signal, and the principal shaft rotation angle interval corresponding to two adjacent impact peaks is calculated. ; Order interval conversion: converting angular intervals Convert to equivalent order interval The order interval is independent of rotational speed and only reflects the periodicity of the impact. The conversion formula is: (7); TIRFE entropy calculation: Statistical analysis of all order intervals within a time window. probability distribution TIRFE is calculated using the information entropy formula: (8); Healthy units: Impact signals are random (no fixed fault source). Dispersed distribution Uniform, with a high TIRFE entropy value (typically greater than 3.5); Faulty unit: The cracked blade generates a fixed impact every revolution. Focusing on specific values ​​(such as (corresponding to one impact per rotation). When concentrated, the TIRFE entropy value decreases significantly (typically less than 2.0); This feature is 5 to 8 times more sensitive to blade cracks than conventional vibration features, making it a core feature for early fault identification.

[0020] Step 4: Intelligent Diagnosis and Early Warning Based on Digital Twins and Transfer Learning To address the challenges of scarce runaway critical fault samples and the difficulty in training deep learning models in practical engineering, this invention combines digital twins and transfer learning to construct a diagnostic model and designs a dynamic hierarchical early warning mechanism, as detailed below: Step 4.1: Generation of fault simulation data driven by digital twin A parametric finite element model (digital twin) of a mixed-flow turbine generator unit in a hydropower station is constructed. This model includes the geometric parameters, material properties (such as elastic modulus and Poisson's ratio), and dynamic characteristics of key components such as the volute, runner, guide vanes, and main shaft, and can simulate the acoustic response under different operating conditions. Set simulation variables, including fault type (blade crack, cavitation), fault location (runner blades 1~16), fault size (crack length 0.5~5cm), and runaway speed (1.6~2.4 times rated speed). Simulation data output: Through dynamic simulation, the acoustic fingerprint signals under different fault conditions are calculated to generate a massive simulation dataset (sample size ≥ 100,000 sets) containing "acoustic fingerprint data - fault label (type / location / size)" to provide data support for model pre-training.

[0021] Step 4.2: Construction of Transfer Learning Diagnostic Model A one-dimensional convolutional neural network (1D-CNN) is used as the fault classifier, and transfer learning is utilized to address the few-shot problem. The training process is as follows: Pre-training phase: The simulation dataset generated by the digital twin is input into the 1D-CNN to train the network to learn the deep abstract features of fault soundprints during the flyaway critical process (such as the changing trend of TIRFE entropy, etc.). (Abnormal fluctuation patterns) enable the network to have preliminary fault identification capabilities; Fine-tuning phase: Collect a small amount of on-site measured health / fault data (sample size ≥ 500 groups), fine-tune the parameters of the pre-trained 1D-CNN, correct the difference between simulation data and actual data, and improve the diagnostic accuracy of the model to over 95%. Diagnostic output: The results extracted in step 3 The model trained with TIRFE features outputs diagnostic results for fault type (cracks / cavitation) and fault severity (mild / moderate / severe).

[0022] Step 4.3: Dynamic hierarchical early warning and fault location Dynamic threshold setting: The average value is calculated based on the TIRFE entropy value of the unit's recent (e.g., 30 days) health status. with standard deviation Set dynamic early warning thresholds: (9); In the formula, The sensitivity coefficient (set according to the importance of the unit, with a value of 1.5~2.5) and the dynamic threshold can adapt to the influence of environmental changes (such as water temperature and head fluctuations); Step 4.4, Tiered Early Warning Mechanism: Level 1 Warning (Early Fault Alert): When the TIRFE value is below the dynamic threshold for three consecutive sampling periods and shows a downward trend, and the vibration amplitude is still within the allowable range, the system issues an early warning to alert maintenance personnel to potential faults. Level 2 Warning (Emergency Warning): When the TIRFE value is lower than " "or When the value exceeds twice the healthy average, the fault is determined to have progressed to the middle or late stage, and the system issues an emergency warning and recommends shutdown for maintenance. Fault location: Combining the composite acoustic sensor array deployed in step one, the Time Difference of Arrival (TDOA) technology is used: the time difference of the same fault acoustic signal arriving at different sensors is calculated, and the location of the fault acoustic emission source is determined by geometric positioning algorithm based on the three-dimensional coordinates of the sensors. Finally, the fault blade number is output (e.g., "Rotor Blade No. 8"), and the positioning error is less than 0.5 blade spacing.

[0023] The acoustic signature-based early warning method for runaway critical faults in mixed-flow turbine units of hydropower stations provided by this invention has the following beneficial effects: 1. This invention captures high-frequency AE sensor signals (100-300kHz) and TIRFE entropy characteristics to identify the initiation and propagation of micro-cracks in blades 30-60 minutes before macroscopic vibrations are triggered by a fault, achieving true "advanced warning" and effectively overcoming the lag problem of existing speed and vibration sensor monitoring methods.

[0024] 2. The order ratio domain blind source separation (OD-BSS) technology and TIRFE entropy feature proposed in this invention greatly improve the signal-to-noise ratio and accuracy of fault identification in strong noise background (fault identification accuracy rate of over 95%), while effectively suppressing strong periodic noise related to rotational speed (noise reduction rate ≥ 85%). Furthermore, the TIRFE entropy feature is sensitive to the regularity of fault impact and is not affected by hydraulic or mechanical noise.

[0025] 3. Based on composite acoustic sensor array and TDOA positioning technology, this invention can locate the fault to a specific impeller blade with a positioning error of less than 0.5 blade spacing, solving the problem of existing technology that "only knows the fault, but does not know the location", and providing a clear direction for maintenance work.

[0026] 4. This invention generates massive amounts of simulation data through digital twin technology and combines it with transfer learning to achieve efficient model training. It does not rely on a large number of field fault samples, thus breaking through the technical bottleneck of "scarcity of fault samples" in engineering, ensuring the generalization ability of the model, and can be adapted to different models of mixed-flow units.

[0027] 5. The sensor array deployment of this invention does not require modification of the core structure of the unit. The adaptive sampling and dynamic threshold design can adapt to different operating scenarios with different heads and speeds. The early warning results are intuitive (tiered early warning + location information), which facilitates rapid response by operation and maintenance personnel and provides conditions for large-scale promotion and application.

[0028] 6. This invention addresses the fly-through transition process by arranging a composite array of high-frequency acoustic emission sensors and broadband acoustic pressure sensors at key locations such as the volute inlet and the 45° section, and embedding rotational speed pulse signals to provide a synchronization reference. This establishes an adaptive sampling strategy, solving the problem that fixed sampling strategies cannot effectively capture dynamic features.

[0029] 7. Based on speed synchronization order tracking, this invention resamples the acoustic signal from the time domain to the angular domain and performs order domain blind source separation and noise reduction processing, which overcomes the shortcomings of conventional filtering methods that lose a lot of information and improves the accuracy and effectiveness of signal processing.

[0030] 8. This invention enhances the ability to characterize fault features by extracting the normalized acoustic energy features of rotational speed and the entropy features of transient impact repetition frequency, making fault identification more sensitive and accurate.

[0031] 9. This invention utilizes digital twin technology to establish a parameterized finite element model of the unit structure to generate simulated voiceprint data. Combined with transfer learning, it achieves efficient training and fine-tuning of the model, thereby improving the reliability and engineering applicability of the model.

[0032] 10. This invention uses the extracted feature input diagnostic model to perform fault simulation driven by digital twins, and uses the transfer learning diagnostic model to achieve intelligent diagnosis. At the same time, it establishes an advanced early warning mechanism that can issue early warnings in time before a fault occurs, ensuring the safe operation of the unit.

[0033] 11. This invention reduces data processing volume and time through an adaptive sampling strategy and efficient signal processing technology, improving the real-time performance and efficiency of fault identification, and providing strong support for the rapid response and maintenance of the unit.

[0034] 12. By combining advanced voiceprint recognition technology and digital twin simulation technology, this invention effectively solves the problems of lag, insufficient sensitivity, inability to locate, and small sample size in existing technologies, providing a strong guarantee for the safe and stable operation of mixed-flow units in hydropower stations. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the technical process of the method of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments: Example 1 like Figure 1 As shown in the figure, this embodiment provides a method for early identification and warning of runaway critical faults using acoustic fingerprints in hydropower station mixed-flow turbine generator units. Taking a hydropower station equipped with a mixed-flow turbine generator unit as an example, it faces the risk of runaway during operation. To identify runaway critical faults in advance and to ensure the safe and stable operation of the unit, the acoustic fingerprint early identification and warning method proposed in this invention is adopted. The specific implementation method and steps are as follows: Step 1: Deployment of acoustic sensor arrays and synchronous signal acquisition for the fly-out transition process Step 1.1, Sensor System Construction A composite array consisting of high-frequency acoustic emission sensors (AE, main frequency 100kHz~300kHz) and broadband acoustic pressure sensors (20Hz~100kHz) is deployed at the inlet of the spiral casing, the 45° section, the nose of the turbine, and the area near the guide vanes and runner on the turbine top cover. The high-frequency AE sensor is used to capture high-frequency burst signals such as blade crack propagation, while the broadband acoustic pressure sensor is used to monitor mid-to-low frequency continuous noise such as cavitation vortex bands and water flow impact.

[0037] Step 1.2: Synchronization Signal Acquisition Rotational speed pulses from the key phase sensor are synchronously embedded into all sensor signals. Each pulse represents one revolution of the shaft, providing a strictly synchronized rotational speed phase reference for all acoustic signals. The calculation formula is: (1); In the formula, for The rotational speed phase reference (unit: rad) at any given moment is used to provide a synchronous rotational speed phase reference for all acoustic signals; for Instantaneous angular velocity at a given moment (unit: rad / s); Pi (approximately 3.1416). Current time (in seconds); Let be the integration variable (unit: s), representing any time from 0 to t; for The instantaneous rotational speed at time τ (unit: r / min) is the number of times the unit rotates per minute at time τ.

[0038] Step 1.3, Adaptive Sampling Strategy The system monitors the rate of change of rotational speed in real time. When this value exceeds a preset threshold (determined through experimental calibration, indicating the start of the flyaway transition process), the system automatically switches to an equal-angle incremental sampling mode, with the sampling interval changing from a fixed axis rotation angle. (e.g., 0.1°) determines this. Each key phase pulse triggers one data block acquisition, ensuring a fixed angular period for the rotation of the corresponding axis for each data block. The preset threshold is determined through experimental calibration based on the rated speed, maximum head, and speed regulation system characteristics of the mixed-flow turbine unit in the hydropower station.

[0039] Step 2: Preprocessing and Enhancement of Acoustic Text Signal Based on Rotational Speed ​​Synchronous Order Tracking Step 2.1: Rotational speed synchronization and angular domain resampling Using the rotational speed pulse signal obtained in step 1 time-domain acoustic signals Resampling to corner domain This makes periodic noise (such as rotation frequency, blade passage frequency) appear as a stable signal in the angular domain, and fault characteristics (such as crack propagation) appear as a non-periodic signal.

[0040] Step 2.2, Blind Source Separation in the Order Ratio Domain Diagonal domain signal Performing a short-time angular domain Fourier transform (STFT) yields the order spectrum of the voiceprint. ,in For order ratio, This is an index for the corner domain block. On the order spectrum, periodic noise sources (such as rotational frequency, blade passage frequency and its harmonics) will be concentrated on a few clear order lines, while fault acoustic emission signals and random noise are widely distributed.

[0041] Periodic noise is separated using an order ratio masking algorithm: the kurtosis of the order ratio spectrum is calculated. The formula is: (2); In the formula, Indicates order ratio The corresponding kurtosis value is used to distinguish between periodic noise and non-periodic fault signals; the periodic component... Much higher than the random component; This represents the expectation operator, used to calculate the statistical average of the variables within the parentheses; It represents the order ratio spectral amplitude (unit: V·s or Pa·s), which characterizes the acoustic energy distribution of different order ratios and different corner blocks; The order ratio, which is the ratio of the signal frequency to the unit's operating frequency, is used to eliminate the influence of speed changes on frequency analysis. This is a corner domain block index used to mark different corner domain data segments; order ratio All The mean amplitude (unit: V·s or Pa·s); For all orders of Ω The standard deviation of the amplitude (unit: V·s or Pa·s) characterizes The degree of dispersion.

[0042] The noise source is separated using an order-ratio masking algorithm to generate a binary mask. Its definition is: (3); in, The kurtosis threshold is used to extract periodic noise components through this mask. The angular domain noise signal is obtained through inverse transformation. : (4); Then, the enhanced fault voiceprint signal is obtained: (5).

[0043] when Greater than the preset threshold At that time, a binary mask is generated. Otherwise, it is 0; The threshold for determining kurtosis (calibrated experimentally, typically 3-5) is used. This is a ratio mask matrix used to filter out periodic noise components in the ratio spectrum. The kurtosis threshold... By statistically analyzing the order spectrum kurtosis distribution during the runaway transition of healthy units, we can ensure the effective separation of periodic noise and fault signals.

[0044] Step 3: Extraction and Fusion of Voiceprint Features Sensitive to Faults During Flight Step 3.1, Normalized sound energy characteristics of rotational speed Calculate the acoustic energy of the enhanced signal in the 150~250kHz frequency band. And normalize the rotational speed: (6); In the formula, Represents the normalized acoustic energy at rotational speed (unit: or This eliminates the influence of speed variation on acoustic energy and is used to characterize the intensity of fault-related acoustic energy. This represents the acoustic energy of the enhanced signal in a specific frequency band (150~250kHz) (unit: or It is obtained by integrating the square of the amplitude of the acoustic signal; This indicates the current instantaneous speed of the unit (unit: r / min); The rotational speed compensation index (range 1.5~2.0, determined experimentally) is used to compensate for the nonlinear growth relationship of sound energy with rotational speed.

[0045] Step 3.2, Transient Impact Repetition Frequency Entropy (TIRFE): Step 3.2.1: Analyze the enhanced angular domain signal Envelope signal is obtained by envelope demodulation. , The enhanced back-angle domain acoustic signal output from step 2 (Unit: V or Pa) The envelope signal (unit: V or Pa) is used to highlight the transient impact component in the acoustic signal; Step 3.2.2: Detect the envelope signal in the angular domain The peak value of the impact is calculated, and the angular interval between adjacent peak values ​​is determined. ; For the first The angular interval (unit: rad) between adjacent impact peaks characterizes the temporal (angular) distribution characteristics of the impact signal; Step 3.2.3: Convert the angular interval into an equivalent order interval. : (7); In the formula, For the first The order interval (dimensionless) corresponding to the group angle interval is used to eliminate the influence of rotational speed on the impact interval; 2π is the circumferential angle (unit: rad). Step 3.2.4: Statistical probability distribution of the order interval within a time window The formula for calculating the TIRFE entropy is: (8); In the fault state, the TIRFE entropy value is significantly lower than that in the healthy state, where, For order interval The probability distribution (dimensionless, ranging from 0 to 1) represents the frequency of occurrence of different order ratio intervals.

[0046] Step 4: Intelligent Diagnosis and Early Warning Based on Digital Twins and Transfer Learning Step 4.1: Fault Simulation Driven by Digital Twin A parametric finite element model of the unit was established to simulate the dynamic response of blade cracks at different locations and sizes under runaway speeds, generating massive amounts of simulated acoustic signature data and fault labels (such as crack location and size). The parametric finite element model is a numerical model constructed based on the actual structural dimensions and material properties (such as elastic modulus and density) of the unit, and the fault labels include crack location (such as runner blade number) and crack size (such as length and depth).

[0047] Step 4.2: Construction of Transfer Learning Model A one-dimensional convolutional neural network (1D-CNN) was used as the classifier. The network was pre-trained using simulation data to learn deep abstract patterns of fault acoustic signatures during runaway operations. The pre-trained model was then fine-tuned using a small amount of field-tested data to improve its generalization ability and diagnostic accuracy. The 1D-CNN is a deep learning network suitable for time-series signals (such as acoustic signatures), extracting local signal features through convolutional kernels. Pre-training involved optimizing the initial network parameters using massive amounts of simulation data, while fine-tuning involved adjusting the network parameters using a small amount of field data to adapt to actual working conditions.

[0048] Step 4.3, Advanced Early Warning Mechanism The mechanism for early warning includes dynamic threshold setting and tiered early warning. Dynamic threshold... The calculation formula is: (9); In the formula, The dynamic warning threshold (dimensionless) represents the characteristics of TIRFE and is used to determine whether a fault has occurred. This represents the statistical mean (dimensionless) of TIRFE under recent health conditions, calculated by collecting TIRFE data during the healthy operation of the unit. This is a sensitivity coefficient (dimensionless, ranging from 1 to 3, adjusted according to the unit's operational stability requirements), used to regulate the early warning sensitivity. The statistical standard deviation (dimensionless) of TIRFE under recent health conditions represents the range of fluctuation of TIRFE under health conditions. When the TIRFE value remains below the dynamic threshold and shows a downward trend, a Level 1 warning (early fault indication) is issued; when the characteristic value exceeds a preset larger threshold (calibrated experimentally, typically [value missing]), a Level 1 warning is issued; when the characteristic value exceeds a preset larger threshold (typically determined experimentally), a Level 1 warning is issued. When this occurs, a Level 2 alarm (emergency warning) will be issued and a shutdown for maintenance will be recommended.

[0049] Step 4.4, Fault Location The approximate location of the faulty blade is initially determined by using time-difference-of-arrival (TDOA) technology of acoustic arrays, combined with the deployment location information of composite acoustic sensor arrays.

[0050] This method utilizes high-frequency acoustic emission signals and TIRFE entropy characteristics to achieve early warning of runaway critical faults, solving the problems of lag, insufficient sensitivity, and difficulty in localization in existing technologies. The order-domain blind source separation and digital twin simulation technology improve the accuracy of fault identification under strong noise backgrounds, ensuring the safe and stable operation of the unit.

[0051] Example 2 In another preferred embodiment, based on Embodiment 1 above, this embodiment provides a method for early identification and warning of runaway critical faults using acoustic signatures in a hydropower station mixed-flow turbine generator unit. The application is based on an HL220-LJ-500 mixed-flow turbine generator unit in a hydropower station. The unit's rated parameters are: rated speed 300 r / min, maximum head 120 m, 16 runner blades, and rated power 50 MW. The effectiveness of the method is verified, and the specific implementation process is as follows: 1. Implementation Preparation (1) Equipment selection: High-frequency acoustic emission (AE) sensor: PAC R15α model, main frequency 100-300kHz, sensitivity 80dB; Wideband sound pressure sensor: B&K 4948 model, frequency range 20Hz-100kHz, dynamic range 140dB; Key phase sensor: Keyence IV2 photoelectric sensor with pulse resolution of 0.01° is selected; Data acquisition card: NICDAQ-9178 is selected, with a maximum sampling rate of 2MS / s and a synchronization accuracy of ±1μs.

[0052] (2) Experimental scenario setting: The simulation unit operates under the condition of "sudden load shedding + speed control system failure". The guide vane control circuit is forcibly closed through the speed governor control cabinet, triggering the runaway transition process, and monitoring the critical stage when the speed increases from 300 r / min to 580 r / min (1.93 times the rated speed).

[0053] 2. Implementation Steps Step 1: Deployment of acoustic sensor array and synchronous signal acquisition Step 1.1, Array Deployment: Four AE sensors are installed at the inlet of the volute (1 point), the 45° section of the volute (2 points), and the nose of the volute (1 point). Four broadband acoustic pressure sensors are installed on the top cover of the turbine, near the guide vanes (2 points) and near the edge of the runner (2 points). The key phase sensor is installed on the non-drive end of the spindle and aligned with the engraving line on the shaft end (1 pulse per revolution).

[0054] Step 1.2, Speed ​​Pulse Synchronization and Adaptive Sampling: The rotational speed pulse signal of the embedded key phase sensor is used to calculate the rotational speed phase reference according to formula (1). ; The preset speed change rate threshold is 50 r / min² (calibrated through unit no-load test; exceeding this value determines the transition to runaway). When the rotational speed increases from 300 r / min to 320 r / min The system automatically switches to isotropic sampling. It collects 3600 data points per revolution to avoid spectrum ambiguity.

[0055] Step 2: Voiceprint signal preprocessing and enhancement Step 2.1, Time-domain to angular domain resampling: use The acoustic time-domain signal (sampling rate 1 MS / s) is resampled into an angular domain signal. This ensures that periodic noises such as the main shaft rotation frequency (5Hz) and blade passage frequency (80Hz, 16×5Hz) present as stable sinusoidal signals in the angular domain.

[0056] Step 2.2, Blind Source Separation in the Order Ratio Domain: The diagonal domain signal was subjected to a short-time angular domain Fourier transform using a Hanning window (window length 180 angle points) to obtain the order ratio spectrum. ; The order spectral kurtosis of the flight process of a healthy aircraft was statistically analyzed, and the following parameters were set. (Covering 99.2% of the periodic noise kurtosis values); Generate binary mask Periodic noise was extracted and removed, and the signal-to-noise ratio of the enhanced fault acoustic signal was improved from 2dB to 17dB, and the crack AE signal (180kHz) was clearly distinguishable.

[0057] Step 3: Extraction of Fault-Sensitive Voiceprint Features Step 3.1, Normalized sound energy based on rotational speed ( ): Calculate the acoustic energy of the enhanced signal in the 150-250kHz frequency band. Through experimental calibration ; In a healthy state, Stable at When a 0.8cm crack was pre-formed on blade #12, Rise to The increase was 53.3%.

[0058] Step 3.2, Transient Impact Repetition Frequency Entropy (TIRFE): After demodulating the enhanced signal envelope, it was detected that the cracked blade generated one impact per revolution. ), (1 / 16, corresponding to the blade spacing); Under healthy conditions, TIRFE is 3.9~4.1; under fault conditions, TIRFE drops to 1.6, and the entropy value decreases significantly, which is consistent with the characteristics of faults.

[0059] Step 4: Intelligent Diagnosis and Early Warning Step 4.1, Digital Twin and Model Training: A parametric finite element model of the unit was established using ANSYS, and the fly-out sound patterns of 16 blades with 0.5~5cm cracks were simulated, generating 120,000 sets of simulation data. The model employs a 1D-CNN (2 convolutional layers + 2 pooling layers + 1 fully connected layer), pre-trained with simulated data, and then fine-tuned with 600 sets of real-world test data (300 healthy sets and 300 faulty sets), achieving a diagnostic accuracy of 96.5%.

[0060] Step 4.2, Tiered Early Warning and Location: Set dynamic threshold sensitivity coefficient Recent health TIRFE average Standard deviation , ; When the TIRFE value remains below 2.8 for three consecutive sampling periods (12s), the system issues a Level 1 warning (identifying crack initiation 45 minutes in advance); when Rise to At that time, a level-two warning will be issued; Using Time Difference of Array (TDOA) positioning with AE sensor array, the time difference (maximum difference 0.2ms) of the impact signal reaching the four AE sensors was calculated. Combined with the sensor coordinates, the faulty blade was located as number 12, with a positioning error of 0.3 blade spacing (approximately 8mm).

[0061] 3. Verification of Implementation Results After shutdown and maintenance, a 0.9cm crack was found at the water outlet edge of blade No. 12 (12.5% ​​deviation from the simulation preset of 0.8cm). This verifies that the method of the present invention can achieve advanced identification (45 minutes in advance), accurate positioning (error <10mm) and strong noise interference resistance (signal-to-noise ratio improved by 15dB) of the critical fault of flight, which fully meets the requirements of engineering applications.

[0062] In the preferred embodiment, the sensing system in step 1 is a composite acoustic sensor array, including a high-frequency acoustic emission sensor (AE) and a broadband acoustic pressure sensor. The main frequency range of the high-frequency acoustic emission sensor is 100kHz~300kHz, used to capture high-frequency burst signals of blade crack propagation; the frequency range of the broadband acoustic pressure sensor is 20Hz~100kHz, used to monitor cavitation vortex bands and low-frequency continuous noise from water flow impact. The above settings enable full-band monitoring of the blade's operating status, with high-frequency and low-frequency signals complementing each other, effectively separating crack propagation and water flow interference signals. Through multi-channel synchronous acquisition and feature fusion algorithms, the crack location can be accurately located and the degree of damage can be assessed, improving the reliability of turbine operation.

[0063] In a preferred embodiment, the composite acoustic sensor array is positioned at the inlet of the volute, the 45° section, the nose end, and the area near the guide vanes and runner on the turbine top cover. This configuration effectively captures vibration and noise signals from different areas inside the turbine. Through multi-location collaborative monitoring, it comprehensively and accurately reflects the equipment's operating status, providing rich and reliable data support for fault diagnosis and performance evaluation, and ensuring the stable and efficient operation of the turbine.

[0064] In the preferred embodiment, the rotational speed pulse signal in step 1 comes from a key phase sensor, with each pulse representing one revolution of the shaft, providing a strictly synchronized rotational speed phase reference for all acoustic signals. The above settings ensure that the acoustic signal acquisition accurately corresponds to the shaft rotation state, avoiding data distortion caused by phase deviation. Step 2 uses a multi-channel synchronous acquisition module to align the acoustic sensor signal with the rotational speed pulse signal in time, and sets the sampling frequency to more than 10 times the maximum shaft speed.

[0065] In a preferred embodiment, the adaptive sampling strategy in step 1 is implemented as follows: the system monitors the rate of change of rotational speed in real time. When this value exceeds a preset threshold, it automatically switches to the equal-angle incremental sampling mode, with the sampling interval determined by a fixed axis rotation angle. It is determined that each key phase pulse triggers one data block acquisition, ensuring that each data block corresponds to a fixed angle period of axis rotation, wherein the fixed axis rotation angle... The value is set to 0.1°; this setting effectively improves the stability and accuracy of data acquisition, especially suitable for operating conditions with large speed fluctuations. The system is also equipped with a dynamic calibration module, which can automatically adjust the sampling parameters according to actual speed changes, avoiding data distortion caused by sudden speed changes and further enhancing the system's environmental adaptability.

[0066] In the preferred embodiment, the preset threshold is determined through experimental calibration based on the rated speed, maximum head, and speed regulation system characteristics of the mixed-flow turbine unit in the hydropower station. This setting ensures stable operation of the mixed-flow turbine unit under different operating conditions and prevents equipment failure caused by speed fluctuations exceeding safe limits. The experimental calibration process requires comprehensive consideration of various factors, and through repeated testing and data analysis, the most reasonable preset threshold is ultimately determined.

[0067] In the preferred embodiment, the specific process of resampling the acoustic signal from the time domain to the angular domain in step 2 is as follows: using the rotational speed phase reference obtained in step 1... All acoustic time-domain signals Convert to angular domain signal This design ensures that periodic noise components synchronized with or multiplied by the rotational speed appear as stable periodic signals, while fault characteristics unrelated to rotational speed appear as aperiodic signals. This effectively separates periodic noise from aperiodic fault characteristics, reducing noise interference. Angular domain resampling ensures the signal is uniformly distributed within the rotational period, facilitating subsequent spectral analysis or time-frequency analysis to accurately extract fault characteristic frequency components, thus improving the accuracy and reliability of fault diagnosis.

[0068] In a preferred embodiment, the implementation process of order ratio domain blind source separation in step 2 includes: diagonal domain signal Perform a short-time angular domain Fourier transform to obtain the order spectrum of the voiceprint. ,in For order ratio, As a corner domain block index, periodic noise sources are concentrated on clear order lines of the order spectrum, while fault acoustic emission signals and random noise are widely distributed. The above settings can accurately distinguish between periodic noise and fault signal characteristics. Then, an independent component analysis algorithm is used to further separate the order spectrum signal, extract the fault acoustic emission signal components, and remove random noise interference, thereby providing clear and accurate signal basis for subsequent fault diagnosis.

[0069] In a preferred embodiment, the kurtosis of the order spectrum is calculated. Identify periodic components, when Greater than the preset threshold At that time, a binary mask is generated. Otherwise, it is 0; where, The threshold for determining kurtosis (calibrated experimentally, typically 3-5) is used. The order ratio mask matrix is ​​used to filter out periodic noise components in the order ratio spectrum. The above settings can accurately locate periodic noise in the order ratio spectrum, effectively filter out interference through binary masking, and improve signal purity. Experimental verification shows that this method can improve the signal-to-noise ratio by 15%-20% and significantly enhance the reliability of subsequent feature extraction.

[0070] In a preferred embodiment, the noise source is separated using an order-ratio masking algorithm to generate a binary mask. , The kurtosis threshold is used to extract periodic noise components through this mask. The angular domain noise signal is obtained through inverse transformation. Then, the enhanced fault voiceprint signal is obtained. The above settings can effectively improve the clarity and recognizability of the fault voiceprint signal and reduce background noise interference. Experimental verification shows that this scheme can stably separate noise sources and accurately extract fault features under different operating conditions, providing a reliable basis for subsequent fault diagnosis.

[0071] In a preferred embodiment, the kurtosis threshold By statistically analyzing the order spectrum kurtosis distribution during the runaway transition of healthy units, effective separation of periodic noise and fault signals is ensured. These settings improve fault diagnosis accuracy to over 95%, while keeping the false alarm rate below 3%. This adaptive threshold adjustment mechanism dynamically optimizes based on unit operating conditions, and actual testing has verified that it maintains signal separation efficiency of over 88% even under variable speed conditions.

[0072] In the preferred embodiment, the rotational speed normalized acoustic energy characteristics in step 3... The calculation process is as follows: First, calculate the acoustic energy of the enhanced signal in the 150~250kHz frequency band. Then, rotational speed normalization is performed. These settings eliminate the influence of different rotational speeds on the acoustic energy characteristics, making the characteristics more comparable. Specifically, the calculated acoustic energy is divided by the baseline energy value at the corresponding rotational speed to obtain the normalized acoustic energy characteristics, providing reliable data support for subsequent accurate analysis.

[0073] In the preferred embodiment, the early warning mechanism in step 4 includes dynamic threshold setting and tiered warning. When the TIRFE value is continuously lower than the dynamic threshold and shows a downward trend, a level 1 warning (early fault indication) is issued; when the characteristic value exceeds a preset higher threshold, a level 2 alarm (emergency warning) is issued and shutdown for maintenance is recommended. The above settings can effectively improve the timeliness and accuracy of equipment fault warning and reduce the risk of unplanned downtime. The dynamic threshold is automatically adjusted according to the equipment's historical operating data to adapt to different operating conditions. The tiered warning mechanism ensures that maintenance personnel prioritize handling high-risk faults and optimize resource allocation.

[0074] In a preferred embodiment, the fault location in step 4 is achieved using Time Difference of Origin (TDOA) technology for acoustic arrays. Combined with the deployment location information of the composite acoustic sensor array, the approximate location of the faulty blade is initially determined. TDOA technology refers to calculating the time difference between the arrival of the faulty acoustic signal at different sensors, combining this with the sensor array coordinates, and using geometric positioning algorithms (such as hyperbolic positioning) to invert the location of the fault source. This setup significantly improves the accuracy and efficiency of fault location, and is particularly suitable for complex noise environments. By optimizing the sensor array layout, the fault area can be further narrowed down. Furthermore, by combining machine learning algorithms for in-depth analysis of the time difference data, multi-source interference signals can be effectively distinguished.

[0075] In the preferred embodiment, the digital twin-driven fault simulation in step 4 simulates the dynamic response of blade cracks at different locations and sizes under runaway speeds using a parametric finite element model, generating simulated acoustic signature data and corresponding fault labels. The parametric finite element model is a numerical model constructed based on the actual structural dimensions and material properties (such as elastic modulus and density) of the unit. The fault labels include crack location (such as runner blade number) and crack size (such as length and depth). These settings accurately reproduce the dynamic characteristics of blade cracks under extreme operating conditions, providing high-fidelity data support for subsequent acoustic signature recognition model training. By comparing the spectral characteristics of simulated and measured acoustic signatures, early warning and location of crack faults can be achieved, with the error controlled within 3%.

[0076] In the preferred embodiment, the transfer learning model in step 4 uses a one-dimensional convolutional neural network (1D-CNN) as the classifier. It is first pre-trained using simulation data and then fine-tuned using real-world measurement data. 1D-CNN is a deep learning network suitable for time-series signals (such as voiceprint signals), extracting local signal features through convolutional kernels. Pre-training refers to optimizing the initial network parameters using massive amounts of simulation data, while fine-tuning refers to adjusting the network parameters using a small amount of real-world data to adapt to actual working conditions. This setup effectively improves the model's generalization ability in complex environments and reduces reliance on real-world measurement data. Pre-training with simulation data allows the model to quickly converge and capture key features, while the fine-tuning stage further optimizes the model's adaptability to interference factors such as actual noise and signal attenuation, ensuring classification accuracy.

[0077] In summary, the acoustic signature-based early warning method for runaway critical faults in mixed-flow turbine units of hydropower stations proposed in this invention is a powerful innovation addressing the numerous shortcomings of traditional monitoring technologies under runaway conditions. Traditional monitoring primarily relies on speed and vibration sensors, which not only provide delayed warnings when runaway critical faults occur, making it difficult to issue timely alerts in the early stages of the fault, but also lack sensitivity to weak fault signals such as early crack initiation and cavitation intensification at high frequencies and low energy levels, often failing to capture these crucial information. Furthermore, it is inadequate in fault location, struggling to accurately pinpoint the specific component where the fault occurs, greatly hindering subsequent maintenance and troubleshooting. This invention was developed precisely to address these pain points, aiming to provide a more reliable guarantee for the safe and stable operation of mixed-flow turbine units in hydropower stations.

[0078] This invention boasts numerous technological highlights, forming a complete and efficient solution. In the early warning and identification phase, a 100-300kHz high-frequency AE sensor is applied for the first time to blade crack monitoring. Combined with TIRFE entropy characteristics, it can identify the initiation and propagation of micro-cracks in blades 30-60 minutes in advance, achieving true early warning and giving personnel sufficient time to take countermeasures. The innovative order-domain blind source separation (OD-BSS) technology, through speed synchronization and angular domain analysis, effectively suppresses strong periodic noise related to speed. Simultaneously, the TIRFE entropy characteristics are only sensitive to the regularity of fault impacts and are unaffected by hydraulic or mechanical noise, greatly improving the signal-to-noise ratio and accuracy of fault identification in strong noise environments. The uniquely extracted speed-normalized acoustic energy characteristics and transient impact repetition frequency entropy characteristics further enhance the characterization ability of fault features, making fault identification more accurate. In terms of positioning, based on the composite acoustic sensor array and TDOA positioning technology, the fault can be accurately located to the specific runner blade with a positioning error of less than 0.5 blade spacing, which greatly shortens the maintenance time and improves the unit's operating efficiency and safety.

[0079] Furthermore, this invention also excels in data and model construction and in handling special operating conditions. It utilizes digital twins to generate massive amounts of simulation data, combined with transfer learning to achieve efficient model training. This eliminates the need for a large number of field fault samples, overcoming the technical bottleneck of "scarcity of fault samples" in engineering and making it adaptable to different models of mixed-flow turbines. The established intelligent diagnosis and early warning mechanism automates and intelligently processes fault feature extraction, model training, and early warning, improving the efficiency and real-time performance of fault monitoring. For the flyaway transition process, a composite array of high-frequency acoustic emission sensors and broadband acoustic pressure sensors is deployed at key locations such as the volute inlet and the 45° section, with embedded speed pulse signals providing synchronous reference. An adaptive sampling strategy is established, fully considering the special characteristics of this process, accurately capturing dynamic features and providing strong support for early fault detection and accurate diagnosis. This demonstrates a deep understanding and innovative response to specific engineering scenarios, exhibiting high technical level and application value.

Claims

1. A method for early identification and warning of runaway critical fault sound print of Francis turbine unit of hydropower station, characterized in that, The method comprises the following steps: Step 1: acoustic sensor array layout and synchronous signal acquisition for runaway transition process, constructing a sensing system suitable for speed variation, embedding a speed pulse signal and establishing an adaptive sampling strategy; Step 2: acoustic fingerprint signal preprocessing and enhancement based on speed synchronization order ratio tracking, resampling the acoustic signal from the time domain to the angle domain, and realizing noise separation and fault signal enhancement through order ratio domain blind source separation; Step 3: extraction and fusion of fault-sensitive acoustic fingerprint features in the runaway process, calculation of speed-normalized acoustic energy features and transient impact repetition frequency entropy TIRFE features; Step 4: intelligent diagnosis and early warning based on digital twinning and transfer learning, inputting the extracted features into a diagnostic model pre-trained by digital twinning simulation data and fine-tuned by on-site measured data, and realizing early fault identification, hierarchical warning and fault location by combining dynamic thresholds.

2. The runaway critical fault acoustic fingerprint early identification and warning method for a hydroelectric power station Francis turbine unit according to claim 1, characterized in that: The sensing system in step 1 is a composite acoustic sensor array, including a high-frequency acoustic emission sensor AE and a broadband acoustic pressure sensor. The high-frequency acoustic emission sensor has a frequency range of 100 kHz to 300 kHz, which is used to capture high-frequency burst signals caused by blade crack propagation. The broadband acoustic pressure sensor has a frequency range of 20 Hz to 100 kHz, which is used to monitor cavitation vortex bands and low-frequency continuous noise caused by water flow impact.

3. The runaway critical fault acoustic fingerprint early identification and warning method for a hydroelectric power station Francis turbine unit according to claim 2, characterized in that: The layout positions of the composite acoustic sensor array include the inlet of the volute, the 45° cross section, the nose end, and the area near the guide vane and runner of the turbine top cover.

4. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 1, characterized in that: The rotational speed pulse signal in step 1 comes from a key-phasor sensor, each pulse represents one revolution of the shaft, providing a strict rotational speed phase reference for all acoustic signals The formula is: （1）； wherein is the rotational speed phase reference at the instant of time; is the instantaneous angular velocity at the instant of time; is the current time; is is the instantaneous rotational speed at the instant of time.

5. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 4, characterized in that, The adaptive sampling strategy in step 1 is implemented as follows: the system monitors the speed variation rate in real time, and when the value exceeds a preset threshold, it automatically switches to an equal-angle incremental sampling mode, with the sampling interval determined by a fixed shaft rotation angle. Each key phase pulse triggers data block collection, ensuring that each data block corresponds to a fixed angle period of shaft rotation.

6. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 5, characterized in that: The preset threshold is determined by experiments based on the rated speed, maximum water head and speed regulation system characteristics of the hydroelectric station Francis turbine unit.

7. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 1, characterized in that, The specific process of resampling the acoustic signal from time domain to angle domain in step 2 is: using the speed phase reference obtained in step 1 Convert all acoustic time domain signals Into angle domain signals Make the periodic noise components synchronized with the speed or multiplied by the frequency appear as stable periodic signals, and the fault features irrelevant to the speed appear as non-periodic signals.

8. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 7, characterized in that, The implementation process of the step 2 includes: performing short-time angular domain Fourier transform on the diagonal domain signal Performing short-time angular domain Fourier transform on the diagonal domain signal, obtaining the order ratio spectrum of the voiceprint, and the periodic noise source is concentrated on the clear order ratio line of the order ratio spectrum, and the fault acoustic emission signal and the random noise are widely distributed.

9. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 8, characterized in that: by calculating a kurtosis of the order spectrum identifying periodic components, kurtosis The formula for calculating the kurtosis is: （2）； wherein denotes the desired operator; denotes the order ratio spectrum amplitude; is the order ratio; is the angular domain block index; is the order ratio all of the amplitude mean; is the order ratio all of the amplitude standard deviation.

10. The runaway critical fault acoustic fingerprint early identification and warning method for a hydroelectric power station Francis turbine unit according to claim 9, characterized in that: The noise sources are separated by an order masking algorithm to generate binary masks defined as: （3）； wherein is a kurtosis threshold, periodic noise components are extracted by the mask obtained by inverse transform : （4）； Then, the enhanced fault acoustic fingerprint signal is obtained: （5）。 11. The runaway critical fault acoustic fingerprint early identification and warning method for a hydroelectric power station Francis turbine unit according to claim 10, characterized in that: the kurtosis threshold The kurtosis threshold is determined by the statistical distribution of the order spectrum kurtosis of the health unit flight transients, ensuring effective separation of periodic noise and fault signals.

12. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 1, characterized in that, The step 3 in the rotational speed normalization sound energy characteristic The calculation process is that the sound energy of the enhanced signal in the 150-250 kHz frequency band is calculated first Then, the rotational speed normalization processing is performed, and the normalization formula is: （6）； wherein represents the normalized sound energy of the rotational speed; represents the sound energy of the enhanced signal in a specific frequency band; represents the current unit instantaneous rotational speed; represents the rotational speed compensation index.

13. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 1, characterized in that, The calculation steps of the transient impact repetition frequency entropy TIRFE feature in step 3 include: Step 3.1: envelope demodulation of the enhanced angle domain signal to obtain an envelope signal; Step 3.2: Detecting the impulsive peaks of the envelope signal in the angular domain, calculating the angular separation between adjacent peaks ; Step 3.3: Convert the angular intervals to equivalent fractional intervals : （7）； Step 3.4: Calculate the distribution probability of the statistical order ratio interval in a time window Calculate the TIRFE entropy value, the formula is: （8）； The TIRFE entropy value in the fault state is significantly lower than that in the healthy state.

14. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 1, characterized in that: In step 4, the combination of digital twinning and transfer learning is as follows: Step 4.1: establish a parameterized finite element model of the unit structure, simulate the dynamic response of blade cracks of different sizes at different positions under runaway speed, and generate a large amount of simulation acoustic fingerprint data and corresponding fault labels; Step 4.2: use a one-dimensional convolutional neural network as a classifier, pre-train the network with simulation data, and learn the deep abstract patterns of fault acoustic fingerprints in the runaway process; Step 4.3: fine-tune the pre-trained model with a small amount of on-site measured data to improve the model's generalization ability and diagnostic accuracy.

15. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 1, characterized in that, The mechanism of early warning in step 4 includes dynamic threshold setting and hierarchical early warning, and the dynamic threshold The calculation formula is: （9）； wherein, and are the mean and standard deviation of TIRFE in the near-term health state, respectively, is the sensitivity coefficient; when the TIRFE value is continuously lower than the dynamic threshold and shows a downward trend, a first-level early warning is issued; when the feature value exceeds the preset larger threshold, a second-level alarm is issued and maintenance is recommended.

16. The runaway critical fault acoustic fingerprint early identification and warning method for hydroelectric station Francis turbine unit according to claim 1, characterized in that: In step 4, fault location is realized through time difference positioning TDOA technology of the acoustic array, combined with the layout position information of the composite acoustic sensor array, to preliminarily determine the approximate direction of the fault blade.

Images