Turbine fault online diagnosis method and device based on voiceprint recognition, storage medium and computer equipment

Abstract

The application discloses a steam turbine fault online diagnosis method and device based on voiceprint recognition, a storage medium and a computer device, and relates to the technical field of steam turbine fault diagnosis. The method comprises the following steps: collecting wide-frequency sound pressure signals of a steam turbine during operation through a non-contact microphone array arranged at a target position of the steam turbine; performing frame processing on the wide-frequency sound pressure signals to obtain a plurality of signal segments, and removing noise segments and non-steady-state impact segments from the plurality of signal segments to obtain a plurality of effective signal segments; for each effective signal segment, extracting frequency domain features, time-frequency features and high-order statistical quantity features constituting a voiceprint spectrum from the effective signal segment, generating a fusion feature vector according to the fusion feature vector, inputting the fusion feature vector into a pre-trained fault classification model, and outputting the confidence of each fault type corresponding to the effective signal segment; and determining the confidence of each fault type corresponding to each effective signal segment of the same target position within a preset time window to generate a fault online diagnosis result of the target position.

Description

Technical Field

[0001] This application relates to the field of steam turbine fault diagnosis and condition monitoring technology, and in particular to a method and device for online diagnosis of steam turbine faults based on voiceprint recognition, a storage medium, and a computer device. Background Technology

[0002] With the deepening implementation of the "dual carbon" goals, my country's energy structure is rapidly transforming towards a cleaner and lower-carbon direction. The role of coal-fired power units is shifting from traditional baseload power sources to flexible regulation power sources, with frequent start-ups and shutdowns, deep peak shaving, and rapid load changes becoming the norm for their operation. This dramatic change in operating mode presents more severe challenges to the safe and stable operation of steam turbines. Core components such as turbine blades, rotors, and valves are subjected to alternating thermal and mechanical stresses, leading to frequent failures such as crack propagation, loosening, and leakage. Unplanned shutdowns can cause significant economic losses or even safety accidents. Therefore, how to achieve real-time perception and early warning of the operating status of key steam turbine components has become a key technical issue for ensuring the safe and economical operation of units under the new power system.

[0003] Currently, turbine fault diagnosis mainly relies on the following methods: vibration monitoring, which involves installing acceleration sensors in bearing housings, bearing bushes, and other locations to collect vibration signals and analyze their spectral characteristics, enabling the identification of typical faults such as rotor imbalance, misalignment, and oil film oscillation; thermodynamic parameter analysis, which involves monitoring the changing trends of parameters such as pressure, temperature, flow rate, and power in various sections of the unit during operation, which can identify abnormalities such as scaling in the flow path, blade wear, and internal leakage in valves; and non-destructive testing, which utilizes ultrasonic, eddy current, and magnetic particle methods to conduct offline testing on key components during unit shutdown and maintenance, enabling the direct detection of structural defects such as cracks. However, these existing technologies all have inherent limitations in practical applications. Vibration monitoring is limited in high-temperature and high-pressure areas because the sensors need to be installed in contact, and it is not sensitive to early micro-cracks. The signal is easily affected by structural acoustic interference, making fault location difficult. Thermodynamic parameter analysis has a lag in response, and parameter changes are often macroscopic manifestations after the fault has developed to a certain stage. It is not accurate enough for diagnosing early local faults such as blade micro-cracks and valve micro-leaks. Non-destructive testing requires offline operation after shutdown, which cannot achieve online continuous monitoring. This not only affects power generation efficiency but may also miss sudden faults that occur between two maintenance intervals, making it difficult to meet the diagnostic requirements of "early, online, non-contact, and highly reliable" turbine faults in deep peak shaving scenarios. Summary of the Invention

[0004] In view of this, this application provides a method and device for online fault diagnosis of steam turbines based on voiceprint recognition, a storage medium, and a computer device. It collects broadband sound pressure signals from a non-contact microphone array located around the steam turbine and performs voiceprint recognition, eliminating the need for contact installation of sensors on high-temperature and high-pressure components. This fundamentally eliminates the impact on equipment safety and integrity, overcoming the limitation of vibration monitoring methods in high-temperature and high-pressure areas. Simultaneously, by extracting frequency domain features, time-frequency features, and higher-order statistical features and fusing them into a fused feature vector, it can capture weak abnormal signs even when vibration and thermodynamic parameters have not yet changed significantly, thus improving the accuracy of turbine fault diagnosis. It exhibits extremely high sensitivity to early faults such as microcracks and minor valve leaks, compensating for the insensitivity of vibration monitoring to early microcracks and the lag in response of thermodynamic parameter analysis. By comprehensively judging the confidence of each effective signal segment within a preset time window at the same target location, the false alarm rate is significantly reduced, preventing normal steam flow fluctuations under varying load conditions from being misjudged as impacts and thus eliminated, effectively ensuring the reliability of diagnostic results across the entire operating range. The entire acquisition, processing, and diagnostic process is completed online continuously while the unit is running, without the need for shutdown, achieving early, online, non-contact, and highly reliable diagnosis of turbine faults in deep peak-shaving scenarios.

[0005] According to one aspect of this application, an online fault diagnosis method for steam turbines based on voiceprint recognition is provided, comprising: A non-contact microphone array is installed at a target location on the steam turbine to collect broadband sound pressure signals during the turbine's operation. The target location includes each cylinder and connecting component of the steam turbine, and the microphone array includes multiple microphones with different sensitivities. The broadband sound pressure signal is framed to obtain multiple signal segments, and noise segments and non-steady-state impact segments are removed from the multiple signal segments to obtain multiple effective signal segments; For each valid signal segment, frequency domain features, time-frequency features, and higher-order statistical features constituting the acoustic signature are extracted from the valid signal segment. Based on the frequency domain features, time-frequency features, and higher-order statistical features, a fused feature vector is generated. The fused feature vector is input into a pre-trained fault classification model, and the confidence level of each fault type corresponding to the valid signal segment is output. For each valid signal segment corresponding to each fault type at the same target location, a comprehensive judgment is made within a preset time window to generate an online fault diagnosis result for the target location.

[0006] According to another aspect of this application, an online fault diagnosis device for steam turbines based on voiceprint recognition is provided, comprising: The signal acquisition module is used to acquire broadband sound pressure signals during the operation of the steam turbine by using a non-contact microphone array set at the target part of the steam turbine. The target part includes each cylinder and connecting component of the steam turbine, and the microphone array includes multiple microphones with different sensitivities. The frame segmentation processing module is used to perform frame segmentation processing on the broadband sound pressure signal to obtain multiple signal segments, and to remove noise segments and non-steady-state impact segments from the multiple signal segments to obtain multiple effective signal segments. The feature extraction module is used to extract frequency domain features, time-frequency features, and higher-order statistical features that constitute the acoustic signature for each valid signal segment. Based on the frequency domain features, time-frequency features, and higher-order statistical features, a fused feature vector is generated. The fused feature vector is input into a pre-trained fault classification model, and the confidence level of each fault type corresponding to the valid signal segment is output. The fault diagnosis module is used to comprehensively determine the confidence level of each fault type corresponding to each effective signal segment of the same target part within a preset time window, and generate the online fault diagnosis result of the target part.

[0007] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described online turbine fault diagnosis method based on voiceprint recognition.

[0008] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described online turbine fault diagnosis method based on voiceprint recognition.

[0009] By employing the above technical solutions, this application provides a method and device for online fault diagnosis of steam turbines based on voiceprint recognition, a storage medium, and a computer device. This method collects broadband sound pressure signals from a non-contact microphone array positioned around the steam turbine and performs voiceprint recognition. It eliminates the need for contact installation of sensors on high-temperature and high-pressure components, fundamentally eliminating the impact on equipment safety and integrity, and overcoming the limitations of vibration monitoring methods in high-temperature and high-pressure areas. Furthermore, by extracting frequency domain features, time-frequency features, and higher-order statistical features and fusing them into a fused feature vector, it can capture weak abnormal signs even when vibration and thermodynamic parameters have not yet changed significantly. It exhibits extremely high sensitivity to early faults such as blade microcracks and valve microleakage, compensating for the insensitivity of vibration monitoring methods to early microcracks and the lag in response of thermodynamic parameter analysis methods. By comprehensively judging the confidence of each effective signal segment within a preset time window at the same target location, the false alarm rate is significantly reduced, preventing normal steam flow fluctuations under varying load conditions from being misjudged as impacts and thus eliminated, effectively ensuring the reliability of diagnostic results across the entire operating range. The entire acquisition, processing, and diagnostic process is completed online continuously while the unit is running, without the need for shutdown, achieving early, online, non-contact, and highly reliable diagnosis of turbine faults in deep peak-shaving scenarios.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating an online turbine fault diagnosis method based on voiceprint recognition provided in an embodiment of this application is shown. Figure 2 This illustration shows a schematic diagram of the process for generating online fault diagnosis results according to an embodiment of this application; Figure 3 This illustration shows a structural schematic diagram of an online turbine fault diagnosis device based on voiceprint recognition provided in an embodiment of this application; Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] This embodiment provides an online fault diagnosis method for steam turbines based on voiceprint recognition, such as... Figure 1 As shown, the method includes: Step 101: Acquire broadband sound pressure signals during turbine operation using a non-contact microphone array located at the target location of the turbine. The target location includes each cylinder and connecting component of the turbine, and the microphone array includes multiple microphones with different sensitivities.

[0014] Step 102: Perform frame segmentation on the broadband sound pressure signal to obtain multiple signal segments, and remove noise segments and non-steady-state impact segments from the multiple signal segments to obtain multiple effective signal segments.

[0015] Step 103: For each valid signal segment, extract the frequency domain features, time-frequency features, and higher-order statistical features that constitute the acoustic signature from the valid signal segment. Generate a fused feature vector based on the frequency domain features, the time-frequency features, and the higher-order statistical features. Input the fused feature vector into a pre-trained fault classification model and output the confidence level of each fault type corresponding to the valid signal segment.

[0016] Step 104: For each valid signal segment corresponding to each fault type at the same target location, a comprehensive judgment is made within a preset time window to generate an online fault diagnosis result for the target location.

[0017] This application provides an online fault diagnosis method for steam turbines based on voiceprint recognition. First, it collects wideband boost signals during steam turbine operation. Instead of placing a single microphone at a single location on the steam turbine, this application uses non-contact microphone arrays at each cylinder and connecting component. Cylinders include high-pressure, intermediate-pressure, and low-pressure cylinders, while connecting components include main steam valves and regulating valves connected to the cylinders via flanges, bolts, and steam pipes. Each array contains multiple microphones with different sensitivities, aiming to cover a wide frequency range from low to high frequencies. "Non-contact" is used because contact sensors cannot be installed on the steam turbine itself, especially in high-temperature and high-pressure areas. The microphones only need to be positioned 1-2 meters outside the equipment to collect signals without affecting equipment safety. Simultaneously, continuous sampling at a frequency of at least 44.1 kHz satisfies the Nyquist theorem, capturing the high-frequency acoustic components generated during steam turbine operation and providing sufficient information for subsequent analysis.

[0018] The raw broadband sound pressure signal acquired cannot be directly used for analysis because it is mixed with periodic power frequency interference from various oil pumps and other rotating machinery on site, random impulse noise in the environment, and unsteady-state impact segments generated during the unit's operation (such as valve opening and closing). These need to be removed before feature extraction. Specifically, the continuous broadband sound pressure signal is first divided into multiple segments according to time windows, allowing the computer to process each segment. For each signal segment, filtering and denoising methods are used to initially purify the signal. Then, endpoint detection is used to automatically identify which segments are valid sound signals and which are noise or impact segments that need to be removed. Finally, what remains are relatively clean, valid signal segments that can truly reflect the turbine's operating state. The filtering and denoising methods can include bandpass filters to remove ultra-high frequency circuit noise and adaptive filters to suppress periodic background noise. Endpoint detection involves calculating the short-time energy and short-time zero-crossing rate of each segment.

[0019] Next, the effective signal segments are converted into characteristic parameters that can characterize the device status, which is to construct the acoustic signature. This step extracts features in parallel from three dimensions: First, frequency domain features. By performing a Fast Fourier Transform (FFT) on each signal segment to obtain the power spectrum, the energy of the blade passing frequency (BPF, i.e., the number of blades multiplied by the rotational frequency) and its harmonics, as well as the energy distribution of the sidebands near the BPF, are extracted. When a blade cracks, its stiffness decreases, its natural frequency changes, and sideband modulation occurs near the BPF. Second, time-frequency features, namely Mel-frequency cepstral coefficients (MFCC), simulate the auditory characteristics of the human ear. The power spectrum is mapped onto the Mel scale before extracting the coefficients. It is very sensitive to the texture details of sound. When a valve leaks steam, high-pressure gas is ejected through a tiny gap, generating high-frequency turbulent noise. Its MFCC texture is significantly different from normal steam flow sound. Third, higher-order statistical features, namely the bispectrum and its slices, are obtained by calculating the third-order cumulant of the signal. The bispectrum can effectively suppress Gaussian noise and preserve the phase information of the signal. It is specifically used to detect nonlinear phase coupling phenomena. The presence of cracks will cause the system to exhibit nonlinear characteristics, presenting specific peaks in the bispectrum. The three elements characterize the equipment status from three different physical dimensions: mechanical cycle signal, turbulent texture, and nonlinear coupling. After fusion, they form a fusion feature vector containing rich information, which is then input into a pre-trained deep convolutional neural network (CNN) or support vector machine (SVM) classification model. The model outputs the confidence level of the signal segment corresponding to various types of faults.

[0020] If the confidence scores of each valid signal segment are used individually, they are easily susceptible to false alarms due to momentary interference in the field. Therefore, the final step does not directly use the confidence score of a single frame as the final conclusion. Instead, it aggregates the confidence scores of all valid signal segments of the same target location within a time window for a comprehensive judgment. The final diagnostic result is a conclusion based on a comprehensive consideration of all evidence within the time window, which is more reliable than single-frame output.

[0021] By applying the technical solution of this embodiment, a non-contact microphone array installed around the turbine is used to collect broadband sound pressure signals and perform voiceprint recognition. This eliminates the need for physical sensors on high-temperature, high-pressure components, fundamentally eliminating any impact on equipment safety and integrity, and overcoming the limitations of vibration monitoring methods in high-temperature, high-pressure areas. Furthermore, by extracting frequency domain features, time-frequency features, and higher-order statistical features and fusing them into a fused feature vector, it can capture subtle abnormal signs even before significant changes in vibration and thermodynamic parameters, thus detecting early faults such as microcracks in blades and minor leaks in valves. It has extremely high sensitivity, making up for the insensitivity of vibration monitoring methods to early micro-cracks and the lag in response of thermodynamic parameter analysis methods. By comprehensively judging the confidence of each effective signal segment within a preset time window at the same target location, the false alarm rate is greatly reduced, and normal steam flow fluctuations under variable load conditions are not mistakenly judged as impacts and rejected, effectively ensuring the reliability of diagnostic results across the entire operating range. The entire acquisition, processing and diagnosis process is completed online continuously while the unit is running, without the need for shutdown, realizing early, online, non-contact and highly reliable diagnosis of turbine faults in deep peak shaving scenarios.

[0022] Optionally, in this embodiment of the application, step 102, "performing frame segmentation of the broadband sound pressure signal to obtain multiple signal segments," includes: calculating the short-time energy mean and the short-time zero-crossing rate mean of the broadband sound pressure signal; determining the signal-to-noise ratio (SNR) level of the broadband sound pressure signal based on the short-time energy mean and the short-time zero-crossing rate mean; determining the corresponding frame duration and frame overlap rate based on the SNR level; performing frame segmentation of the broadband sound pressure signal using the frame duration and the frame overlap rate; and adding a Hamming window to each signal segment obtained after frame segmentation to obtain multiple signal segments.

[0023] In this embodiment, the signal-to-noise ratio (SNR) of the original broadband sound pressure signal is not constant. Under stable load, the background noise of the steam turbine is relatively stable, resulting in a good acoustic environment. However, during start-up, shutdown, or load changes, interference such as steam flow noise and valve operation noise significantly increases, leading to a decrease in signal quality. To adapt to this dynamically changing acoustic environment, this embodiment assesses the quality of the current signal before framing. Specifically, it calculates the average short-time energy and average short-time zero-crossing rate of the signal segment. The short-time energy reflects the average intensity of the signal within the time window, and the short-time zero-crossing rate reflects the frequency with which the signal waveform crosses zero levels. Combining these two values ​​allows for a rough estimate of the SNR level of the current signal interval, i.e., the relative strength of the effective signal components and the background noise components. A high SNR means that fault characteristics are clearly identifiable, while a low SNR means that the characteristics are submerged in strong background noise, requiring different processing strategies in subsequent stages.

[0024] After determining the current signal-to-noise ratio (SNR) level, the framing parameters are dynamically selected accordingly. If the current SNR is high, it indicates a good acoustic environment and that the effective signal components dominate. In this case, a longer frame length and a lower overlap rate can be used for framing. Longer frames provide higher frequency resolution, helping to accurately capture subtle frequency shifts in mechanical cycle signals, while a lower overlap rate reduces inter-frame redundant calculations and improves overall processing efficiency. Conversely, if the current SNR is low, it indicates strong background noise, and the fault signal may be partially masked. In this case, a shorter frame length and a higher overlap rate can be used. Shorter frames improve temporal resolution, enabling more sensitive capture of transient fault signals (such as the weak impact response caused by blade cracks), while a higher overlap rate increases inter-frame redundancy, allowing subsequent feature extraction to enhance weak features and suppress random noise through information complementarity between adjacent frames. This adaptive strategy ensures that the most favorable analysis unit for fault identification is obtained under different signal quality conditions.

[0025] After framing, a Hamming window can be applied to each signal segment. The Hamming window is a widely used window function in signal processing. Its core function is to smoothly weight the data at both ends before performing a Fast Fourier Transform (FFT), causing the signal amplitude to gradually decay to near zero at the beginning and end of the frame, thereby suppressing spectral leakage introduced by truncation. After windowing, the beginning and end of each frame are smoothly transitioned, spectral leakage is effectively suppressed, and the subsequently extracted frequency domain features are therefore purer and more reliable.

[0026] This application's embodiments employ short frames combined with high overlap rates at low signal-to-noise ratios (SNR), achieving a better balance between temporal resolution and noise suppression. This allows early, weak fault signals that might otherwise be masked by strong background noise to still be captured, enhancing the algorithm's robustness in harsh acoustic environments. At high SNRs, long frames combined with low overlap rates are used, significantly reducing redundant computation and substantially improving online processing efficiency. This enables the system to handle denser acquisition channels or shorter analysis cycles with the same computing resources. The selection of all framing parameters is automatically completed based entirely on the current signal quality, requiring no manual intervention. It can adaptively adjust when switching between different operating conditions, working with subsequent stages to ensure stable and reliable diagnostic performance across all operating conditions.

[0027] Optionally, in this embodiment, step 102, "removing noise segments and non-steady-state impulse segments from the plurality of signal segments to obtain a plurality of valid signal segments," includes: for each signal segment, extracting the spectral flatness of the signal segment; when the spectral flatness exceeds a preset flatness threshold, determining the signal segment as a noise segment and removing it; for each remaining signal segment not determined as a noise segment, calculating the short-time energy change rate and short-time zero-crossing rate change rate of the remaining signal segment in the time domain; when the short-time energy change rate of the remaining signal segment exceeds a first change rate threshold and the short-time zero-crossing rate change rate exceeds a second change rate threshold, determining the remaining signal segment as a candidate impulse segment; detecting the peak factor and impulse factor of the time-domain waveform of the candidate impulse segment; when the peak factor exceeds a peak factor threshold and the impulse factor exceeds an impulse factor threshold, confirming the candidate impulse segment as a non-steady-state impulse segment and removing it; and using the signal segments after removal as valid signal segments.

[0028] In this embodiment, the signal segments obtained after framing and windowing still contain a large amount of data that should not be included in subsequent diagnostic analysis. One type is pure noise segments, such as background noise collected when there is no effective sound source on site, and the other type is non-steady-state impact segments, such as instantaneous impact sounds generated by valve switching operations. Both need to be removed before feature extraction. Specifically, the first step is to remove pure noise segments using spectral flatness. Spectral flatness is calculated by taking the ratio of the geometric mean to the arithmetic mean of the signal power spectrum. When the spectral flatness of a certain signal exceeds a preset flatness threshold, it indicates that the segment lacks meaningful mechanical feature components and is essentially pure noise, which can be directly removed without entering the subsequent diagnostic process.

[0029] After spectral flatness screening, the remaining signal segments may still contain some high-energy impulse interferences. These impulses manifest as dramatic, instantaneous energy jumps in the time domain and may overlap with certain fault characteristics in the frequency domain, making them difficult to distinguish using spectral flatness alone. In this case, the short-time energy change rate and short-time zero-crossing rate change rate are calculated for each remaining signal segment. When both exceed their respective rate-of-change thresholds, the segment is marked as a candidate impulse segment, meaning it possesses the time-domain characteristics of an impulse, but cannot yet be definitively confirmed.

[0030] For signals marked as candidate impact segments, further confirmation is needed to verify whether their waveform morphology truly conforms to the characteristics of mechanical impact. At this point, two waveform morphology indicators are introduced: the peak factor and the impulse factor. The peak factor is the ratio of the peak value to the effective value (root mean square value) of the signal's time-domain waveform, reflecting the steepness of the waveform; the impulse factor is the ratio of the peak value to the absolute average value, reflecting the presence of short-duration high-energy pulses in the signal. Mechanical impact manifests in the time domain as a sharp, instantaneous rise and fall of pulses, and its peak factor and impulse factor will be significantly higher than those of normal steam flow fluctuation signals. Therefore, only when both indicators simultaneously exceed their respective thresholds is the segment ultimately confirmed as a non-steady-state impact segment and eliminated.

[0031] The embodiments of this application conduct step-by-step checks from three different dimensions: frequency domain morphology, time domain fluctuation amplitude, and time domain waveform steepness. This effectively avoids normal steam flow disturbances being misjudged as impacts and ensures that genuine mechanical impacts are not missed. As a result, the quality of the final retained effective signal segments is significantly better than that of single-level screening schemes, providing a cleaner and more reliable data foundation for subsequent feature extraction and fault diagnosis.

[0032] Optionally, extracting higher-order statistical features from the effective signal segment includes: centering the effective signal segment to remove the DC component and obtain a zero-mean signal segment; calculating the third-order cumulant of the zero-mean signal segment and performing a two-dimensional Fourier transform on the third-order cumulant to obtain a bispectral image of the effective signal segment; determining at least one feature extraction region on the bispectral image based on the blade passage frequency and its harmonics in the effective signal segment, wherein the feature extraction region corresponds to the nonlinear phase coupling frequency band between the blade passage frequency and its harmonics; extracting bispectral slices along the diagonal direction within the feature extraction region, calculating the amplitude integral or peak amplitude of the bispectral slices, and using the calculation result as the higher-order statistical features.

[0033] In this embodiment, before extracting higher-order statistical features, the effective signal segment is first centered by subtracting its DC component from the original signal. The DC component is essentially the average value of the signal on the time axis; it reflects the static offset of the signal but does not carry fault-related dynamic information. In subsequent third-order cumulant calculations, the DC component generates cumulant peaks near the zero-frequency point, interfering with the observation of the true nonlinear coupling features. By removing the DC component to obtain a zero-mean signal segment, subsequent cumulant calculations can focus on the signal's fluctuation components, thereby eliminating the interference of static offset on bispectral analysis and ensuring that the extracted features are purely dynamic nonlinear information.

[0034] After centering, bispectral analysis is performed on the zero-mean signal segment. First, the third-order cumulant of the segment is calculated, and then a two-dimensional Fourier transform is performed on the cumulant to obtain the bispectral graph. The bispectral graph can detect nonlinear phase coupling phenomena in signals: when a system exhibits nonlinearity (such as nonlinear stiffness caused by blade cracks), two frequency components can couple nonlinearly to generate a third frequency component (whose frequency is the sum or difference of the two), which appears as a peak at a definite location on the bispectral graph; linear systems or simple Gaussian noise do not produce this coupling structure and cannot form a corresponding peak in the bispectral domain. It is this characteristic that allows the bispectral graph to specifically capture the nonlinear features introduced by faults such as blade cracks, complementing both frequency domain features and time-frequency features.

[0035] Next, the actual blade passing frequencies and their harmonics in the current signal segment are identified. Then, one or more feature extraction regions are delineated on the bispectral graph centered on these frequency positions. This is because the turbine speed changes under varying load conditions, and the blade passing frequencies drift accordingly. If a fixed frequency range is used for feature extraction, the BPF will drift out of the fixed range under varying load, and the features will be lost. Dynamic tracking ensures that nonlinear coupling information related to blade faults can be accurately captured at any speed.

[0036] After determining the feature extraction region, bispectral slices are extracted along the diagonal direction within this region. The diagonal direction refers to the direction where frequencies f1 and f2 are equal. Secondary phase coupling exhibits a peak distribution along the frequency diagonal (satisfying f1 + f2 = f3) in the frequency domain of the third-order cumulant. Extracting slices along this direction best characterizes the intensity distribution of nonlinear coupling. After slice extraction, the amplitude integral or peak amplitude of the slice curve is calculated, and the result is used as a higher-order statistical feature of the effective signal segment.

[0037] In this embodiment, the dynamic positioning feature extraction region solves the problem of fixed-band failure caused by BPF drift under variable load conditions, ensuring adaptability across all operating conditions. The dimensionality reduction strategy of diagonal slicing with amplitude integration / peak amplitude condenses the two-dimensional bispectral image into one-dimensional feature values, preserving core nonlinear coupling information while reducing the input dimensionality of subsequent classification models. This embodiment enables the final output high-order statistical features to accurately reflect the weak nonlinear characteristics caused by faults such as blade cracks under strong background noise, effectively supporting early fault identification.

[0038] In the embodiments of this application, optionally, as shown... Figure 2 As shown, step 104 includes: Step 104-1: Obtain the confidence sequence of various faults corresponding to each effective signal segment of the same target location within a preset time window.

[0039] Step 104-2: For each type of fault, calculate the mean confidence level of the corresponding confidence level sequence, the number of consecutive segments with confidence levels exceeding the first confidence level threshold, and the rate of increase of confidence level along the time axis.

[0040] Step 104-3: When the average confidence level of any type of fault exceeds the second confidence level threshold, the number of consecutive segments exceeds the number of segments threshold, and the upward trend rate exceeds the trend rate threshold, it is determined that the target part has the type of fault, wherein the second confidence level threshold is greater than the first confidence level threshold.

[0041] Step 104-4: Based on the judgment results of various faults, generate online fault diagnosis results for the target part.

[0042] In this embodiment, the fault classification model outputs a set of confidence levels independently for each valid signal segment, such as corresponding to various faults like blade cracks, vibration faults, and valve leaks. In other words, each frame has a judgment on "what fault it currently resembles." However, the confidence level of a single frame is unreliable. Occasional noise spikes or brief steam flow fluctuations can cause the confidence level of a particular fault in a frame to rise instantaneously. If alarms were triggered independently frame by frame, a large number of false alarms would occur. Therefore, this embodiment summarizes the confidence levels of all valid signal segments for the same target location within a preset time window according to fault type, forming a confidence level sequence corresponding to each type of fault. This sequence reflects the complete change trajectory of the confidence level for that type of fault within the time window, providing a data foundation for subsequent multi-dimensional comprehensive judgment.

[0043] Next, statistics were calculated from three different dimensions. The first dimension was the mean confidence level; the second dimension was the number of consecutive segments with confidence levels exceeding the first confidence threshold; and the third dimension was the rate of increase in confidence level over time. Some interference may persist, causing the mean confidence level to be higher. However, the source of interference differs from the actual fault; its confidence level usually fluctuates randomly rather than consistently exceeding the threshold. Therefore, the number of consecutive segments exceeding the first confidence threshold can be counted to determine whether the fault symptoms have temporal continuity. The rate of increase in confidence level over time reflects the direction of change in confidence level within a time window. When a real fault occurs, the voiceprint characteristics often show a gradual deterioration, and the confidence level will increase over time. While persistent interference or stable noise may occasionally cause the confidence level to be higher, it will not show a continuously increasing evolutionary direction.

[0044] After obtaining the three statistics, a decision-making strategy based on the simultaneous fulfillment of three conditions is adopted. A fault of a certain type is determined to have occurred in the target area only when the mean confidence level of a certain type of fault exceeds the second confidence threshold, the number of consecutive segments exceeding the threshold exceeds the segment count threshold, and the upward trend rate exceeds the trend rate threshold. The second confidence threshold is greater than the first confidence threshold. The first confidence threshold is used to determine whether a single frame is worthy of attention, with a lower threshold to ensure sensitivity; the second confidence threshold is used to determine whether the overall mean is high enough, with a higher threshold to ensure the reliability of the overall diagnostic tendency.

[0045] Compared to traditional single-frame judgment or simple majority voting, this application's embodiments construct a triple gating mechanism of "mean compliance + continuous compliance + trend compliance," which significantly reduces the false alarm rate through three progressive layers. At the same time, the judgment is based on the confidence statistics of all valid signal segments within the time window, without relying on the output of any single frame, thus ensuring the stability of the diagnostic results. In particular, the introduction of the upward trend rate can issue an early warning when the fault is still in the early climbing stage, truly achieving the purpose of early identification.

[0046] Optionally, after step 104, the method further includes: pushing the online diagnostic results of the target location fault, each fault type and its corresponding confidence level to the human-machine interface for display, and simultaneously displaying the acoustic signature of the target location on the human-machine interface; when the confidence level of any fault type exceeds a preset alarm threshold, displaying alarm information on the human-machine interface, and automatically capturing and saving the original broadband sound pressure signal and the corresponding fused feature vector within a preset time before and after the alarm time, marking them as associated alarm records, wherein the alarm information includes the fault type, occurrence time, confidence level and the suggested inspection location corresponding to the fault type.

[0047] In this embodiment, after generating the online fault diagnosis results, the online fault diagnosis results, the confidence levels corresponding to various faults, and the acoustic signature of the target location are all pushed to the human-machine interface for display. The online fault diagnosis results provide a clear fault type judgment, the confidence levels allow maintenance personnel to understand the reliability of this judgment, and the acoustic signature visually displays the acoustic signature pattern of the current signal. The combination of these three elements allows maintenance personnel not only to know "what the problem is," but also "why this conclusion was reached," providing a complete visual basis for subsequent review and maintenance decisions.

[0048] When the confidence level of a certain type of fault exceeds the preset alarm threshold, it indicates that the symptoms of the fault are significant enough to warrant proactive alerting maintenance personnel. At this point, an alarm message is displayed on the human-machine interface, containing key information such as the fault type, occurrence time, confidence level, and suggested inspection locations corresponding to that fault type. Simultaneously, a data retention mechanism is activated, automatically capturing and saving the original broadband sound pressure level signal and its corresponding fused feature vector within a preset timeframe before and after the alarm. This saved data is then marked as an associated alarm record for that specific alarm. The original broadband sound pressure level signal retains the most complete acoustic information, allowing for future review or reanalysis; the fused feature vector retains the feature data upon which the model made its judgment, facilitating subsequent investigation to determine if the model misjudged due to feature anomalies. The combination of these two elements provides a traceable and reviewable complete data package for the alarm event, enabling maintenance personnel to retrieve the record for verification, determine the accuracy of the alarm, confirm the existence of the fault, and thus decide on subsequent maintenance strategies.

[0049] This application embodiment simultaneously displays the acoustic signature spectrum while showing the diagnostic results, enabling maintenance personnel to intuitively understand the fault criteria and enhancing the interpretability and credibility of the diagnostic results. The alarm information is associated with the suggested inspection parts corresponding to the fault, providing maintenance personnel with a clear direction for handling and shortening the reaction time from alarm to maintenance decision. The original broadband sound pressure signal and fused feature vector before and after the alarm are automatically retained, establishing a traceable and complete data archive for each alarm, which is convenient for maintenance personnel to review and verify afterward. It can also accumulate real case samples for continuous optimization of the fault classification model, enabling the diagnostic system to have a closed-loop capability from problem discovery and problem tracing to continuous optimization.

[0050] Optionally, after the associated alarm record is marked, the method further includes: when the fault type is blade crack, outputting an ultrasonic phased array flaw detection verification command, and using the flaw detection verification result and the associated alarm record of the blade crack as new training samples to update the fault classification model based on the new training samples; when the fault type is vibration fault, outputting a verification command to review the vibration monitoring system data, and using the review result and the associated alarm record of the vibration fault as new training samples to update the fault classification model based on the new training samples; when the fault type is valve leakage, outputting a verification command for on-site confirmation by infrared thermal imaging or a stethoscope, and using the confirmation result and the associated alarm record of the valve leakage as new training samples to update the fault classification model based on the new training samples.

[0051] In this embodiment, when a blade crack is detected at the target location and an ultrasonic phased array flaw detection verification command is output, the corresponding stage of the blade can be subjected to ultrasonic phased array flaw detection during shutdown maintenance in response to this command. After the flaw detection is completed, the verification results (crack presence / absence, crack size, etc.) are returned to the system. The system pairs the "fusion feature vector at the alarm time" with the "flaw detection verification result" and adds it to the training set as a new training sample with a definite conclusion. Ultrasonic phased array is chosen here instead of ordinary ultrasonic flaw detection because turbine blades are located inside the cylinder block where space is limited. Traditional single-crystal probes require mechanical scanning, which is difficult to meet the coverage requirements in narrow spaces. Phased array can achieve large-area coverage through electronic scanning, making it more suitable for in-situ rapid detection applications.

[0052] When a vibration fault is detected in a target area, a completely independent verification method is not directly provided. Instead, a verification command to check the vibration monitoring system data is output, which is fundamentally different from checking for blade cracks or valve leaks. Steam turbine units typically have vibration monitoring systems (such as bearing vibration sensors). When a vibration fault is reported, the synchronous data from the vibration monitoring system should exhibit corresponding abnormal characteristics. Upon receiving this command, historical data from the vibration monitoring system for the same period can be retrieved to check for vibration characteristics corresponding to imbalance, misalignment, or oil film oscillation. After verification, the verification results (whether the vibration characteristics match the audible alarm, the degree of deviation, etc.) are returned to the system, forming new training samples to update the fault classification model.

[0053] When a valve leak is detected at the target location, a verification command is output for on-site confirmation using infrared thermal imaging or a stethoscope. Infrared thermal imaging detects the temperature distribution on the valve and pipe surfaces to determine if a high-temperature gas leak exists (an abnormally high-temperature area will appear near the leak point). A stethoscope involves maintenance personnel placing one end of a metal probe against the valve surface and the other end close to their ear, listening to the sound of internal steam flow to determine if a leak is present. Both are commonly used on-site valve leak detection methods in power plants. After on-site confirmation, the result (leakage / no leakage) is returned to the system and used as new training samples to update the model.

[0054] This application's embodiments associate three types of verification methods with corresponding fault types and uniformly transform the verification results into training samples. This enables the fault classification model to perform incremental learning using the verification feedback from each alarm, overcoming the shortcomings of traditional systems where the model becomes fixed after deployment and cannot learn from new data. This achieves online self-optimization capabilities of "deployment as continuous evolution." Furthermore, this closed-loop mechanism also makes this application's embodiments more adaptable to deep peak-shaving scenarios.

[0055] Furthermore, as Figure 1 In terms of specific implementation, this application provides an online turbine fault diagnosis device based on voiceprint recognition, such as... Figure 3 As shown, the device includes: The signal acquisition module is used to acquire broadband sound pressure signals during the operation of the steam turbine by using a non-contact microphone array set at the target part of the steam turbine. The target part includes each cylinder and connecting component of the steam turbine, and the microphone array includes multiple microphones with different sensitivities. The frame segmentation processing module is used to perform frame segmentation processing on the broadband sound pressure signal to obtain multiple signal segments, and to remove noise segments and non-steady-state impact segments from the multiple signal segments to obtain multiple effective signal segments. The feature extraction module is used to extract frequency domain features, time-frequency features, and higher-order statistical features that constitute the acoustic signature for each valid signal segment. Based on the frequency domain features, time-frequency features, and higher-order statistical features, a fused feature vector is generated. The fused feature vector is input into a pre-trained fault classification model, and the confidence level of each fault type corresponding to the valid signal segment is output. The fault diagnosis module is used to comprehensively determine the confidence level of each fault type corresponding to each effective signal segment of the same target part within a preset time window, and generate the online fault diagnosis result of the target part.

[0056] Optionally, the frame processing module is used to: Calculate the short-time average energy and short-time average zero-crossing rate of the broadband sound pressure signal, and determine the signal-to-noise ratio level of the broadband sound pressure signal based on the short-time average energy and the short-time average zero-crossing rate; Based on the signal-to-noise ratio level, the corresponding frame duration and frame overlap rate are determined. The broadband sound pressure signal is then framed using the frame duration and frame overlap rate. Hamming windows are then added to each signal segment obtained after framing to obtain multiple signal segments.

[0057] Optionally, the frame processing module is further configured to: For each signal segment, the spectral flatness of the signal segment is extracted. When the spectral flatness exceeds a preset flatness threshold, the signal segment is determined to be a noise segment and is removed. For each remaining signal segment not identified as a noise segment, the short-time energy change rate and short-time zero-crossing rate change rate of the remaining signal segment in the time domain are calculated. When the short-time energy change rate of the remaining signal segment exceeds a first change rate threshold and the short-time zero-crossing rate change rate exceeds a second change rate threshold, the remaining signal segment is identified as a candidate impulse segment. The peak factor and impulse factor of the time domain waveform of the candidate impulse segment are detected. When the peak factor exceeds a peak factor threshold and the impulse factor exceeds an impulse factor threshold, the candidate impulse segment is confirmed as an unsteady impulse segment and is eliminated. The signal segments that have been removed are considered valid signal segments.

[0058] Optionally, the feature extraction module is used for: The effective signal segment is centered to remove the DC component, resulting in a zero-mean signal segment; Calculate the third-order cumulant of the zero-mean signal segment, and perform a two-dimensional Fourier transform on the third-order cumulant to obtain the bispectral map of the effective signal segment. Based on the blade passage frequency and its harmonics in the effective signal segment, at least one feature extraction region on the bispectral image is determined, and the feature extraction region corresponds to the nonlinear phase coupling frequency band between the blade passage frequency and its harmonics. Bispectral slices are extracted along the diagonal direction within the feature extraction region. The amplitude integral or peak amplitude of the bispectral slices is calculated, and the calculation result is used as the higher-order statistical feature.

[0059] Optionally, the fault diagnosis module is used for: Obtain the confidence sequence of various faults corresponding to each valid signal segment of the same target location within a preset time window; For each type of fault, calculate the mean confidence level of the corresponding confidence level sequence, the number of consecutive segments with confidence levels exceeding the first confidence level threshold, and the rate of increase of confidence level along the time axis. When the average confidence level of any type of fault exceeds the second confidence level threshold, the number of consecutive segments exceeds the number of segments threshold, and the upward trend rate exceeds the trend rate threshold, it is determined that the target part has the type of fault, wherein the second confidence level threshold is greater than the first confidence level threshold. Based on the judgment results of various faults, online fault diagnosis results for the target part are generated.

[0060] Optionally, the device further includes a result output module; the result output module is used for: After generating the online fault diagnosis result of the target part, the online fault diagnosis result of the target part, each fault type and the corresponding confidence level are pushed to the human-computer interaction interface for display, and the voiceprint spectrum of the target part is displayed synchronously on the human-computer interaction interface. When the confidence level of any fault type exceeds the preset alarm threshold, alarm information is displayed on the human-machine interface, and the original broadband sound pressure signal and the corresponding fused feature vector within a preset time before and after the alarm time are automatically captured and saved and marked as associated alarm records. The alarm information includes the fault type, occurrence time, confidence level and suggested inspection location corresponding to the fault type.

[0061] Optionally, the result output module is further configured to: After the associated alarm record is marked, when the fault type is blade crack, an ultrasonic phased array flaw detection verification command is output, and the flaw detection verification result and the associated alarm record of the blade crack are used as new training samples to update the fault classification model based on the new training samples. When the fault type is a vibration fault, a verification command for reviewing the vibration monitoring system data is output, and the verification result and the associated alarm record of the vibration fault are used as new training samples to update the fault classification model based on the new training samples. When the fault type is valve leakage, an infrared thermal imaging or on-site confirmation command is output, and the confirmation result and the associated alarm record of the valve leakage are used as new training samples to update the fault classification model based on the new training samples.

[0062] It should be noted that other corresponding descriptions of the functional units involved in the online turbine fault diagnosis device based on voiceprint recognition provided in this application embodiment can be found in the following references. Figures 1 to 2 The corresponding descriptions in the method will not be repeated here.

[0063] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 4 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0064] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0065] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0066] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0067] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0068] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for online diagnosis of steam turbine faults based on voiceprint recognition, characterized in that, include: A non-contact microphone array is installed at the target location of the steam turbine to collect broadband sound pressure signals during the operation of the steam turbine. The target location includes each cylinder and connecting component of the steam turbine. The microphone array includes multiple microphones with different sensitivities. The broadband sound pressure signal is framed to obtain multiple signal segments, and noise segments and non-steady-state impact segments are removed from the multiple signal segments to obtain multiple effective signal segments; For each valid signal segment, frequency domain features, time-frequency features, and higher-order statistical features constituting the acoustic signature are extracted from the valid signal segment. Based on the frequency domain features, time-frequency features, and higher-order statistical features, a fused feature vector is generated. The fused feature vector is input into a pre-trained fault classification model, and the confidence level of each fault type corresponding to the valid signal segment is output. For each valid signal segment corresponding to each fault type at the same target location, a comprehensive judgment is made within a preset time window to generate an online fault diagnosis result for the target location.

2. The method according to claim 1, characterized in that, The process of framing the broadband sound pressure signal yields multiple signal segments, including: Calculate the short-time average energy and short-time average zero-crossing rate of the broadband sound pressure signal, and determine the signal-to-noise ratio level of the broadband sound pressure signal based on the short-time average energy and the short-time average zero-crossing rate; Based on the signal-to-noise ratio level, the corresponding frame duration and frame overlap rate are determined. The broadband sound pressure signal is then framed using the frame duration and frame overlap rate. Hamming windows are added to each signal segment obtained after framing to obtain multiple signal segments.

3. The method according to claim 1, characterized in that, The process of removing noise segments and non-steady-state impulse segments from the plurality of signal segments to obtain a plurality of valid signal segments includes: For each signal segment, the spectral flatness of the signal segment is extracted. When the spectral flatness exceeds a preset flatness threshold, the signal segment is determined to be a noise segment and is removed. For each remaining signal segment not identified as a noise segment, the short-time energy change rate and short-time zero-crossing rate change rate of the remaining signal segment in the time domain are calculated. When the short-time energy change rate of the remaining signal segment exceeds a first change rate threshold and the short-time zero-crossing rate change rate exceeds a second change rate threshold, the remaining signal segment is identified as a candidate impulse segment. The peak factor and impulse factor of the time domain waveform of the candidate impulse segment are detected. When the peak factor exceeds a peak factor threshold and the impulse factor exceeds an impulse factor threshold, the candidate impulse segment is confirmed as an unsteady impulse segment and is eliminated. The signal segments that have been removed are considered valid signal segments.

4. The method according to claim 1, characterized in that, Extracting high-order statistical features from the effective signal segment includes: The effective signal segment is centered to remove the DC component, resulting in a zero-mean signal segment; Calculate the third-order cumulant of the zero-mean signal segment, and perform a two-dimensional Fourier transform on the third-order cumulant to obtain the bispectral map of the effective signal segment. Based on the blade passage frequency and its harmonics in the effective signal segment, at least one feature extraction region on the bispectral image is determined, and the feature extraction region corresponds to the nonlinear phase coupling frequency band between the blade passage frequency and its harmonics. Bispectral slices are extracted along the diagonal direction within the feature extraction region. The amplitude integral or peak amplitude of the bispectral slices is calculated, and the calculation result is used as the higher-order statistical feature.

5. The method according to claim 1, characterized in that, The confidence levels of each fault type corresponding to each valid signal segment of the same target location are comprehensively determined within a preset time window to generate an online fault diagnosis result for the target location, including: Obtain the confidence sequence of various faults corresponding to each valid signal segment of the same target location within a preset time window; For each type of fault, calculate the mean confidence level of the corresponding confidence level sequence, the number of consecutive segments with confidence levels exceeding the first confidence level threshold, and the rate of increase of confidence level along the time axis. When the average confidence level of any type of fault exceeds the second confidence level threshold, the number of consecutive segments exceeds the number of segments threshold, and the upward trend rate exceeds the trend rate threshold, it is determined that the target part has the type of fault, wherein the second confidence level threshold is greater than the first confidence level threshold. Based on the judgment results of various faults, online fault diagnosis results for the target part are generated.

6. The method according to claim 1, characterized in that, After generating the online fault diagnosis result for the target location, the method further includes: The online diagnostic results of the target part, each fault type and the corresponding confidence level are pushed to the human-computer interaction interface for display, and the voiceprint spectrum of the target part is displayed synchronously on the human-computer interaction interface. When the confidence level of any fault type exceeds the preset alarm threshold, alarm information is displayed on the human-machine interface, and the original broadband sound pressure signal and the corresponding fused feature vector within a preset time before and after the alarm time are automatically captured and saved and marked as associated alarm records. The alarm information includes the fault type, occurrence time, confidence level and suggested inspection location corresponding to the fault type.

7. The method according to claim 6, characterized in that, After the record is marked as an associated alarm record, the method further includes: When the fault type is blade crack, an ultrasonic phased array flaw detection verification command is output, and the flaw detection verification result and the associated alarm record of the blade crack are used as new training samples to update the fault classification model based on the new training samples. When the fault type is a vibration fault, a verification command for reviewing the vibration monitoring system data is output, and the verification result and the associated alarm record of the vibration fault are used as new training samples to update the fault classification model based on the new training samples. When the fault type is valve leakage, an infrared thermal imaging or on-site confirmation command is output, and the confirmation result and the associated alarm record of the valve leakage are used as new training samples to update the fault classification model based on the new training samples.

8. An online fault diagnosis device for steam turbines based on voiceprint recognition, characterized in that, include: The signal acquisition module is used to acquire broadband sound pressure signals during the operation of the steam turbine by using a non-contact microphone array set at the target part of the steam turbine. The target part includes each cylinder and connecting component of the steam turbine, and the microphone array includes multiple microphones with different sensitivities. The frame segmentation processing module is used to perform frame segmentation processing on the broadband sound pressure signal to obtain multiple signal segments, and to remove noise segments and non-steady-state impact segments from the multiple signal segments to obtain multiple effective signal segments. The feature extraction module is used to extract frequency domain features, time-frequency features, and higher-order statistical features that constitute the acoustic signature for each valid signal segment. Based on the frequency domain features, time-frequency features, and higher-order statistical features, a fused feature vector is generated. The fused feature vector is input into a pre-trained fault classification model, and the confidence level of each fault type corresponding to the valid signal segment is output. The fault diagnosis module is used to comprehensively determine the confidence level of each fault type corresponding to each effective signal segment of the same target part within a preset time window, and generate the online fault diagnosis result of the target part.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images