Water turbine rapid gate working condition abnormity monitoring method based on sound recognition

Abstract

The invention discloses a water turbine fast gate working condition abnormity monitoring method based on sound recognition, and the method comprises the steps: evaluating the working condition of a water turbine fast gate through multi-feature data, which comprises a voiceprint signal, an infrared temperature measurement signal and a vibration signal collected on a fast gate hoist, and a position signal above a fast gate as the features of fault state recognition; and constructing a state evaluation model based on linear probability discrimination, training the feature information, integrating to form a database under different fault types of the water turbine fast gate, and inputting real-time data to obtain the fault type with the highest PLDA ratio, namely the fault type of the to-be-diagnosed feature sequence. Fault features and fault diagnosis of the rapid gate of the water turbine can be extracted in real time.

Description

Technical Field

[0001] The invention relates to the technical field of hydraulic turbine unit fault detection, and in particular to a method for monitoring abnormal working conditions of hydraulic turbine fast gates based on sound recognition. Background Art

[0002] Hydroelectric power generation uses water with potential energy in high places such as rivers and lakes to flow to low places, converting the potential energy contained therein into kinetic energy of the turbine, and then using the turbine as the motive force to drive the generator to generate electricity. A fast working door (hereinafter referred to as the fast door) is set at the water inlet. The fast door is used to close the door quickly to cut off the water flow when there is a penstock accident, the guide vane cannot be closed when the turbine generator unit is shut down, or the unit is overspeeding, etc., to prevent the accident from further expanding; in non-accident situations, such as when the unit is under maintenance, the door is closed slowly to block the water. The gate type is a flat fixed wheel gate (the power station is a flat slide gate), and the opening and closing machinery is a hydraulic opening and closing machine.

[0003] Too frequent state transitions such as starting and stopping are a huge consumption for turbine gates, and at the same time increase the requirements for gate status monitoring. At present, the development of information monitoring technology mainly uses vibration, force, temperature, radiation, electromagnetic and other signals for analysis. Most of the existing industry monitoring methods for the working conditions of turbine fast gates are relatively simple, and the fast gate is not the main monitoring position. Usually, the abnormal vibration of the fast gate is monitored by displacement sensors or vibration sensors, and the abnormal signal is uploaded, and the operator judges the abnormal signal.

[0004] The existing monitoring methods are not targeted and have low accuracy. At the same time, due to its single monitoring method, it can only identify limited abnormal situations and cannot reflect the gate working conditions from multiple angles. For example, if the temperature of the hydraulic system room is abnormal or the sound of the hydraulic system is abnormal, it can only be confirmed on site by the operation and maintenance personnel.

[0005] With the improvement of sound collection and processing technology, some scholars have proposed the concept of using sound to monitor the operating conditions and abnormalities of hydro-turbine generators. Sound is a mechanical wave of electric equipment that radiates energy from vibration to the sound transmission medium. The sound signal contains a large amount of vibration information and is an important indicator for analyzing the operating status of the equipment. When the equipment is operating normally, the different states of the mutual movement between the fuselage and the firmware, between the parts, and between the parts themselves correspond to different sounds emitted by the equipment. When the operating state changes, the sound generated by the hydro-turbine generator also changes accordingly. Therefore, it becomes possible to judge the working condition of the fast gate by detecting sound. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a method for monitoring the abnormal working condition of a turbine rapid gate based on sound recognition, and to monitor all internal faults of the turbine rapid gate through voiceprint recognition technology, so as to realize real-time online monitoring of abnormal sounds; by arranging bone conduction sensors, cameras with thermal infrared temperature measurement functions, acceleration sensors, etc. on the hydraulic hoist, and integrating multimodal data into the same system, real-time online monitoring of water flow size, equipment temperature, and vibration amplitude is realized, so as to solve the problem that the existing monitoring method cannot conduct comprehensive analysis; By arranging position sensors at appropriate locations above the rapid gate, real-time online monitoring of the rapid gate's sliding position and full opening and closing states can be achieved. By conducting deep learning on various data parameters and designing the fault judgment process logic, comprehensive monitoring of abnormal operating conditions of turbine rapid gates can be achieved, greatly improving the recognition rate and accuracy of abnormal operating conditions or fault judgments of rapid gate equipment.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for monitoring abnormal working condition of a turbine fast gate based on sound recognition, comprising: A monitoring system for abnormal working condition of turbine rapid gate based on sound recognition is established, including acquisition unit, data transmission unit, voiceprint analysis unit and comprehensive fault diagnosis unit. The sensor hardware constitutes the acquisition unit, which collects data online in real time and transmits it to the background server through the data transmission unit for comprehensive analysis and judgment. The voiceprint analysis unit is used to judge whether there are known faults or unknown abnormalities in the working condition of turbine rapid gate. The specific type of fault is judged comprehensively based on multi-modal sensor data, and finally the result is output.

[0008] The above-mentioned acquisition unit includes a bone conduction pickup and an acceleration sensor installed on the gate opening and closing machine, an infrared image monitor installed on the wall at the gate opening and closing machine, and a position sensor installed on the top of the gate.

[0009] The above-mentioned data transmission unit includes an edge computing gateway, a switch and a server platform, which transmits the sensor data to the cloud data platform of the background server through a data cable for further analysis.

[0010] The above-mentioned voiceprint analysis unit has a built-in voiceprint recognition algorithm engine. The voiceprint recognition algorithm engine performs weighted dimensionality reduction optimization based on the MFCC feature vector, and applies a vector quantization algorithm to identify the mechanical noise signal outside the cable line, including voiceprint MFCC extraction, UBM calculation, i-vector calculation, and linear probability discriminant analysis.

[0011] The above voiceprint MFCC extraction steps are: Step 1: Pre-emphasis: The medium is the carrier of energy. When the sound source size is constant, the higher the frequency, the more serious the loss of sound energy by the medium. Pre-emphasis can compensate for the loss of high-frequency part to a certain extent and protect the integrity of the sound channel information. Step 2: Framing: Divide the continuous voiceprint signal into multiple short time periods, i.e. frames, for subsequent processing; Step 3: FFT transformation: Perform fast Fourier transform (FFT) on each frame of signal. Through Fourier transform, the waveform data is transformed into the frequency domain. The specific formula of FFT is: ; in, is the time domain signal, is the frequency domain signal, N is the number of sampling points, and j is the imaginary unit; Step 4: Mel coefficient filtering; Mel scale is the cepstrum parameter extracted from the frequency domain, which describes the nonlinear characteristics of the human ear frequency. Its relationship with frequency can be expressed as follows: Mel(f)=2585×log(1+f / 700); Step 5: Endpoint detection: Endpoint detection will detect the final result to ensure the availability of data and process invalid data.

[0012] In the above-mentioned calculation of UBM, the universal background model UBM, namely the Universal Background Model, first collects speech irrelevant to the target, trains a UBM, and then uses the target speech data to adjust the parameters of the UBM through an adaptive algorithm to obtain the target model parameters; the parameters adjusted by the adaptive algorithm are only the mean.

[0013] The above calculation process of i-vector is: The process of solving i-vector is the MAP maximum a posteriori probability estimation process, that is, given the speaker data, the posterior probability distribution P (w|X) P(w|X)P(w|X) of i-vector is estimated, and the mean of the distribution is the i-vector of the speaker. In this process, the EM algorithm process is included to solve the model parameter T matrix, and w is the hidden variable. We get the calculation formula of i-vector: ; Among them, N(u) dimension is a diagonal matrix of CF×CF, and the diagonal block is NcI, (c=1,...,C), is a supervector of CF×1 consisting of all first-order BW statistics F~c.

[0014] The above linear probability discriminant analysis process is: Probabilistic Linear Discriminant Analysis (PLDA) is a channel compensation algorithm, also known as the probabilistic LDA algorithm. PLDA is usually based on I-Vector features and provides channel compensation for them. Assume that the training data speech consists of I speakers’ speech, where each speaker has J different speech segments. Then, we define i Personal j The voice is ; Based on factor analysis, define The generative model is: ; in, μ represents the mean of all training data; F can be regarded as the identity space, which contains information that can be used to represent various speakers; It can be regarded as the identity of a specific speaker, or the position of the speaker in the identity space; G can be regarded as the error space, which contains information that can be used to represent different voice changes of the same speaker; It represents the position in G space; is the final residual noise term, which is used to represent the other parts that have not been explained; this term is a zero-mean Gaussian distribution with a variance of Σ; The goal of PLDA model training is to input a bunch of data , the output can best represent the parameters of the data set θ = [μ,F,G,Σ]; In PLDA, it is necessary to calculate whether two speech sounds are related by features in the speaker space. generated by The likelihood of the generated class is not affected by the difference in the class space. The scoring formula is given below: ; As in the above formula, and are the I-Vector vectors of the two voices. The assumption that the two voices come from the same space is Hs, and the assumption that they come from different spaces is Hd. Among them, p( , | hs) is the likelihood function of the two speech sounds coming from the same space; p( |hd),p |hd) are and By calculating the log-likelihood ratio, we can measure the similarity between two speech sounds. The higher the ratio, the higher the score, and the greater the possibility that the two voices belong to the same speaker; The lower the ratio, the lower the score, and the less likely it is that the two voices belong to the same speaker.

[0015] After the sound pattern model of the turbine rapid gate, the above-mentioned comprehensive fault diagnosis unit determines whether leakage occurs through the parameters transmitted by the bone conduction pickup, thermal infrared thermometer and vibration sensor installed on the hydraulic device of the turbine gate hoist, outputs the results and takes corresponding measures; When an abnormality is found in the voiceprint score of the real-time analysis device, it is compared with the known fault model issued by the edge computing gateway. If it is a known fault, an alarm of the known fault is directly output; if it does not match the known fault model library, an unknown abnormality alarm is output.

[0016] Once an unknown abnormal alarm is reported, the surveillance camera will carefully observe and analyze the alarm point, and at the same time, the operation personnel will be arranged to go to the site to confirm the situation and make corresponding treatment. After the treatment is completed, they will come back to report and then improve the fault model database. Next time, the same fault will be reported as a known fault; After confirming the fault, in addition to taking corresponding measures according to the fault, it is also necessary to register the identified fault and voiceprint data, mark it as a known fault, enter it into the fault model library, and send it to the local control unit. The next time the voiceprint data is abnormal, the known fault will be added for judgment.

[0017] The present invention provides a method for monitoring abnormal working conditions of a turbine rapid gate based on sound recognition, which has the following beneficial effects: 1. Through the installation of various types of acquisition sensors, multi-modal online real-time monitoring of the fast door equipment area can be realized, truly realizing unmanned duty; 2. The bone conduction pickup collects 50Hz-4KHz audio data, which can collect almost all abnormal conditions in the rapid door equipment area. At the same time, it is assisted by temperature, vibration, image, position monitoring and other data. Therefore, this solution can identify all faults in the rapid door equipment area of ​​the equipment, greatly improving the accuracy of abnormal monitoring of the rapid door equipment working condition; 3. Through voiceprint recognition technology, and through comprehensive judgment and analysis of multimodal data such as temperature, vibration, image, and position monitoring, the accuracy of abnormal operating condition monitoring of fast door equipment can be greatly improved, while saving human resources and 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] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the installation position of the collection unit of the present invention; Figure 2 It is a structural schematic diagram of the data transmission module of the present invention; Figure 3 Schematic diagram of UBM training of the present invention. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0020] The comprehensive monitoring method for abnormal multi-modal working conditions of turbine rapid gates based on voiceprint recognition includes four major units: acquisition unit, data transmission unit, voiceprint analysis unit, and comprehensive fault diagnosis unit. As shown in the figure, the sensor hardware constitutes the acquisition unit, which collects data online in real time and transmits it to the background server through the data transmission unit for comprehensive analysis and judgment. The voiceprint analysis unit is used to determine whether there are known faults or unknown abnormalities in the working conditions of the turbine rapid gates, and the specific type of fault is comprehensively judged based on the multi-modal sensor data, and finally the result is output.

[0021] 1. Collection unit The main hardware of the comprehensive monitoring method for multi-modal abnormal working conditions of turbine rapid doors based on voiceprint recognition includes: 1. A thermal infrared camera installed on the wall opposite to the gate hoist; 2. A bone conduction microphone installed on the gate hoist to better collect sound; 3. An acceleration sensor installed on the gate hoist; 4. A position sensor installed on the top of the rapid door; the specific location is as follows: Figure 1 shown.

[0022] The data collection includes: Voiceprint signal: The sound signal generated by the operation of the rapid gate hydraulic device is collected through a bone conduction microphone. The signal contains rich information about the equipment's operating status.

[0023] Position signal: The position sensor collects the position signal of the rapid gate when the rapid gate hydraulic device is running. This signal can be used to determine the operating status of the equipment.

[0024] Vibration data: The vibration signal of the hydraulic device is collected through the acceleration sensor to analyze the mechanical operating status of the equipment.

[0025] Temperature data: The temperature distribution of the hydraulic gate hoist and its surroundings is measured by thermal infrared sensors to detect overheating or abnormal temperature points.

[0026] Image monitoring data: The image of the hydraulic gate hoist and its surrounding environment is captured by the camera to assist in judging changes in the appearance of the equipment and the surrounding environment.

[0027] 2. Data transmission module The data of the corresponding sensors are transmitted to the cloud data platform of the background server through the data line for further analysis. The data transmission system includes edge computing gateways, switches, server platforms, etc. Among them, the voiceprint feature value data is set with an edge computing gateway, and the gateway has the function of initial screening of audio data. Since the thermal infrared camera has two functions of temperature measurement and image monitoring, it is necessary to set up an edge computing gateway to transmit the temperature measurement data separately from the image and video data, which is convenient for the subsequent normalization processing of data; Figure 2 shown.

[0028] 3. Voiceprint analysis unit The voiceprint recognition algorithm engine performs weighted dimensionality reduction optimization based on MFCC feature vectors, and applies vector quantization algorithm to identify mechanical noise signals outside the cable line; including voiceprint MFCC extraction, UBM calculation, i-vector calculation, and linear probability discriminant analysis.

[0029] Voiceprint MFCC extraction process: The first step is pre-emphasis. As a carrier of energy, the medium loses more sound energy with higher frequencies when the size of the sound source is constant. Pre-emphasis can compensate for the loss of high-frequency parts to a certain extent and protect the integrity of the sound channel information. The second step is framing. During the generation process, the speech signal is affected by the state change of the sound-generating object, and the speed of state change is much slower than the speed of sound vibration. Therefore, it can be considered to be short-term stable. After framing, processing each frame signal is equivalent to processing a continuous signal with fixed features, which can reduce the impact of non-steady-state time variation. The third step is FFT transformation; FFT is the fast Fourier transform, which transforms the waveform data into the frequency domain through Fourier transform; the specific formula of FFT is: ; in, is the time domain signal, is the frequency domain signal, N is the number of sampling points, and j is the imaginary unit; The fourth step is Mel coefficient filtering; the Mel scale is the cepstrum parameter extracted from the frequency domain, which describes the nonlinear characteristics of the human ear frequency. Its relationship with the frequency can be approximately expressed by the following formula: Mel(f)=2585×log(1+f / 700); The fifth step of the MFCC algorithm is endpoint detection; endpoint detection will detect the final result to ensure the availability of the data and process the invalid data.

[0030] Calculate UBM: Universal Background Model (UBM) is a model that first collects a large amount of speech that is not related to the target, trains a UBM, and then uses a small amount of target speech data to adjust the parameters of the UBM through an adaptive algorithm to obtain the target model parameters. In general, the only parameter adjusted by the adaptive algorithm is the mean; this approach can reduce the amount of data and parameters in actual use, and facilitates rapid training convergence and decoding calculations on mobile terminals.

[0031] The training of UBM is actually the training of Gaussian mixture model GMM. Assume that we use a large amount of background speech data to train a UBM. The next focus is how to adapt the parameters of UBM to obtain the GMM parameters of the target speaker. The schematic diagram is as follows: Figure 3 shown.

[0032] Calculating i-vector The process of solving i-vector is actually the maximum a posteriori probability MAP estimation process, that is, given the speaker data, the posterior probability distribution P ( w|X ) P(w|X)P(w|X) of i-vector is estimated, and the mean of the distribution is the i-vector of the speaker; this process includes the EM algorithm process - used to solve the model parameter T matrix, and w is actually the hidden variable; in order to obtain the mean of this posterior probability distribution, we need alignment statistics, align means that some frames in the speech data of a certain speaker are attached to, generated from, and come from a certain component Gaussian in UBM. We can get a series of alignment statistics, including zero-order, first-order, and second-order statistics; using soft counting of the probability from a certain Gaussian component, we get Baum-Welch statistics, based on hard counting, the probability from a certain Gaussian component is the largest, that is, 1, and we get Viterbi statistics. Using Baum-Welch statistics, we get the calculation formula of i-vector: ; Among them, N(u) dimension is a diagonal matrix of CF×CF, and the diagonal block is NcI, (c=1,...,C), is a supervector of CF×1 consisting of all first-order BW statistics F~c.

[0033] Linear Probability Discriminant Analysis Probabilistic Linear Discriminant Analysis (PLDA) is also a channel compensation algorithm, also known as the probabilistic form of the LDA algorithm. PLDA is usually based on I-Vector features and provides channel compensation for them. Since the compensation ability of the PLDA algorithm is better than that of LDA, most people prefer PLDA.

[0034] In the field of voiceprint recognition, we assume that the training data speech consists of the speech of I speakers, where each speaker has J different speech segments. Then, we define i Personal j The voice is ; Based on factor analysis, define The generative model is: ; in, μ represents the mean of all training data; F can be regarded as the identity space, which contains information that can be used to represent various speakers; It can be regarded as the identity of a specific speaker, or the position of the speaker in the identity space; G can be regarded as the error space, which contains information that can be used to represent different voice changes of the same speaker; It represents the position in G space; is the final residual noise term, which is used to represent the other parts that have not been explained; this term is a zero-mean Gaussian distribution with a variance of Σ; The goal of PLDA model training is to input a bunch of data , the output can best represent the parameters of the data set θ = [μ,F,G,Σ]; In PLDA, it is necessary to calculate whether two speech sounds are related by features in the speaker space. generated by The likelihood of the generated class is not affected by the difference in the class space. The scoring formula is given below: ; As in the above formula, and are the I-Vector vectors of the two voices. The assumption that the two voices come from the same space is Hs, and the assumption that they come from different spaces is Hd. Among them, p( , | hs) is the likelihood function of the two speech sounds coming from the same space; p( |hd),p |hd) are and By calculating the log-likelihood ratio, we can measure the similarity between two speech sounds. The higher the ratio, the higher the score, and the greater the possibility that the two voices belong to the same speaker; The lower the ratio, the lower the score, and the less likely it is that the two voices belong to the same speaker.

[0035] 4. Comprehensive fault judgment module After the soundprint model of the turbine rapid door is obtained, the parameters transmitted by the bone conduction microphone, thermal infrared thermometer and vibration sensor installed on the hydraulic device of the turbine rapid door are used to determine whether leakage occurs, output the results and take corresponding measures.

[0036] When an abnormality is found in the voiceprint score of the real-time analysis device, it is compared with the known fault model issued by the edge computing gateway. If it is a known fault, an alarm of the known fault is directly output; if it does not match the known fault model library, an unknown abnormality alarm is output.

[0037] Once an unknown abnormal alarm is reported, the surveillance camera will first carefully observe and analyze the alarm point, and at the same time, the operating personnel will be arranged to go to the site to confirm the situation and make corresponding treatment. After the treatment is completed, they will come back to report and then improve the fault model database. Next time, the same fault will be alarmed as a known fault.

[0038] After confirming the fault, in addition to taking corresponding measures according to the fault, it is also necessary to register the identified fault and voiceprint data, mark it as a known fault, enter it into the fault model library, and send it to the local control unit. The next time the voiceprint data is abnormal, add the known fault for judgment. Through long-term operation accumulation, the accumulation of most common fault model libraries has been completed.

Claims

1. A method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition, characterized in that: include: A monitoring system for abnormal working condition of turbine rapid gate based on sound recognition is established, including acquisition unit, data transmission unit, voiceprint analysis unit and comprehensive fault diagnosis unit. The sensor hardware constitutes the acquisition unit, which collects data online in real time and transmits it to the background server through the data transmission unit for comprehensive analysis and judgment. The voiceprint analysis unit is used to determine whether there are known faults or unknown abnormalities in the working condition of turbine rapid gate. The specific type of fault is comprehensively judged by comprehensive multi-modal sensor data, and finally the result is output.

2. According to the method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition as described in claim 1, it is characterized in that: The acquisition unit includes a bone conduction pickup and an acceleration sensor installed on the gate opening and closing machine, an infrared image monitor installed on the wall at the gate opening and closing machine, and a position sensor installed on the top of the gate.

3. According to the method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition as described in claim 2, it is characterized in that: The data transmission unit includes an edge computing gateway, a switch and a server platform, which transmits the sensor data to the cloud data platform of the background server through a data cable for further analysis.

4. According to the method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition as described in claim 3, it is characterized in that: The voiceprint analysis unit has a built-in voiceprint recognition algorithm engine. The voiceprint recognition algorithm engine performs weighted dimensionality reduction optimization based on MFCC feature vectors, and applies a vector quantization algorithm to identify mechanical noise signals outside the cable, including voiceprint MFCC extraction, UBM calculation, i-vector calculation, and linear probability discriminant analysis.

5. A method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition according to claim 4, characterized in that: The voiceprint MFCC extraction steps are: Step 1: Pre-emphasis: The medium is the carrier of energy. When the sound source size is constant, the higher the frequency, the more serious the loss of sound energy by the medium. Pre-emphasis can compensate for the loss of high-frequency part to a certain extent and protect the integrity of the sound channel information. Step 2: Framing: Divide the continuous voiceprint signal into multiple short time periods, i.e. frames, for subsequent processing; Step 3: FFT transformation: Perform fast Fourier transform (FFT) on each frame of signal. Through Fourier transform, the waveform data is transformed into the frequency domain. The specific formula of FFT is: ； in, is the time domain signal, is the frequency domain signal, N is the number of sampling points, and j is the imaginary unit; Step 4: Mel coefficient filtering; Mel scale is the cepstrum parameter extracted from the frequency domain, which describes the nonlinear characteristics of the human ear frequency. Its relationship with frequency can be expressed as follows: Mel(f)=2585×log(1+f / 700); Step 5: Endpoint detection: Endpoint detection will detect the final result to ensure the availability of data and process invalid data.

6. A method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition according to claim 5, characterized in that: In the calculation of UBM, the universal background model UBM, namely the Universal Background Model, first collects speech irrelevant to the target, trains a UBM, and then uses the target speech data to adjust the parameters of the UBM through an adaptive algorithm to obtain the target model parameters; the parameters adjusted by the adaptive algorithm are only the mean.

7. A method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition according to claim 6, characterized in that: The i-vector calculation process is as follows: The process of solving i-vector is the MAP maximum a posteriori probability estimation process, that is, given the speaker data, the posterior probability distribution P (w|X) P(w|X)P(w|X) of i-vector is estimated, and the mean of the distribution is the i-vector of the speaker. In this process, the EM algorithm process is included to solve the model parameter T matrix, and w is the hidden variable. The calculation formula of i-vector is obtained as follows: ； Where N(u) is a diagonal matrix of dimension CF × CF, and the diagonal blocks are N c I , ( c = 1 , . . . , C ) , is a CF × 1 supervector consisting of all first-order BW statistics F~c.

8. A method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition according to claim 7, characterized in that: The linear probability discriminant analysis process is: Probabilistic Linear Discriminant Analysis (PLDA) is a channel compensation algorithm, also known as the probabilistic LDA algorithm. PLDA is usually based on I-Vector features and provides channel compensation for them. Assume that the training data speech consists of I speakers, each of whom has J different speech segments; then, we define i Personal j The voice is ; According to factor analysis, we define The generative model is: ； in, μ represents the mean of all training data; F can be regarded as the identity space, which contains information that can be used to represent various speakers; It can be regarded as the identity of a specific speaker, or the position of the speaker in the identity space; G can be regarded as the error space, which contains information that can be used to represent different voice changes of the same speaker; It represents the position in G space; is the final residual noise term, which is used to represent the other parts that have not been explained; this term is a zero-mean Gaussian distribution with a variance of Σ; The goal of PLDA model training is to input a bunch of data , the output can best represent the parameters of the data set θ = [μ,F,G,Σ]; In PLDA, it is necessary to calculate whether two speech sounds are related by features in the speaker space. generated by The likelihood of the generated class is not affected by the difference in the class space. The scoring formula is given below: ； As in the above formula, and are the I-Vector vectors of the two voices. The assumption that the two voices come from the same space is Hs, and the assumption that they come from different spaces is Hd. Among them, p( , | hs) is the likelihood function of the two speech sounds coming from the same space; p( |hd),p |hd) are and By calculating the log-likelihood ratio, we can measure the similarity between two speech sounds. The higher the ratio, the higher the score, and the greater the possibility that the two voices belong to the same speaker; The lower the ratio, the lower the score, and the less likely it is that the two voices belong to the same speaker.

9. A method for monitoring abnormal working condition of a turbine rapid gate based on sound recognition according to claim 8, characterized in that: The fault comprehensive diagnosis unit determines whether leakage occurs after the sound pattern model of the turbine rapid door through the bone conduction pickup, thermal infrared thermometer, and vibration sensor installed on the hydraulic device of the turbine gate hoist, outputs the result and takes corresponding measures; When the real-time analysis of the device voiceprint score shows an abnormality, it is compared with the known fault model issued by the edge computing gateway. If it is a known fault, the known fault alarm is directly output; if it does not match the known fault model library, an unknown abnormality alarm is output; Once an unknown abnormal alarm is reported, the surveillance camera will carefully observe and analyze the alarm point, and at the same time, the operation personnel will be arranged to go to the site to confirm the situation and make corresponding treatment. After the treatment is completed, they will come back to report and then improve the fault model database. Next time, the same fault will be reported as a known fault; After confirming the fault, in addition to taking corresponding measures according to the fault, it is also necessary to register the identified fault and voiceprint data, mark it as a known fault, enter it into the fault model library, and send it to the local control unit. The next time the voiceprint data is abnormal, the known fault will be added for judgment.

Images