A method for monitoring the working condition of surge pipes of hydropower units based on MFCC and fault model library

Abstract

The present invention relates to a method for monitoring the operating conditions of a surge tank of a hydropower unit based on MFCC and a fault model library. A soundprint sensor is used to collect the soundprint signal of the surge tank. Based on historical data of the soundprint signal, liquid level signal, vibration signal, and pressure pulsation signal of known surge tank faults, a soundprint model library containing typical soundprint features is established for quickly comparing and identifying whether common faults are present under the current operating condition based on the soundprint signal of the current surge tank. Furthermore, a fault model library containing soundprint, liquid level, vibration, and pressure pulsation feature data is constructed for comparing and identifying whether known faults are present under the current operating condition based on the soundprint, liquid level, vibration, and pressure pulsation signals of the current unit. The present invention achieves real-time, online, high-precision, and automatic identification of surge tank faults, significantly improving the efficiency and accuracy of fault diagnosis and reducing downtime and maintenance costs caused by misdiagnosis or missed diagnosis.

Description

Technical Field

[0001] The present invention belongs to the field of safety monitoring of hydropower units, and in particular relates to a method for monitoring the working condition of a surge pipe of a hydropower unit based on MFCC and a fault model library. Background Art

[0002] The surge tank in a hydropower station's roof is a crucial component of the turbine. Its primary function is to balance the accumulated water pressure between the roof and the upper chamber of the runner, reducing the axial hydraulic thrust transmitted to the thrust bearing and ensuring safe and stable operation of the turbine. Vibration of the roof often causes vibration in the surge tank. Long-term effects of this vibration can lead to loosening of the surge tank's bolts, damage to seals, and cracks in welds. Further complicating matters, the water flowing through the surge tank often carries sediment, which, over time, wears the surge tank's inner wall, gradually thinning it. Furthermore, cavitation further exacerbates this damage, particularly at the welds between the surge tank and the roof, or on the surge tank's inner wall, where it can easily form pinholes. All of these factors can lead to surge tank leakage. These leaks not only disrupt the normal operation of the system but, in severe cases, can also threaten the safe and stable operation of the turbine. Currently, addressing surge tank leakage typically requires shutting down the unit and draining the water for repair. This not only complicates and prolongs the repair process, but also significantly impacts production operations.

[0003] Traditional methods for monitoring surge tank leaks rely primarily on manual inspections and periodic checks. Manual inspections identify potential leaks through visual inspection, touch, and listening of the surge tank and its associated equipment. Inspectors require specialized knowledge and experience to make preliminary assessments of the surge tank's operating status based on changes in the equipment's appearance, sound, and temperature. However, this method is affected by the inspector's skill level, accumulated experience, and inspection frequency, making accuracy and timeliness difficult to guarantee. Periodic inspections involve systematic and comprehensive checks of the surge tank at predetermined intervals. These inspections typically include visual inspections, performance tests, and pressure tests to assess the health of the surge tank.

[0004] Regular inspections can improve monitoring accuracy to a certain extent, but they also present some challenges. First, the timing of regular inspections needs to be carefully considered based on a variety of factors, including the equipment's operating conditions and operating environment. If the inspection period is too long, leaks may not be detected promptly; if the period is too short, maintenance costs and workload will increase. Second, regular inspections still cannot monitor the real-time operating conditions of the surge tanks, making it difficult to detect sudden leaks.

[0005] Both manual inspections and regular inspections have relatively long inspection cycles. The frequency of manual inspections is limited by the number of inspectors, their skills, and their work schedules, while regular inspections are constrained by pre-set inspection cycles. These long inspection cycles prevent traditional methods from promptly detecting and addressing water leaks, especially in the early stages of a leak. This delay can lead to the problem worsening due to untimely detection.

[0006] Traditional monitoring methods rely heavily on the experience, skills, and attention of inspectors. Their judgment can be affected by various factors, such as fatigue and lack of concentration, which can lead to missed detections or misjudgments. Furthermore, due to individual differences in perception and judgment, inconsistent detection results can occur between different inspectors.

[0007] In summary, the existing technology lacks a real-time online monitoring method for abnormal operating conditions of a surge tank. Summary of the Invention

[0008] The purpose of the present invention is to address the above problems and provide a method for monitoring the working condition of the surge tank of a hydropower unit based on MFCC and a fault model library, establish a voiceprint model library of common faults containing typical voiceprint features and a fault model library containing multimodal data such as the voiceprint, liquid level, vibration and pressure pulsation signals of the unit, extract feature vectors from the voiceprint signal of the surge tank under the current working condition, and compare them with the common faults in the voiceprint model library to achieve rapid identification of common faults; compare the feature vectors of the voiceprint, liquid level, vibration and pressure pulsation signals under the current working condition with the feature vectors of known faults in the fault model library to achieve automatic identification of known faults that have occurred in the surge tank of the unit top cover; and use multimodal data fusion analysis to improve the accuracy of fault identification and real-time online monitoring of surge tank abnormalities.

[0009] The technical solution of the present invention is a method for monitoring the working condition of a surge pipe of a hydropower unit based on MFCC and a fault model library, comprising the following steps:

[0010] Step 1: Collect the acoustic signal of the equalizing pipe, the liquid level signal at the top cover, the vibration signal of the unit, and the pressure pulsation signal data when the unit is in operation;

[0011] Step 2: Use the Mel-frequency cepstral coefficient method MFCC to extract the feature vector i-Vector from the equalizer tube voiceprint signal;

[0012] Step 2.1: Pre-emphasize the voiceprint signal;

[0013] Step 2.2: Frame and window the pre-emphasized voiceprint signal;

[0014] Step 2.3: Use the Fourier transform method to deconstruct the framed time domain signal into a frequency domain signal;

[0015] Step 2.4: Filter the obtained spectrum through a Mel filter and divide it into multiple Mel filter banks;

[0016] Step 2.5: Endpoint detection, removing invalid data from the signal data;

[0017] Step 2.6: Extract feature vector i-Vector;

[0018] Step 3: Construct a probabilistic linear discriminant analysis model PLDA and train it to obtain the PLDA model parameters;

[0019] Step 4: Using the PLDA model, calculate the similarity between the feature vector i-Vector in step 2 and the feature vectors of typical faults in the voiceprint model library, and score the current working condition of the surge tank;

[0020] Step 5: Preliminarily determine whether the current working condition of the surge tank is normal based on the score of the working condition of the surge tank in step 4. If it is normal, execute step 1 and continue to collect the soundprint, vibration, liquid level and pressure pulsation signal data of the surge tank in the working state; if the working condition of the surge tank is abnormal, execute step 6;

[0021] Step 6: Compare the characteristic data of the soundprint, liquid level, vibration, and pressure pulsation signals under the current working condition with the characteristic data of various faults in the fault model library, calculate the degree of correlation, and determine whether it is a known fault in the fault model library based on the correlation degree. If it is a known fault, issue a fault alarm signal; otherwise, execute step 7;

[0022] Step 7: Based on the liquid level signal data in the top cover obtained in step 1, determine whether the liquid level is normal. If it is normal, proceed to step 8. If it is abnormal, determine whether there is a leak in the top cover or the equalizing pipe based on the image recognition of the top cover;

[0023] Step 8: Based on the vibration signal and pressure pulsation signal data of the unit in the current operating state, determine whether the unit vibration is abnormal. If the unit vibration is abnormal, further determine whether the unit is in a special vibration zone. If so, adjust the unit operating condition. If not, determine whether there are loose bolts, damaged seals, cracked welds, or sand holes and water leakage faults based on image recognition at the top cover and equalizing pipe. If there is no vibration abnormality, proceed to step 9.

[0024] Step 9: Conduct on-site inspection of the unit top cover and pressure-equalizing pipe to determine the fault type, and store the characteristic data of the sound print signal, liquid level signal, vibration signal and pressure pulsation signal of the fault type into the fault model library.

[0025] Preferably, in step 5, before executing step 6, it is first determined and identified whether the fault is a cable breakage based on the voiceprint signal, and whether the fault is the A phase, B phase, or C phase cable based on the position of the voiceprint sensor.

[0026] Preferably, the step 2.6 specifically includes:

[0027] 1) Build and train the Gaussian mixture model-universal background model GMM-UBM;

[0028] 2) Inputting the MFCC feature vector of the voiceprint signal of the surge protector into the Gaussian mixture model-general background model, and obtaining the Gaussian mixture model parameters of the surge protector by the maximum a posteriori probability method;

[0029] 3) Using the universal background model (UBM) and the Baum-Welch statistic, we train the total variation model (TVM). TVM is a low-dimensional subspace model that captures all possible variations in the speech signal.

[0030] 4) Project the Baum-Welch statistic of the equalizer onto the low-dimensional subspace of TVM. The resulting projection vector is the i-Vector of the equalizer.

[0031] Preferably, in step 4, the similarity between the characteristic vector i-Vector in step 2 and the characteristic vector of the typical fault in the voiceprint model library is calculated as follows:

[0032] ;

[0033] In the formula represents the similarity score, The characteristic vector i-Vector representing the voiceprint signal of the equalizing tube, The feature vector representing a typical fault in the voiceprint model library, Expresses the same assumption, that is, and For the same fault type, Indicates different assumptions, that is, and For different fault types; In the same assumption Lower eigenvector and The joint probability density function value of In different assumptions Lower eigenvector The probability density function value of In different assumptions Lower eigenvector The probability density function value of .

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1) This invention constructs a voiceprint model library of common faults containing typical voiceprint features, using Mel frequency

[0036] The cepstral coefficient method extracts feature vectors from the equalizing pipe soundprint signal under the current working condition, and uses the probabilistic linear discriminant analysis model to calculate the similarity between the feature vectors of the soundprint signal under the current working condition and the common faults in the soundprint model library. It identifies and judges whether common faults with typical soundprint characteristics occur under the current working condition, and realizes the online, rapid, and high-precision automatic recognition of common equalizing pipe faults with typical soundprint characteristics.

[0037] 2) The present invention constructs a fault model library of multimodal data such as the unit's soundprint, liquid level, vibration, and pressure pulsation signals. It calculates the degree of correlation between the characteristic vectors of the unit's soundprint, liquid level, vibration, and pressure pulsation signals under the current operating condition and the characteristic vectors of known faults in the fault model library, and determines whether a known fault occurs under the current operating condition. This achieves online, real-time, and high-precision identification and judgment of faults and anomalies in the unit's top cover surge tank.

[0038] 3) After the method of the present invention determines the unknown fault through image recognition, on-site inspection and confirmation, the corresponding unit sound print, liquid level, vibration and pressure pulsation signal characteristic data are incorporated into the fault model library, continuously enriching and improving the content of the fault database, so that the monitoring system based on the present invention can automatically match and identify similar faults in the future more quickly and accurately, significantly improving the efficiency and accuracy of fault diagnosis, and reducing downtime and maintenance costs caused by misdiagnosis or missed diagnosis.

[0039] 4) Based on voiceprint signals, the present invention realizes the rapid and efficient diagnosis of damaged transmission cables of the A, B, and C phases of the unit. At the same time, it sends a cable damage alarm signal to the unit operation and maintenance personnel, facilitating timely treatment measures to avoid affecting the normal power generation of the unit.

[0040] 5) Based on voiceprint recognition technology, the present invention can greatly improve the accuracy and efficiency of identifying abnormal operating conditions of the surge tank through comprehensive judgment and analysis of multimodal data such as images, vibrations, and liquid levels. It can also achieve online real-time monitoring and identification of faults and abnormalities, saving human resources while reducing the time required for fault identification, providing an opportunity for emergency response to unit accidents, and greatly improving the stability of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention will be further described below with reference to the accompanying drawings and examples.

[0042] Figure 1 This is a schematic structural diagram of the unit top cover and the equalizing pipe according to an embodiment of the present invention.

[0043] Figure 2 Schematic diagram of a data acquisition system according to an embodiment of the present invention.

[0044] Figure 3 Schematic diagram of the process of calculating the similarity score of voiceprint signals in an embodiment of the present invention.

[0045] Figure 4 Schematic diagram of the flow of a method for monitoring the working condition of a surge pipe of a hydropower unit according to an embodiment of the present invention.

[0046] Figure 5 Schematic diagram of a hydropower unit surge tank operating condition monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In the embodiment, eight bone conduction microphones 1 are arranged on the wall of the equalizing tube 7, two cameras 2 with thermal infrared temperature measurement function are installed above the equalizing tube 7 and on the inner wall of the top cover 8, two horizontal vibration sensors 3 are set on the inner rib of the top cover 8, two vertical vibration sensors 4 are arranged on the inner rib of the top cover, two water level sensors 5 and two pressure pulsation sensors 6 are installed inside the top cover, one of the pressure pulsation sensors collects the pressure pulsation signal data between the impeller 9 and the movable guide vane 10, and the other pressure pulsation sensor collects the pressure pulsation signal data between the movable guide vane 10 and the fixed guide vane 11, as shown in FIG. Figure 1 shown.

[0048] like Figure 2 As shown, the data from the corresponding sensors is transmitted via a fiber optic channel to the cloud server platform for further analysis and judgment. The network transmission system includes optoelectronic switches, aggregation switches, servers, forward isolation devices, cloud platforms, etc. The voiceprint feature data is equipped with a local control host, which has the function of feature data extraction and model downlink, facilitating the extraction of voiceprint feature data and the judgment of known fault models within the local control host. The thermal infrared camera has temperature measurement and image monitoring functions, and is also equipped with a local control host to facilitate the upload of temperature feature values and video files, and has the function of downlink camera motion control.

[0049] like Figure 3 and Figure 4 As shown in FIG, a method for monitoring the working condition of a surge pipe of a hydropower unit based on MFCC and a fault model library includes:

[0050] Step 1: Collect the acoustic signal of the equalizing pipe, the liquid level signal at the top cover, the vibration signal of the unit, and the pressure pulsation signal data when the unit is in operation;

[0051] Step 2: Use the Mel-frequency cepstral coefficient method MFCC to extract the feature vector i-Vector from the equalizer tube voiceprint signal;

[0052] Step 2.1: Pre-emphasize the voiceprint signal;

[0053] Step 2.2: Frame and window the pre-emphasized voiceprint signal;

[0054] Step 2.3: Use the Fourier transform method to deconstruct the framed time domain signal into a frequency domain signal;

[0055] Step 2.4: Filter the obtained spectrum through a Mel filter and divide it into multiple Mel filter banks;

[0056] Step 2.5: Endpoint detection, removing invalid data from the signal data;

[0057] Step 2.6: Extract feature vector i-Vector;

[0058] Step 3: Construct a probabilistic linear discriminant analysis model PLDA and train it to obtain the PLDA model parameters;

[0059] The probabilistic linear discriminant analysis model PLDA is as follows:

[0060]

[0061] In the formula Indicates the i The target object j Voiceprint sample data; represents the mean of all sample data, reflecting the central tendency of all sample data; F represents the identity space, h i represents the fixed effect part of the voiceprint feature sample data of the i-th target object, that is, the individual characteristics that do not change over time; G represents the error space, w ij It represents the random effect part of the jth sample data of the i-th target object, reflecting the random changes or differences of the target object over time; It represents the noise part of the jth sample data of the i-th target object, representing the random error or unpredictable factors in the observed data.

[0062] Step 4: Using the PLDA model, calculate the similarity between the feature vector i-Vector in step 2 and the feature vector of the typical fault in the voiceprint model library, and score the current working condition of the surge tank. The score calculation formula is:

[0063] ;

[0064] In the formula represents the similarity score, The characteristic vector i-Vector representing the voiceprint signal of the equalizing tube, The feature vector representing a typical fault in the voiceprint model library, Expresses the same assumption, that is, and For the same fault type, Indicates different assumptions, that is, and For different fault types; In the same assumption Lower eigenvector and The joint probability density function value of In different assumptions Lower eigenvector The probability density function value of In different assumptions Lower eigenvector The probability density function value of .

[0065] Step 5: Preliminarily determine whether the current working condition of the surge tank is normal based on the score of the working condition of the surge tank in step 4. If it is normal, execute step 1 and continue to collect the soundprint, vibration, liquid level and pressure pulsation signal data of the surge tank in the working state; if the working condition of the surge tank is abnormal, execute step 6;

[0066] Step 6: Compare the characteristic data of the soundprint, liquid level, vibration, and pressure pulsation signals under the current working condition with the characteristic data of various faults in the fault model library, calculate the degree of correlation, and determine whether it is a known fault in the fault model library based on the correlation degree. If it is a known fault, issue a fault alarm signal; otherwise, execute step 7;

[0067] The calculation formula for the degree of association is:

[0068]

[0069] In the formula The characteristic vector representing the current working state of the surge tank The characteristic vector of the known fault The degree of correlation, is the resolution coefficient;

[0070] The characteristic vector of the known fault is:

[0071] ;

[0072] In the formula Indicates the characteristic data of the smooth pressure tube voiceprint signal. Indicates the characteristic data of the liquid level signal at the top cover, Indicates the characteristic data of the unit's horizontal vibration signal, Indicates the characteristic data of the vertical vibration signal of the unit, Indicates the characteristic data of the pressure pulsation signal between the unit runner and the movable guide vane. Represents characteristic data of pressure pulsation signal between movable guide vanes and fixed guide vanes;

[0073] The eigenvector under the current working condition is:

[0074]

[0075] and Correspondingly, They respectively represent the characteristic data of the sound pattern, liquid level, horizontal vibration, vertical vibration, and pressure pulsation signal under the current working condition.

[0076] Step 7: Based on the liquid level signal data in the top cover obtained in step 1, determine whether the liquid level is normal. If it is normal, proceed to step 8. If it is abnormal, determine whether there is a leak in the top cover or the equalizing pipe based on the image recognition of the top cover;

[0077] Step 8: Based on the vibration signal and pressure pulsation signal data of the unit in the current operating state, determine whether the unit vibration is abnormal. If the unit vibration is abnormal, further determine whether the unit is in a special vibration zone. If so, adjust the unit operating condition. If not, determine whether there are loose bolts, damaged seals, cracked welds, or sand holes and water leakage faults based on image recognition at the top cover and equalizing pipe. If there is no vibration abnormality, proceed to step 9.

[0078] Step 9: Conduct on-site inspection of the unit top cover and pressure-equalizing pipe to determine the fault type, and store the characteristic data of the sound print signal, liquid level signal, vibration signal and pressure pulsation signal of the fault type into the fault model library.

[0079] In the embodiment, before executing step 6, a vector quantization algorithm is first used to determine and identify whether it is a cable damage fault based on the voiceprint feature signal, and determine whether it is a phase A, phase B, or phase C cable fault based on the position of the voiceprint sensor, and send a cable damage alarm signal to the unit operation and maintenance personnel.

[0080] The hydropower unit surge pipe condition monitoring system using the above method is as follows: Figure 5 As shown, it includes an acquisition module, a data network transmission module, a voiceprint training and scoring module, a comprehensive fault judgment module, and a result output execution module.

[0081] The acquisition module is used to collect multimodal data of the unit top cover surge tank under the current operating condition.

[0082] The data network transmission module is used to transmit the collected multimodal data to the cloud server platform.

[0083] The voiceprint training and scoring module is used to calculate the similarity score between the voiceprint signal feature data under the current working condition and the feature data of the voiceprint model library, and identify and judge common faults with typical voiceprint features.

[0084] The comprehensive fault judgment module is used to compare the characteristic data of the unit's sound pattern, liquid level, vibration and pressure pulsation signals under the current operating condition with the fault characteristic data in the fault model library to determine the specific type of fault.

[0085] The result output execution module issues corresponding alarm information or performs corresponding equipment operations based on the faults identified by the voiceprint training and scoring module and the comprehensive fault judgment module.

[0086] The present invention realizes real-time online monitoring of abnormal working conditions of the equalizing pipe through voiceprint recognition technology, filling the industry gap in real-time online monitoring methods for abnormal working conditions of the equalizing pipe; by installing bone conduction microphones, cameras with thermal infrared temperature measurement functions, vibration sensors, water level sensors, pressure pulsation sensors, etc. in appropriate areas, and integrating multimodal data into the same system, comprehensive diagnosis of multimodal abnormal working conditions of the equalizing pipe is realized, greatly improving the timeliness of judgment of abnormal working conditions or faults of the equalizing pipe.

Claims

1. A method for monitoring the working condition of a surge pipe of a hydropower unit based on MFCC and a fault model library, characterized in that: The operating condition monitoring method utilizes a soundprint sensor to collect the soundprint signal of the surge tank. Based on historical data of soundprint signals, liquid level signals, vibration signals, and pressure pulsation signals of known surge tank faults, a soundprint model library of common faults containing typical soundprint features is established. The library is used to quickly compare, identify, and determine whether common faults occur under the current operating condition based on the soundprint signal of the current surge tank. A fault model library is also constructed that contains soundprint, liquid level, vibration, and pressure pulsation feature data. The library is used to compare, identify, and determine whether known faults occur under the current operating condition based on the soundprint, liquid level, vibration, and pressure pulsation signals of the current unit. The working condition monitoring method comprises the following steps: Step 1: Collect the acoustic signal of the equalizing pipe, the liquid level signal at the top cover, the vibration signal of the unit, and the pressure pulsation signal data when the unit is in operation; Step 2: Use the Mel-frequency cepstral coefficient method MFCC to extract the feature vector i-Vector from the equalizer tube voiceprint signal; Step 2.1: Pre-emphasize the voiceprint signal; Step 2.2: Frame the pre-emphasized voiceprint signal; Step 2.3: Use the Fourier transform method to deconstruct the framed time domain signal into a frequency domain signal; Step 2.4: Filter the obtained spectrum through a Mel filter and divide it into multiple Mel filter banks; The filter calculation formula is: ； Where k is the frequency index; m is the serial number of the Mel filter; f is the actual frequency; represents the filter function; Step 2.5: Endpoint detection, removing invalid data from the signal data; Step 2.6: Extract feature vector i-Vector; Step 3: Construct a probabilistic linear discriminant analysis model PLDA and train it to obtain the PLDA model parameters; The probabilistic linear discriminant analysis model PLDA is as follows: ； In the formula Indicates the i The first target object j Voiceprint sample data; represents the mean of all sample data; F represents the identity space, h i represents the fixed effect part of the voiceprint feature sample data of the i-th target object, that is, the individual characteristics that do not change over time; G represents the error space, w ij It represents the random effect part of the jth sample data of the i-th target object, reflecting the random changes or differences of the target object over time; The noise portion of the jth sample data of the i-th target object represents the random error or unpredictable factors in the observed data; Step 4: Using the PLDA model, calculate the similarity between the feature vector i-Vector in step 2 and the feature vectors of typical faults in the voiceprint model library, and score the current working condition of the surge tank; Step 5: Preliminarily determine whether the current working condition of the surge tank is normal based on the score of the working condition of the surge tank in step 4. If it is normal, execute step 1 and continue to collect the soundprint, vibration, liquid level and pressure pulsation signal data of the surge tank in the working state; if the working condition of the surge tank is abnormal, execute step 6; Step 6: Compare the characteristic data of the soundprint, liquid level, vibration, and pressure pulsation signals under the current working condition with the characteristic data of various faults in the fault model library, calculate the degree of correlation, and determine whether it is a known fault in the fault model library based on the correlation degree. If it is a known fault, issue a fault alarm signal; otherwise, execute step 7; Step 7: Based on the liquid level signal data in the top cover obtained in step 1, determine whether the liquid level is normal. If it is normal, proceed to step 8. If it is abnormal, determine whether there is a leak in the top cover or the equalizing pipe based on the image recognition of the top cover; Step 8: Based on the vibration signal and pressure pulsation signal data of the unit in the current operating state, determine whether the unit vibration is abnormal. If the unit vibration is abnormal, further determine whether the unit is in a special vibration zone. If so, adjust the unit operating condition. If not, determine whether there are loose bolts, damaged seals, cracked welds, or sand holes and water leakage faults based on image recognition at the top cover and equalizing pipe. If there is no vibration abnormality, proceed to step 9. Step 9: Conduct on-site inspection of the unit top cover and pressure-equalizing pipe to determine the fault type, and store the characteristic data of the sound print signal, liquid level signal, vibration signal and pressure pulsation signal of the fault type into the fault model library.

2. The method for monitoring the working condition of the surge tank of a hydropower unit according to claim 1, characterized in that: In step 5, before executing step 6, first determine and identify whether it is a cable damage fault based on the voiceprint signal, and determine whether it is a phase A, phase B or phase C cable fault based on the position of the voiceprint sensor, and issue a cable damage alarm signal.

3. The method for monitoring the working condition of the surge pipe of a hydropower unit according to claim 1 or 2, characterized in that: The step 2.6 specifically includes: 1) Build and train the Gaussian mixture model-universal background model GMM-UBM; 2) Inputting the MFCC feature vector of the voiceprint signal of the surge protector into the Gaussian mixture model-general background model, and obtaining the Gaussian mixture model parameters of the surge protector by the maximum a posteriori probability method; 3) Using the universal background model (UBM) and the Baum-Welch statistic, we train the total variation model (TVM). TVM is a low-dimensional subspace model that captures all possible variations in the speech signal. 4) Project the Baum-Welch statistic of the equalizer onto the low-dimensional subspace of TVM. The resulting projection vector is the i-Vector of the equalizer.

4. The method for monitoring the working condition of the surge pipe of a hydropower unit according to claim 3, characterized in that: In step 4, the similarity between the characteristic vector i-Vector in step 2 and the characteristic vector of the typical fault in the voiceprint model library is calculated as follows: ； In the formula represents the similarity score, The characteristic vector i-Vector representing the voiceprint signal of the equalizing tube, The feature vector representing a typical fault in the voiceprint model library, Expresses the same assumption, that is, and For the same fault type, Indicates different assumptions, that is, and For different fault types; In the same assumption Lower eigenvector and The joint probability density function value of In different assumptions Lower eigenvector The probability density function value of In different assumptions Lower eigenvector The probability density function value of .

5. The method for monitoring the working condition of the surge pipe of a hydropower unit according to claim 4, characterized in that: In step 6, the degree of correlation is used to determine whether the fault is a known fault in the fault model library. The calculation formula for the degree of correlation is: ； In the formula The characteristic vector representing the current working state of the surge tank The characteristic vector of the known fault The degree of correlation, is the resolution coefficient; The characteristic vector of the known fault is: ； In the formula Indicates the characteristic data of the smooth pressure tube voiceprint signal. Indicates the characteristic data of the liquid level signal at the top cover, Indicates the characteristic data of the unit's horizontal vibration signal, Indicates the characteristic data of the vertical vibration signal of the unit, Indicates the characteristic data of the pressure pulsation signal between the unit runner and the movable guide vane, Represents the characteristic data of the pressure pulsation signal between the movable guide vanes and the fixed guide vanes.

Images