Voiceprint monitoring and diagnosing method for water-turbine generator set

Abstract

A water-turbine generator set voiceprint monitoring and diagnosing method comprises the steps that a multi-channel voiceprint collecting system is adopted, a plurality of voiceprint sensors are arranged at key parts of a water-turbine generator set, and voiceprint signals at different positions are obtained; performing time-frequency analysis on the pre-processed voiceprint signal by using wavelet transform, and extracting voiceprint features to obtain a voiceprint feature vector; if the matching degree of the voiceprint feature vector and the features in the fault feature library exceeds a preset threshold value, determining the fault type and the fault degree; classifying the determined fault type and fault degree by adopting a support vector machine algorithm to obtain a fault diagnosis result; a fault diagnosis result is optimized through a deep learning network, the diagnosis precision is improved, and a final fault diagnosis report is obtained; performing association analysis on the fault diagnosis report and the equipment operation parameters to generate an equipment operation state evaluation report; according to the equipment operation state evaluation report, an equipment maintenance plan is made, and intelligent maintenance of the water-turbine generator set is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of hydro-generator equipment monitoring, and in particular relates to a method for monitoring and diagnosing soundprints of a hydro-generator set. Background Art

[0002] During the long-term operation of hydro-turbine generator sets, due to the impact of water flow, mechanical vibration, electromagnetic interference and other factors, its internal structure and components are prone to wear, looseness, falling off and other problems, resulting in the gradual deterioration of the equipment operation status and even causing serious failures and accidents. At present, the monitoring of hydro-turbine generator sets mainly relies on regular manual inspections and routine measurement of parameters such as vibration and temperature, but these methods have lags and limitations, making it difficult to timely detect and diagnose potential equipment failures.

[0003] Sound is an important information carrier that reflects the operating status of equipment. Different fault types and degrees will produce sounds with different characteristics. However, the sound generated by the operation of the hydro-turbine generator set is complex and changeable, mixed with multiple components such as water flow noise, mechanical noise, electromagnetic noise, and the sound of different parts is obviously different. A single sound sensor is difficult to fully collect and identify. In addition, the installation environment of the hydro-turbine generator set is harsh, with strong electromagnetic interference, high humidity, vibration and other factors, which puts higher requirements on voiceprint collection and transmission. How to achieve all-round, multi-angle, and high-precision voiceprint collection and diagnosis of hydro-turbine generator sets in a complex environment is a technical problem that needs to be solved urgently. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for monitoring and diagnosing the soundprint of a hydro-turbine generator set, by arranging multiple soundprint sensors at key locations to ensure that the operating status of the equipment is fully captured from different angles and positions. The multi-channel data acquisition system can synchronously obtain the soundprint signals of each sensor, ensuring the consistency and accuracy of the data.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for monitoring and diagnosing soundprints of a hydro-turbine generator set, comprising the following steps: S1: A multi-channel voiceprint acquisition system is used to arrange multiple voiceprint sensors at key locations of the hydro-generator set to obtain voiceprint signals at different locations; S1.1: According to the structural characteristics of the hydro-generator set, determine the key monitoring locations and install the soundprint sensors at the key monitoring locations; S1.2: Use a multi-channel data acquisition system to synchronously acquire the voiceprint signals collected by each sensor to obtain the initial voiceprint data; Assume that sensors are arranged at five key parts of the hydro-turbine generator set, such as the bearing, runner room, and stator. The sensor collection frequency of each part is 100Hz, and the collection time is 10s. Then the total number of data points collected is (N = 5×100×10 = 5000), which is the length of the initial voiceprint data. S1.3: Preprocess the initial voiceprint data to obtain standardized voiceprint data; the preprocessing includes denoising and normalization; S1.4: Extract the time-frequency domain features of the voiceprint data, including spectrum, cepstrum and Mel-frequency cepstrum coefficient (MFCC), and obtain the voiceprint feature vector; For example, when using the MFCC algorithm, the calculation process is as follows: ; in: represents the mth Mel-frequency cepstrum coefficient, , ; is the preprocessed voiceprint signal sequence, n is the sequence index, and its value range is 0 to ,in ; is the Hamming window function, which is used to perform windowing on the signal to reduce spectrum leakage. Its form is: , ; is the frequency index after discrete Fourier transform (DFT), assuming that ,here To represent different frequency components; is a constant used for the cosine transform, usually equal to ; S1.5: Based on historical fault cases, establish a corresponding relationship database between voiceprint features and fault types to form a voiceprint fault diagnosis model; S1.6: Use machine learning algorithms, such as support vector machines (SVMs), to perform fault diagnosis on the current voiceprint features to determine whether there is a fault and the type of fault; S1.7: If the diagnosis results show that there is a fault, the cause of the fault shall be comprehensively analyzed in combination with other monitoring data such as vibration and temperature, and a corresponding maintenance plan shall be formulated.

[0006] S2: Pre-process the collected voiceprint signal through a preset noise filtering algorithm to remove interference components such as water flow noise, mechanical noise and electromagnetic noise; S2.1: Obtain the collected original voiceprint signal as the input signal for noise filtering; S2.2: According to the preset noise type, determine the water flow noise and location, use the wavelet decomposition method to perform multi-scale decomposition on the input soundprint signal to obtain wavelet coefficients of different frequency bands; S2.4: According to the energy distribution characteristics of each frequency band, determine whether there are interference components of water flow noise, mechanical noise and electromagnetic noise; S2.5: If it is determined that there is interference noise, threshold processing is performed on the wavelet coefficients of the corresponding frequency band to remove the noise interference; S2.6: Perform wavelet reconstruction on the processed wavelet coefficients to obtain the voiceprint signal after removing noise interference; S2.7: Output the denoised voiceprint signal as an input signal for subsequent voiceprint recognition to improve recognition accuracy.

[0007] S3: Use wavelet transform to perform time-frequency analysis on the preprocessed voiceprint signal, extract voiceprint features, and obtain a voiceprint feature vector; S3.1: Obtain the preprocessed voiceprint signal, use the wavelet transform algorithm to perform time-frequency analysis on it, extract the characteristic parameters of the voiceprint signal according to the time-frequency analysis results, and construct the voiceprint feature vector; then use the support vector machine algorithm to classify the voiceprint feature vector, train and generate a voiceprint recognition model; obtain the voiceprint signal to be identified, preprocess and extract features to obtain the voiceprint feature vector to be identified; input the voiceprint feature vector to be identified into the voiceprint recognition model to perform identity authentication; if the identity authentication is passed, the user is allowed to perform subsequent operations; otherwise, the authentication failure message is prompted; continuously optimize the voiceprint recognition algorithm and model parameters to improve the system's security protection capabilities and user experience.

[0008] S4: According to the pre-established fault feature library, the extracted voiceprint feature vector is matched with the features in the fault feature library to determine whether there is a fault; S4.1: According to the pre-established fault feature library, the voiceprint signal is obtained and preprocessed, and the voiceprint feature vector is extracted using the Mel Frequency Cepstral Coefficient (MFCC) algorithm; S4.2: For the extracted voiceprint feature vector, the dynamic time warping (DTW) algorithm is used to match it with the features in the fault feature library, and the similarity between the feature vectors is calculated; at the same time, in order to measure the matching degree between the voiceprint feature vector and the features in the fault feature library, the Euclidean distance formula can also be used: ; Specific parameters: is the currently extracted voiceprint feature vector, dimension ; is a feature vector in the fault feature library, with the same dimension ; and The eigenvectors are and The i-th component of ; Representation vector and The smaller the Euclidean distance between them, the more similar the two feature vectors are; S4.3: According to the set similarity threshold, determine whether the voiceprint feature vector matches the feature in the fault feature library. If the similarity is greater than the threshold, determine that the corresponding fault exists; S4.4: For the matched fault features, obtain the corresponding fault type and fault location information, and determine the specific situation of the fault; S4.5: If the voiceprint feature vector matches multiple fault features, the most likely fault type and location are determined based on the degree of similarity; S4.6: Correlate the fault judgment result with the voiceprint signal to obtain fault diagnosis information corresponding to the voiceprint signal for subsequent fault analysis and maintenance; S4.7: Continuously collect voiceprint signals during equipment operation, regularly update the fault feature library, and improve the accuracy and coverage of fault diagnosis.

[0009] S5: If the matching degree between the voiceprint feature vector and the feature in the fault feature library exceeds a preset threshold, the fault type and fault degree are determined; S5.1: Acquire the sound signal during the operation of the device, and pre-process the sound signal, including filtering and noise reduction operations, to obtain the pre-processed sound signal; S5.2: For the preprocessed sound signal, the Mel Frequency Cepstral Coefficient (MFCC) algorithm is used to extract the voiceprint features to obtain the voiceprint feature vector; S5.3: According to the pre-established fault feature library, obtain feature vectors of various faults as reference feature vectors; S5.4: Calculate the matching degree between the voiceprint feature vector and the reference feature vector, using a metric such as Euclidean distance or cosine similarity to obtain a matching degree value; S5.5: Determine whether the matching value exceeds a preset threshold value. If it exceeds the threshold value, determine the fault type corresponding to the voiceprint feature; S5.6: Based on the determination of the fault type, the fault degree is judged according to the eigenvalue of the voiceprint feature vector by using the support vector machine (SVM) algorithm to obtain the fault degree level; For example, when using the SVM algorithm to judge the degree of fault, taking linear classification as an example, the formula is: ; Specific parameters: It is a classification decision function, which judges the sample according to the output symbol The category to which it belongs, so as to classify the fault type and degree; is the Lagrange multiplier obtained by solving the optimization problem and used to determine the weights of the support vectors; is a training sample The corresponding category labels, for example, the fault severity is divided into mild, moderate, and severe, respectively. , 0, 1 represents; Representation vector and The inner product operation is used to measure the similarity between samples; It is a bias term, which is also determined through the training process and determines the classification decision boundary together with the support vector; is the number of samples in the training set; S5.7: The fault type and fault severity level are used as output results and provided to the subsequent fault diagnosis and maintenance decision-making process to provide a basis for the operation and maintenance management of the equipment.

[0010] S6: Using the support vector machine algorithm, classify the determined fault type and fault degree to obtain the fault diagnosis result; S6.1: Obtain various parameter data during the operation of the equipment, pre-process the data, remove noise and outliers, and obtain a standardized data set; S6.2: Based on the prior knowledge of fault type and degree, key parameters reflecting fault characteristics are extracted from the data set to construct feature vectors; S6.2: Use the support vector machine algorithm to train the feature vector and obtain the optimal classification decision boundary by optimizing the model parameters; S6.3: When a device fails, obtain the current operating parameters, construct a feature vector, and input it into the trained support vector machine model; S6.4: The support vector machine model determines the fault type and degree to which the feature vector belongs based on the classification decision boundary and outputs the diagnosis result; S6.5: If the confidence of the diagnosis result is lower than the preset threshold, the support vector machine model is updated and optimized online through the incremental learning method; S6.6: Feedback the diagnostic results to the equipment control system and take appropriate maintenance or fault-tolerant measures according to the type and severity of the fault to ensure the reliable operation of the equipment.

[0011] S7: Through the deep learning network, the fault diagnosis results are optimized to improve the diagnosis accuracy and obtain the final fault diagnosis report; S7.1: Based on the business scenarios of deep learning and fault diagnosis, a convolutional neural network is used as the basic network model, and a network training data set is constructed to improve the diagnostic accuracy. S7.2: Clean and label the original fault diagnosis data through data preprocessing and data enhancement to obtain high-quality training samples; S7.3: Use the cross-validation method to divide the training data into a training set and a test set, and set appropriate evaluation indicators, such as accuracy, precision, and recall; S7.4: Select appropriate optimization algorithms, such as stochastic gradient descent and Adam, according to the complexity of the network model and the size of the training data, and adjust hyperparameters, such as learning rate and regularization coefficient; S7.5: During the training process, monitor the performance of the model on the validation set to determine whether overfitting or underfitting occurs, and take appropriate measures, such as early stopping and regularization. S7.6: Based on the evaluation results on the test set, fine-tune and optimize the model until the expected diagnostic accuracy is achieved; S7.7: Apply the trained deep learning model to the actual fault diagnosis scenario, generate the final diagnosis report, and continuously collect user feedback to iteratively optimize the model.

[0012] S8: Correlate and analyze the fault diagnosis report with the equipment operation parameters to generate an equipment operation status assessment report; S8.1: Obtain the operating parameter data and fault diagnosis report data of the equipment, and pre-process the operating parameter data, including data cleaning and normalization, to obtain standardized operating parameter data; S8.2: Perform text mining and key information extraction on the fault diagnosis report data to extract key information such as fault type, fault time and fault cause; S8.3: Based on the standardized operating parameter data and the extracted key fault information, an association rule mining algorithm is used to mine the association rules between the operating parameters and the fault types to obtain a parameter-fault association matrix; S8.4: Based on the parameter-fault association matrix, the support vector machine algorithm is used to train the equipment operation status evaluation model, the input is the real-time operation parameters of the equipment, and the output is the operation status score of the equipment; S8.5: During the operation of the equipment, the equipment operation parameters are collected in real time and input into the trained equipment operation status evaluation model to obtain the real-time operation status score of the equipment; S8.6: Based on the historical operating status score data of the equipment, use the time series analysis algorithm to predict the operating status trend of the equipment in the future and determine whether the equipment has the risk of failure; S8.7: Automatically generate an equipment operation status assessment report based on the equipment's real-time operation status score and future operation status prediction results. The report content includes information such as the equipment's current operation status, future operation status prediction, and fault risk prompts.

[0013] S9: Develop equipment maintenance plans based on equipment operation status assessment reports to achieve intelligent maintenance of hydro-generator sets; S9.1: The operating status data of the turbine generator set is obtained through sensors, including parameters such as speed, vibration and temperature, and these data are transmitted to the central monitoring system for real-time monitoring and analysis: (1) Various types of sensors are arranged at key locations of the hydro-turbine generator set, such as speed sensors, vibration sensors, and temperature sensors, to collect real-time operating status data of the equipment; (2) Using wired or wireless communication technologies (such as industrial Ethernet, Wi-Fi, or Zigbee) to transmit the collected data to a central monitoring system or cloud server; (3) Store the received data in a database for subsequent processing and long-term storage; the database should have efficient data reading and writing performance and support rapid query and analysis of large-scale data; (4) The monitoring interface displays various operating parameters of the hydro-generator set in real time, helping operators to understand the operating status of the equipment in a timely manner; S9.2: The status assessment module uses the support vector machine algorithm to assess the current operating status of the hydro-generator set based on the pre-established health status model to determine whether the equipment is in a healthy state; S9.3: If the status assessment result indicates that the equipment is abnormal, the abnormal information is transmitted to the fault prediction module, which uses a convolutional neural network algorithm to predict the type and time of possible faults of the equipment based on historical operation data and current status data; S9.4: Based on the results of condition assessment and fault prediction, the intelligent maintenance decision module uses a decision tree algorithm to select the optimal maintenance strategy from the preset maintenance strategy library and generate a detailed equipment maintenance plan, including maintenance time, maintenance content, and required resources; S9.5: Send the generated equipment maintenance plan to the maintenance execution system, which automatically schedules maintenance tasks according to the plan and assigns the tasks to corresponding execution units; S9.6: During the maintenance task, the equipment status and maintenance progress are monitored in real time through the Internet of Things technology, and the data is fed back to the intelligent maintenance decision module for dynamic adjustment of maintenance strategies; S9.7: After the maintenance task is completed, the equipment's operating status data and maintenance records are updated, and the data is stored in the knowledge base for optimizing the health status model and maintenance strategy library to achieve continuous improvement and intelligent upgrading of equipment maintenance.

[0014] The present invention can achieve the following beneficial effects: 1. Multiple voiceprint sensors are placed at key locations to ensure that the operating status of the equipment is fully captured from different angles and positions. The multi-channel data acquisition system can synchronously obtain the voiceprint signals of each sensor to ensure the consistency and accuracy of the data.

[0015] 2. The preset noise filtering algorithm removes water flow noise, mechanical noise and electromagnetic noise, significantly improving the quality of the voiceprint signal and reducing the impact of interference on the diagnosis results. The initial voiceprint data is preprocessed by denoising and normalizing to obtain standardized data, which improves the reliability of subsequent feature extraction and analysis.

[0016] 3. Extracting multiple time-frequency domain features including spectrum, cepstrum and Mel-frequency cepstrum coefficients (MFCC) enhances the representation capability of feature vectors and helps to more accurately identify different types of fault modes.

[0017] 4. Based on historical fault cases, a corresponding relationship database between voiceprint features and fault types is established to form an intelligent fault diagnosis model, which improves the accuracy and efficiency of diagnosis. Machine learning algorithms such as support vector machines (SVM) are used to classify fault types and degrees, realizing automated and intelligent fault diagnosis.

[0018] 5. Combined with other monitoring data such as vibration and temperature, the fault cause is comprehensively analyzed to provide more comprehensive fault diagnosis information. According to the diagnosis results, corresponding maintenance plans are formulated to effectively guide maintenance work, reduce downtime, and improve equipment availability.

[0019] 6. Continue to collect voiceprint signals during equipment operation and regularly update the fault feature library to ensure the accuracy and coverage of the diagnostic system. Optimize fault diagnosis results through deep learning networks and iterate and improve the model based on user feedback to maintain the advancement and adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is a flow chart of the method of the present invention in Example 1; Figure 2 This is a data transmission diagram of the present invention in Example 1; Figure 3 This is a framework diagram of the soundprint monitoring and diagnostic system for a hydro-turbine generator set of the present invention in Example 1. DETAILED DESCRIPTION

[0021] Example: The preferred solution is Figures 1 to 3 As shown, a method for monitoring and diagnosing soundprints of a hydro-turbine generator set comprises the following steps: S1. Use a multi-channel voiceprint acquisition system and arrange multiple voiceprint sensors at key locations of the hydro-turbine generator set to obtain voiceprint signals at different locations.

[0022] According to the structural characteristics of the hydro-generator set, key monitoring locations are determined and voiceprint sensors are installed at these locations. A multi-channel data acquisition system is used to synchronously acquire the voiceprint signals collected by each sensor to obtain the initial voiceprint data. The initial voiceprint data is preprocessed, including denoising and normalization, to obtain standardized voiceprint data. The time-frequency domain features of the voiceprint data, including spectrum, cepstrum, MFCC, etc., are extracted to obtain the voiceprint feature vector. According to historical fault cases, a corresponding relationship library between voiceprint features and fault types is established to form a voiceprint fault diagnosis model. Machine learning algorithms such as support vector machines (SVM) are used to perform fault diagnosis on the current voiceprint features to determine whether there is a fault and the type of fault. If the diagnosis result shows that there is a fault, the cause of the fault is comprehensively analyzed in combination with other monitoring data such as vibration and temperature, and a corresponding maintenance plan is formulated.

[0023] For example, according to the structural characteristics of the hydro-generator set, it can be determined that the key monitoring parts mainly include the turbine water inlet, guide vane, runner, tailwater pipe, generator stator, rotor, etc. YD-SY-01 soundprint sensors are installed at these locations, with a sampling frequency of 52kHz and a dynamic range of 120dB. The NI PXIe-6358 16-channel synchronous data acquisition system is used to synchronously collect the soundprint signals collected by each sensor, and the data length of each sampling point is 24 bits. The initial soundprint data is subjected to wavelet denoising and maximum and minimum value normalization. The Daubechies4 wavelet is used for denoising, the decomposition layer number is 5, and the normalized data range is [0,1]. The time-frequency domain features of the soundprint data are extracted, including the spectrum features based on fast Fourier transform, the cepstrum features based on linear prediction analysis, and the MFCC features based on Mel frequency, etc., to form a 72-dimensional soundprint feature vector. According to historical fault cases, the K-means clustering algorithm is used to cluster the soundprint features and obtain six typical fault modes, namely turbine cavitation, abnormal guide vane opening, runner breakage, tailwater pipe blockage, generator stator winding inter-turn short circuit, rotor imbalance, etc. A corresponding relationship library between soundprint features and fault types is established to form a soundprint fault diagnosis model. The SVM support vector machine algorithm is used to diagnose the current soundprint features. The radial basis kernel function is used, the penalty factor C=10, and the kernel function parameter g=1. The diagnostic accuracy can reach more than 95%. If the diagnosis result shows that there is a fault, the Bayesian network algorithm is used to comprehensively analyze the cause of the fault in combination with other monitoring data such as vibration and temperature, and a probability distribution diagram of the fault cause is generated. According to the cause of the fault, a corresponding maintenance plan is formulated. For example, the turbine cavitation fault can be solved by optimizing the turbine operating conditions and increasing the tailwater pipe pressure. The abnormal guide vane opening fault can be solved by calibrating the guide vane opening sensor and checking the guide vane opening control system.

[0024] S2. Pre-process the collected voiceprint signal through a preset noise filtering algorithm to remove interference components of water flow noise, mechanical noise and electromagnetic noise.

[0025] The collected original voiceprint signal is obtained as the input signal for noise filtering. According to the preset noise type, the filtering algorithm for water flow noise, mechanical noise and electromagnetic noise is determined. The input voiceprint signal is decomposed at multiple scales using the wavelet decomposition method to obtain wavelet coefficients of different frequency bands. According to the energy distribution characteristics of each frequency band, it is determined whether there are interference components of water flow noise, mechanical noise and electromagnetic noise. If it is determined to contain interference noise, threshold processing is performed on the wavelet coefficients of the corresponding frequency band to remove noise interference. Wavelet reconstruction is performed on the processed wavelet coefficients to obtain the voiceprint signal after removing noise interference. The denoised voiceprint signal is output as the input signal for subsequent voiceprint recognition to improve recognition accuracy.

[0026] For example, first, the collected original voiceprint signal is preprocessed, including pre-emphasis, framing, windowing, etc., to obtain the input signal. According to the preset frequency range of water flow noise, mechanical noise and electromagnetic noise, the wavelet basis function and decomposition layer number for different noise types are determined. For example, for water flow noise, the db4 wavelet basis function can be selected to decompose 5 layers; for mechanical noise, the sym8 wavelet basis function can be selected to decompose 8 layers; for electromagnetic noise, the coif3 wavelet basis function can be selected to decompose 10 layers. The input signal is decomposed at multiple scales using wavelet transform to obtain the wavelet coefficients of each frequency band. By calculating the energy value of the wavelet coefficients of each frequency band, it is determined whether there is interference component. If the energy value exceeds the preset threshold, such as the energy of the water flow noise frequency band exceeds 20% of the total energy, it is determined to contain corresponding interference. For noisy frequency bands, a soft threshold or a hard threshold is set according to the noise type, and the wavelet coefficients are threshold processed to remove noise interference. For example, the water flow noise uses the heursure soft threshold with a threshold size of 2; the mechanical noise uses the sqtwolog hard threshold with a threshold size of 8; the electromagnetic noise uses the minimaxi soft threshold with a threshold size of 6. Finally, the processed wavelet coefficients are reconstructed by wavelet to obtain the denoised voiceprint signal, which is used as the input of voiceprint recognition and compared with the template in the voiceprint template library to identify the speaker. Experiments show that the voiceprint denoising method can effectively suppress the interference of water flow noise, mechanical noise and electromagnetic noise, and the average signal-to-noise ratio is improved by 5dB, and the recognition accuracy is improved by 12%.

[0027] S3. Use wavelet transform to perform time-frequency analysis on the preprocessed voiceprint signal, extract voiceprint features, and obtain a voiceprint feature vector.

[0028] Get the pre-processed voiceprint signal and use the wavelet transform algorithm to perform time-frequency analysis. According to the time-frequency analysis results, extract the characteristic parameters of the voiceprint signal and construct the voiceprint feature vector. Use the support vector machine algorithm to classify the voiceprint feature vector and train to generate the voiceprint recognition model. Get the voiceprint signal to be identified, pre-process and extract features to obtain the voiceprint feature vector to be identified. Input the voiceprint feature vector to be identified into the voiceprint recognition model for identity authentication. If the identity authentication is passed, the user is allowed to perform subsequent operations; otherwise, the authentication failure message is prompted. Continuously optimize the voiceprint recognition algorithm and model parameters to improve the system's security protection capabilities and user experience.

[0029] For example, the acquired voiceprint signal is first preprocessed, and the Butterworth filter is used to reduce noise on the signal. The order of the filter is 4 and the cutoff frequency is 8kHz. Then the preprocessed voiceprint signal is subjected to wavelet transform, and the Daubechies4 wavelet basis function is selected to decompose the signal into 4 layers to obtain the time-frequency representation matrix. According to the time-frequency analysis results, the characteristic parameters of the voiceprint signal are extracted, including the fundamental frequency, the resonance peak frequency, the zero-crossing rate, etc., and a 36-dimensional voiceprint feature vector is constructed. The support vector machine algorithm is used to train the voiceprint feature vector, and the radial basis kernel function is selected. The penalty factor C is set to 10, and the voiceprint recognition model is obtained by training. When the user requests voiceprint authentication, the system prompts the user to say the specified text, obtain the voiceprint signal to be identified, preprocess it and extract features, and obtain the 36-dimensional voiceprint feature vector to be identified. The feature vector is input into the voiceprint recognition model for classification. If the recognition result matches the user's registered voiceprint by more than 95%, the identity authentication is determined to be passed, and the user is allowed to perform subsequent operations; otherwise, the authentication fails and the user request is rejected. The system continuously collects users’ voiceprint data, regularly retrains the voiceprint recognition model, optimizes model parameters, and improves recognition accuracy. At the same time, encryption algorithms are used to protect users’ voiceprint data to prevent leakage and improve system security.

[0030] S4. According to a pre-established fault feature library, the extracted voiceprint feature vector is matched with the features in the fault feature library to determine whether a fault exists.

[0031] According to the pre-established fault feature library, the voiceprint signal is obtained and preprocessed, and the voiceprint feature vector is extracted using the Mel Frequency Cepstral Coefficient (MFCC) algorithm. For the extracted voiceprint feature vector, the dynamic time warping (DTW) algorithm is used to match the features in the fault feature library and calculate the similarity between the feature vectors. According to the set similarity threshold, it is determined whether the voiceprint feature vector matches the features in the fault feature library. If the similarity is greater than the threshold, it is determined that the corresponding fault exists. For the matched fault features, the corresponding fault type and fault location information are obtained to determine the specific situation of the fault. If the voiceprint feature vector matches multiple fault features, the most likely fault type and location are determined based on the size of the similarity. The fault judgment result is associated with the voiceprint signal to obtain the fault diagnosis information corresponding to the voiceprint signal for subsequent fault analysis and maintenance. Continuously collect voiceprint signals during equipment operation, regularly update the fault feature library, and improve the accuracy and coverage of fault diagnosis.

[0032] For example, in the fault diagnosis process, the acquired voiceprint signal is first preprocessed, and the voiceprint feature vector is extracted using the 10th-order Mel-Frequency Cepstrum Coefficient (MFCC) algorithm, with each frame length of 25 milliseconds and a frame shift of 10 milliseconds. Then the extracted voiceprint feature vector is matched with the features in the pre-established fault feature library using the dynamic time warping (DTW) algorithm, and the Euclidean distance between the feature vectors is calculated as the similarity measure, with the similarity threshold set to 8. When the similarity between the voiceprint feature vector and a feature in the fault feature library is greater than 8, it is determined that there is a corresponding fault, and the fault type and fault location information corresponding to the fault feature are obtained. If the similarity between the voiceprint feature vector and multiple fault features is greater than 8, the fault feature with the greatest similarity is selected to determine the most likely fault type and location. The fault judgment result is associated with the voiceprint signal, and a fault diagnosis report corresponding to the voiceprint signal is generated, including information such as the fault type, fault location, and fault occurrence time, which is used to guide subsequent fault analysis and maintenance work. During the operation of the equipment, voiceprint signals are continuously collected and the fault feature library is updated every hour. By adding new fault samples and optimizing the feature extraction algorithm, the accuracy and coverage of fault diagnosis are continuously improved to achieve intelligent operation and maintenance of the equipment.

[0033] S5. If the matching degree between the voiceprint feature vector and the feature in the fault feature library exceeds a preset threshold, the fault type and fault degree are determined.

[0034] Acquire the sound signal during the operation of the equipment, preprocess the sound signal, including filtering, noise reduction and other operations, and obtain the preprocessed sound signal. For the preprocessed sound signal, use the Mel Frequency Cepstral Coefficient (MFCC) algorithm to extract the voiceprint features and obtain the voiceprint feature vector. According to the pre-established fault feature library, obtain the feature vectors of various faults as reference feature vectors. Calculate the matching degree between the voiceprint feature vector and the reference feature vector, and use the Euclidean distance or cosine similarity and other measurement methods to obtain the matching degree value. Determine whether the matching degree value exceeds the preset threshold. If it exceeds the threshold, determine the fault type corresponding to the voiceprint feature. On the basis of determining the fault type, use the support vector machine (SVM) algorithm to determine the fault degree according to the feature value of the voiceprint feature vector and obtain the fault degree level. Provide the fault type and fault degree level as output results to the subsequent fault diagnosis and maintenance decision process, providing a basis for the operation and maintenance management of the equipment.

[0035] For example, when acquiring sound signals during the operation of the device, an audio acquisition device with a sampling frequency of 41kHz and a bit depth of 16bit can be used to acquire the sound signals in real time. When preprocessing the acquired sound signals, a Butterworth low-pass filter can be used, with a cutoff frequency of 5kHz and an attenuation rate of 12dB / oct, to effectively filter out high-frequency noise interference. At the same time, a spectral subtraction algorithm is used to reduce the noise of the signal, with a noise estimation window length of 10ms, an overlap rate of 50%, and a noise reduction coefficient of 8, to improve the signal-to-noise ratio of the signal. When extracting voiceprint features, the MFCC algorithm is used, with the number of Mel filter groups set to 24, the order of cepstral coefficients set to 12, the frame length set to 25ms, and the frame shift set to 10ms, to extract the voiceprint feature vector. According to the pre-established fault feature library, the feature vectors of typical faults such as bearing faults, gear faults, and motor faults are obtained as reference feature vectors. When calculating the matching degree between the voiceprint feature vector and the reference feature vector, the Euclidean distance measurement method is used, and the matching degree threshold is set to 85. When the matching degree exceeds the threshold, the fault type corresponding to the voiceprint feature is determined. When judging the degree of fault, the voiceprint feature vector is classified by SVM algorithm, the Gaussian kernel function parameter γ is set to 1, the penalty factor C is set to 10, and the fault degree is divided into three levels: mild, moderate, and severe. Finally, the fault type and fault degree level are used as output results to provide a quantitative basis for equipment status monitoring and fault diagnosis, and optimize equipment operation and maintenance strategies. Through the combined application of sound signal analysis and machine learning algorithms, early warning and accurate diagnosis of equipment failures are achieved, reducing unplanned equipment downtime and improving equipment operation reliability.

[0036] S6. Use the support vector machine algorithm to classify the determined fault type and fault degree to obtain the fault diagnosis result.

[0037] Obtain the data of various parameters during the operation of the equipment, pre-process the data, remove noise and outliers, and obtain a normalized data set. According to the prior knowledge of the fault type and degree, extract the key parameters reflecting the fault characteristics from the data set and construct the feature vector. Use the support vector machine algorithm to train the feature vector, and obtain the optimal classification decision boundary by optimizing the model parameters. When the equipment fails, obtain the current operating parameters, construct the feature vector, and input it into the trained support vector machine model. The support vector machine model determines the fault type and degree to which the feature vector belongs based on the classification decision boundary and outputs the diagnosis result. If the confidence of the diagnosis result is lower than the preset threshold, the support vector machine model is updated and optimized online through the incremental learning method. Feedback the diagnosis result to the equipment control system, and take corresponding maintenance or fault-tolerant measures according to the fault type and degree to ensure the reliable operation of the equipment.

[0038] For example, during the operation of the equipment, the temperature, pressure, vibration and other parameter data are collected by sensors with a sampling frequency of 1kHz. The collected data is preprocessed, and the median filter algorithm is used to remove random noise. The data that exceeds the normal range by 20% is regarded as anomalies and removed to obtain a normalized data set. According to historical fault cases and expert experience, 10 key parameters such as temperature, pressure, and vibration amplitude are selected to construct feature vectors. The LIBSVM toolbox is used to select the RBF kernel function, and the model parameters c and g are optimized through grid search and cross-validation. The optimal parameters c=10, g=1 are obtained to train the classification model. When the equipment operation parameters are abnormal, such as temperature exceeding 80℃, pressure below 2MPa, and vibration amplitude greater than 5mm, a feature vector is constructed to input the trained model. The model determines that the feature vector belongs to a bearing fault, the fault degree is moderate, and the diagnostic confidence is 92%. If the confidence is lower than 90%, incremental learning is automatically triggered, and the model is fine-tuned and updated using the newly collected fault data to improve the diagnostic accuracy. The diagnostic results are reported to the equipment control system. The control system adjusts the equipment operating parameters according to the preset fault handling logic, such as reducing the speed and increasing the lubricating oil flow. At the same time, a work order is generated to notify maintenance personnel to replace the bearing to ensure the safe operation of the equipment.

[0039] S7. Through the deep learning network, the fault diagnosis results are optimized, the diagnostic accuracy is improved, and the final fault diagnosis report is obtained.

[0040] According to the business scenarios of deep learning and fault diagnosis, convolutional neural network is used as the basic network model. Aiming at the goal of improving diagnostic accuracy, a network training data set is constructed. The original fault diagnosis data is cleaned and labeled through data preprocessing, data enhancement, etc. to obtain high-quality training samples. The training data is divided into training set, validation set and test set by cross-validation method, and appropriate evaluation indicators such as accuracy, precision, recall, etc. are set. According to the complexity of the network model and the scale of training data, appropriate optimization algorithms such as stochastic gradient descent and Adam are selected, and hyperparameters such as learning rate and regularization coefficient are adjusted. During the training process, by monitoring the performance of the model on the validation set, it is determined whether overfitting or underfitting occurs, and corresponding measures such as early stopping and regularization are taken. According to the evaluation results on the test set, the model is fine-tuned and optimized until the expected diagnostic accuracy is achieved. The trained deep learning model is applied to the actual fault diagnosis scenario to generate the final diagnosis report, and user feedback is continuously collected to iteratively optimize the model.

[0041] For example, in the business scenario of fault diagnosis, the original fault diagnosis data is first preprocessed, the Gaussian filter algorithm is used to denoise the data, and then the Z-score normalization method is used to standardize the data to improve the data quality. Then, the original data set is expanded to 50,000 samples through data enhancement techniques such as rotation, translation, and scaling to improve the generalization ability of the model. Next, the 5-fold cross-validation method is used to divide the data set into training set, validation set, and test set in a ratio of 8:1:1, and the accuracy, precision, recall rate, and F1 value are selected as evaluation indicators. Considering the large data scale, the Adam optimization algorithm is selected, and the learning rate is set to 001, the regularization coefficient is set to 01, and the batch size is set to 128. During the training process, by drawing the loss curves of the training set and the validation set, it is found that the model has overfitting in the 50th epoch, so the early stopping strategy is adopted to stop training when the loss of the validation set no longer decreases. After 100 epochs of training, the model achieved an accuracy of 95% on the test set. Finally, the trained model is deployed to the production environment to predict real-time fault diagnosis data and generate a diagnosis report. At the same time, through log analysis and user feedback collection, the performance of the model is continuously monitored and optimized to continuously improve diagnostic accuracy and user experience.

[0042] S8. Correlate and analyze the fault diagnosis report with the equipment operation parameters to generate an equipment operation status assessment report.

[0043] The operating parameter data and fault diagnosis report data of the equipment are obtained, and the operating parameter data are preprocessed, including data cleaning, normalization and other operations, to obtain standardized operating parameter data. Text mining and key information extraction are performed on the fault diagnosis report data to extract key information such as fault type, fault time, and fault cause. According to the standardized operating parameter data and the extracted key fault information, the association rule mining algorithm is used to mine the association rules between the operating parameters and the fault type to obtain the parameter-fault association matrix. According to the parameter-fault association matrix, the support vector machine algorithm is used to train the equipment operation status evaluation model, with the input being the real-time operating parameters of the equipment and the output being the equipment operation status score. During the operation of the equipment, the equipment operation parameters are collected in real time and input into the trained equipment operation status evaluation model to obtain the real-time operation status score of the equipment. According to the historical operation status score data of the equipment, the time series analysis algorithm is used to predict the operation status trend of the equipment in the future period of time to determine whether the equipment has the risk of failure. According to the real-time operation status score of the equipment and the future operation status prediction results, the equipment operation status evaluation report is automatically generated, and the report content includes information such as the current operation status of the equipment, the future operation status prediction, and the fault risk prompt.

[0044] For example, first, obtain the operating parameter data and fault diagnosis report data from the equipment. The operating parameter data includes the temperature, pressure, speed, etc. of the equipment. Each parameter is collected once per second for 1 hour, totaling 3600 sets of data. The fault diagnosis report data is in text format and contains information such as the time, type, and cause of the equipment's historical faults. Clean the operating parameter data to remove abnormal values, such as data whose temperature exceeds the rated operating temperature range of the equipment. Then normalize the data and map the data of each parameter to between 0 and 1 to eliminate the dimension effect. Next, perform text mining on the fault diagnosis report data. First, perform Chinese word segmentation, then use the TF-IDF algorithm to extract keywords. According to the frequency and order of the keywords, extract key information such as fault type, fault time, and fault cause. Then, use the Apriori association rule mining algorithm to mine the association rules between the operating parameters as the antecedent and the fault type as the posterior. Set the minimum support to 05 and the minimum confidence to 8, obtain the top 10 association rules with the highest confidence, and generate the parameter-fault association matrix. Then, according to the parameter-fault association matrix, the support vector machine (SVM) algorithm is used to train the equipment operation status assessment model. The association matrix is ​​used as the training sample of SVM, the operation parameters are used as the feature vector, the fault type is used as the label, and the RBF kernel function is used to optimize the penalty coefficient C of SVM and the kernel function parameter gamma through grid search to obtain the optimal SVM model for real-time equipment operation status assessment. During the operation of the equipment, the equipment operation parameters are collected every 30 seconds and input into the trained SVM model to obtain the current operation status score of the equipment. The operation status score data within the past month is used as a time series, and the ARIMA model is used for trend prediction to predict the operation status trend of the equipment within the next day. When the predicted score is lower than 6, it is judged that the equipment has the risk of failure and needs early warning. Finally, according to the current operation status score of the equipment and the operation status prediction results within the next day, the equipment operation status assessment report is automatically generated. The report content includes the current operation parameter values ​​of the equipment, the operation status score, the operation status prediction curve within the next 24 hours, the fault risk prompt, etc., which are presented in the form of charts and text descriptions, so that the equipment management personnel can intuitively understand the equipment operation status and take preventive maintenance measures in advance.

[0045] S9. Develop equipment maintenance plans based on the equipment operation status assessment report to achieve intelligent maintenance of the hydro-turbine generator set.

[0046] The operating status data of the hydro-generator set is obtained through sensors, including parameters such as speed, vibration, and temperature, and the data is transmitted to the status assessment module. The status assessment module uses the support vector machine algorithm to evaluate the current operating status of the hydro-generator set according to the pre-established health status model to determine whether the equipment is in a healthy state. If the status assessment result shows that the equipment is abnormal, the abnormal information is passed to the fault prediction module, which uses the convolutional neural network algorithm to predict the type and time of possible faults of the equipment based on historical operation data and current status data. According to the results of status assessment and fault prediction, the intelligent maintenance decision module uses the decision tree algorithm to select the optimal maintenance strategy from the preset maintenance strategy library and generate a detailed equipment maintenance plan, including maintenance time, maintenance content, and required resources. The generated equipment maintenance plan is sent to the maintenance execution system, which automatically schedules maintenance tasks according to the plan and assigns tasks to corresponding maintenance personnel or robots. In the process of executing maintenance tasks, the equipment status and maintenance progress are monitored in real time through the Internet of Things technology, and the data is fed back to the intelligent maintenance decision module for dynamic adjustment of maintenance strategies. After the maintenance task is completed, the equipment's operating status data and maintenance records are updated, and the data is stored in the knowledge base to optimize the health status model and maintenance strategy library, thereby achieving continuous improvement and intelligent upgrading of equipment maintenance.

[0047] For example, the operating status data of the hydro-turbine generator set is collected through vibration sensors, temperature sensors, etc., with a sampling frequency of 1kHz, and the data is transmitted to the status assessment module through industrial Ethernet. The status assessment module adopts the support vector machine algorithm and uses the radial basis kernel function to establish a health status model by training 80 sets of historical data. The model accuracy rate reaches 95%, and the current operating status of the hydro-turbine generator set is evaluated in real time. If the vibration data exceeds 5mm / s or the temperature data exceeds 80℃, the equipment is judged to be abnormal. After the abnormal information is transmitted to the fault prediction module, the module uses the convolutional neural network algorithm to analyze the operating data of the equipment in the past 3 months, extract vibration and temperature characteristics, and predict that the probability of bearing failure in the equipment within the next week is 80%. According to the above analysis results, the intelligent maintenance decision module adopts the C5 decision tree algorithm, combines factors such as equipment importance, failure risk, and maintenance cost, selects the optimal strategy from 200 preset maintenance strategies, and generates a detailed equipment maintenance plan, including replacing bearings within the next 3 days, and conducting a comprehensive inspection of the equipment. It is estimated that 4 maintenance personnel will be required and it will take 6 hours. The IoT technology is used to monitor the equipment status and maintenance progress in real time. If the abnormal status of the equipment deteriorates or the maintenance progress is delayed, the intelligent maintenance decision module is triggered to adjust the maintenance strategy, such as increasing the number of maintenance personnel or shortening the maintenance cycle. After the maintenance task is completed, the equipment operation status data and maintenance records are updated and stored in the knowledge base. The health status model and maintenance strategy library are optimized through machine learning algorithms, such as using genetic algorithms to adaptively adjust the maintenance strategy parameters, improve the adaptability and effectiveness of the strategy, and achieve continuous improvement and intelligent upgrades in equipment maintenance.

[0048] 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 method for monitoring and diagnosing soundprints of a hydro-turbine generator set, characterized in that The following steps are involved: S1: A multi-channel voiceprint acquisition system is used to arrange multiple voiceprint sensors at key locations of the hydro-generator set to obtain voiceprint signals at different locations; S2: Pre-process the collected voiceprint signal through a preset noise filtering algorithm to remove interference components such as water flow noise, mechanical noise and electromagnetic noise; S3: Use wavelet transform to perform time-frequency analysis on the preprocessed voiceprint signal, extract voiceprint features, and obtain a voiceprint feature vector; S4: According to the pre-established fault feature library, the extracted voiceprint feature vector is matched with the features in the fault feature library to determine whether there is a fault; S5: If the matching degree between the voiceprint feature vector and the feature in the fault feature library exceeds a preset threshold, the fault type and fault degree are determined; S6: Using the support vector machine algorithm, classify the determined fault type and fault degree to obtain the fault diagnosis result; S7: Through the deep learning network, the fault diagnosis results are optimized to improve the diagnosis accuracy and obtain the final fault diagnosis report; S8: Correlate and analyze the fault diagnosis report with the equipment operation parameters to generate an equipment operation status assessment report; S9: Develop equipment maintenance plans based on the equipment operation status assessment report to achieve intelligent maintenance of the hydro-turbine generator set.

2. A method for monitoring and diagnosing soundprints of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S1 are as follows: S1.1: According to the structural characteristics of the hydro-generator set, determine the key monitoring locations and install the soundprint sensors at the key monitoring locations; S1.2: Use a multi-channel data acquisition system to synchronously acquire the voiceprint signals collected by each sensor to obtain the initial voiceprint data; S1.3: Preprocess the initial voiceprint data to obtain standardized voiceprint data; the preprocessing includes denoising and normalization; S1.4: Extract the time-frequency domain features of the voiceprint data. The time-frequency domain features include spectrum, cepstrum and Mel-frequency cepstrum coefficients to obtain the voiceprint feature vector: When using the MFCC algorithm, the calculation process is as follows: ； in: represents the mth Mel-frequency cepstrum coefficient, , ; is the preprocessed voiceprint signal sequence, n is the sequence index, and its value range is 0 to ,in ; is the Hamming window function, which is used to perform windowing on the signal to reduce spectrum leakage. Its form is: ， ； is the frequency index after discrete Fourier transform. Assume that , To represent different frequency components; is a constant used for the cosine transform, usually equal to ; S1.5: Based on historical fault cases, establish a corresponding relationship database between voiceprint features and fault types to form a voiceprint fault diagnosis model; S1.6: Use machine learning algorithms to perform fault diagnosis on the current voiceprint features to determine whether there is a fault and the type of fault; S1.7: If the diagnosis results show that there is a fault, combine other monitoring data such as vibration and temperature to comprehensively analyze the cause of the fault and formulate a corresponding maintenance plan.

3. A method for monitoring and diagnosing soundprints of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S2 are as follows: S2.1: Obtain the collected original voiceprint signal as the input signal for noise filtering; S2.2: According to the preset noise type, determine the water flow noise and location, use the wavelet decomposition method to perform multi-scale decomposition on the input soundprint signal to obtain wavelet coefficients of different frequency bands; S2.4: According to the energy distribution characteristics of each frequency band, determine whether there are interference components of water flow noise, mechanical noise and electromagnetic noise; S2.5: If it is determined that there is interference noise, threshold processing is performed on the wavelet coefficients of the corresponding frequency band to remove the noise interference; S2.6: Perform wavelet reconstruction on the processed wavelet coefficients to obtain the voiceprint signal after removing noise interference; S2.7: Output the denoised voiceprint signal as an input signal for subsequent voiceprint recognition to improve recognition accuracy.

4. A method for monitoring and diagnosing soundprints of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S3 are as follows: Obtain the preprocessed voiceprint signal, use the wavelet transform algorithm to perform time-frequency analysis on it, extract the characteristic parameters of the voiceprint signal according to the time-frequency analysis results, and construct the voiceprint feature vector; then use the support vector machine algorithm to classify the voiceprint feature vector, and train to generate a voiceprint recognition model; obtain the voiceprint signal to be identified, preprocess and extract features to obtain the voiceprint feature vector to be identified; input the voiceprint feature vector to be identified into the voiceprint recognition model to perform identity authentication; if the identity authentication is passed, the user is allowed to perform subsequent operations; otherwise, the authentication failure message is prompted; continuously optimize the voiceprint recognition algorithm and model parameters to improve the system's security protection capabilities and user experience.

5. The method for monitoring and diagnosing soundprint of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S4 are as follows: S4.1: According to the pre-established fault feature library, the voiceprint signal is obtained and preprocessed, and the voiceprint feature vector is extracted using the Mel frequency cepstral coefficient algorithm; S4.2: For the extracted voiceprint feature vector, the dynamic time warping algorithm is used to match it with the features in the fault feature library, and the similarity between the feature vectors is calculated; at the same time, in order to measure the matching degree between the voiceprint feature vector and the features in the fault feature library, the Euclidean distance formula can also be used: ； Specific parameters: is the currently extracted voiceprint feature vector, dimension ; is a feature vector in the fault feature library, with the same dimension ; and The eigenvectors are and The i-th component of ; Representation vector and The smaller the Euclidean distance between them, the more similar the two feature vectors are; S4.3: According to the set similarity threshold, determine whether the voiceprint feature vector matches the feature in the fault feature library. If the similarity is greater than the threshold, determine that the corresponding fault exists; S4.4: For the matched fault features, obtain the corresponding fault type and fault location information, and determine the specific situation of the fault; S4.5: If the voiceprint feature vector matches multiple fault features, the most likely fault type and location are determined based on the degree of similarity; S4.6: Correlate the fault judgment result with the voiceprint signal to obtain fault diagnosis information corresponding to the voiceprint signal for subsequent fault analysis and maintenance; S4.7: Continuously collect voiceprint signals during equipment operation, regularly update the fault feature library, and improve the accuracy and coverage of fault diagnosis.

6. A method for monitoring and diagnosing soundprints of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S5 are as follows: S5.1: Acquire the sound signal during the operation of the device, and pre-process the sound signal, including filtering and noise reduction operations, to obtain the pre-processed sound signal; S5.2: For the preprocessed sound signal, the voiceprint features are extracted using the Mel frequency cepstral coefficient algorithm to obtain the voiceprint feature vector; S5.3: According to the pre-established fault feature library, obtain feature vectors of various faults as reference feature vectors; S5.4: Calculate the matching degree between the voiceprint feature vector and the reference feature vector, using a metric such as Euclidean distance or cosine similarity to obtain a matching degree value; S5.5: Determine whether the matching value exceeds a preset threshold value. If it exceeds the threshold value, determine the fault type corresponding to the voiceprint feature; S5.6: On the basis of determining the fault type, the fault degree is judged according to the eigenvalue of the voiceprint feature vector by using the support vector machine algorithm to obtain the fault degree level; When using the SVM algorithm to judge the degree of fault, taking linear classification as an example, the formula is: ； Specific parameters: It is a classification decision function, which judges the sample according to the output symbol The category to which it belongs, so as to classify the fault type and degree; is the Lagrange multiplier obtained by solving the optimization problem and used to determine the weight of the support vector; is a training sample The corresponding category labels are divided into mild, moderate and severe fault levels, respectively. , 0, 1 represents; Representation vector and The inner product operation is used to measure the similarity between samples; It is a bias term, which is also determined through the training process and determines the classification decision boundary together with the support vector; is the number of samples in the training set; S5.7: The fault type and fault severity level are used as output results and provided to the subsequent fault diagnosis and maintenance decision-making process to provide a basis for the operation and maintenance management of the equipment.

7. A method for monitoring and diagnosing soundprints of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S6 are as follows: S6.1: Obtain various parameter data during the operation of the equipment, pre-process the data, remove noise and outliers, and obtain a standardized data set; S6.2: Based on the prior knowledge of fault type and degree, key parameters reflecting fault characteristics are extracted from the data set to construct feature vectors; S6.2: Use the support vector machine algorithm to train the feature vector and obtain the optimal classification decision boundary by optimizing the model parameters; S6.3: When a device fails, obtain the current operating parameters, construct a feature vector, and input it into the trained support vector machine model; S6.4: The support vector machine model determines the fault type and degree to which the feature vector belongs based on the classification decision boundary and outputs the diagnosis result; S6.5: If the confidence of the diagnosis result is lower than the preset threshold, the support vector machine model is updated and optimized online through the incremental learning method; S6.6: Feedback the diagnostic results to the equipment control system and take appropriate maintenance or fault-tolerant measures according to the type and severity of the fault to ensure the reliable operation of the equipment.

8. The method for monitoring and diagnosing soundprint of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S7 are as follows: S7.1: Based on the business scenarios of deep learning and fault diagnosis, a convolutional neural network is used as the basic network model, and a network training data set is constructed to improve the diagnostic accuracy. S7.2: Clean and label the original fault diagnosis data through data preprocessing and data enhancement to obtain high-quality training samples; S7.3: Use the cross-validation method to divide the training data into a training set and a test set, and set appropriate evaluation indicators, including accuracy, precision, and recall; S7.4: Select an appropriate optimization algorithm and adjust hyperparameters based on the complexity of the network model and the size of the training data; S7.5: During the training process, monitor the performance of the model on the validation set to determine whether overfitting or underfitting occurs, and take appropriate measures; S7.6: Based on the evaluation results on the test set, fine-tune and optimize the model until the expected diagnostic accuracy is achieved; S7.7: Apply the trained deep learning model to the actual fault diagnosis scenario, generate the final diagnosis report, and continuously collect user feedback to iteratively optimize the model.

9. A method for monitoring and diagnosing soundprints of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S8 are as follows: S8.1: Obtain the operating parameter data and fault diagnosis report data of the equipment, and pre-process the operating parameter data, including data cleaning and normalization operations, to obtain standardized operating parameter data; S8.2: Perform text mining and key information extraction on the fault diagnosis report data to extract key information including fault type, fault time and fault cause; S8.3: Based on the standardized operating parameter data and the extracted key fault information, an association rule mining algorithm is used to mine the association rules between the operating parameters and the fault types to obtain a parameter-fault association matrix; S8.4: Based on the parameter-fault association matrix, the support vector machine algorithm is used to train the equipment operation status evaluation model, the input is the real-time operation parameters of the equipment, and the output is the operation status score of the equipment; S8.5: During the operation of the equipment, the equipment operation parameters are collected in real time and input into the trained equipment operation status evaluation model to obtain the real-time operation status score of the equipment; S8.6: Based on the historical operating status score data of the equipment, use the time series analysis algorithm to predict the operating status trend of the equipment in the future and determine whether the equipment has the risk of failure; S8.7: Automatically generate an equipment operation status assessment report based on the equipment's real-time operation status score and future operation status prediction results. The report content includes the equipment's current operation status, future operation status prediction, and fault risk warning information.

10. A method for monitoring and diagnosing soundprints of a hydro-generator set according to claim 1, characterized in that: The sub-steps of S9 are as follows: S9.1: Obtain the operating status data of the turbine generator set through sensors, including speed, vibration and temperature parameters, and transmit these data to the central monitoring system for real-time monitoring and analysis: (1) Various types of sensors are arranged at key locations of the hydro-turbine generator set, including speed sensors, vibration sensors, and temperature sensors, to collect real-time operating status data of the equipment; (2) Using wired or wireless communication technology, the collected data is transmitted to a central monitoring system or cloud server; (3) Store the received data in a database for subsequent processing and long-term storage; The database should have efficient data reading and writing performance and support fast query and analysis of large-scale data; (4) The monitoring interface displays various operating parameters of the hydro-generator set in real time, helping operators to understand the operating status of the equipment in a timely manner; S9.2: The status assessment module uses the support vector machine algorithm to assess the current operating status of the hydro-generator set based on the pre-established health status model to determine whether the equipment is in a healthy state; S9.3: If the status assessment result indicates that the equipment is abnormal, the abnormal information is transmitted to the fault prediction module, which uses a convolutional neural network algorithm to predict the type and time of possible faults of the equipment based on historical operation data and current status data; S9.4: Based on the results of condition assessment and fault prediction, the intelligent maintenance decision module uses a decision tree algorithm to select the optimal maintenance strategy from the preset maintenance strategy library and generate a detailed equipment maintenance plan, including maintenance time, maintenance content, and required resources; S9.5: Send the generated equipment maintenance plan to the maintenance execution system, which automatically schedules maintenance tasks according to the plan and assigns the tasks to corresponding execution units; S9.6: During the maintenance task, the equipment status and maintenance progress are monitored in real time through the Internet of Things technology, and the data is fed back to the intelligent maintenance decision module for dynamic adjustment of maintenance strategies; S9.7: After the maintenance task is completed, the equipment's operating status data and maintenance records are updated, and the data is stored in the knowledge base for optimizing the health status model and maintenance strategy library to achieve continuous improvement and intelligent upgrading of equipment maintenance.

Images