A method for de-noising audio frequency of power generation equipment in thermal power plants based on SGN

Abstract

The present invention relates to an audio denoising method for power generation equipment in a thermal power plant based on SGN, comprising the following steps: step S1: collecting audio signals during the operation of the transfer equipment in the power plant through a microphone array, recorded as sampling signals; step S2: performing Laplace transform on the sampling signals, and extracting signal features in the complex frequency domain; step S3: sending the complex frequency domain features to an improved SGN network for denoising, and the improved SGN network uses large-scale context information through a hierarchical cascade network structure to achieve more refined data denoising, remove the energy component of the noise, and restore the original audio signal. The method of the present invention can effectively denoise the sound signals collected by the power generation equipment in the thermal power plant, and the denoised audio signals can meet the general requirements of audio diagnosis, and the average signal-to-noise ratio obtained is better than other similar methods. At the same time, when processing the audio signals, the method has a stronger suppression effect on the noise, and can retain more audio detail information.

Description

Technical Field

[0001] The invention relates to an audio frequency denoising method for power generation equipment in a thermal power plant based on SGN, belonging to the technical field of fault diagnosis and intelligent detection. Background Art

[0002] The essence of sound is vibration. According to the Dirichlet condition, the sound heard by the human ear is actually formed by the superposition of sinusoidal vibrations at different frequencies. Usually, the time domain waveform of sound can only represent the change of amplitude over time. Therefore, the sound feature analysis must be continued in the frequency domain. By analyzing the spectrum diagram of the transfer equipment of the thermal power plant with different fault levels, it can be found that when the equipment is running smoothly, the audio signals are mostly stable signals in the low-frequency range, and as the fault occurs and expands, high-frequency non-stationary components gradually appear in the audio. By analyzing the high-frequency components, the current operating status of the equipment can be evaluated.

[0003] Using the audio signal of the transfer equipment during operation as the basis for fault diagnosis can capture the fault status at the earliest stage of the fault. However, the production area of ​​the thermal power plant is noisy, and it is difficult to collect audio data of specific equipment. It is easily disturbed by environmental noise, which has a great impact on fault diagnosis. Therefore, it is very important to denoise the collected original audio signal. By suppressing the noise signal and retaining the useful audio signal details, the fault warning of the transfer equipment can be realized. Summary of the invention

[0004] The technical problem to be solved by the present invention is that in the production area of ​​a thermal power plant, the noise signal is very complex, and the audio signal of the running of the transfer equipment will be drowned by the environmental noise, making it difficult to obtain useful audio information for equipment fault diagnosis.

[0005] The present invention provides a method for audio denoising of thermal power plant power generation equipment based on SGN to solve the above-mentioned problem. By studying and designing an audio signal denoising method, the sound signal collected by the thermal power plant power generation equipment is denoised. The denoised audio signal can meet the general requirements of audio diagnosis, and the average signal-to-noise ratio obtained is better than that of other similar methods. At the same time, when processing the audio signal, the method has a stronger suppression effect on noise and can retain more audio detail information.

[0006] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0007] A method for denoising the sound frequency of power generation equipment in a thermal power plant based on SGN, the method comprising the following steps:

[0008] Step S1: collecting audio signals during the operation of the transfer equipment in the power plant through a microphone array, and recording them as sampling signals;

[0009] Step S2: Perform Laplace transform on the sampled signal and extract signal features in the complex frequency domain;

[0010] Step S3: The complex frequency domain features are sent to the improved SGN network for denoising. The improved SGN network uses a hierarchical and cascaded network structure to use large-scale context information to achieve more refined data denoising, remove the energy component of the noise, and restore the original audio signal.

[0011] The working principle is: by analyzing the audio signal, usually the time domain waveform of the sound can only represent the change of amplitude over time, so the sound feature analysis must be continued in the frequency domain. In view of the noisy environment in the production area of ​​the thermal power plant, the collected audio signals have a lot of noise interference and cannot be directly used for fault diagnosis. The present invention proposes an audio denoising method for thermal power plant power generation equipment based on SGN. It first performs Laplace transform on the signal to extract the frequency domain features, and then sends it to the improved SGN network for denoising to remove the noise component. The improved SGN network is divided into a top-level subnetwork, a middle-level subnetwork, and a bottom-level subnetwork from top to bottom. A more sophisticated data denoising process is achieved by utilizing large-scale context information through a hierarchical and cascaded network structure. The method of the present invention has a stronger suppression effect on the noise in the audio signal, while retaining more audio detail information.

[0012] Furthermore, in step S3, sending the complex frequency domain features to the improved SGN network for denoising includes the following steps:

[0013] S31: performing a shuffle operation on the sampled signal to obtain a multi-resolution input audio signal;

[0014] S32: sending audio signals of different resolutions into convolutional layers of different sizes for processing;

[0015] S33: From top to bottom, through the SGN top-level sub-network, SGN middle-level sub-network, SGN bottom-level sub-network, and optimize the number of activations and pooling layers, the signal is denoised; Specifically:

[0016] S331: The steps of SGN top-level sub-network are as follows:

[0017] 1) Assume that the original size of the input data is M×N×C. After a shuffle / 2 operation, the data size is transformed into M / 2×N / 2×4C, that is, the resolution of the data is reduced and its input channels are increased;

[0018] 2) Assume that the number of input channels of the i-th layer is c i , the self-guided network is f i (·), then the number of feature matrices obtained at the Kth layer can be expressed as ck =2 k ×c 0

[0019] 3) After the top-level sub-network features are extracted, the data is shuffled × 2 to make it equal to the next level network to achieve the connection of feature data.

[0020] S332: The steps of the SGN middle-layer subnetwork are as follows:

[0021] 1) The SGN middle-layer sub-network adds a data connection layer based on the top-layer structure to concatenate the features extracted from the top layer with the features of the current layer. Assuming that the current layer is k+1 and the previous layer is k, then f k+1 (·) The obtained feature map is shuffled by 2 and combined with f k (·) features are concatenated. Through the connection operation, the formula for the transformation of the mapping quantity is: k →(c k +c k+1 / 4);

[0022] 2) Adding a pooling layer to reduce the feature map restores the feature map to c k .

[0023] S333: The steps of SGN underlying sub-network are as follows:

[0024] 1) The guidance method merged with the middle layer is adopted. Since there is a global residual between the input data and the final restored data, the SGN bottom layer sub-network uses m Conv+ReLu to process the feature map;

[0025] 2) Send the processed data into the convolutional layer to generate the estimation result.

[0026] S334: Optimize the activation function. Use the h-Swish function as the activation function in the convolution layer of the bottom sub-network of SGN, and the ReLu function is still used as the activation function in other layers. This can improve the network accuracy and further reduce the system overhead. The h-Swish activation function expression is

[0027] S335: Optimize the pooling layer. In order to ensure the recovery of signal detail information, the network deletes the pooling layer and replaces it with a convolutional layer with a step size of 2, thereby reducing the impact of feature loss on detail recovery.

[0028] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0029] 1. The present invention is a method for denoising an audio signal collected in a noisy environment, which can have a strong noise suppression effect;

[0030] 2. The present invention is a method for audio denoising of power generation equipment in a thermal power plant based on SGN. The method can retain more detailed information of the original audio signal while denoising the audio signal. The denoised audio signal can meet the general requirements of audio diagnosis, and the average signal-to-noise ratio after denoising is better than other similar methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flow chart of a method for audio denoising of power generation equipment in a thermal power plant based on SGN;

[0032] Figure 2 This is a diagram of measured data of audio frequency information of a transfer equipment obtained in a thermal power plant in an embodiment of the present invention;

[0033] Figure 3 This is the improved SGN network structure diagram;

[0034] Figure 4 This is a comparison diagram of the original waveform, the noisy waveform, and the denoised waveform when denoising the measured signal in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0036] Example

[0037] like Figures 1 to 4 As shown, the present invention is a method for denoising the sound frequency of power generation equipment in a thermal power plant based on SGN, comprising the following steps:

[0038] Step S1: collecting audio signals during the operation of the transfer equipment in the power plant through a microphone array, and recording them as sampling signals;

[0039] Step S2: Perform Laplace transform on the sampled signal and extract signal features in the complex frequency domain;

[0040] Step S3: The complex frequency domain features are sent to the improved SGN network for denoising. The improved SGN network uses a hierarchical and cascaded network structure to use large-scale context information to achieve more refined data denoising, remove the energy component of the noise, and restore the original audio signal.

[0041] The working principle of this embodiment is:

[0042] By analyzing the audio signal, it is found that the time domain waveform of the sound can only represent the change of amplitude over time under normal circumstances, so the sound feature analysis must be continued in the frequency domain. In view of the noisy environment in the production area of ​​the thermal power plant, the collected audio signals have a lot of noise interference and cannot be directly used for fault diagnosis. The present invention proposes an audio denoising method for thermal power plant power generation equipment based on SGN. It first performs Laplace transform on the signal to extract the frequency domain features, and then sends it to the improved SGN network for denoising to remove the noise component. The improved SGN network is divided into a top-level subnetwork, a middle-level subnetwork, and a bottom-level subnetwork from top to bottom. A more sophisticated data denoising process is achieved by utilizing large-scale context information through a hierarchical and cascaded network structure. The method of the present invention has a stronger suppression effect on the noise in the audio signal, while retaining more audio detail information.

[0043] like Figure 2 The figure shows the audio information of the transfer equipment obtained in a thermal power plant. It can be seen that the operating sound of the transfer equipment is drowned out by the environmental noise and it is difficult to obtain useful audio information.

[0044] Specifically, the improved SGN network denoising process includes the following steps:

[0045] S31: performing a shuffle operation on the sampled signal to obtain a multi-resolution input audio signal;

[0046] S32: sending audio signals of different resolutions into convolutional layers of different sizes for processing;

[0047] S33: From top to bottom, through the SGN top-level sub-network, SGN middle-level sub-network, SGN bottom-level sub-network, and optimize the activation function and pooling layer to reduce the noise of the signal;

[0048] The method of obtaining multi-resolution input through shuffle operation in this method is compared with downsampling. It can be seen that compared with downsampling, shuffle reduces the resolution while retaining the detail information. Each layer of SGN can adaptively extract features at different scales, thereby better completing the denoising task. Table 1 shows the comparison of denoising effects on the unified data set when using shuffle and downsampling:

[0049] Table 1 Comparison of denoising effects between shuffle and downsampling (σ=50)

[0050] Average signal-to-noise ratio Downsampling 18.7 Shuffle 19.1

[0051] like Figure 3The figure shows the basic structure of the improved SGN network, where the input layer is used to input data and shuffled data; the output layer is used to output processed data; the data connection layer is used to concatenate the data of this layer with the data output of the previous level; and the convolution layer is used for feature extraction.

[0052] S331: The steps of SGN top-level sub-network are as follows:

[0053] 1) Assume that the original size of the input data is M×N×C. After a shuffle / 2 operation, the data size is transformed into M / 2×n / 2×4C, that is, the resolution of the data is reduced and its input channels are increased;

[0054] 2) Assume that the number of input channels of the i-th layer is c i , the self-guided network is f i (·), then the number of feature matrices obtained at the Kth layer can be expressed as c k =2 k ×c 0

[0055] 3) After the top-level sub-network features are extracted, the data is shuffled × 2 to make it equal to the next level network to achieve the connection of feature data.

[0056] S332: The steps of the SGN middle-layer subnetwork are as follows:

[0057] 1) The SGN middle-layer sub-network adds a data connection layer based on the top-layer structure to concatenate the features extracted from the top layer with the features of the current layer. Assuming that the current layer is k+1 and the previous layer is k, then f k+1 (·) The obtained feature map is shuffled by 2 and combined with f k (·) features are concatenated. Through the connection operation, the formula for the transformation of the mapping quantity is: k →(c k +c k+1 / 4);

[0058] 2) Adding a pooling layer to reduce the feature map restores the feature map to c k .

[0059] S333: The steps of SGN underlying sub-network are as follows:

[0060] 1) The guidance method merged with the middle layer is adopted. Since there is a global residual between the input data and the final restored data, the SGN bottom layer sub-network uses m Conv+ReLu to process the feature map;

[0061] 2) Send the processed data into the convolutional layer to generate the estimation result.

[0062] S334: Optimize the activation function. Use the h-Swish function as the activation function in the convolution layer of the bottom sub-network of SGN, and the ReLu function is still used as the activation function in other layers. This can improve the network accuracy and further reduce the system overhead. The h-Swish activation function expression is

[0063] S335: Optimize the pooling layer. In order to ensure the recovery of signal detail information, the network deletes the pooling layer and replaces it with a convolutional layer with a step size of 2, thereby reducing the impact of feature loss on detail recovery.

[0064] Figure 4 The figure is a comparison diagram of the original waveform, the noisy waveform, and the denoised waveform when the measured signal is denoised in an embodiment of the present invention. It can be seen from the figure that the SGN algorithm can restore the signal submerged by noise and retain the detail information almost without distortion.

[0065] Through the test of 100 sets of data, the average signal-to-noise ratio of the audio information after denoising reaches 19.7, which meets the basic requirements of audio analysis. Table 2 shows the comparison of the signal-to-noise ratio of the restored signal between the SGN algorithm and other denoising algorithms. It can be seen from the table that the proposed algorithm is significantly better than other comparison algorithms in terms of restoring the signal-to-noise ratio, and is about 6% higher than the optimal MemNet algorithm.

[0066] Table 2 Signal-to-noise ratio comparison (dB)

[0067] This algorithm Unet DnCNN MemNet Actual Dataset 19.7 18.29 18.43 18.46

[0068] It can be seen that the denoising effect of the method for audio denoising of thermal power plant power generation equipment based on SGN in the present invention is better than other similar technologies and has achieved a better denoising effect.

[0069] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific implementation method of the present invention and does not limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for denoising the sound frequency of power generation equipment in a thermal power plant based on SGN, characterized in that: The method comprises the following steps: Step S1: collecting audio signals during the operation of the transfer equipment in the power plant through a microphone array, and recording them as sampling signals; Step S2: Perform Laplace transform on the sampled signal and extract signal features in the complex frequency domain; Step S3: sending the complex frequency domain features to the improved SGN network for denoising. The improved SGN network uses a hierarchical and cascaded network structure to achieve more refined data denoising, remove the energy component of the noise, and restore the original audio signal. The step S3 involved in sending the complex frequency domain features into the improved SGN network for denoising includes the following steps: S31: performing a shuffle operation on the sampled signal to obtain a multi-resolution input audio signal; S32: sending audio signals of different resolutions into convolutional layers of different sizes for processing; S33: From top to bottom, through the SGN top-level sub-network, SGN middle-level sub-network, SGN bottom-level sub-network, and optimize the activation function and pooling layer to reduce the noise of the signal; The steps of step S33 specifically include: S331: The steps of SGN top-level sub-network are as follows: 1) Assume that the original size of the input data is M×N×C. After a shuffle / 2 operation, the data size is transformed into M / 2×N / 2×4C, that is, the resolution of the data is reduced and its input channels are increased; 2) Assume that the number of input channels of the i-th layer is c i , the self-guided network is f i (·), then the number of feature matrices obtained at the Kth layer is expressed as c k =2 k ×c 0 ; 3) After the top-level sub-network features are extracted, the data is shuffled × 2 to make it equal to the next level network to achieve the connection of feature data; S332: The steps of the SGN middle-layer subnetwork are as follows: 1) The SGN middle-layer sub-network adds a data connection layer based on the top-level structure to concatenate the features extracted from the top layer with the features of the current layer; assuming that the current layer is the k+1 layer and the previous layer is the k layer, then f k+1 (·) The obtained feature map is shuffled by 2 and combined with f k (·) features are concatenated; through the connection operation, the formula for the transformation of the mapping quantity obtained is c k →(c k +c k+1 / 4); 2) Adding a pooling layer to reduce the feature map restores the feature map to c k ; S333: The steps of SGN underlying sub-network are as follows: 1) The guidance method merged with the middle layer is adopted. Since there is a global residual between the input data and the final restored data, the SGN bottom layer sub-network uses m Conv+ReLu to process the feature map; 2) Send the processed data to the convolutional layer to generate the estimated result; S334: Optimize the activation function. Use the h-Swish function as the activation function in the convolutional layer of the bottom sub-network of SGN, and the ReLu function is still used as the activation function in other layers. This improves the network accuracy and further reduces the system overhead. The h-Swish activation function expression is: S335: Optimize the pooling layer. In order to ensure the recovery of signal detail information, the network deletes the pooling layer and replaces it with a convolutional layer with a step size of 2, so as to reduce the impact of feature loss on detail recovery.

Images