Method for monitoring abnormal working condition of multi-dimensional excitation rectification system based on sound recognition

Abstract

A multi-dimensional excitation rectification system working condition abnormity monitoring method based on voice recognition comprises the steps that a multi-dimensional excitation rectification system working condition abnormity monitoring system based on voice recognition is established, the monitoring system comprises an acquisition unit, a data transmission unit, a data analysis module and a result output module, and sensor hardware forms the acquisition unit; data are collected online in real time and transmitted to a background server through a data transmission unit for comprehensive analysis and judgment, and whether known faults or unknown abnormities exist in the operation condition of the generator excitation rectification system or not is judged through a data analysis module. Monitoring devices such as a bone conduction pickup and an infrared temperature measurement camera are installed at proper positions of an excitation system, the operation condition of the excitation rectification system is monitored by integrating the reactive power of a generator, stator and rotor current, voltage and the conduction angle of a silicon controlled rectifier, and the purpose of monitoring the abnormity of the operation condition of the excitation rectification system of the generator is achieved. And the stability of generator grid-connected operation and a power system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of generator fault detection, and in particular to a method for monitoring abnormal working conditions of a multi-dimensional excitation rectifier system based on sound recognition. Background Art

[0002] The excitation system is the power supply and its auxiliary equipment for supplying the excitation current to the synchronous generator. It is generally composed of two main parts: the excitation power unit and the excitation regulator. The excitation power unit provides the excitation current to the synchronous generator rotor; and the excitation regulator controls the output of the excitation power unit according to the input signal and the given regulation criteria. Its functions include: maintaining the generator terminal voltage at a given value, and when the generator load changes, the terminal voltage is constant by adjusting the strength of the magnetic field; reasonably distributing the reactive power distribution between the parallel running units; and improving the stability of the power system, including static stability, transient stability and dynamic stability.

[0003] As the core component of the generator, the excitation system plays an important role not only in the stable operation of the generator, but also in the stability of the entire power system. Therefore, it is also very important to monitor the abnormal working conditions of the excitation rectifier system. However, the current monitoring of the excitation system mainly relies on the generator monitoring system. But when the excitation system fails, the monitoring can only issue an alarm, but cannot reflect the specific type of fault, which affects the timely handling of the fault or the investigation of the cause of the accident. In order to solve the above problems, the present invention provides a multi-dimensional excitation rectifier system working condition abnormality monitoring method based on sound recognition. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a multi-dimensional excitation rectifier system operating condition abnormality monitoring method based on sound recognition. By installing bone conduction pickups, infrared temperature measurement cameras and other monitoring devices at appropriate positions in the excitation system, the reactive power of the generator, stator and rotor currents, voltages, and conduction angles of thyristors are comprehensively considered to monitor the operating conditions of the excitation rectifier system of the generator, thereby improving the grid-connected operation of the generator and the stability of the power system.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A multi-dimensional excitation rectifier system operating condition abnormality monitoring method based on sound recognition, comprising: A multi-dimensional excitation rectifier system operating condition abnormality monitoring system based on sound recognition is established. The monitoring system includes an acquisition unit, a data transmission unit, a data analysis module and a result output module. 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 data analysis module is used to determine whether there are known faults or unknown abnormalities in the operating conditions of the generator excitation rectifier system.

[0006] The above-mentioned acquisition unit includes a bone conduction pickup installed inside the excitation power cabinet, at the fan outlet on the top of the excitation power cabinet, and on the outer shell of the excitation power cabinet, and an infrared image monitor installed on the wall above the control panel cabinet.

[0007] The above-mentioned data analysis module performs weighted dimensionality reduction optimization based on the Mel-frequency cepstral coefficient MFCC feature vector, applies the vector quantization algorithm to identify the mechanical noise signal outside the cable line, and extracts the feature vector i-vector of the excitation rectifier system. The probability linear discriminant analysis PLDA based on the I-Vector feature calculates the likelihood ratio to determine the fault type of the excitation rectifier system.

[0008] In the above data analysis module, the process of weighted dimensionality reduction optimization based on Mel-frequency cepstral coefficient MFCC feature vector is as follows: Step 1: pre-emphasis; Step 2: Frame division; Step 3: FFT transformation; FFT is the fast Fourier transform, which transforms the waveform data into the frequency domain through Fourier transform; Since it is usually difficult to see the characteristics of the signal from the transformation in the time domain, it is usually converted into energy distribution in the frequency domain for observation. Different energy distributions can represent the characteristics of different speech. Therefore, after multiplying the Hamming window, each frame must be subjected to a fast Fourier transform to obtain the energy distribution in the spectrum. The spectrum of each frame after frame division and windowing is obtained by fast Fourier transform. The spectrum of the speech signal is squared modulo to obtain the power spectrum of the speech signal. Step 4: Mel coefficient filtering; Mel scale is the cepstrum parameter extracted from the frequency domain, and 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.

[0009] The above process of obtaining the feature vector i-vector is: Extract a more compact vector, also called total factors, from the mean supervector GMM, i.e. mean supervectors3; Use GMM-UBM to get the Baum-Welch statistic UBM, which is the universal background model; Input the mfcc of the material data into GMM-UBM, train it, and put the mean of each Gaussian classification into the same vector, which is called the mean supervector m; input the MFCC vector of the target into GMM-UBM, and adaptively obtain the GMM parameter M of the excitation cabinet through the MAP algorithm based on the original UBM; Use DNN to get the Baum-Welch statistic and extract the i-vector: To find out whether a speech sample Y comes from the excitation cabinet S, there are two hypotheses: H0: Y comes from the model marker SH1: Y does not come from the model marker S So the Likelihood Ratio LR formula is: LR = p(Y|H0) / p(Y|H1); log(LR)=log(p(Y|H0))-log(p(Y|H1)); Among them, p() is the likelihood function from the space; when LR>θ, θ is a given threshold, we accept H1, otherwise we accept H0.

[0010] The above-mentioned probabilistic linear discriminant analysis PLDA process is: Assume that the training data audio is composed of audio from I devices, where each device has J different audio segments; then, define the jth audio of the ith device as ; 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 devices; It can be regarded as the identity of a specific device, or the position of the device in the identity space; G can be regarded as the error space, which contains information that can be used to represent different audio changes of the same device; 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 , output the parameters θ = [μ, F, G, Σ] that can best represent the data set; In PLDA, calculate whether two audios are represented by features in device space generated by The likelihood of the generated class is calculated without considering the difference in the intra-class space; the scoring formula is given below: ; 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 Likelihood functions from different spaces; 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 more likely it is that the two audios belong to the same device; The lower the ratio, the lower the score, and the less likely it is that the two audios belong to the same device.

[0011] The above-mentioned result output module collects and processes the voiceprint information of the excitation cabinet in real time, extracts characteristic values ​​to measure the similarity and score of the audio to determine whether the excitation system has a fault and the type of fault.

[0012] The specific situation of extracting the score of the characteristic value in the above result output module to judge the fault and fault type is as follows: If the score is normal and there is no alarm in the generator monitoring system, the excitation system is operating normally; If the score is abnormal and the generator monitoring system has an alarm, it is compared with the data in the training PLDA model, and combined with the stator and rotor voltage, current, temperature, reactive power, etc. of the generator excitation system, it is judged as a normal event or a known fault, and the result is output on the cloud platform or an alarm is issued and the fault type is output; If the score is abnormal and the degree of comparison with the known data in the database is low, the operation and maintenance personnel are required to use image monitoring or on-site to check the operating status of the equipment and enter the fault or event type into the database to complete the database.

[0013] The present invention provides a method for monitoring abnormal working conditions of a multi-dimensional excitation rectifier system based on sound recognition, which has the following beneficial effects: 1. Through real-time monitoring of sound, intelligent monitoring of the excitation system is realized to avoid further expansion of accidents when failures occur.

[0014] 2. The signals transmitted by the comprehensive excitation system make the monitoring and judgment of the excitation system more accurate and complete. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the installation position of the acquisition module of the present invention; Figure 3 It is a structural schematic diagram of the data transmission module of the present invention; Figure 4 It is a processing flow diagram of the data analysis module; Figure 5 It is a schematic diagram of the data comparison diagnosis process of the present invention.

[0016] In the picture: 1 bone conduction microphone, 2 infrared image monitor. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0018] Example: A multi-dimensional excitation rectifier system operating condition abnormality monitoring method based on sound recognition, the multi-dimensional excitation rectifier system operating condition abnormality monitoring method based on sound recognition comprises the following steps: 1. Collection module The main hardware of the multi-dimensional excitation rectifier system working condition abnormality monitoring method based on sound recognition includes: 1. Bone conduction pickup * 3, located inside the excitation power cabinet, at the fan outlet on the top of the excitation power cabinet, and on the shell of the excitation power cabinet; 2. Infrared image monitoring instrument, above the wall around the control panel cabinet; the specific location is as follows Figure 2 shown.

[0019] 2. Data transmission module The data collected by the corresponding sensors, such as sound, temperature, image, etc., are transmitted to the server through the optical fiber channel and the stator and rotor voltage, current, reactive power, thyristor conduction angle, etc. collected by the generator local control unit LCU. The corresponding data is further integrated, analyzed and calculated in the server. The data acquisition module includes switches, servers, and cloud platforms, such as Figure 3 shown.

[0020] 3. Data analysis module In terms of sound processing, this method uses Mel-frequency cepstral coefficients, or MFCC feature vectors for weighted dimensionality reduction optimization, applies vector quantization algorithm to identify mechanical noise signals outside the cable, and extracts the feature vector i-vector of the excitation rectifier system. Based on the I-Vector feature, the probability linear discriminant analysis PLDA calculates the likelihood ratio to determine the fault type of the excitation rectifier system. The process is as follows: Figure 4 shown.

[0021] 3.1 MFCC algorithm 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 conversion; FFT is the fast Fourier transform, which converts the waveform data into the frequency domain through Fourier transform; 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.

[0022] 3.2. Get i-vector i-vector is a more compact vector extracted from mean supervectors3, also called total factors. This supervector represents the model label and is developed based on the concept of joint factor analysis (JFA).

[0023] GMM-UBM is used to obtain the Baum-Welch statistic UBM, that is, the universal background model to make up for the lack of target speech data. A large amount of other similar data is collected for broad speech feature extraction, which can have a preconceived impression and make feature capture targeted.

[0024] UBM is also a GMM in essence. The mfcc of the material data is input into GMM-UBM, trained, and the mean of each Gaussian classification is put into the same vector, which is called the mean supervector m. The MFCC vector of the target is input into GMM-UBM, and the GMM parameter M of the excitation cabinet is adaptively obtained through the MAP algorithm based on the original UBM.

[0025] Use DNN to get the Baum-Welch statistic and extract the i-vector: Perform the corresponding comparison; if we want to find out whether a speech sample Y comes from the excitation cabinet S, there are two hypotheses: H0: Y comes from the model marker SH1: Y does not come from the model marker S So the Likelihood Ratio LR formula is: LR = p(Y|H0) / p(Y|H1); log(LR)=log(p(Y|H0))-log(p(Y|H1)); Among them, p() is the likelihood function from the space; when LR>θ, θ is a given threshold, we accept H1, otherwise we accept H0.

[0026] 3.3 Probabilistic Linear Discriminant Analysis Probabilistic Linear Discriminant Analysis (PLDA) is also a channel compensation algorithm, also known as the probabilistic form of LDA algorithm. Assume that the training data audio consists of audio from I devices, where each device has J different audio segments. Then, define the jth audio of the i-th device as ; 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 devices; It can be regarded as the identity of a specific device, or the position of the device in the identity space; G can be regarded as the error space, which contains information that can be used to represent different audio changes of the same device; 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 , output the parameters θ = [μ, F, G, Σ] that can best represent the data set; In PLDA, calculate whether two audios are represented by features in device space generated by The likelihood of the generated class is calculated without considering the difference in the intra-class space; the scoring formula is given below: ; 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 Likelihood functions from different spaces; 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 more likely it is that the two audios belong to the same device; The lower the ratio, the lower the score, and the less likely it is that the two audios belong to the same device.

[0027] 4. Result output module By collecting and processing the voiceprint information of the excitation cabinet in real time and extracting characteristic values ​​to measure the similarity and score of the audio, it is possible to determine whether the excitation system has a fault and the type of fault.

[0028] If the score is normal and there is no alarm in the generator monitoring system, the excitation system is operating normally; If the score is abnormal and the generator monitoring system has an alarm, it is compared with the data in the training PLDA model, and combined with the stator and rotor voltage, current, temperature, reactive power, etc. of the generator excitation system, it is judged as a normal event or a known fault, and the result is output on the cloud platform or an alarm is issued and the fault type is output; If the score is abnormal and the comparison with the known data in the database is low, the operation and maintenance personnel need to use image monitoring or on-site to check the operating status of the equipment and enter the fault or event type into the database to improve the database. The process is as follows: Figure 5 shown.

[0029] Through the multi-dimensional comprehensive judgment of voiceprints and the continuous improvement of the database, the abnormal condition monitoring of the excitation system will be more complete, the judgment of abnormal conditions will be more accurate, and the further expansion of accidents when failures occur can be further avoided.

Claims

1. A multi-dimensional excitation rectifier system operating condition abnormality monitoring method based on sound recognition, characterized in that: include: A multi-dimensional excitation rectifier system operating condition abnormality monitoring system based on sound recognition is established. The monitoring system includes an acquisition unit, a data transmission unit, a data analysis module and a result output module. 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 data analysis module is used to determine whether there are known faults or unknown abnormalities in the operating conditions of the generator excitation rectifier system.

2. According to the method for monitoring abnormal working condition of a multi-dimensional excitation rectifier system based on sound recognition as described in claim 1, it is characterized in that: The acquisition unit includes a bone conduction pickup installed inside the excitation power cabinet, at the fan outlet on the top of the excitation power cabinet and on the outer shell of the excitation power cabinet, and an infrared image monitor installed above the wall of the control panel cabinet.

3. According to the method for monitoring abnormal working condition of a multi-dimensional excitation rectifier system based on sound recognition as described in claim 2, it is characterized in that: The data analysis module performs weighted dimensionality reduction optimization based on the Mel-frequency cepstral coefficient MFCC feature vector, applies a vector quantization algorithm to identify the mechanical noise signal outside the cable, and extracts the feature vector i-vector of the excitation rectifier system. The probability linear discriminant analysis PLDA based on the I-Vector feature calculates the likelihood ratio to determine the fault type of the excitation rectifier system.

4. According to the method for monitoring abnormal working condition of a multi-dimensional excitation rectifier system based on sound recognition as described in claim 3, it is characterized in that: In the data analysis module, the process of weighted dimensionality reduction optimization based on Mel-frequency cepstral coefficient MFCC feature vector is as follows: Step 1: pre-emphasis; Step 2: Frame division; Step 3: FFT transformation; FFT is the fast Fourier transform, which transforms the waveform data into the frequency domain through Fourier transform; Step 4: Mel coefficient filtering; Mel scale is the cepstrum parameter extracted from the frequency domain, and 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.

5. According to the method for monitoring abnormal working condition of a multi-dimensional excitation rectifier system based on sound recognition as described in claim 4, it is characterized in that: The process of obtaining the characteristic vector i-vector is as follows: Extract a more compact vector, also called total factors, from the mean supervector GMM, i.e. mean supervectors3; Use GMM-UBM to get the Baum-Welch statistic UBM, which is the universal background model; Input the mfcc of the material data into GMM-UBM, train it, and put the mean of each Gaussian classification into the same vector, which is called the mean supervector m; input the MFCC vector of the target into GMM-UBM, and adaptively obtain the GMM parameter M of the excitation cabinet through the MAP algorithm based on the original UBM; Use DNN to get the Baum-Welch statistic and extract the i-vector: To find out whether a speech sample Y comes from the excitation cabinet S, there are two hypotheses: H0: Y comes from the model marker SH1: Y does not come from the model marker S So the Likelihood Ratio LR formula is: LR = p(Y|H0) / p(Y|H1); log(LR)=log(p(Y|H0))-log(p(Y|H1)); Among them, p() is the likelihood function from the space; when LR>θ, θ is a given threshold, we accept H1, otherwise we accept H0.

6. According to the method for monitoring abnormal working condition of a multi-dimensional excitation rectifier system based on sound recognition as described in claim 5, it is characterized in that: The probabilistic linear discriminant analysis PLDA process is: Assume that the training data audio is composed of audio from I devices, where each device has J different audio segments; then, define the jth audio of the ith device as ; 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 devices; It can be regarded as the identity of a specific device, or the position of the device in the identity space; G can be regarded as the error space, which contains information that can be used to represent different audio changes of the same device; 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 , output the parameters θ = [μ, F, G, Σ] that can best represent the data set; In PLDA, calculate whether two audios are represented by features in device space generated by The likelihood of the generated class is calculated without considering the difference in the intra-class space; the scoring formula is given below: ； 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 more likely it is that the two audios belong to the same device; The lower the ratio, the lower the score, and the less likely it is that the two audios belong to the same device.

7. The method for monitoring abnormal working condition of a multi-dimensional excitation rectifier system based on sound recognition according to claim 6 is characterized in that: The result output module collects and processes the voiceprint information of the excitation cabinet in real time, extracts characteristic values ​​to measure the similarity and score of the audio to determine whether the excitation system has a fault and the type of fault.

8. According to the method for monitoring abnormal working condition of a multi-dimensional excitation rectifier system based on sound recognition as described in claim 7, it is characterized in that: The specific situation of extracting the score of the characteristic value in the result output module to judge the fault and the fault type is as follows: If the score is normal and there is no alarm in the generator monitoring system, the excitation system is operating normally; If the score is abnormal and the generator monitoring system has an alarm, it is compared with the data in the training PLDA model, and combined with the stator and rotor voltage, current, temperature, reactive power, etc. of the generator excitation system, it is judged as a normal event or a known fault, and the result is output on the cloud platform or an alarm is issued and the fault type is output; If the score is abnormal and the degree of comparison with the known data in the database is low, the operation and maintenance personnel are required to use image monitoring or on-site to check the operating status of the equipment and enter the fault or event type into the database to complete the database.

Images