Multi-dimensional hydropower station control panel cabinet working condition abnormity monitoring method based on sound recognition

Abstract

A multi-dimensional hydropower station control panel cabinet working condition abnormity monitoring method based on sound recognition is characterized in that a sound pickup, a mutual inductor, an infrared image monitor, a switch node sensor and the like are installed at different positions of a control panel cabinet to acquire multi-dimensional data including voiceprint, voltage, current, temperature, images and switch displacement signals; the voiceprint signals are subjected to pre-emphasis, framing, FFT transformation, Mel frequency filter bank application, MFCC feature extraction and the like, and feature vectors are extracted by using UBM, GMM and MAP algorithms. A PLDA algorithm is adopted for data analysis and fault judgment, multi-dimensional data comprehensive scoring is fused, the fault type is determined, alarm information and processing measures are output, and the system has the advantages of comprehensive multi-dimensional data acquisition, advanced signal processing technology, scientific and accurate fault judgment, high practicability and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of abnormal monitoring of control panels in hydropower stations, and in particular relates to a multi-dimensional abnormal monitoring method for operating conditions of control panels in hydropower stations based on sound recognition. Background Art

[0002] In large hydropower stations, control panels are usually assembled in closed or semi-closed metal cabinets with controllers, relays, circuit breakers, power supplies and electrical components according to electrical wiring requirements. They are used to control and monitor the operating status of electrical equipment to ensure the normal operation and safety of the equipment.

[0003] As an important monitoring and control equipment in a hydropower station, the "health" status of the control panel itself plays a key role in the stable operation of the entire hydro-turbine generator set. However, the traditional control panel monitoring method can only rely on manual inspections or image monitoring inspections, which requires manual judgment of the operating conditions of the control panel, and cannot monitor the control panel itself and its interior for a long time. With the development of intelligence and informatization of power equipment, sound detection as a fast, accurate, non-contact detection method has gradually become an important technical means in the power industry. In order to solve the above problems, it is urgently necessary to provide a multi-dimensional hydropower station control panel 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 hydropower station control panel cabinet operating condition abnormality monitoring method based on sound recognition, which can provide accurate fault information and effective processing suggestions for hydropower station staff, improve the efficiency of fault processing and the reliability of equipment.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A multi-dimensional hydropower station control panel cabinet abnormality monitoring method based on sound recognition, the steps are: Step 1: Collect data; Step 2: transmitting the real-time collected data of the control cabinet to the server, wherein the data includes voiceprint, voltage, current, temperature, image and switch position change signal; Step 3: Data processing and feature extraction; Step 3.1, pre-emphasize the collected voiceprint signal to compensate for the loss of the high-frequency part; Step 3.2, process the voiceprint signal in frames to reduce the impact of non-steady-state time variation; Step 3.3, convert the framed voiceprint signal into the frequency domain through fast Fourier transform (FFT); Step 3.4, applying the Mel frequency filter bank on the frequency domain signal; Step 3.5, calculate the logarithmic energy of each Mel filter output; Step 3.6, extract Mel frequency cepstral coefficient (MFCC) features; Step 3.7: Detect and process invalid data to ensure data availability; Step 3.8, extract feature vector (i-vector) using universal background model (UBM) and Gaussian mixture model (GMM); Step 3.9, adaptively adjust the Gaussian mixture model (GMM) parameters through the maximum a posteriori probability (MAP) algorithm; Step 4: Perform data analysis and fault diagnosis; Step 5: Determine the fault result.

[0006] Preferably, the sub-steps of step 1 are as follows: Step 1.1: Install the microphone to collect voiceprint signals: Bone conduction microphones are installed inside, on the top and on the outer shell of the control cabinet to collect voiceprint signals at different positions. , , Respectively represent the soundprint signals collected from the inside, top and shell of the control cabinet; the soundprint signal can be expressed as time The function of , , ; Step 1.2: Install the transformer to collect voltage and current data: Install voltage transformer and current transformer at the power coil of control panel cabinet to collect voltage and current data; Indicates the collected voltage data. Indicates the collected current data; Step 1.3: Install infrared image monitor to collect temperature and images: Install infrared image monitors on the walls around the control cabinet to monitor ambient temperature and images; Indicates the collected temperature data. Represents the collected image data; Step 1.4: Install the switch node sensor to detect the switch state change: Install a switch node sensor at the three-phase switch of the control panel cabinet to detect the change of switch status; Indicates the switch state, with a value of 0 or 1, indicating the switch is open or closed, respectively.

[0007] Preferably, in step 2, the transmitted data can be represented as a vector; .

[0008] Preferably, in step 3.1, the collected sound signal is pre-emphasized to compensate for the loss of the high-frequency part and to protect the integrity of the sound channel information; The pre-emphasis formula is: , in is the original voiceprint signal, is the pre-emphasized voiceprint signal, is the pre-emphasis coefficient, usually between 0.9 and 1.

[0009] The collected sound signal is pre-emphasized to compensate for the loss of the high-frequency part and protect the integrity of the sound channel information.

[0010] Preferably, in step 3.2, The voiceprint signal is processed in frames to reduce the impact of non-steady-state time variation, so that each frame signal remains relatively stable in a short period of time; Assume that the length of each frame is sampling points, the frame shift is sampling points; The first The voiceprint signal of a frame can be expressed as .

[0011] Represents the original voiceprint signal, which is the time Function of Indicates the frame number; Indicates the length of each frame, that is, the number of sampling points contained in each frame; Indicates The voiceprint signal of the frame is obtained from the original voiceprint signal A section taken from the time i Start to Finish.

[0012] The sound signal is processed in frames to reduce the impact of non-steady-state time variations, so that each frame signal remains relatively stable in a short period of time.

[0013] Preferably, in step 3.3, the framed voiceprint signal is converted to the frequency domain by fast Fourier transform (FFT); the fast Fourier transform formula is: , Indicates the frame number; represents the frequency index; Represents a time index; Indicates the length of each frame, that is, the number of sampling points contained in each frame; Indicates The frame voiceprint signal has a time index of The value at time; Indicates The representation of the frame voiceprint signal in the frequency domain, that is, the frequency index is The value at time; is a natural constant; Represents an imaginary unit.

[0014] The framed signal is converted to the frequency domain through the Fast Fourier Transform (FFT) for further frequency domain analysis; the Fast Fourier Transform (FFT) converts the time domain signal into the frequency domain signal; this step converts the sound signal from the waveform representation on the time axis to the spectrum representation on the frequency axis; through the Fast Fourier Transform (FFT), the energy distribution of the signal at different frequencies can be observed, which is very important for subsequent feature extraction; the spectrum diagram of each frame represents the energy distribution of the frame at different frequencies; The Mel frequency scale is designed based on the nonlinear perception of frequency by the human ear. Applying a Mel frequency filter bank to a frequency domain signal can better simulate the sensitivity of the human ear to different frequencies. Calculating the logarithmic energy of each Mel filter output helps compress the dynamic range and reduce the energy difference between high and low frequencies.

[0015] Preferably, in step 3.4, applying a Mel frequency filter bank to the frequency domain signal can better simulate the sensitivity of the human ear to different frequencies; the conversion formula between the Mel frequency and the actual frequency is: , in is the Mel frequency, is the actual frequency; For the Mel filter at frequency The output after the Mel filter bank is ; Extract Mel Frequency Cepstral Coefficient (MFCC) features to reflect the nonlinear frequency perception of the human ear and extract more representative acoustic features; the MFCC feature vector of each frame usually contains 12-13 coefficients plus an energy coefficient to form a 13-dimensional feature vector; In step 3.5, the logarithmic energy is calculated for the output of each Mel filter. This step helps to compress the dynamic range and reduce the energy difference between high and low frequencies. The logarithmic energy calculation formula is: ; In step 3.6, the Mel frequency cepstral coefficient (MFCC) feature is extracted to reflect the nonlinear perception of the human ear to frequency and extract more representative acoustic features; the MFCC calculation formula is: , in For the The first frame of the voiceprint signal MFCC coefficients, is the number of Mel filters, It usually takes a value between 12 and 13, plus an energy coefficient, to form a 13-dimensional eigenvector; i-vector is a more compact vector extracted from the GMM mean supervector and is used to represent the characteristics of the control cabinet; In step 3.7, invalid data is detected and processed to ensure data availability; whether the data is valid can be determined by setting a threshold or using statistical methods; A low-dimensional feature vector, usually several hundred dimensions, can effectively represent the acoustic characteristics of the control cabinet; In step 3.8, i-vector is a more compact vector extracted from the GMM mean supervector, which is used to represent the characteristics of the control cabinet; is the GMM model, where is the number of Gaussian mixtures, For the The weights of the Gaussian components, For the Gaussian distribution, is the mean vector, is the covariance matrix; by training UBM and GMM, the mean supervector can be obtained ,in is the feature dimension; then through the total variation matrix Map the mean supervector to the i-vector space, that is ,in is i-vector, is the average i-vector, is a low-dimensional random vector; In step 3.9, the Gaussian mixture model (GMM) parameters are adaptively adjusted through the maximum a posteriori probability (MAP) algorithm to obtain the characteristic vector (i-vector) of the control cabinet; the update formula of the MAP algorithm is: , , in and are the updated mean and covariance matrices, and are the old mean and covariance matrices, For the The number of samples of Gaussian components, is the smoothing parameter, and are the prior mean and covariance matrices.

[0016] Preferably, the sub-steps of step 4 are: Step 4.1. Calculate the log-likelihood ratio using the PLDA algorithm: The PLDA algorithm is used to calculate the log-likelihood ratio of two audios to determine whether the audios come from the same device. The PLDA model can be expressed as ,in is the eigenvector, mu is the global mean vector, is the factor loading matrix, is a latent variable, representing the identity or status of the device, is the noise term; let and is the feature vector of the two audios, then the log-likelihood ratio calculation formula is: ; and is the feature vector of two audios; is a latent variable, indicating the identity or status of the device; Indicates that given the latent variable In the case of , the audio feature vector is probability; Indicates that given the latent variable In the case of , the audio feature vector is probability; is a latent variable The probability distribution of and They represent different latent variables corresponding to the two audios; Indicates that given the latent variable In the case of , the audio feature vector is probability; Indicates that given the latent variable In the case of , the audio feature vector is probability; and The latent variables and The probability distribution of .

[0017] The PLDA algorithm is used to calculate the log-likelihood ratio of two audios to determine whether the audios come from the same device; Step 4.2, determine the audio similarity to judge the fault; By comparing the log-likelihood ratio, the similarity of the audio is determined, thereby judging whether the control panel cabinet has a fault; Assume that the training data audio is composed of audio from multiple devices, define the audio generation model for each device; For the The audio generation model of the device , is the number of devices; Through the training data, the output can best represent the parameters of the data set ;in is the global mean vector, is the factor loading matrix, is a diagonal matrix representing the variance of different devices, is the noise covariance matrix; Calculate the log-likelihood ratio of two audios to determine whether they come from the same device or the same state; Step 4.3, fusion decision: Extract MFCC features from the sound signal and obtain the feature vector of the control cabinet through i-vector extraction; Combine MFCC features and i-vector features to form a multi-dimensional feature vector; ,in is the MFCC feature vector, is the i-vector feature vector; The combined feature vectors were analyzed using the PLDA model to calculate the log-likelihood ratio; The log-likelihood ratio of the PLDA model is comprehensively scored with other sensor data to form a comprehensive fault score; ,in is the log-likelihood ratio score for the PLDA model, Score the voltage data, Score the current data, Score the temperature data, Score the switch status, is the weight coefficient; Set a threshold ,If the comprehensive score exceeds the threshold, it is judged as a fault; If the comprehensive score is close to the threshold, it will be combined with the data from the infrared image monitor for secondary confirmation; Determine the specific fault type based on the comprehensive score and known fault data in the database.

[0018] Preferably, the sub-steps of step 5 are: After completing the data analysis and fault judgment of the control cabinet, the results need to be judged and output in detail; Step 5.1: Determine the confidence value: Confidence value Indicates the degree of certainty of whether the control cabinet has failed; through a comprehensive score With threshold The difference is calculated, that is ;if , indicating that the comprehensive score exceeds the threshold, it can be preliminarily determined that the control panel cabinet may have a fault; if , it indicates that the comprehensive score is within the normal range and the control cabinet may be in normal operation; Step 5.2, determine the fault type: If the confidence value , then further determine the specific fault type; analyze and judge based on the data characteristics of different sensors and the historical fault database; Overheating fault judgment: If the temperature data Continuously above the set overheating temperature threshold , and there is no downward trend for a period of time. At the same time, the voiceprint signal may show features such as increased high-frequency noise and abnormal vibration of specific frequencies. Combined with the possible local overheating and discoloration of the device in the image data, it is judged as an overheating fault; Short circuit fault judgment: When the current data The voltage suddenly increases and exceeds the normal operating current range. There may be fluctuations or drops, strong arc discharge sounds may appear in the voiceprint signal, and the switch state may change abnormally, which may be judged as a short circuit fault; Mechanical failure judgment: If abnormal mechanical friction sound, impact sound, etc. appear in the voiceprint signal, and such sound persists for a period of time, and abnormal displacement or looseness of equipment parts may be observed in the image data, it is judged as a mechanical failure; Step 5.3: Output alarm information: If it is judged to be abnormal, an alarm message is output; the alarm message should include the time when the fault occurred, the specific fault type and the severity of the fault; for example: "[time], a [fault type] fault occurred in the control panel cabinet of the hydropower station, and the severity was [high / medium / low];" At the same time, the fault type and recommended treatment measures are displayed on the cloud platform; for different fault types, corresponding treatment suggestions are provided so that the staff can take correct measures for maintenance and treatment in time; For overheating failures, the recommended treatment measures may include checking whether the cooling system is working properly, cleaning the cooling channel, reducing the equipment load, etc. In case of short circuit, it is recommended to cut off the power supply immediately, check whether there is a short circuit point in the circuit connection, replace the damaged electrical components, etc.; For mechanical failures, it is recommended to check whether the mechanical parts of the equipment are loose or worn, and tighten or replace damaged parts.

[0019] A multi-dimensional hydropower station control panel cabinet operating condition abnormality monitoring system based on sound recognition adopts the multi-dimensional hydropower station control panel cabinet operating condition abnormality monitoring method based on sound recognition.

[0020] The present invention can achieve the following beneficial effects: 1. By installing bone conduction pickups, mutual inductors, infrared image monitors and switch node sensors at different locations of the control panel, multi-dimensional data such as voiceprints, voltage, current, temperature, images and switch position change signals can be collected, which can comprehensively reflect the working status of the control panel from multiple angles and improve the accuracy and reliability of fault detection.

[0021] 2. Different types of data can complement each other and make up for the limitations of a single data type. For example, voiceprint signals can detect some mechanical faults that are difficult to detect through electrical parameters, while voltage and current data can reflect the operating status of the electrical system, temperature and image data can provide information about heat dissipation and appearance changes, and switch displacement signals can monitor changes in switch status.

[0022] 3. Pre-emphasize the voiceprint signal to compensate for the loss of the high-frequency part, protect the integrity of the vocal channel information, and improve the accuracy of subsequent feature extraction.

[0023] 4. The voiceprint signal is processed in frames and converted to the frequency domain through fast Fourier transform (FFT), which reduces the impact of non-steady-state time variation and makes each frame signal relatively stable in a short time, facilitating frequency domain analysis and observing the energy distribution of the signal at different frequencies, providing richer information for subsequent feature extraction.

[0024] 5. Use the PLDA algorithm to calculate the log-likelihood ratio of two audios to determine whether the audios come from the same device. This can effectively determine the similarity of the audios and thus determine whether the control panel cabinet has a fault. The algorithm takes into account potential variables such as the identity or status of the device, improving the accuracy of fault judgment.

[0025] 6. Combine MFCC features and i-vector features to form a multi-dimensional feature vector, and use the PLDA model to analyze the combined feature vector. At the same time, the log-likelihood ratio of the PLDA model is comprehensively scored with other sensor data to form a comprehensive fault score. This fusion decision-making method fully utilizes the advantages of multi-dimensional data and improves the accuracy and reliability of fault judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a flow chart of the present invention; Figure 2 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0027] The preferred solution is Figure 1 to Figure 2 As shown, a multi-dimensional hydropower station control panel cabinet operating condition abnormality monitoring method based on sound recognition, the steps are: Step 1: Data collection: 1.1. Install the microphone to collect voiceprint signals: When installing bone conduction pickups inside, on top, and on the outside of a control cabinet, you need to consider the sensitivity, frequency response range, and anti-interference capabilities of the pickups. For example, choosing a pickup with high sensitivity can better capture weak soundprint signals, while a pickup with a wide frequency response range can collect richer sound information.

[0028] For the collected voiceprint signal , , , the characteristics of the soundprint signals at different locations can be further analyzed. For example, the soundprint signal inside the control cabinet may contain more operating sounds of electrical equipment, the soundprint signal on the top may be affected by factors such as the cooling fan, and the soundprint signal on the outer shell may be interfered by the surrounding environmental noise. By analyzing the soundprint signals at different locations, the working status of the control cabinet can be judged more accurately.

[0029] 1.2. Install transformers to collect voltage and current data: The voltage transformer and current transformer should be installed in a position that ensures accurate measurement of the voltage and current at the power coil of the control panel cabinet. During installation, it is necessary to ensure that the wiring of the transformer is correct and firm to avoid measurement errors.

[0030] For the collected voltage data and current data , can monitor the changes in voltage and current in real time. For example, when the control panel cabinet fails, the voltage and current may fluctuate abnormally. By analyzing the voltage and current data, potential fault hazards can be discovered in time.

[0031] 1.3. Install infrared image monitor to collect temperature and images: The infrared image monitor should be installed in a position that covers the key areas around the control panel cabinet to accurately monitor the ambient temperature and obtain clear images. During the installation process, it is necessary to pay attention to the parameters such as the viewing angle, resolution and frame rate of the monitor to ensure that high-quality images and temperature data can be obtained.

[0032] For the collected temperature data , you can set different temperature thresholds and issue an early warning when the temperature exceeds the threshold. Through analysis, we can observe the changes in the appearance of the control panel cabinet, such as whether there are overheated areas, whether there are any damaged parts, etc.

[0033] 1.4. Install switch node sensors to detect switch status changes: The installation of the switch node sensor should ensure that it can accurately detect the state changes of the three-phase switch of the control panel cabinet. During the installation process, attention should be paid to the sensitivity and reliability of the sensor to avoid false alarms or missed alarms.

[0034] For switch status , can monitor the switch's open and closed status in real time, and issue an alarm in time when the switch status changes abnormally. For example, when a switch is accidentally opened or closed, it may mean that the control panel cabinet has a fault.

[0035] Step 2: Data transmission: Data transmission is the real-time transmission of collected data to the server for subsequent processing and analysis. During the data transmission process, data can be transmitted to the server by wired or wireless means. If the wired method is used, high-speed communication protocols such as Ethernet can be used to ensure the stability and reliability of data transmission. If the wireless method is used, communication technologies such as WiFi, Bluetooth or ZigBee can be used, but attention should be paid to issues such as signal interference and transmission distance.

[0036] Transmitted data vector Each element in the data needs to be accurately encoded and transmitted. For example, for voiceprint signals, audio encoding format can be used for transmission; for image data, image compression format can be used for transmission; for voltage, current and temperature data, digital signals can be used for transmission.

[0037] On the server side, a corresponding data receiving and storage system needs to be established to process and analyze the transmitted data. The server can use a high-performance computer or server cluster to ensure that a large amount of data can be processed quickly.

[0038] Step 3: Data processing and feature extraction: 3.1. Voiceprint signal pre-emphasis processing: In practical applications, the pre-emphasis factor The value of needs to be adjusted according to the specific characteristics of the voiceprint signal. If the high-frequency part of the voiceprint signal has a large loss, it can be appropriately increased. If the high frequency part of the voiceprint signal is relatively good, you can reduce it appropriately. The value of .

[0039] Voiceprint signal after pre-emphasis processing Further analysis can be performed to observe the enhancement effect of the high-frequency part. For example, the energy change of the high-frequency part before and after pre-emphasis processing can be compared by plotting the spectrum of the voiceprint signal.

[0040] 3.2. Voiceprint signal frame processing: The length of each frame and frame shift The value of needs to be adjusted according to the characteristics of the voiceprint signal and the requirements of subsequent processing. If the voiceprint signal changes slowly, the frame length can be appropriately increased. , reduce frame shift ; If the voiceprint signal changes rapidly, the frame length can be appropriately reduced , increase the frame shift .

[0041] Voiceprint signal after frame processing Further analysis can be performed to observe the stability of each frame signal. For example, the stability of each frame signal can be evaluated by calculating statistics such as the energy, mean, and variance of each frame signal.

[0042] 3.3 Fast Fourier Transform (FFT): In practical applications, the fast Fourier transform (FFT) algorithm can be implemented using efficient FFT libraries, such as FFTW (Fastest Fourier Transform in the West), etc. These libraries can greatly improve the calculation speed of FFT and meet the requirements of real-time monitoring.

[0043] Voiceprint signal converted to frequency domain through FFT You can further analyze the signal to see how the signal is distributed at different frequencies. For example, you can plot a spectrum to see the signal's main frequency, harmonics, and noise.

[0044] 3.4. Mel frequency filter bank: Conversion formula between Mel frequency and actual frequency The parameters in can be adjusted according to the auditory characteristics of the human ear. For example, the conversion relationship between the Mel frequency and the actual frequency in different frequency ranges can be determined through experiments to better simulate the sensitivity of the human ear to different frequencies.

[0045] Output after passing through the Mel filter bank You can further analyze and observe the response of different Mel filters. For example, you can plot the frequency response curve of the Mel filter and evaluate the selectivity of each filter for different frequencies.

[0046] 3.5. Logarithmic energy calculation: Logarithmic energy calculation formula The logarithmic function in can adopt different logarithmic forms such as natural logarithm or common logarithm. In practical applications, the appropriate logarithmic form can be selected according to the characteristics of the data and the requirements of subsequent processing.

[0047] The calculated logarithmic energy You can further analyze the signal to see how the energy is distributed. For example, you can plot a histogram of the logarithmic energy to assess how uniform the energy distribution of the signal is at different frequencies.

[0048] 3.6. Extract Mel frequency cepstral coefficient (MFCC) features: MFCC calculation formula Parameters in and The value of needs to be adjusted according to specific application requirements. Generally speaking, The larger the value of , the higher the dimension of the MFCC feature vector, and the more frequency information can be extracted; The value of is usually between 12-13, plus an energy coefficient, forming a 13-dimensional eigenvector.

[0049] The extracted MFCC feature vector can be further analyzed to observe its ability to represent the voiceprint signal. For example, the stability and discrimination of the MFCC feature vector can be evaluated by calculating the mean, variance, and correlation of the MFCC feature vector.

[0050] 3.7. Detect and process invalid data: In practical applications, a variety of methods can be used to detect invalid data. For example, a threshold can be set, and when the data exceeds the threshold range, it is considered invalid data; statistical methods such as mean, variance, and standard deviation can also be used, and when the deviation of the data from the mean exceeds a certain range, it is considered invalid data.

[0051] Different processing methods can be used for the detected invalid data. For example, the invalid data can be directly deleted; or the invalid data can be estimated and repaired by using interpolation or regression methods.

[0052] 3.8. Extract feature vector (i-vector) using universal background model (UBM) and Gaussian mixture model (GMM): When training UBM and GMM models, you need to select appropriate training data and parameters. For example, you can select a large amount of voiceprint data from different devices as training data to improve the generalization ability of the model; you can adjust the number of Gaussian mixtures. , weight , mean vector and the covariance matrix etc. to optimize the performance of the model.

[0053] Through the total variation matrix In the process of mapping the mean supervector to the i-vector space, it is necessary to select an appropriate total variation matrix The influence of different total variation matrices on the i-vector eigenvector can be determined through experiments to select the optimal total variation matrix.

[0054] 3.9. Adaptively adjust the Gaussian mixture model (GMM) parameters through the maximum a posteriori probability (MAP) algorithm: When using the MAP algorithm to update the parameters of the Gaussian mixture model, it is necessary to select an appropriate smoothing parameter and the prior mean , covariance matrix The effects of different parameter values ​​on model performance can be determined experimentally to select the optimal parameter values.

[0055] Updated mean and covariance matrices and Further analysis can be performed to observe the convergence of the model. For example, the log-likelihood function value of the model can be calculated to evaluate whether the performance of the model improves with the increase in the number of iterations.

[0056] Step 4: Data analysis and fault diagnosis: 4.1. Calculate the log-likelihood ratio using the PLDA algorithm: In practical applications, the parameters of the PLDA model can be estimated using methods such as maximum likelihood estimation or Bayesian estimation. These methods can estimate the global mean vector based on the training data. , factor loading matrix , latent variables and the noise term Parameter value.

[0057] When calculating the log-likelihood ratio of two audios, you need to select a suitable audio pair for comparison. You can select audio collected by the same device at different times, or audio collected by different devices, to evaluate the model's ability to distinguish.

[0058] 4.2. Determine the audio similarity to judge the fault: Define the audio generation model for each device When using the audio signal processing method, different probability distribution models can be used, such as Gaussian distribution, mixed Gaussian distribution or hidden Markov model, etc. Select a suitable model according to the actual situation to better describe the audio characteristics of the device.

[0059] Output parameters through training data In the process of training, it is necessary to select appropriate optimization algorithms and training parameters. You can use optimization algorithms such as gradient descent, Newton method or quasi-Newton method, and adjust parameters such as learning rate and number of iterations to improve the training efficiency and accuracy of the model.

[0060] 4.3. Fusion Decision: When extracting MFCC features from sound signals and obtaining feature vectors of control cabinets through i-vector extraction, it is necessary to select appropriate feature extraction methods and parameters. Parameters such as the dimension of MFCC features and the length of i-vector can be adjusted to optimize the performance of feature vectors.

[0061] When combining MFCC features and i-vector features, different combination methods can be used, such as concatenation, weighted summation, or fusion methods in deep learning, etc. Select the appropriate combination method according to the actual situation to improve the distinguishing ability of the feature vector.

[0062] When using the PLDA model to analyze the combined feature vectors, it is necessary to select appropriate PLDA model parameters and scoring methods. The parameters of the PLDA model can be adjusted, such as the dimension of the factor loading matrix, the distribution of latent variables, etc.; different scoring methods can be used, such as log-likelihood ratio, cosine similarity or Euclidean distance, to evaluate the similarity of feature vectors.

[0063] When comprehensively scoring the log-likelihood ratio of the PLDA model with other sensor data, it is necessary to select an appropriate weight coefficient. The importance of different sensor data can be determined through experiments, and the weight coefficients can be adjusted to improve the accuracy of the comprehensive score.

[0064] Setting Thresholds When the fault is detected, it needs to be adjusted according to the actual situation. The appropriate threshold range can be determined by analyzing historical fault data to improve the accuracy of fault judgment.

[0065] Step 5: Fault result judgment: Fault result judgment is to determine whether the control panel cabinet has a fault based on the results of data analysis and fault judgment, and output corresponding alarm information and processing measures. The following is a detailed embodiment of each sub-step of step 5.

[0066] 5.1. Determine the confidence value: When calculating the confidence value When and threshold The calculation is accurate. The reliability of the result can be ensured through multiple calculations and verifications.

[0067] According to the confidence value When judging whether the control panel cabinet may have a fault, it is necessary to set reasonable judgment criteria. When , it is judged that there may be a fault; when At the same time, the judgment standard can be adjusted according to the actual situation to improve the accuracy of fault judgment.

[0068] 5.2. Determine the fault type When judging the fault type, it is necessary to comprehensively consider the data characteristics of different sensors and the historical fault database. A fault diagnosis model can be established, the sensor data can be input into the model, and the judgment result of the fault type can be output.

[0069] For overheat fault judgment, the overheat temperature threshold can be set Adjustments are made based on the actual working conditions of the control cabinet and equipment specifications. At the same time, the accuracy of overheating fault judgment can be improved by analyzing the characteristics of the voiceprint signal, such as the increase in high-frequency noise, abnormal vibration of specific frequencies, and local overheating and discoloration of the equipment in the image data.

[0070] For short circuit fault judgment, the normal working current range can be set. If it exceeds this range, it is judged that there may be a short circuit fault. At the same time, it can be combined with voltage data The fluctuation or decline of the signal, the strong arc discharge sound in the voiceprint signal, and the abnormal change of the switch state can improve the accuracy of short-circuit fault judgment.

[0071] For mechanical fault diagnosis, we can determine whether there is a mechanical fault by analyzing abnormal mechanical friction sound, impact sound and other features in the voiceprint signal, as well as abnormal displacement and looseness of equipment components in the image data. At the same time, we can establish a mechanical fault diagnosis model and improve the accuracy of fault diagnosis through machine learning and other methods.

[0072] 5.3. Output alarm information When outputting alarm information, it is necessary to ensure the accuracy and timeliness of the information. The alarm information should include the time when the fault occurred, the specific fault type, and the severity of the fault. SMS, email, sound and light alarm, etc. can be used to notify relevant personnel in a timely manner.

[0073] When displaying fault types and recommended treatment measures on the cloud platform, it is necessary to ensure that the information is clear and easy to understand. For different fault types, corresponding treatment suggestions are provided so that the staff can take the correct measures for repair and treatment in time. The treatment suggestions should be specific and feasible, and be able to guide the staff to troubleshoot and repair the faults.

[0074] A multi-dimensional hydropower station control panel cabinet operating abnormality monitoring system based on sound recognition is a comprehensive system that integrates data collection, transmission, processing, analysis and fault judgment functions. The following is a detailed embodiment of the system.

[0075] 1. System architecture: The system can adopt a distributed architecture, consisting of multiple sensor nodes, data transmission networks and servers. The sensor nodes are responsible for collecting data such as the voiceprint, voltage, current, temperature, image and switch position signal of the control cabinet, and transmit the data to the server through the data transmission network. The server is responsible for processing, analyzing and fault judgment of the data, and outputting alarm information and treatment measures.

[0076] The data transmission network can be wired or wireless, such as Ethernet, WiFi, Bluetooth or ZigBee communication technologies. Choose the appropriate communication technology according to the actual situation to ensure the stability and reliability of data transmission.

[0077] 2. Sensor selection and installation: The selection of sensors should be based on the monitoring needs and the characteristics of the control cabinet. For example, for voiceprint signal acquisition, you can choose a bone conduction pickup with high sensitivity and a wide frequency response range; for voltage and current data acquisition, you can choose a voltage transformer and current transformer with high accuracy and good stability; for temperature and image data acquisition, you can choose an infrared image monitor with high resolution and fast response speed; for switch state detection, you can choose a switch node sensor with high reliability and good sensitivity.

[0078] The installation position of the sensor should ensure that the working condition data of the control panel can be accurately collected. For example, the bone conduction pickup should be installed in different locations such as the inside, top and shell of the control panel to collect voiceprint signals from different directions; the voltage transformer and current transformer should be installed at the power coil of the control panel to accurately measure the voltage and current data; the infrared image monitor should be installed above the wall around the control panel to cover the key area and monitor the ambient temperature and image; the switch node sensor should be installed at the three-phase switch of the control panel to detect the change of the switch state.

[0079] 3. Data processing and analysis software: The system needs to be equipped with special data processing and analysis software to realize the functions of processing the collected data, feature extraction, data analysis and fault diagnosis. The software can adopt modular design, including data acquisition module, data transmission module, data processing module, feature extraction module, data analysis module and fault diagnosis module.

[0080] The data processing module can perform preprocessing on the collected data, such as filtering, denoising, normalization and other operations, to improve the quality and stability of the data. The feature extraction module can use the above-mentioned voiceprint signal processing method to extract acoustic features such as MFCC features and i-vector features, as well as features of sensor data such as voltage, current, and temperature. The data analysis module can use machine learning, deep learning and other methods to analyze and model the extracted features to determine whether the control panel cabinet has a fault. The fault judgment module can output alarm information and treatment measures based on the results of data analysis.

[0081] 4. System performance evaluation and optimization: In order to ensure the performance and reliability of the system, it is necessary to evaluate and optimize the system. The actual hydropower station control panel cabinet can be used for testing, a large amount of operating data can be collected, and the system's fault detection accuracy, false alarm rate, missed alarm rate and other performance indicators can be evaluated.

[0082] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A multi-dimensional hydropower station control cabinet operating condition abnormality monitoring method based on sound recognition, characterized in that The following steps are involved: Step 1: Collect data; Step 2: transmitting the real-time collected data of the control cabinet to the server, wherein the data includes voiceprint, voltage, current, temperature, image and switch position change signal; Step 3: Data processing and feature extraction; Step 3.1, pre-emphasize the collected voiceprint signal to compensate for the loss of the high-frequency part; Step 3.2, process the voiceprint signal in frames to reduce the impact of non-steady-state time variation; Step 3.3, convert the framed voiceprint signal into the frequency domain through fast Fourier transform; Step 3.4, applying the Mel frequency filter bank on the frequency domain signal; Step 3.5, calculate the logarithmic energy of each Mel filter output; Step 3.6, extracting Mel frequency cepstral coefficient features; Step 3.7: Detect and process invalid data to ensure data availability; Step 3.8, extract feature vectors using the general background model and Gaussian mixture model; Step 3.9, adaptively adjust the Gaussian mixture model parameters through the maximum a posteriori probability algorithm; Step 4: Perform data analysis and fault diagnosis; Step 5: Determine the fault result.

2. According to the multi-dimensional hydropower station control cabinet operating condition abnormality monitoring method based on sound recognition according to claim 1, it is characterized by: The sub-steps of step 1 are as follows: Step 1.1: Install the microphone to collect voiceprint signals: Bone conduction microphones are installed inside, on the top and on the outer shell of the control cabinet to collect voiceprint signals at different positions. , , Respectively represent the soundprint signals collected from the inside, top and shell of the control cabinet; the soundprint signal can be expressed as time The function of , , ; Step 1.2: Install the transformer to collect voltage and current data: Install voltage transformer and current transformer at the power coil of control panel cabinet to collect voltage and current data; Indicates the collected voltage data. Indicates the collected current data; Step 1.3: Install infrared image monitor to collect temperature and images: Install infrared image monitors on the walls around the control cabinet to monitor ambient temperature and images; Indicates the collected temperature data. Represents the collected image data; Step 1.4: Install the switch node sensor to detect the switch state change: Install a switch node sensor at the three-phase switch of the control panel cabinet to detect the change of switch status; Indicates the switch state, with a value of 0 or 1, indicating the switch is open or closed, respectively.

3. The method for monitoring abnormal working conditions of a multi-dimensional hydropower station control cabinet based on sound recognition according to claim 2 is characterized by: In step 2, the transmitted data can be represented as a vector; 。 4. The method for monitoring abnormal working conditions of a multi-dimensional hydropower station control cabinet based on sound recognition according to claim 3 is characterized by: In step 3.1, the collected sound signal is pre-emphasized to compensate for the loss of the high-frequency part and to protect the integrity of the sound channel information; The pre-emphasis formula is: ， in is the original voiceprint signal, is the pre-emphasized voiceprint signal, is the pre-emphasis coefficient.

5. The method for monitoring abnormal working conditions of a multi-dimensional hydropower station control cabinet based on sound recognition according to claim 4 is characterized in that: In step 3.2, the voiceprint signal is processed in frames to reduce the impact of non-steady-state time variation, so that each frame signal remains relatively stable in a short time; Assume that the length of each frame is sampling points, the frame shift is sampling points; The first The voiceprint signal of a frame can be expressed as: ； Represents the original voiceprint signal, which is the time Function of Indicates the frame number; Indicates the length of each frame, that is, the number of sampling points contained in each frame; Indicates The voiceprint signal of the frame is obtained from the original voiceprint signal A section taken from time i Start to Finish.

6. The method for monitoring abnormal working conditions of a multi-dimensional hydropower station control cabinet based on sound recognition according to claim 5 is characterized by: In step 3.3, the framed voiceprint signal is converted to the frequency domain through fast Fourier transform; the fast Fourier transform formula is: ， Indicates the frame number; represents the frequency index; Represents a time index; Indicates the length of each frame, that is, the number of sampling points contained in each frame; Indicates The frame voiceprint signal has a time index of The value at time; Indicates The representation of the frame voiceprint signal in the frequency domain, that is, the frequency index is The value at time; is a natural constant; Represents an imaginary unit.

7. The method for monitoring abnormal working conditions of a control panel cabinet of a hydropower station in multiple dimensions based on sound recognition according to claim 6 is characterized by: In step 3.4, applying the Mel frequency filter bank to the frequency domain signal can better simulate the sensitivity of the human ear to different frequencies; the conversion formula between Mel frequency and actual frequency is: ， in is the Mel frequency, is the actual frequency; For the Mel filter at frequency The output after the Mel filter bank is ; In step 3.5, the logarithmic energy is calculated for the output of each Mel filter; the logarithmic energy calculation formula is: ； In step 3.6, the Mel frequency cepstral coefficient features are extracted to reflect the nonlinear perception of the human ear to the frequency and extract more representative acoustic features; The MFCC calculation formula is: ， in For the The first frame of the voiceprint signal MFCC coefficients, is the number of Mel filters, It usually takes a value between 12 and 13, plus an energy coefficient, to form a 13-dimensional eigenvector; In step 3.7, invalid data is detected and processed to ensure data availability; whether the data is valid can be determined by setting a threshold or using statistical methods; In step 3.8, the feature vector is a more compact vector extracted from the GMM mean supervector, which is used to represent the characteristics of the control cabinet; let is the GMM model, where is the number of Gaussian mixtures, For the The weights of the Gaussian components, For the Gaussian distribution, is the mean vector, is the covariance matrix; by training UBM and GMM, the mean supervector can be obtained ,in is the feature dimension; Then the total variation matrix Map the mean supervector to the feature vector space, that is, ,in is the feature vector, is the average eigenvector, is a low-dimensional random vector; In step 3.9, the Gaussian mixture model parameters are adaptively adjusted through the maximum a posteriori probability algorithm to obtain the characteristic vector of the control cabinet; the update formula of the MAP algorithm is: ， ， in and are the updated mean and covariance matrices, and are the old mean and covariance matrices, For the The number of samples of Gaussian components, is the smoothing parameter, and are the prior mean and covariance matrices.

8. The method for monitoring abnormal working conditions of a control panel cabinet of a hydropower station in multiple dimensions based on sound recognition according to claim 1 is characterized by: The sub-steps of step 4 are: Step 4.

1. Calculate the log-likelihood ratio using the PLDA algorithm: The PLDA algorithm is used to calculate the log-likelihood ratio of two audios to determine whether the audios come from the same device; the PLDA model can be expressed as ,in is the eigenvector, mu is the global mean vector, is the factor loading matrix, is a latent variable, representing the identity or status of the device, is the noise term; let and is the feature vector of the two audios, then the log-likelihood ratio calculation formula is: ； In the formula, and is the feature vector of two audios; is a latent variable, indicating the identity or status of the device; Indicates that given the latent variable In the case of , the audio feature vector is probability; Indicates that given the latent variable In the case of , the audio feature vector is probability; is a latent variable The probability distribution of and They represent different latent variables corresponding to the two audios; Indicates that given the latent variable In the case of , the audio feature vector is probability; Indicates that given the latent variable In the case of , the audio feature vector is probability; and The latent variables and The probability distribution of Step 4.2, determine the audio similarity to judge the fault; By comparing the log-likelihood ratio, the similarity of the audio is determined, thereby judging whether the control panel cabinet has a fault; Assume that the training data audio is composed of audio from multiple devices, define the audio generation model for each device; For the The audio generation model of the device , is the number of devices; Through the training data, the output can best represent the parameters of the data set ;in is the global mean vector, is the factor loading matrix, is a diagonal matrix representing the variance of different devices, is the noise covariance matrix; Calculate the log-likelihood ratio of two audios to determine whether they come from the same device or the same state; Step 4.3, fusion decision: Extract MFCC features from the sound signal, and obtain the feature vector of the control cabinet through feature vector extraction; Combine the MFCC features and the feature vector of the control cabinet to form a multi-dimensional feature vector; ,in is the MFCC feature vector, is the characteristic vector of the control cabinet; The combined feature vectors were analyzed using the PLDA model to calculate the log-likelihood ratio; The log-likelihood ratio of the PLDA model is comprehensively scored with other sensor data to form a comprehensive fault score; ,in is the log-likelihood ratio score for the PLDA model, Score the voltage data, Score the current data, Score the temperature data, Score the switch status, is the weight coefficient; Set a threshold ,If the comprehensive score exceeds the threshold, it is judged as a fault; If the comprehensive score is close to the threshold, it will be combined with the data from the infrared image monitor for secondary confirmation; Determine the specific fault type based on the comprehensive score and known fault data in the database.

9. The method for monitoring abnormal working conditions of a control panel cabinet of a hydropower station in multiple dimensions based on sound recognition according to claim 1 is characterized by: The sub-steps of step 5 are: Step 5.1: Determine the confidence value: Confidence value Indicates the degree of certainty of whether the control cabinet has failed; through a comprehensive score With threshold The difference is calculated, that is ;if , indicating that the comprehensive score exceeds the threshold, it can be preliminarily determined that the control panel cabinet may have a fault; if , it indicates that the comprehensive score is within the normal range and the control cabinet may be in normal operation; Step 5.2, determine the fault type: If the confidence value , then further determine the specific fault type; analyze and judge based on the data characteristics of different sensors and the historical fault database; Overheating fault judgment: If the temperature data Continuously above the set overheating temperature threshold , and there is no downward trend for a period of time. At the same time, the voiceprint signal may show features such as increased high-frequency noise and abnormal vibration of specific frequencies. Combined with the possible local overheating and discoloration of the device in the image data, it is judged as an overheating fault; Short circuit fault judgment: When the current data The voltage suddenly increases and exceeds the normal operating current range. There may be fluctuations or drops, strong arc discharge sounds may appear in the voiceprint signal, and the switch state may change abnormally, which may be judged as a short circuit fault; Mechanical failure judgment: If abnormal mechanical friction sound, impact sound, etc. appear in the voiceprint signal, and such sound persists for a period of time, and abnormal displacement or looseness of equipment parts may be observed in the image data, it is judged as a mechanical failure; Step 5.3: Output alarm information: If it is judged to be abnormal, an alarm message is output; the alarm message should include the time when the fault occurred, the specific fault type and the severity of the fault; At the same time, the fault type and recommended treatment measures are displayed on the cloud platform; for different fault types, corresponding treatment suggestions are provided so that the staff can take correct measures for maintenance and treatment in time; For overheating failures, the recommended treatment measures may be to check whether the cooling system is working properly, clean the cooling channel, and reduce the equipment load; In case of short circuit, it is recommended to cut off the power supply immediately, check whether there is a short circuit point in the circuit connection, and replace the damaged electrical components; For mechanical failures, it is recommended to check whether the mechanical parts of the equipment are loose or worn, and tighten or replace damaged parts.

10. A multi-dimensional hydropower station control panel cabinet operating abnormality monitoring system based on sound recognition, characterized by: A multi-dimensional hydropower station control panel cabinet operating condition abnormality monitoring method based on sound recognition according to any one of claims 1 to 9 is adopted.

Images