Generator rotor abnormal sound detection method and system
By processing the acoustic data of the hydro-generator using Log-Mel and CQT transforms, training a neural network model, and replacing the sensor category, the accuracy and flexibility issues of anomaly detection in hydro-generator units were resolved, enabling real-time monitoring and precise anomaly detection of the generator status.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SANXIA JINSHAJIANG YUNCHUAN HYDROPOWER DEV CO LTD
- Filing Date
- 2024-03-07
- Publication Date
- 2026-04-24
AI Technical Summary
Existing technologies are insufficient to effectively extract high signal-to-noise ratio fault signals during the operation of hydro-generator units, and traditional methods struggle to achieve accurate anomaly detection under complex background noise conditions.
Log-Mel transform and CQT transform are used to process the sound data of the hydro-generator to generate a spectrogram. An initial neural network model is trained and the last layer is replaced with sensors of multiple categories. The feature vectors are trained using a Gaussian mixture model to generate a classification model for anomaly detection.
It improves the accuracy and flexibility of abnormal sound detection of generator rotor, enhances the ability to classify different types of sounds, and realizes real-time monitoring and accurate anomaly detection of generator status.
Smart Images
Figure CN121922151A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning technology, and in particular to a method and system for detecting abnormal sounds from a generator rotor. Background Technology
[0002] In hydropower station operation, the turbine generator is one of the core pieces of equipment, and its stable operation is crucial for the normal operation of the hydropower station. However, due to the harsh working environment and long operating time of the turbine generator, various abnormalities are prone to occur, such as bearing damage, imbalance, and overheating. If these abnormalities are not detected and handled in a timely manner, they will seriously affect the safe and stable operation of the turbine generator and the entire hydropower station. Therefore, designing an efficient and accurate method for detecting abnormalities in turbine generators is of great significance for the safe operation of hydropower stations.
[0003] In existing technologies, hydro-generator units generate strong and complex background noise during operation, covering almost the entire frequency domain. Traditional signal processing methods such as wavelets and EMD cannot effectively extract fault signals with high signal-to-noise ratios. When there are multiple fault types, it is necessary to study the discrimination method for each case, greatly increasing the research difficulty. Traditional methods can utilize feature vectors extracted from neural networks using CQT and STFT, but the classification methods involved cannot be used for anomaly detection. Summary of the Invention
[0004] In view of the problems existing in existing methods for detecting abnormal sounds from generator rotors, this invention is proposed. Therefore, the problem to be solved by this invention is how to provide a method and system for detecting abnormal sounds from generator rotors.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0006] In a first aspect, the present invention provides a method for detecting abnormal sounds in a generator rotor, comprising: acquiring sound data of a hydro-generator; performing Log-Mel transform and CQT transform on the sound data to obtain a spectrogram; training an initial neural network model; replacing the last layer of the initial neural network model with sensors of multiple categories; using the spectrogram as training samples to obtain multiple fused feature vectors; training the sensors in the initial neural network model based on the multiple fused feature vectors to obtain a classification model corresponding to the sensors; using the classification model as an anomaly detection model for detecting hydro-generator anomalies; and calculating performance indicators to evaluate the detection results.
[0007] As a preferred embodiment of the generator rotor abnormal sound detection method of the present invention, the method of acquiring sound data of the hydro-generator includes acquiring sound data of the hydro-generator through a hydro-generator sound acquisition device; the sound acquisition device includes multiple sound acquisition units, a network relay, and a data server; the sound acquisition unit includes a microphone and a signal processing circuit, the microphone acquires the sound signal of the hydro-generator, the signal processing circuit performs signal conditioning, analog-to-digital conversion and digital signal processing on the sound signal, and the processed digital signal is packaged using the TCP / IP protocol.
[0008] As a preferred embodiment of the generator rotor abnormal sound detection method of the present invention, the Log-Mel transform includes the following steps: taking the normal operating time of the generator as the total length, segmenting the sound data, and using a preset window length and a preset shift length to control the size and overlap of each data window; for each data window, performing Log-Mel transform using a first transform parameter; using FFT to convert the time-domain signal of the sound data into a frequency-domain signal; dividing the frequency axis into several sub-bands according to the Mel scale, and summing the squares of the signals in each sub-band; then taking the logarithm of the energy value of each sub-band and arranging them in chronological order to generate a first spectrogram.
[0009] In a preferred embodiment of the generator rotor abnormal sound detection method of the present invention, the CQT transformation includes the following steps: performing CQT transformation on each data window using second transformation parameters to generate a second spectrogram; the relevant calculation formulas are as follows.
[0010]
[0011]
[0012] Round up
[0013] f k =2 k / b f min k = 0, 1, ..., K-1
[0014] And there are
[0015] f k+1 -f k =f k (2 1 / b -1)
[0016] Right now:
[0017] f k =2 k / b f min k = 0, 1, ..., K-1
[0018]
[0019]
[0020] in, It is of length N k The Hanning window, where Q is a constant factor in the CQT transform, and k is the frequency index of the CQT spectrum; f min For the lowest frequency to be processed, f k δ represents the frequency of the k-th component, and b is the number of spectral lines contained within an octave; f N represents the frequency bandwidth at frequency f, i.e., the frequency resolution. k It is the window length that varies with frequency, f s Indicates the sampling frequency.
[0021] As a preferred embodiment of the generator rotor abnormal sound detection method of the present invention, wherein: obtaining the classification model corresponding to the sensor includes training a Gaussian Mixture Model (GMM) by fusing the feature vectors of the sensors to obtain a GMM model for each sensor; the known sample set is D = {x1, x2, ... x...} m Cluster the samples into k classes, where the samples follow a Gaussian mixture distribution:
[0022]
[0023] In the formula, α i Let represent the mixing coefficient, i.e., the probability of selecting the i-th mixture component; initialize the model parameters α of the Gaussian mixture distribution. i μ i And ∑i; calculate the observed data x j The posterior probability generated from the mixture components, i.e., the observed data x j The probability p(z) generated by the i-th sub-model j =i|x j ), denoted as γ ji The relevant calculation formula is:
[0024]
[0025] The new model parameters are calculated using the following formulas:
[0026]
[0027]
[0028]
[0029] Repeat the training process with the new model parameters until the stopping condition is met; then, for each sample, perform training according to λ. j=argmaxγ ij , where i∈{1,2,…,k} are assigned to the corresponding clusters. Each sample is divided into clusters according to the probability from the sub-model, and finally k clusters are obtained.
[0030] As a preferred embodiment of the generator rotor abnormal sound detection method of the present invention, the step of obtaining multiple fused feature vectors includes generating a Log-Mel first feature vector based on a first spectrogram, generating a CQT second feature vector based on a second spectrogram, fusing the first feature vector and the second feature vector to obtain a fused feature vector; and training a Gaussian mixture model (GMM) on each sensor in the initial neural network model based on the multiple fused feature vectors to obtain a classification model corresponding to each sensor.
[0031] In a preferred embodiment of the generator rotor abnormal sound detection method of the present invention, the performance indicators include accuracy, precision, and recall calculated based on model prediction results, and the relevant calculation formulas are as follows:
[0032]
[0033]
[0034]
[0035] In the formula, AC represents accuracy, TP represents true positives, TN represents true negatives, FP represents false positives, FN represents false negatives, PR represents precision, and RE represents recall.
[0036] Secondly, the present invention provides a generator rotor abnormal sound detection system, comprising: a first acquisition module for acquiring sound data of a hydro-generator; a transformation module for performing Log-Mel transform and CQT transform on the sound data to obtain a spectrogram; a second acquisition module for acquiring a pre-trained initial neural network model and replacing the last layer of the initial neural network model with multiple categories of sensors; a transformation module for using the spectrogram as training samples and converting it into multiple fused feature vectors; and a training module for training each sensor in the initial neural network model based on the multiple fused feature vectors to obtain a classification model corresponding to each sensor, and using the classification models corresponding to each sensor as an anomaly detection model for detecting hydro-generator abnormalities.
[0037] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a generator rotor abnormal sound detection method.
[0038] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements the steps of a generator rotor abnormal sound detection method.
[0039] The beneficial effects of this invention are as follows: By collecting sound data from hydro-generators, real-time monitoring of generator operating status is achieved. This improves the system's perception of generator status and enhances the accuracy of anomaly detection. Sound data is processed using Log-Mel and CQT transforms, providing richer and more discriminative features for subsequent neural network models, thus enhancing the ability to identify abnormal sounds. By training the neural network and replacing the final layer with sensors of multiple categories, the neural network model can be customized to adapt to the characteristics of different sensors, improving the model's accuracy in classifying different types of sounds. The system is more flexible and adaptable. Effective fusion of information from different sensors improves the model's robustness and contributes to more accurate anomaly detection. The system's sensitivity to anomalies in different generator parts is increased, improving detection accuracy. Feedback evaluation results help improve the model, continuously enhancing the system's performance and reliability. Attached Figure Description
[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 This is a flowchart of a method for detecting abnormal sounds from a generator rotor.
[0042] Figure 2 A flowchart illustrating the training and application of a generator rotor abnormal sound detection classification model.
[0043] Figure 3 This is a structural diagram of the device for generating the anomaly detection model. Detailed Implementation
[0044] To make the above-mentioned objects, features, and advantages of the present invention more readily understood, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0047] Example 1
[0048] Reference Figure 1 This is the first embodiment of the present invention, which provides a method for detecting abnormal sounds from a generator rotor, including:
[0049] S1: Acquire the sound data of the hydro-generator, perform Log-Mel transform and CQT transform on the sound data to obtain the spectrogram;
[0050] Specifically, sound data from the hydro-generator can be acquired using a sound acquisition device. This device includes multiple sound acquisition units, a network relay, and a data server. Each sound acquisition unit comprises a microphone and a signal processing circuit. The multiple sound acquisition units are distributed within a preset range on the stator base. The sound acquisition units are used to: acquire sound signals from the hydro-generator using microphones; perform signal conditioning, analog-to-digital conversion, and digital signal processing on the sound signals using the signal processing circuit; package the processed digital signals using the TCP / IP protocol; the network relay is used to forward the packaged data; and the data server is used to receive and store the packaged data.
[0051] The microphone is used to collect the sound signal from the hydroelectric generator. A microphone is a sensor that converts sound signals into electrical signals. The signal processing circuit performs signal conditioning, analog-to-digital conversion, and digital signal processing on the sound signal collected by the microphone. This includes operations such as filtering, amplification, and gain control to ensure signal quality and meet the requirements of subsequent processing. TCP / IP packetization encapsulates the processed digital signal using the TCP / IP protocol for transmission over a network.
[0052] The network relay is responsible for receiving packetized data from the sound acquisition unit, forwarding the data, and sending it to the designated destination. The data server receives the packetized data from the network relay, stores and manages it for subsequent data analysis, processing, and use. Through the coordinated operation of these components, the sound acquisition device can acquire, process, transmit, and store sound data from the hydro-generator. This provides basic data support for subsequent data analysis, fault diagnosis, performance monitoring, and other applications.
[0053] Using the normal operating time of the generator as the total segmentation length, the sound data of the hydro-generator is segmented according to the preset window length and preset movement length to obtain multiple data windows. Based on the preset first transformation parameter, a Log-Mel transform is performed on each data window to generate a first spectrogram. Based on the preset second transformation parameter, a CQT transform is performed on each data window to generate a second spectrogram.
[0054] First, the data window is segmented. The audio data is segmented based on the generator's normal operating time. Preset window lengths and preset movement lengths are used to control the size and overlap of each data window.
[0055] For example, the total length can be divided into several windows of equal length, with each window having a preset length. By controlling the overlap between adjacent windows using a preset movement length, the data windows can have continuity and a certain degree of repeatability, thus better reflecting the generator's sound characteristics.
[0056] For each data window, a Log-Mel transformation is performed using preset first transformation parameters.
[0057] First, the time-domain signal is converted into a frequency-domain signal using FFT.
[0058] Then, the frequency axis is divided into several sub-bands according to the Mel scale, and the squares of the signals in each sub-band are summed. The logarithm of the energy value of each sub-band is then taken and arranged in chronological order to generate the first spectrogram.
[0059] Specifically, the data window is first segmented using parameters of a data window length of 2048 and a data window movement of 512. The audio data is then segmented according to the window size and movement length to obtain multiple data windows.
[0060] Perform an FFT transform on each data window to convert the time-domain signal to a frequency-domain signal. Apply a window function (such as the Hanning window) to each data window to reduce frequency domain leakage. Padded the number of sampling points in each data window to 2048; if insufficient, zero-padding is used.
[0061] Sixty-four Mel filters are created to divide the signal into different subbands in the frequency domain. The center frequency of each filter is distributed according to the Mel scale to ensure higher resolution in the low-frequency and high-frequency regions. The FFT-transformed spectrum is filtered through each Mel filter to obtain the energy value of each filter. The logarithm of the energy value of each filter is taken and normalized to the range of 0-1 to represent the energy distribution of the corresponding frequency band.
[0062] To capture temporal correlations and dynamic changes, each data window is expanded to the left and right by a context window of length 1000. Each data window is then expanded to a length of 2048 + 2 * 1000 = 4048 to obtain more contextual information.
[0063] The amplitude spectrum of the expanded window is indexed with an exponent of 2 to amplify smaller amplitude values and highlight larger amplitude differences.
[0064] The indexed amplitude spectrum is converted to a logarithmic scale to obtain the Log-Mel spectral features.
[0065] Features represent the energy distribution of a sound signal at different frequencies. Grayscale image / sample generation: Each Log-Mel spectral feature is treated as a sample, forming a grayscale image. Each sample is 1000×64×1 in size, representing 64 Mel filters and the temporal distribution of the energy values of each filter.
[0066] Through the above steps, a Log-Mel transform is performed on each data window, and the first spectrogram is generated for subsequent analysis and processing.
[0067] For each data window, a CQT transformation is performed using preset second transformation parameters. The CQT transformation can improve the time resolution of low-frequency signals while maintaining frequency resolution. Through the CQT transformation, the time-domain signal is converted into complex amplitude values at different frequencies, generating a second spectrogram.
[0068] Log-Mel transform and CQT transform are performed on each data window to generate a first spectrogram and a second spectrogram. These are used for subsequent applications such as feature extraction, classification, and fault diagnosis.
[0069] The second spectrogram is generated by performing a CQT transformation on each data window using the second transformation parameters. The relevant calculation formula is as follows.
[0070]
[0071]
[0072] Round up
[0073] fk =2 k / b f min k = 0, 1, ..., K-1
[0074] And there are
[0075] f k+1 -f k =f k (2 1 / b -1)
[0076] Right now:
[0077] f k =2 k / b f min k = 0, 1, ..., K-1
[0078]
[0079]
[0080] in, It is of length N k The Hanning window, where Q is a constant factor in the CQT transform, and k is the frequency index of the CQT spectrum; f min For the lowest frequency to be processed, f k δ represents the frequency of the k-th component, and b is the number of spectral lines contained within an octave; f N represents the frequency bandwidth at frequency f, i.e., the frequency resolution. k It is the window length that varies with frequency, f s Indicates the sampling frequency. N k The value is related to the value of k. The parameters used in the transformation are: data window length = 2048, data window shift = 512, context length = 1000, and K is 64. Assuming the lowest frequency being processed is fmin, fk represents the frequency of the k-th component, and b is the number of spectral lines contained in an octave, for example, b = 36, which means there are 36 spectral lines in each octave.
[0081] S2: Train the initial neural network model and replace the last layer of the initial neural network model with sensors of multiple categories;
[0082] Specifically, the pre-trained weights are run in the initial neural network model until a preset stopping condition is met.
[0083] Specifically, you can first obtain a pre-trained ResNet model (initial neural network model) and load pre-trained weights. At the end of the model, remove the original softmax layer and add a new fully connected layer with 6 output nodes (representing 6 sensor categories), using softmax as the activation function. Freeze all previous layers and train only the newly added fully connected layer. Train the newly added fully connected layer using training data until the training conditions are met. Fine-tune the entire network by unfreezing some of the previous convolutional and batch normalization layers (fine-tuning all previous layers). Then, train the entire network with a smaller learning rate until the training conditions are met. Finally, predict on the test set and calculate performance metrics such as accuracy, precision, and recall. The relevant calculation formulas are as follows:
[0084]
[0085]
[0086]
[0087] In the formula, AC represents accuracy, TP represents true positives, TN represents true negatives, FP represents false positives, FN represents false negatives, PR represents precision, and RE represents recall.
[0088] S3: Using spectrograms as training samples, multiple fused feature vectors are obtained;
[0089] Specifically, a Log-Mel first feature vector is generated based on each first spectrogram. Then, a CQT second feature vector is generated based on each second spectrogram. Finally, the first and second feature vectors are fused to obtain a fused feature vector.
[0090] The first and second feature vectors are 128-dimensional vectors, and the fused feature vector is a 256-dimensional vector. The first feature vector is based on Log-Mel, the second feature vector is based on CQT, and these two feature vectors will be fused into a single 256-dimensional feature vector.
[0091] S4: Based on multiple fused feature vectors, the sensors in the initial neural network model are trained to obtain the classification model corresponding to the sensors. The classification model is used as the anomaly detection model for detecting hydro-generators, and the performance indicators are calculated to evaluate the detection results.
[0092] Specifically, the feature vectors fused from the sensors are used to train a GMM to obtain a GMM model for each sensor.
[0093] Given a sample set D = {x1, x2, ..., x...} mCluster the samples into k classes, where the samples follow a Gaussian mixture distribution:
[0094]
[0095]
[0096] In the formula, α i This represents the mixing coefficient, i.e., the probability of selecting the i-th mixture component;
[0097] Initialize the model parameters α of the Gaussian mixture distribution i μ i and ∑i;
[0098] Calculate the observed data x j The posterior probability generated from the mixture components, i.e., the observed data x j The probability p(z) generated by the i-th sub-model j =i|x j ), denoted as γ ji The relevant calculation formula is:
[0099]
[0100] The new model parameters are calculated using the following formulas:
[0101]
[0102]
[0103]
[0104] Repeat the training process with the new model parameters until the stopping condition is met.
[0105] Each sample is divided according to λ j =arg maxγ ij , where i∈{1,2,…,k} are assigned to the corresponding clusters. Each sample is divided into clusters according to the probability from the sub-model, and finally k clusters are obtained.
[0106] The sound data of the hydro-generator is measured in real time. The sound data is then converted to obtain a feature vector with the same form as the fused feature vector. This feature vector is then input into the anomaly detection model to obtain a test score. If the test score is less than a preset threshold, it is determined that the hydro-generator has an anomaly.
[0107] Feature extraction is performed on the collected sound data to obtain a feature vector with the same form as the fused feature vector. Commonly used feature extraction methods include Short-Time Fourier Transform (STFT) and Mel-frequency coefficients (MFCC). These features can capture information such as the frequency, energy, and spectral pattern of the sound. The extracted feature vector is then input into the model for testing. Based on the output of the anomaly detection model, a corresponding test score is obtained. The test score is compared with a pre-set threshold. If the test score is less than the threshold, it can be determined that the hydroelectric generator is malfunctioning; otherwise, it indicates normal operation.
[0108] Furthermore, this embodiment also provides a generator rotor abnormal sound detection system, including: a first acquisition module 310, used to acquire sound data of a hydro-generator; a transformation module 320, used to perform Log-Mel transformation and CQT transformation on the sound data to obtain a spectrogram; a second acquisition module 330, used to acquire a pre-trained initial neural network model and replace the last layer of the initial neural network model with multiple categories of sensors; a transformation module 340, used to use the spectrogram as training samples and convert it into multiple fused feature vectors; and a training module 350, used to train each sensor in the initial neural network model based on the multiple fused feature vectors to obtain a classification model corresponding to each sensor, and use the classification models corresponding to each sensor as an anomaly detection model for detecting hydro-generator abnormalities.
[0109] The first acquisition module is specifically used for: acquiring the sound data of the hydro-generator through a sound acquisition device, wherein the sound acquisition device includes multiple sound acquisition units, a network relay, and a data server. Each sound acquisition unit includes a microphone and a signal processing circuit. The multiple sound acquisition units are distributed within a preset range on the stator base. The sound acquisition units are used to acquire the sound signals of the hydro-generator through the microphone; the signal processing circuit performs signal conditioning, analog-to-digital conversion, and digital signal processing on the sound signals; the processed digital signals are packaged using the TCP / IP protocol, and the network relay is used to forward the packaged data; the data server is used to receive and store the packaged data.
[0110] The transformation module is specifically used for: using the normal operating time of the generator as the total segmentation length, segmenting the sound data of the hydro-generator according to the preset window length and preset moving length to obtain multiple data windows; performing Log-Mel transformation on each data window based on the preset first transformation parameters, and generating a first spectrogram; and performing CQT transformation on each data window based on the preset second transformation parameters, and generating a second spectrogram.
[0111] The second acquisition module is also used to: run the pre-trained weights in the initial neural network model until a preset stopping condition is met.
[0112] The conversion module is specifically used for: generating a Log-Mel first feature vector based on each first spectrogram; generating a CQT second feature vector based on each second spectrogram; and fusing the first and second feature vectors to obtain a fused feature vector. The first and second feature vectors are 128-dimensional vectors, and the fused feature vector is a 256-dimensional vector.
[0113] The training module is specifically used to train a Gaussian Mixture Model (GMM) for each sensor in the initial neural network model based on multiple fused feature vectors, so as to obtain a classification model corresponding to each sensor.
[0114] The training module is also used to: measure the sound data of the hydro-generator in real time; convert the sound data to obtain a feature vector in the same form as the fused feature vector, and then input it into the anomaly detection model to obtain a test score; if the test score is less than a preset threshold, it is determined that the hydro-generator is abnormal.
[0115] This embodiment also provides a computer device applicable to the method for detecting abnormal sounds from a generator rotor, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement all or part of the steps of the method described in the above embodiments of the present invention.
[0116] This embodiment also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, it performs the method in any optional implementation of the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0117] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.
[0118] As shown above, by collecting sound data from the hydro-generator, real-time monitoring of the generator's operating status is achieved. This data is used to capture abnormal sounds from the generator rotor, providing accurate data for subsequent anomaly detection. This improves the system's perception of the generator's state and enhances the accuracy of anomaly detection. Sound data is processed using Log-Mel and CQT transforms. This achieves a conversion from the time domain to the frequency domain, resulting in a more representative spectrogram, providing richer and more discriminative features for subsequent neural network models. This improves the model's ability to abstract generator sound features and enhances the ability to identify abnormal sounds. By training the neural network and replacing the final layer with sensors of multiple categories, the neural network model is customized to adapt to the characteristics of different sensors. This improves the model's accuracy in classifying different types of sounds. This makes the system more flexible and adaptable, providing a better foundation for training subsequent classification models. Multiple fusion feature vectors are generated from multiple sensors, achieving effective fusion of information from different sensors. This improves the model's comprehensive understanding of overall sound features. This enhances the model's robustness, provides a more comprehensive grasp of the generator's state, and contributes to more accurate anomaly detection. By training sensor classification models using fused feature vectors, personalized anomaly detection model training was achieved for each sensor. Each sensor provides its own unique anomaly detection capability. This improves the system's sensitivity to anomalies in different generator parts, enhances detection accuracy, and allows for model improvement through feedback evaluation results, continuously improving system performance and reliability.
[0119] Example 2
[0120] Reference Figure 2 - Figure 3 This is the second embodiment of the present invention, which provides a method for detecting abnormal sounds from a generator rotor. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculations and simulation experiments.
[0121] Table 1 Comparison of Technical Features
[0122]
[0123] As shown above, by collecting sound data from the hydro-generator, real-time monitoring of the generator's operating status is achieved. This data is used to capture abnormal sounds from the generator rotor, providing accurate data for subsequent anomaly detection. This improves the system's perception of the generator's state and enhances the accuracy of anomaly detection. Sound data is processed using Log-Mel and CQT transforms. This achieves a conversion from the time domain to the frequency domain, resulting in a more representative spectrogram, providing richer and more discriminative features for subsequent neural network models. This improves the model's ability to abstract generator sound features and enhances the ability to identify abnormal sounds. By training the neural network and replacing the final layer with sensors of multiple categories, the neural network model is customized to adapt to the characteristics of different sensors. This improves the model's accuracy in classifying different types of sounds. This makes the system more flexible and adaptable, providing a better foundation for training subsequent classification models. Multiple fusion feature vectors are generated from multiple sensors, achieving effective fusion of information from different sensors. This improves the model's comprehensive understanding of overall sound features. This enhances the model's robustness, provides a more comprehensive grasp of the generator's state, and contributes to more accurate anomaly detection. By training sensor classification models using fused feature vectors, personalized anomaly detection model training was achieved for each sensor. Each sensor provides its own unique anomaly detection capability. This improves the system's sensitivity to anomalies in different generator parts, enhances detection accuracy, and allows for model improvement through feedback evaluation results, continuously improving system performance and reliability.
[0124] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for detecting abnormal sounds from a generator rotor, characterized in that: include, Acquire the sound data of the hydro-generator, and perform Log-Mel transform and CQT transform on the sound data to obtain the spectrogram; Train an initial neural network model, and then replace the last layer of the initial neural network model with sensors of multiple categories; Using spectrograms as training samples, multiple fused feature vectors are obtained; Based on multiple fused feature vectors, the sensors in the initial neural network model are trained to obtain the classification model corresponding to the sensors. The classification model is used as the anomaly detection model for detecting hydro-generators, and the performance indicators are calculated to evaluate the detection results.
2. The generator rotor abnormal sound detection method as described in claim 1, characterized in that: The acquisition of sound data from the hydro-generator includes acquiring sound data from the hydro-generator through a sound acquisition device. The sound acquisition device includes multiple sound acquisition units, a network relay, and a data server. The sound acquisition unit includes a microphone and a signal processing circuit. The microphone acquires the sound signal from the hydro-generator, and the signal processing circuit performs signal conditioning, analog-to-digital conversion, and digital signal processing on the sound signal. The processed digital signal is then packaged using the TCP / IP protocol.
3. The generator rotor abnormal sound detection method as described in claim 2, characterized in that: The Log-Mel transform includes the following steps. The normal operating time of the generator is used as the total length. The sound data is segmented, and the size and overlap of each data window are controlled by preset window length and preset movement length. For each data window, a Log-Mel transformation is performed using the first transformation parameter; Use FFT to convert the time-domain signal of the audio data into a frequency-domain signal; The frequency axis is divided into several sub-bands according to the Mel scale, and the squares of the signals in each sub-band are summed. Then, take the logarithm of the energy value of each subband and arrange them in chronological order to generate the first spectrogram.
4. The generator rotor abnormal sound detection method as described in claim 3, characterized in that: The CQT transformation includes the following steps: The second transformation parameter is used to perform a CQT transformation on each data window to generate a second spectrogram. The relevant calculation formula is as follows. Round up f k =2 k / b f min k=0,1,…,K-1 And there are f k+1 -f k =f k (2 1 / b -1) Right now: f k =2 k / b f min k=0,1,…,K-1 in, It is of length N k The Hanning window, where Q is a constant factor in the CQT transform, and k is the frequency index of the CQT spectrum; f min For the lowest frequency to be processed, f k δ represents the frequency of the k-th component, and b is the number of spectral lines contained within an octave; f N represents the frequency bandwidth at frequency f, i.e., the frequency resolution. k It is the window length that varies with frequency, f s Indicates the sampling frequency.
5. The generator rotor abnormal sound detection method as described in claim 4, characterized in that: The process of obtaining the classification model corresponding to the sensor includes training the GMM by fusing the feature vectors of the sensors to obtain the GMM model for each sensor. Given a sample set D = {x1, x2, ..., x...} m Cluster the samples into k classes, where the samples follow a Gaussian mixture distribution: In the formula, α i This represents the mixing coefficient, i.e., the probability of selecting the i-th mixture component; Initialize the model parameters α of the Gaussian mixture distribution i μ i and Σi; Calculate the observed data x j The posterior probability generated from the mixture components, i.e., the observed data x j The probability p(z) generated by the i-th sub-model j =i|x j ), denoted as γ ji The relevant calculation formula is: The new model parameters are calculated using the following formulas: Repeat the training process with the new model parameters until the stopping condition is met. Each sample is divided according to λ j =arg maxγ ij , where i∈{1,2,…,k} are assigned to the corresponding clusters. Each sample is divided into clusters according to the probability from the sub-model, and finally k clusters are obtained.
6. The generator rotor abnormal sound detection method as described in claim 5, characterized in that: The process of obtaining multiple fused feature vectors includes generating a Log-Mel first feature vector based on a first spectrogram, generating a CQT second feature vector based on a second spectrogram, and fusing the first and second feature vectors to obtain a fused feature vector. Based on multiple fused feature vectors, a Gaussian Mixture Model (GMM) is trained on each sensor in the initial neural network model to obtain a classification model corresponding to each sensor.
7. The generator rotor abnormal sound detection method as described in claim 6, characterized in that: The performance metrics include accuracy, precision, and recall, calculated based on the model prediction results. The relevant calculation formulas are as follows: In the formula, AC represents accuracy, TP represents true positives, TN represents true negatives, FP represents false positives, FN represents false negatives, PR represents precision, and RE represents recall.
8. A generator rotor abnormal sound detection system, based on the generator rotor abnormal sound detection method according to any one of claims 1 to 7, characterized in that: include, The first acquisition module is used to acquire the sound data of the hydro-generator; The transformation module is used to perform Log-Mel transform and CQT transform on the sound data to obtain the spectrogram; The second acquisition module is used to acquire a pre-trained initial neural network model and replace the last layer of the initial neural network model with sensors of multiple categories; The conversion module is used to take spectrograms as training samples and convert them into multiple fused feature vectors; The training module is used to train each sensor in the initial neural network model based on multiple fused feature vectors to obtain a classification model corresponding to each sensor, and to use the classification models corresponding to each sensor as an anomaly detection model for detecting hydro-generators.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the generator rotor abnormal sound detection method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the generator rotor abnormal sound detection method according to any one of claims 1 to 7.