Transform-based water turbine sound monitoring method

Abstract

The invention discloses a water turbine sound monitoring method based on transform. The method comprises the steps of 1, data collection, wherein sound data are collected from a historical monitoring system of a water turbine and related equipment; step 2, data processing: preprocessing the collected sound data; 3, Fourier transform processing: converting the collected time sequence sound signals into a spectrogram by using Fourier transform; step 4, performing pre-training by using an Audio Spectrogram Transform model architecture, so that the model learns, identifies and analyzes different types of sound data, and captures the characteristics of the sound data; in the model training process, the gradient of a loss function about each weight is calculated, and then the weights are adjusted by using the gradients; meanwhile, model hyper-parameters are adjusted; according to the invention, through continuous sound monitoring, real-time tracking of equipment conditions is realized, problems can be found earlier, maintenance can be carried out in advance, and the service life of the equipment is prolonged.

Description

Technical Field

[0001] The invention relates to the technical field of water turbine operation status monitoring and maintenance, in particular to a water turbine sound monitoring method based on transformer. Background Art

[0002] At present, when monitoring and maintaining the operating status of turbines, the operating data indicators of relevant equipment parts are usually analyzed regularly to judge the operating status of turbines, and the equipment parts are checked one by one. This traditional regular inspection method finds problems late and cannot find problems immediately, resulting in delayed maintenance and shortening the life of equipment. In addition, traditional visual or physical inspection methods are difficult to work under severe weather conditions. Summary of the invention

[0003] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a transformer-based turbine sound monitoring method, which can realize real-time tracking of equipment status through continuous sound monitoring, so as to discover problems earlier, carry out maintenance in advance and extend the life of the equipment.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: a turbine sound monitoring method based on transformer, which comprises the following steps:

[0005] Step 1: Data acquisition: Collect sound data from the historical monitoring system of the turbine and related equipment;

[0006] Step 2: Data processing: pre-process the collected sound data;

[0007] Step 3: Fourier transform processing: Use Fourier transform to convert the collected time series sound signal into a spectrogram; this process converts the sound data into a format that can be processed by the Transformer model, making it easier for the model to capture frequency-related features;

[0008] Step 4: Use the Audio Spectrogram Transformer model architecture for pre-training to allow the model to learn to recognize and parse different types of sound data and capture their characteristics; during the model training process, calculate the gradient of the loss function with respect to each weight, and then use these gradients to adjust the weights; at the same time, adjust the model hyperparameters to improve the model's prediction capabilities and realize sound fault monitoring of turbine operation.

[0009] Furthermore, in step 1, the sound data collected from the historical monitoring system of the turbine and related equipment includes the sound during normal operation and the abnormal sound during potential failure, and the collected sound data is combined with the relevant monitoring and performance data.

[0010] Furthermore, the process of combining the collected acoustic data with relevant monitoring and performance data includes: matching the acoustic data of a specific date and time with turbine efficiency data, water flow rate or equipment failure records at the same point in time.

[0011] Furthermore, in step 2, the process of preprocessing the collected sound data includes: first, cleaning to remove outliers to reduce errors; then, using a Gaussian filter to denoise the cleaned data to improve data quality; and finally, normalization processing to label the processed sound data with corresponding faults, abnormalities, and normal conditions.

[0012] Furthermore, the step 4 specifically includes the following steps:

[0013] Step 4.1: Build the Audio Spectrogram Transformer model.

[0014] Step 4.2: Use the sound data collected from the faulty equipment and the sound data of normal operation in the past few years to form a data set, and divide the processed data set into a training set, a validation set, and a test set according to a certain ratio. The training set and the validation set are used to train the parameters of the model, and the test set is used to test the final effect of the model recognition. At the same time, the model architecture and hyperparameters are continuously adjusted; after continuous training and optimization, the accurate recognition of the sound fault defects of the turbine is finally achieved, and the number of various defects is counted to provide data support for later maintenance and repair;

[0015] Step 4.3, set the sampling rate and sampling time of the input data. After each set of data is processed by Fourier transform, select the main characteristic energy, peak frequency and bandwidth, and set the time step. The four types of annotations include "normal operation", "efficiency reduction", "bearing wear" and "speed too high";

[0016] Step 4.4: The input data passes through the input layer, transformer architecture layer, and output layer of the Audio Spectrogram Transformer model in turn to obtain preliminary prediction results. This is the forward propagation process of model training. Then, the parameters in the Audio Spectrogram Transformer model are trained through the backpropagation process, and the hyperparameters are manually adjusted.

[0017] Step 4.5: During the back propagation process, it is necessary to compare the predicted results with the actual results, and then adjust the model weights and parameters in reverse. The loss function can achieve the comparison between the two, and cross entropy is selected as the loss function.

[0018] Step 4.6: Through continuous training and optimization, that is, repeating forward propagation and backward propagation, a complete turbine sound fault recognition model is finally obtained. Through this model, the sound of the turbine can be detected at any time, and its faults and problems can be discovered in time and alarms can be issued to assist operating personnel in management, while providing support and assistance for the intelligent management of the hydropower station.

[0019] Furthermore, the step 4.1 specifically includes the following steps:

[0020] The Audio Spectrogram Transformer model consists of an input layer, a transformer architecture layer, and an output layer. In the model, each spectrogram is divided into multiple small segments, and each segment is mapped into a vector of a fixed size. The self-attention mechanism in the Transformer layer calculates the dot product between these vectors to obtain attention scores, which are then normalized by the softmax function and finally used to weight the transformation vectors of the same input. This process is formalized as follows:

[0021]

[0022] Where Q, K, and V are query, key, and value matrices, respectively. k is the dimension of the key vector.

[0023] Furthermore, in step 4.5, cross entropy is selected as the loss function, and its formula is:

[0024]

[0025] Among them, L is the loss function, x i Represents the i-th element of the true result, y i Indicates the probability that the predicted result x belongs to the i-th category;

[0026] The result of the loss function is used to adjust the model weights and parameters through the optimizer. Adam is selected as the optimizer. The Adam optimizer is an adaptive learning rate optimization algorithm with the formula:

[0027] l=l0×γ [α / β]

[0028] l is the decayed learning rate, l0 is the initial learning rate, α is the current number of iterations, β is the decay step size, [] is rounded down, and γ is a positive number less than 1.

[0029] Beneficial effects of the present invention:

[0030] 1. The present invention uses the Transformer model to more accurately identify subtle sound changes; this model uses the self-attention mechanism to analyze audio data, and can distinguish which sound features are the most critical, so it performs well in detecting subtle changes; by collecting audio data of turbine-related equipment, monitoring the sound features of equipment such as turbines and generators, and identifying abnormal sounds that may indicate equipment failure; after being trained with a large amount of data, the Transformer model can automatically classify different types of damage and status; this method can automatically identify what type of problem it is, reducing the need to rely on manual judgment, improving processing speed and reducing errors; this model structure supports rapid processing of large amounts of data, is suitable for real-time monitoring of the operating status of each device, can not only identify problems immediately, but also predict possible problems in the future, and help technicians prepare in advance.

[0031] 2. The present invention is based on the application of Transformer sound recognition technology, and applies the Transformer architecture to the analysis and processing of sound data, especially converting audio signals into spectrograms and using them to train deep learning models, so as to achieve higher recognition accuracy and efficiency; the present invention uses a self-attention mechanism to dynamically weight key features in sound data, which can improve the monitoring and diagnosis capabilities of the status of turbines and other energy equipment; the present invention uses Fourier transform to convert time series sound signals into spectrograms. This preprocessing step is the key to using visual Transformers for audio analysis; through pre-training and fine-tuning technology, combined with large-scale audio data sets for model training, the model shows excellent performance in specific energy facility monitoring applications.

[0032] 3. The present invention collects sound data to identify early signs of failure of turbine-related equipment, such as gear wear or bearing problems, and can monitor the operating status of the turbine in real time and detect failures more accurately in a faster and more precise manner. Compared with conventional physical detection methods, sound monitoring can provide non-invasive and continuous fault monitoring, thereby warning of possible failures earlier and reducing maintenance costs and downtime.

[0033] 4. The present invention realizes real-time tracking of equipment status through continuous sound monitoring, which can detect problems earlier and carry out maintenance in advance, thus extending the life of the equipment, and the entire process can be carried out under adverse weather conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of a transformer-based turbine sound monitoring method;

[0035] Figure 2 It is a schematic diagram of the algorithm of the model of the present invention;

[0036] Figure 3 Schematic diagram of the Transformer architecture in the model of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0038] Example 1: Figures 1 to 3 As shown, a transformer-based turbine sound monitoring method comprises the following steps:

[0039] Step 1: First, collect a large amount of sound data from the historical monitoring system of the turbine and related equipment. This data includes but is not limited to the sound during normal operation and abnormal sound during potential failures.

[0040] Step 2: Preprocess the collected sound data. First, clean it to remove outliers and reduce errors. Then use a Gaussian filter to denoise the cleaned data to improve data quality. Finally, normalize it and label the processed sound data with relevant faults, abnormalities, and normal conditions.

[0041] Step 3: Use Fourier transform to convert the collected time series sound signal into a spectrogram (audiospectrum). This step converts the sound data into a format that can be processed by the Transformer model, making it easier for the model to capture frequency-related features.

[0042] Fourier transform represents any function as a weighted sum of a series of sine and cosine functions with different frequencies and phases. Through Fourier transform, we can decompose complex signals into simple frequency components and then analyze the characteristics of the signal. Its main formula is:

[0043]

[0044] Where f(ω) is the representation of the function f(t) in the frequency domain, ω is the angular frequency, and e -iωt is a complex exponential function. This formula represents the conversion of the time domain function f(t) to the frequency domain f(ω).

[0045] Step 4: Use the Audio Spectrogram Transformer model architecture for pre-training, so that the model can learn to recognize and parse different types of sound data and capture their features; during the model training process, the gradient of the loss function with respect to each weight is calculated, and then these gradients are used to adjust the weights; at the same time, the model hyperparameters are adjusted to improve the model's prediction ability and realize the sound fault monitoring of the turbine operation. The specific steps include the following:

[0046] Step 4.1: Build the AST sound fault recognition model, which mainly includes the input layer, transformer architecture layer and output layer. In the AST model, each spectrogram is divided into multiple small fragments (patches), and each fragment is mapped into a vector of a fixed size. The self-attention mechanism in the Transformer layer calculates the dot product between these vectors to obtain the attention score, and then these scores are normalized by the softmax function, and finally used to weight the transformation vectors of the same input. This process can be formalized as:

[0047]

[0048] Where Q, K, and V are query, key, and value matrices, respectively. k is the dimension of the key vector.

[0049] Step 4.2: Use the sound data collected from the faulty equipment and the sound data of normal operation in the past three years to form a data set, and divide the processed data set (the sound data after Fourier transform processing and its annotation) into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set and the validation set are used to train the parameters of the model, and the test set is used to test the final effect of the model recognition. At the same time, the model architecture and hyperparameters are continuously adjusted. The network structure of the turbine sound fault recognition model is as follows: Figure 2 As shown in the figure, after continuous training and optimization, we can finally achieve accurate identification of turbine sound fault defects, and count the number of various defects, providing data support and scientific guarantee for subsequent maintenance and repairs.

[0050] Step 4.3: Set the sampling rate of the input data to 48kHz, the sampling time to 10 seconds, and a total of 500 groups of data. After each group of data is processed by Fourier transform, the main characteristic energy, peak frequency and bandwidth are selected. At the same time, the time step is set to 100. The input data size is 15000×500000 (500 groups of data, each group of data size is 30×100), and the annotations include "normal operation", "reduced efficiency", "bearing wear" and "excessive speed". Therefore, the output size is 4×500.

[0051] Step 4.4: The input data passes through the input layer, transformer architecture layer, and output layer in the AST model in turn to obtain preliminary prediction results. This is the forward propagation process of model training. However, the prediction results obtained at this time are obviously impossible to achieve accurate recognition. Therefore, it is necessary to train the parameters in the AST model through the back-propagation process and manually adjust the hyperparameters.

[0052] Step 4.5: During the back propagation process, it is necessary to compare the predicted results with the actual results, and then adjust the model weights and parameters in reverse. The loss function can achieve the comparison between the two. The cross entropy is selected as the loss function, and its main formula is:

[0053]

[0054] Among them, L is the loss function, x i Represents the i-th element of the true result, y i It represents the probability that the predicted result x belongs to the i-th category.

[0055] The result of the loss function is used to adjust the model weights and parameters through the optimizer. The present invention selects Adam as the optimizer. The Adam optimizer is an adaptive learning rate optimization algorithm, and its main formula is:

[0056] l=l0×γ [α / β]

[0057] l is the decayed learning rate, l0 is the initial learning rate, α is the current number of iterations, β is the decay step size, [] is rounded down, and γ is usually a positive number less than 1.

[0058] Step 4.6: Through continuous training and optimization, that is, repeating forward propagation and backward propagation, a complete turbine sound fault recognition model is finally obtained. Through this model, the sound of the turbine can be detected at any time, and its faults and problems can be discovered in time and alarms can be issued to assist operating personnel in management, while providing support and assistance for the intelligent management of the hydropower station.

[0059] 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 transformer-based turbine sound monitoring method, characterized in that: It includes the following steps: Step 1: Data acquisition: Collect sound data from the historical monitoring system of the turbine and related equipment; Step 2: Data processing: pre-process the collected sound data; Step 3: Fourier transform processing: Use Fourier transform to convert the collected time series sound signal into a spectrogram; this process converts the sound data into a format that can be processed by the Transformer model, making it easier for the model to capture frequency-related features; Step 4: Use the Audio Spectrogram Transformer model architecture for pre-training to allow the model to learn to recognize and parse different types of sound data and capture their characteristics; during the model training process, calculate the gradient of the loss function with respect to each weight, and then use these gradients to adjust the weights; at the same time, adjust the model hyperparameters to improve the model's prediction capabilities and realize sound fault monitoring of turbine operation.

2. The transformer-based water turbine sound monitoring method according to claim 1, characterized in that: In step 1, the sound data collected from the historical monitoring system of the turbine and related equipment include the sound during normal operation and the abnormal sound during potential failure, and the collected sound data is combined with the related monitoring and performance data.

3. The transformer-based water turbine sound monitoring method according to claim 1, characterized in that: The process of combining the collected acoustic data with relevant monitoring and performance data involves matching acoustic data from a specific date and time with turbine efficiency data, water flow rate, or equipment failure records at the same point in time.

4. The transformer-based water turbine sound monitoring method according to claim 1, characterized in that: In step 2, the process of preprocessing the collected sound data includes: first, cleaning to remove outliers to reduce errors; then, using a Gaussian filter to denoise the cleaned data to improve data quality; and finally, normalization to label the processed sound data with corresponding faults, abnormalities, and normal conditions.

5. The transformer-based water turbine sound monitoring method according to claim 1, characterized in that: The step 4 specifically comprises the following steps: Step 4.1: Build the Audio Spectrogram Transformer model. Step 4.2: Use the sound data collected from the faulty equipment and the sound data of normal operation in the past few years to form a data set, and divide the processed data set into a training set, a validation set, and a test set according to a certain ratio. The training set and the validation set are used to train the parameters of the model, and the test set is used to test the final effect of the model recognition. At the same time, the model architecture and hyperparameters are continuously adjusted; after continuous training and optimization, the accurate recognition of the sound fault defects of the turbine is finally achieved, and the number of various defects is counted to provide data support for later maintenance and repair; Step 4.3, set the sampling rate and sampling time of the input data. After each set of data is processed by Fourier transform, select the main characteristic energy, peak frequency and bandwidth, and set the time step. There are four types of annotations, including "normal operation", "efficiency reduction", "bearing wear" and "speed too high"; Step 4.4: The input data passes through the input layer, transformer architecture layer, and output layer of the Audio Spectrogram Transformer model in turn to obtain preliminary prediction results. This is the forward propagation process of model training. Then, the parameters in the Audio Spectrogram Transformer model are trained through the backpropagation process, and the hyperparameters are manually adjusted. Step 4.5: During the back propagation process, it is necessary to compare the predicted results with the actual results, and then adjust the model weights and parameters in reverse. The loss function can achieve the comparison between the two, and cross entropy is selected as the loss function. Step 4.6: Through continuous training and optimization, that is, repeating forward propagation and backward propagation, a complete turbine sound fault recognition model is finally obtained. Through this model, the sound of the turbine can be detected at any time, and its faults and problems can be discovered in time and alarms can be issued to assist operating personnel in management, while providing support and assistance for the intelligent management of the hydropower station.

6. The transformer-based water turbine sound monitoring method according to claim 5, characterized in that: The step 4.1 specifically includes the following steps: The Audio Spectrogram Transformer model consists of an input layer, a transformer architecture layer, and an output layer. In the model, each spectrogram is divided into multiple small segments, and each segment is mapped into a vector of a fixed size. The self-attention mechanism in the Transformer layer calculates the dot product between these vectors to obtain attention scores, which are then normalized by the softmax function and finally used to weight the transformation vectors of the same input. This process is formalized as follows: Where Q, K, and V are query, key, and value matrices, respectively. k is the dimension of the key vector.

7. The transformer-based water turbine sound monitoring method according to claim 5, characterized in that: In step 4.5, cross entropy is selected as the loss function, and its formula is: Among them, L is the loss function, x i Represents the i-th element of the true result, y i Indicates the probability that the predicted result x belongs to the i-th category; The result of the loss function is used to adjust the model weights and parameters through the optimizer. Adam is selected as the optimizer. The Adam optimizer is an adaptive learning rate optimization algorithm with the formula: l=l0×γ [α / β] l is the decayed learning rate, l0 is the initial learning rate, α is the current number of iterations, β is the decay step size, [] is rounded down, and γ is a positive number less than 1.

Images