Accident distributing valve multi-mode working condition abnormity monitoring method based on sound recognition

Abstract

The invention discloses an accident distributing valve multi-mode working condition abnormity monitoring method based on voice recognition. The method comprises the steps that liquid level data, audio data, vibration analog quantity data, temperature measurement image data, active power data and guide vane opening degree data are collected through a comprehensive data collection module; voiceprint recognition related data acquired by the comprehensive data acquisition module is transmitted to a cloud server platform through a network transmission module; identifying and scoring the audio data through a voiceprint scoring module; the fault comprehensive judgment module is used for carrying out fault judgment, the problems that in the prior art, remote monitoring of the accident distributing valve cannot be carried out, fault detection depends on manual on-site checking, and the efficiency and accuracy of fault recognition are low are solved, deep learning is carried out on all data parameters, and fault judgment process logic design is carried out. The fault comprehensive monitoring of the accident distributing valve is realized, and the recognition rate and the accuracy of the abnormal working condition or fault judgment of the accident distributing valve equipment are greatly improved.

Description

Technical Field

[0001] The invention belongs to the technical field of fault diagnosis of hydropower stations, and in particular relates to a multi-modal abnormal working condition monitoring method of an accident pressure distribution valve based on sound recognition. Background Art

[0002] The working principle of the emergency pressure distribution valve is: when the speed governor loses control (such as the guide vane feedback is disconnected), or due to other reasons, the unit automatically judges or manually gives an emergency shutdown order, the monitoring will immediately trip the generator outlet switch and excitation switch command to dump the load, and at the same time give a signal to the emergency pressure distribution valve (it can also be operated manually). After this valve is actuated, the opening chamber of the relay is connected to the return oil tank, and the closing chamber of the relay is directly connected to the pressure oil. In this way, the water guide mechanism closes the guide vane under the action of the heavy hammer (without going through the speed governor). It can be seen that the combination of the emergency pressure distribution valve and the heavy hammer is equivalent to the emergency closing valve (fast valve) in the vertical unit.

[0003] Common faults of the accident pressure distribution method include internal and external leakage of the valve, valve core stuck, operating mechanism failure, accident pressure distribution valve refusal to move, frequent twitching, etc. The main reason for installing an accident pressure distribution valve in a large hydropower station is: when the speed governor fails to work properly and loses control, a device is needed to bypass the speed governor (not through the speed governor) to close the guide vane of the unit and shut it down to ensure the safety of the unit and the power plant. This device is the accident pressure distribution valve, which is used in combination with the heavy hammer of the water guide mechanism to quickly close the guide vane and stop the unit. Once the accident pressure distribution method fails, the valve body leaks internally, the valve core gets stuck, and the unit is at risk of being unable to shut down; in severe cases, the accident pressure distribution valve will refuse to move, resulting in major accidents such as the unit being unable to shut down and losing control.

[0004] The abnormal monitoring methods of the existing technology often directly detect faults, such as appearance quality inspection: first, perform an appearance inspection on the accident pressure-matching valve to check whether there is obvious damage, corrosion, cracks or other abnormal conditions. At the same time, check whether the valve markings are clear to ensure that it is the correct model and specification; manual operation inspection: try to manually operate the pressure-matching valve to check whether it can be opened and closed smoothly. During operation, pay attention to whether there are abnormal sounds or increased resistance; sealing performance inspection: perform a pressure test on the pressure-matching valve to check whether its sealing performance meets the requirements. Usually, a certain pressure is applied when the valve is closed, and observe whether there is leakage; action time test: check whether the response speed of the pressure-matching valve meets the design requirements. This usually involves simulating an overspeed situation and observing the time from when the pressure-matching valve receives the signal to when it actually cuts off the oil circuit; pressure release performance test: check the performance of the pressure-matching valve during pressure release to ensure that it can quickly and effectively reduce the pressure in the oil circuit; the existing technology has the following disadvantages: Existing technologies cannot achieve remote comprehensive judgment. When equipment fails, it is necessary for operation and maintenance personnel to conduct inspections or professional departments to conduct on-site experimental inspections. For example, sealing performance inspections, manual operation inspections, and action time tests require manual on-site inspections, which are time-consuming and laborious. There are also some external failures such as appearance quality inspections, which require operation and maintenance personnel to actively patrol and observe; therefore, it is necessary to design a multi-modal abnormal operating condition monitoring method for accident pressure distribution valves based on sound recognition to solve the above problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a multi-modal abnormal operating condition monitoring method for an accident pressure distributing valve based on sound recognition. The method is used to solve the problems that the prior art cannot perform remote monitoring of the accident pressure distributing valve, fault detection relies on manual on-site verification, and the efficiency and accuracy of fault identification are low. By performing deep learning on various data parameters and logical design of the fault judgment process, comprehensive monitoring of accident pressure distributing valve faults is achieved, which greatly improves the recognition rate and accuracy of abnormal operating conditions or fault judgment of the accident pressure distributing valve equipment.

[0006] In order to achieve the above technical effects, the technical solution adopted by the present invention is: A method for monitoring abnormal multi-modal working conditions of an accident pressure distribution valve based on sound recognition comprises the following steps: S1, collect voiceprint recognition related data through the comprehensive data acquisition module, the comprehensive data acquisition module includes a liquid level sensor, a bone conduction pickup, a vibration sensor and an infrared temperature measurement camera; the voiceprint recognition related data includes liquid level data, audio data, vibration analog data, temperature measurement image data, active power data and guide vane opening data; S2, transmits the voiceprint recognition related data collected by the comprehensive data collection module to the cloud server platform through the network transmission module; S3, identifying and scoring the audio data through the voiceprint scoring module; S4, performing fault determination through a comprehensive fault determination module.

[0007] Preferably, in step S1, a liquid level sensor is arranged at the oil collecting tank and the oil pressure tank to collect oil level data of the oil collecting tank and the oil pressure tank; a bone conduction pickup is arranged at the housing of the accident pressure distributing valve to collect audio data; a vibration sensor is arranged at the housing of the accident pressure distributing valve to collect vibration analog data of the housing of the accident pressure distributing valve; an infrared temperature measuring camera is arranged on a wall on one side of the accident pressure distributing valve to collect temperature measurement image data of the operation process of the accident pressure distributing valve; active power data and guide vane opening data are collected by connecting to the turbine unit operation management system; and the temperature measurement image data includes temperature data and image data corresponding to the temperature.

[0008] Furthermore, two liquid level sensors, one bone conduction pickup, and one vibration sensor are provided; the bone conduction pickup adopts the ZKHY-IVOD-001 industrial diagnostic pickup with a frequency response of 20-20kHz; the vibration sensor adopts the DSP vibration sensor with a measurement range of 0-50g.

[0009] Preferably, in step S2, the network transmission module includes a switch module composed of a photoelectric switch and a convergence switch; the active power data, oil level data, guide vane opening data and swing analog data are respectively connected to the convergence switch through the photoelectric switch, the temperature measurement image data and audio data are extracted with characteristic values ​​through the on-site control unit, and the temperature measurement image data and audio data are connected to the convergence switch through the photoelectric switch; the convergence switch is connected to the cloud server platform.

[0010] Furthermore, the PC device is connected to the cloud server platform after passing through the network isolation device, and the cloud server platform communicates with the mobile terminal through the router; the aggregation switch also communicates with the local server for data storage.

[0011] Preferably, in step S3, the audio data is identified and scored by the voiceprint scoring module, including in the local control unit, and the audio data is weighted dimension reduction optimized by the MFCC algorithm: First, the audio data is pre-emphasized, and then the pre-emphasized audio data is framed; the framed audio data is subjected to fast Fourier transform to convert the waveform data into frequency, and then the Mel coefficient filtering is performed; finally, endpoint detection is performed and the final MFCC result is tested to ensure the availability of the data and process the invalid data.

[0012] Furthermore, in the Mel coefficient filtering process, the Mel scale is a cepstrum parameter extracted from the frequency domain, which describes the nonlinear characteristics of the human ear frequency. Its relationship with the frequency is: Mel(f)=2585×log(1+f / 700); Mel(f) represents the Mel scale.

[0013] Preferably, in step S3, the audio data is identified and scored by the voiceprint scoring module, including extracting i-vector processing from the audio data after weighted dimensionality reduction optimization in the local control unit, and the specific method is: Use the GMM-UBM model to obtain the Baum-Welch statistic UBM; The obtained MFCC vector is input into the GMM-UBM for training, and the target MFCC vector is input into the GMM-UBM model. The GMM parameter M of the target device is adaptively obtained through the MAP algorithm based on the original UBM. Process the baum-welch statistic through the DNN neural network and extract the i-vector: Determine whether the audio sample Y comes from device S: Assume that H0 indicates that Y comes from device S, and H1 indicates that Y does not come from device S, then the likelihood formula is: LR = p(Y|H0) / p(Y|H1); log(LR)=log(p(Y|H0))-log(p(Y|H1)); Where p represents the likelihood function; when LR>θ, the result H1 is retained, otherwise the result H0 is retained; θ is a predetermined threshold.

[0014] Preferably, in step S3, the audio data is identified and scored by the voiceprint scoring module, including performing PLDA discriminant analysis in the cloud server platform, and the specific method is: The sound data is used as training audio and divided into audio of I devices. Each device has J different audio segments. The jth audio of the i-th device is defined as X ij ; According to factor analysis, define X ij The generative model is: ; Where μ represents the mean of all training data; F is the identity space, which contains information that can be used to represent various devices; h i is the identity of a specific device or the position of the device in the identity space; G is the error space, which contains information that can be used to represent different audio changes of the same device; w ij represents the position of the device in the G space; ϵij is the final residual noise term; this term is a zero-mean Gaussian distribution with a variance of Σ; Calculate whether two audios are composed of the feature h in the device space i Generate or calculate two audio frequencies i The generated likelihood is used to obtain the score value; the specific formula is: ; In the formula, score represents the score value; n1 and n2 are the I-Vector vectors of the two audios respectively. The assumption that the two audios come from the same space is Hs, and the assumption that they come from different spaces is Hd. Among them, p(n1,n2|hs) is the likelihood function of the two audios coming from the same space; p(n1|hd) and p(n2|hd) are the likelihood functions of n1 and n2 coming from different spaces respectively; the similarity of the two audios is calculated by calculating the log-likelihood ratio: The ratio, the score, and the possibility that the two audios belong to the same device are positively correlated. The higher the ratio, the higher the score, and the greater the possibility that the two audios belong to the same device. The lower the ratio, the lower the score, and the less likely the two audios belong to the same device. A normal value interval is set for the score value, and the score value falling within the normal value interval is judged as normal, otherwise it is judged as abnormal.

[0015] Furthermore, in step S3, the collected audio data is divided into a training set and a test set, and the test set is further divided into a registration set and a verification set. i-vectors are extracted from the three sets respectively, and the data of the training set is used to train the PLDA discriminant analysis model, and the data of the test set is used for scoring.

[0016] Preferably, in step S4, the method for performing fault determination by the comprehensive fault determination module is: S401, the real-time audio data collected by the bone conduction pickup is scored by the voiceprint scoring module, and a judgment is made according to the score value. If the score value is normal, the accident pressure distribution valve equipment is judged to be normal; if the score value is abnormal, the fault type is judged; S402, perform fault type determination, compare the score value with the database, determine whether it is a known fault or an unknown fault, if it is compared with the historical data record in the database, it is determined to be a known fault and the determination result is output; if it is not compared with the historical data record in the database, it is determined to be an unknown fault, and enter step S403 to perform unknown fault type determination; S403, executing unknown fault type determination: The fault is judged to be internal or external through the liquid level data, vibration simulation data, active power data, guide vane opening data and temperature data collected by the temperature measurement camera. If it is judged to be an internal fault, the professional team is directly notified to conduct on-site inspection to confirm the fault. If it is judged to be an external fault, the temperature measurement camera is used to conduct inspections to locate the fault location, and the professional team is notified to conduct on-site inspection to confirm the fault. S404, after confirming the fault, the fault is marked and sent down, and updated into the scoring database of step S402.

[0017] The beneficial effects of the present invention are as follows: The present invention realizes multi-modal online real-time monitoring of the accident pressure distribution valve equipment area by installing various types of acquisition sensors, and truly realizes unmanned duty; the present invention adopts bone conduction pickup to collect 50Hz-4KHz audio data, which can collect almost all abnormal conditions in the accident pressure distribution valve equipment area, and is also assisted by pressure, vibration, flow, switch value and other data. Therefore, this solution can identify all faults in the equipment accident pressure distribution valve equipment area, and greatly improve the accuracy of abnormal monitoring of the accident pressure distribution valve equipment working condition; through voiceprint recognition technology, and through comprehensive judgment and analysis of multi-modal data such as flow, vibration, pressure, image, switch value, etc., it can greatly improve the accuracy of abnormal monitoring of the accident pressure distribution valve equipment working condition, and save human resources while reducing the time required for fault identification, providing an opportunity for emergency handling of power system accidents and greatly improving the stability of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 It is a schematic diagram of the layout structure of each device in the comprehensive data acquisition module in an embodiment of the present invention; Figure 3 A schematic diagram of a data transmission path of a network transmission module in an embodiment of the present invention; Figure 4 Schematic diagram of the process of the voiceprint scoring module in an embodiment of the present invention; Figure 5 The figure is a flow chart of comprehensive fault determination in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Embodiment 1: like Figure 1 As shown, a method for monitoring abnormal multi-modal working conditions of an accident pressure distribution valve based on sound recognition includes the following steps: S1, collect voiceprint recognition related data through the comprehensive data acquisition module, the comprehensive data acquisition module includes a liquid level sensor, a bone conduction pickup, a vibration sensor and an infrared temperature measurement camera; the voiceprint recognition related data includes liquid level data, audio data, vibration analog data, temperature measurement image data, active power data and guide vane opening data; S2, transmits the voiceprint recognition related data collected by the comprehensive data collection module to the cloud server platform through the network transmission module; S3, identifying and scoring the audio data through the voiceprint scoring module; S4, performing fault determination through a comprehensive fault determination module.

[0020] Embodiment 2: like Figure 2As shown, in step S1, the liquid level sensor is arranged at the oil collecting tank and the oil pressure tank to collect the oil level data of the oil collecting tank and the oil pressure tank; the bone conduction pickup is arranged at the housing of the accident pressure distributing valve to collect the audio data; the vibration sensor is arranged at the housing of the accident pressure distributing valve to collect the vibration analog data of the housing of the accident pressure distributing valve; the infrared temperature measuring camera is arranged on the wall on one side of the accident pressure distributing valve to collect the temperature measurement image data of the operation process of the accident pressure distributing valve; the active power data and the guide vane opening data are collected by connecting to the turbine unit operation management system; the temperature measurement image data includes temperature data and image data corresponding to the temperature.

[0021] Furthermore, two liquid level sensors, one bone conduction pickup, and one vibration sensor are provided; the bone conduction pickup adopts the ZKHY-IVOD-001 industrial diagnostic pickup with a frequency response of 20-20kHz; the vibration sensor adopts the DSP vibration sensor with a measurement range of 0-50g.

[0022] Embodiment three: like Figure 3 As shown, in step S2, the network transmission module includes a switch module composed of an optoelectronic switch and an aggregation switch; the active power data, oil level data, guide vane opening data and swing analog data are respectively connected to the aggregation switch through the optoelectronic switch, the temperature measurement image data and the audio data are extracted with characteristic values ​​through the local control unit, and the temperature measurement image data and the audio data are connected to the aggregation switch through the optoelectronic switch; the aggregation switch is connected to the cloud server platform.

[0023] Furthermore, the PC device is connected to the cloud server platform after passing through the network isolation device, and the cloud server platform communicates with the mobile terminal through the router; the aggregation switch also communicates with the local server for data storage.

[0024] Embodiment 4: like Figure 4 As shown, in step S3, the audio data is identified and scored by the voiceprint scoring module, including in the local control unit, and the audio data is weighted and optimized by MFCC algorithm: First, the audio data is pre-emphasized, and then the pre-emphasized audio data is framed; the framed audio data is subjected to fast Fourier transform to convert the waveform data into frequency, and then the Mel coefficient filtering is performed; finally, endpoint detection is performed and the final MFCC result is tested to ensure the availability of the data and process the invalid data.

[0025] Furthermore, as a carrier of energy, when the size of the sound source is constant, the higher the frequency, the more serious the loss of sound energy by the medium, and pre-emphasis can compensate for the loss of the high-frequency part to a certain extent, thereby protecting the integrity of the sound channel information; the audio signal is affected by the state changes of the sound-emitting object during the generation process, and the speed of state change is much slower than the speed of sound vibration, so it can be considered to be short-term stable. Processing each frame of the signal after framing is equivalent to processing a continuous signal with fixed features, thereby reducing the impact of non-steady-state time-varying signals.

[0026] Furthermore, in the Mel coefficient filtering process, the Mel scale is a cepstrum parameter extracted from the frequency domain, which describes the nonlinear characteristics of the human ear frequency. Its relationship with the frequency is: Mel(f)=2585×log(1+f / 700); Mel(f) represents the Mel scale.

[0027] Preferably, in step S3, the audio data is identified and scored by the voiceprint scoring module, including extracting i-vector processing from the audio data after weighted dimensionality reduction optimization in the local control unit, and the specific method is: Use the GMM-UBM model to obtain the Baum-Welch statistic UBM; The obtained MFCC vector is input into the GMM-UBM for training, and the target MFCC vector is input into the GMM-UBM model. The GMM parameter M of the target device is adaptively obtained through the MAP algorithm based on the original UBM. Process the baum-welch statistic through the DNN neural network and extract the i-vector: Determine whether the audio sample Y comes from device S: Assume that H0 indicates that Y comes from device S, and H1 indicates that Y does not come from device S, then the likelihood formula is: LR = p(Y|H0) / p(Y|H1); log(LR)=log(p(Y|H0))-log(p(Y|H1)); Where p represents the likelihood function; when LR>θ, the result H1 is retained, otherwise the result H0 is retained; θ is a predetermined threshold.

[0028] Preferably, in step S3, the audio data is identified and scored by the voiceprint scoring module, including performing PLDA discriminant analysis in the cloud server platform, and the specific method is: The sound data is used as training audio and divided into audio of I devices. Each device has J different audio segments. The jth audio of the i-th device is defined as X ij ; According to factor analysis, define X ijThe generative model is: ; Where μ represents the mean of all training data; F is the identity space, which contains information that can be used to represent various devices; h i is the identity of a specific device or the position of the device in the identity space; G is the error space, which contains information that can be used to represent different audio changes of the same device; w ij represents the position of the device in the G space; ϵij is the final residual noise term; this term is a zero-mean Gaussian distribution with a variance of Σ; Calculate whether two audios are composed of the feature h in the device space i Generate or calculate two audio frequencies i The generated likelihood is used to obtain the score value; the specific formula is: ; In the formula, score represents the score value; n1 and n2 are the I-Vector vectors of the two audios respectively. The assumption that the two audios come from the same space is Hs, and the assumption that they come from different spaces is Hd. Among them, p(n1,n2|hs) is the likelihood function of the two audios coming from the same space; p(n1|hd) and p(n2|hd) are the likelihood functions of n1 and n2 coming from different spaces respectively; the similarity of the two audios is calculated by calculating the log-likelihood ratio: The ratio, the score, and the possibility that the two audios belong to the same device are positively correlated. The higher the ratio, the higher the score, and the greater the possibility that the two audios belong to the same device. The lower the ratio, the lower the score, and the less likely the two audios belong to the same device. A normal value interval is set for the score value, and the score value falling within the normal value interval is judged as normal, otherwise it is judged as abnormal.

[0029] Furthermore, in step S3, the collected audio data is divided into a training set and a test set, and the test set is further divided into a registration set and a verification set. i-vectors are extracted from the three sets respectively, and the data of the training set is used to train the PLDA discriminant analysis model, and the data of the test set is used for scoring.

[0030] Embodiment five: like Figure 5 As shown, the method for performing fault judgment through the comprehensive fault judgment module is: S401, the real-time audio data collected by the bone conduction pickup is scored by the voiceprint scoring module, and a judgment is made according to the score value. If the score value is normal, the accident pressure distribution valve equipment is judged to be normal; if the score value is abnormal, the fault type is judged; S402, perform fault type determination, compare the score value with the database, determine whether it is a known fault or an unknown fault, if it is compared with the historical data record in the database, it is determined to be a known fault and the determination result is output; if it is not compared with the historical data record in the database, it is determined to be an unknown fault, and enter step S403 to perform unknown fault type determination; S403, executing unknown fault type determination: The fault is judged to be internal or external through the liquid level data, vibration simulation data, active power data, guide vane opening data and temperature data collected by the temperature measurement camera. If it is judged to be an internal fault, the professional team is directly notified to conduct on-site inspection to confirm the fault. If it is judged to be an external fault, the temperature measurement camera is used to conduct inspections to locate the fault location, and the professional team is notified to conduct on-site inspection to confirm the fault. S404, after confirming the fault, the fault is marked and sent down, and updated into the scoring database of step S402.

Claims

1. A multi-modal abnormality monitoring method for an accident pressure distribution valve based on sound recognition, characterized in that: The following steps are involved: S1, collect voiceprint recognition related data through the comprehensive data acquisition module, the comprehensive data acquisition module includes a liquid level sensor, a bone conduction pickup, a vibration sensor and an infrared temperature measurement camera; the voiceprint recognition related data includes liquid level data, audio data, vibration analog data, temperature measurement image data, active power data and guide vane opening data; S2, transmits the voiceprint recognition related data collected by the comprehensive data collection module to the cloud server platform through the network transmission module; S3, identifying and scoring the audio data through the voiceprint scoring module; S4, performing fault determination through a comprehensive fault determination module.

2. The method for monitoring abnormal multi-modal working conditions of an accident pressure distribution valve based on sound recognition according to claim 1 is characterized in that: In step S1, liquid level sensors are arranged at the oil collecting tank and the oil pressure tank to collect oil level data of the oil collecting tank and the oil pressure tank; a bone conduction microphone is arranged at the housing of the accident pressure distributing valve to collect audio data; a vibration sensor is arranged at the housing of the accident pressure distributing valve to collect vibration analog data of the housing of the accident pressure distributing valve; an infrared temperature measuring camera is arranged on the wall on one side of the accident pressure distributing valve to collect temperature measurement image data of the operation process of the accident pressure distributing valve; active power data and guide vane opening data are collected by connecting to the turbine unit operation management system.

3. The method for monitoring abnormal multi-modal working conditions of an accident pressure distribution valve based on sound recognition according to claim 1 is characterized in that: In step S2, the network transmission module includes a switch module composed of an optoelectronic switch and an aggregation switch; active power data, oil level data, guide vane opening data and swing analog data are respectively connected to the aggregation switch through the optoelectronic switch, and characteristic values ​​of the temperature measurement image data and audio data are extracted through the local control unit, and the temperature measurement image data and audio data are connected to the aggregation switch through the optoelectronic switch; the aggregation switch is connected to the cloud server platform.

4. The method for monitoring abnormal multi-modal working conditions of an accident pressure distribution valve based on sound recognition according to claim 3 is characterized in that: In step S3, the audio data is identified and scored by the voiceprint scoring module, including in the local control unit, and the audio data is weighted and optimized by MFCC algorithm: First, the audio data is pre-emphasized, and then the pre-emphasized audio data is framed; the framed audio data is fast Fourier transformed to convert the waveform data into frequency, and then the Mel coefficient filtering is performed; Finally, endpoint detection is performed to detect the final MFCC results to ensure the availability of the data and process the invalid data.

5. The method for monitoring abnormal multi-modal working conditions of an accident pressure distribution valve based on sound recognition according to claim 4 is characterized in that: In step S3, the audio data is identified and scored by the voiceprint scoring module, including extracting i-vector processing from the audio data after weighted dimensionality reduction optimization in the local control unit. The specific method is as follows: Use the GMM-UBM model to obtain the Baum-Welch statistic UBM; The obtained MFCC vector is input into the GMM-UBM for training, and the target MFCC vector is input into the GMM-UBM model. The GMM parameter M of the target device is adaptively obtained through the MAP algorithm based on the original UBM. Process the baum-welch statistic through the DNN neural network and extract the i-vector: Determine whether the audio sample Y comes from device S: Assume that H0 indicates that Y comes from device S, and H1 indicates that Y does not come from device S, then the likelihood formula is: LR = p(Y|H0) / p(Y|H1); log(LR)=log(p(Y|H0))-log(p(Y|H1)); Where p represents the likelihood function; when LR>θ, the result H1 is retained, otherwise the result H0 is retained; θ is a predetermined threshold.

6. The method for monitoring abnormal multi-modal working conditions of an accident pressure distribution valve based on sound recognition according to claim 5 is characterized in that: In step S3, the audio data is identified and scored by the voiceprint scoring module, including performing PLDA discriminant analysis in the cloud server platform. The specific method is as follows: The sound data is used as training audio and divided into audio of I devices. Each device has J different audio segments. The jth audio of the i-th device is defined as X ij ; According to factor analysis, define X ij The generative model is: ； Where μ represents the mean of all training data; F is the identity space, which contains information that can be used to represent various devices; h i is the identity of a specific device or the position of the device in the identity space; G is the error space, which contains information that can be used to represent different audio changes of the same device; w ij represents the position of the device in the G space; ϵij is the final residual noise term; this term is a zero-mean Gaussian distribution with a variance of Σ; Calculate whether two audios are composed of the feature h in the device space i Generate or calculate two audio frequencies by h i The generated likelihood is used to obtain the score value; the specific formula is: ； In the formula, score represents the score value; n1 and n2 are the I-Vector vectors of the two audios respectively. The assumption that the two audios come from the same space is Hs, and the assumption that they come from different spaces is Hd. Among them, p(n1,n2|hs) is the likelihood function of the two audios coming from the same space; p(n1|hd) and p(n2|hd) are the likelihood functions of n1 and n2 coming from different spaces respectively; the similarity of the two audios is calculated by calculating the log-likelihood ratio: There is a positive correlation between the ratio, the score, and the possibility that the two audios belong to the same device; A normal value interval is set for the score value, and the score value falling within the normal value interval is judged as normal, otherwise it is judged as abnormal.

7. The method for monitoring abnormal multi-modal working conditions of an emergency pressure distribution valve based on sound recognition according to claim 6 is characterized in that: In step S4, the method for performing fault determination by the comprehensive fault determination module is as follows: S401, scoring the real-time audio data collected by the bone conduction microphone through the voiceprint scoring module, and making a judgment based on the scoring value. If the scoring value is normal, it is judged that the accident pressure distribution valve equipment is normal; If the scoring value is abnormal, the fault type is determined; S402, performing fault type determination, comparing the score value with the database to determine whether it is a known fault or an unknown fault. If it is compared with the historical data record in the database, it is determined to be a known fault and the determination result is output; If there is no historical data record in the database, it is determined to be an unknown fault, and the process goes to step S403 to perform unknown fault type determination; S403, executing unknown fault type determination: The fault is judged to be internal or external through the liquid level data, vibration simulation data, active power data, guide vane opening data and temperature data collected by the temperature measurement camera. If it is judged to be an internal fault, the professional team is directly notified to conduct on-site inspection to confirm the fault. If it is judged to be an external fault, the temperature measurement camera is used to conduct inspections to locate the fault location, and the professional team is notified to conduct on-site inspection to confirm the fault. S404, after confirming the fault, the fault is marked and sent down, and updated into the scoring database of step S402.

Images