Diagnostic method for fault sound signal of rotating machine

Abstract

The invention discloses a method for diagnosing a fault sound signal of a rotating machine, and the method comprises the following specific steps: collecting a sound signal of a waterwheel room, and sequentially carrying out the framing and windowing, and FFT processing, thereby obtaining the amplitude and phase of the signal; the method comprises the following steps: calculating a noise power spectrum estimated value, a signal-to-noise ratio, an over-reduction factor and a denoised signal power spectrum of a waterwheel room sound signal after FFT of a current frame, recovering the signal through IFFT, carrying out pre-emphasis processing on the recovered signal to obtain a denoised clean signal, decomposing the denoised clean signal into IMF components through a CEEMDAN method, calculating a residual component, making a data set, and obtaining the waterwheel room sound signal. And inputting the data set into the trained CNN-TCN neural network model for fault identification. The method is high in fault diagnosis accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery state fault diagnosis methods, and in particular relates to a method for diagnosing rotating machinery fault sound signals. Background Art

[0002] With the rapid development of modern acoustic and artificial intelligence technologies, the importance and practicality of sound signals in the condition monitoring and fault diagnosis of mechanical equipment have become increasingly prominent. As an important information carrier reflecting the operating status of equipment, acoustic characteristics have shown broad application prospects in the operation and maintenance of important equipment such as energy and military due to their advantages such as non-contact measurement, strong real-time performance, and rich information. Among common mechanical equipment, rotating machinery is the most important type of mechanical equipment. It is the core mechanism in the fields of electric power, aerospace, ships and vehicles, such as large generator sets and main shafts of vehicles and ships. Since rotating machinery often has the characteristics of high integration, large loads, high-speed operation, and long-term operation, once a failure occurs, it is often accompanied by huge economic losses or threats to personal safety. Therefore, rotating machinery has extremely high requirements for reliability and safety.

[0003] Rotating machinery fault diagnosis methods based on acoustic signals offer the following key advantages. First, because rotating machinery operates continuously, it is difficult to directly measure its vibration or temperature signals using conventional contact sensors. In practice, sensors are typically deployed on fixed components within the rotating machinery system, effectively constituting an indirect measurement method. While non-contact vibration sensors, such as laser, capacitive, and eddy-current sensors, have made some progress, they still suffer from limitations such as slow response speed, high cost, and complex commissioning and installation. In rotating machinery fault diagnosis applications, vibration signals collected by accelerometers remain the primary information source. One of the greatest advantages of acoustic signals is their non-contact measurement characteristics. They are extremely easy to install and deploy, highly adaptable to various scenarios, and can directly capture fault information from key areas of rotating machinery. Second, rotating machinery is typically a shaft system integrating multiple components, resulting in complex operating conditions and failure modes. Acoustic sensors can typically capture the entire acoustic spectrum of a given area, unconstrained by complex mechanical structures. This allows for large-scale monitoring, making them suitable for monitoring large, integrated electromechanical equipment such as thermal and hydropower units. For a specific mechanical rotor system, only a few acoustic sensors are needed to collect all the information. These sensors are typically only used extensively in scenarios requiring acoustic positioning. Furthermore, compared to vibration signals, sound offers a certain level of early warning capability. Many devices emit abnormal sounds before experiencing noticeable vibration anomalies. Acoustic signals can capture these early abnormal acoustic signals, enabling earlier fault warnings. For example, when cracks, leakage, or overloads occur, mechanical equipment often emits various unusual and high-frequency noises before structural damage occurs. Therefore, acoustic signals have broad application prospects in the condition monitoring and fault diagnosis of rotating machinery.

[0004] Since rotating machinery is often accompanied by strong background noise interference during operation, the acquisition and analysis of acoustic signals face great challenges. Noise interference can cause signal distortion, thus affecting the accuracy of fault diagnosis. Summary of the Invention The purpose of the present invention is to provide a method for diagnosing sound signals of rotating machinery faults, so as to solve the problem that the existing methods have low accuracy in fault diagnosis.

[0005] The technical solution adopted by the present invention is a method for diagnosing sound signals of rotating machinery faults, and the specific steps are as follows: collecting waterwheel chamber sound signals, performing frame segmentation and windowing, and FFT processing in sequence to obtain the amplitude and phase of the signal; calculating the noise power spectrum estimation value, signal-to-noise ratio, over-subtraction factor, and denoised signal power spectrum of the waterwheel chamber sound signal after FFT of the current frame, restoring the signal through IFFT, pre-emphasis processing on the restored signal to obtain a denoised clean signal, decomposing the denoised clean signal into IMF components through the CEEMDAN method, and calculating the residual components to create a data set, and inputting the data set into a trained CNN-TCN neural network model for fault identification.

[0006] The present invention is also characterized in that: Please follow the steps below to implement it: Step 1: An acoustic sensor is set in the waterwheel chamber, and a sound signal of the waterwheel chamber is collected by the acoustic sensor. The collected sound signal of the waterwheel chamber is subjected to frame and window processing, and the processed sound signal is subjected to FFT, and the amplitude and phase of the sound signal of the waterwheel chamber after FFT are calculated; Step 2, calculating the noise power spectrum estimation value of the waterwheel chamber sound signal after FFT of the current frame; Step 3, calculating the signal-to-noise ratio based on the noise power spectrum estimate; Step 4, calculating the over-subtraction factor by the signal-to-noise ratio; Step 5, calculating the power spectrum of the denoised signal according to the over-subtraction factor; Step 6: Perform IFFT on the power spectrum of the final denoised signal obtained in step 5 and the phase to recover the signal, and perform pre-emphasis processing on the recovered signal to obtain a clean denoised signal; Step 7: Use the CEEMDAN method to remove the noise from the clean signal Decompose into IMF components and calculate the residual components; Step 8: Divide the IMF components obtained in step 7 into a training set, a test set, and a validation set; Step 9: Input the training set into the CNN-TCN neural network model for training to obtain the trained CNN-TCN neural network model; input the validation set into the trained CNN-TCN neural network model for verification and optimization to obtain the trained CNN-TCN neural network model; input the test set into the trained CNN-TCN neural network model for fault identification.

[0007] In step 1, the expression for frame splitting and windowing is: (1) In formula (1), Indicates the processed i Frame waterwheel room sound signal; Indicates the collected sound signal of the waterwheel room; N Indicates the frame length; R represents frame shift, R=N / 2; Indicates that the time index n Relative to the i Starting point of the sound signal in the frame waterwheel room iR Offset to get the window function w The relative index of , where the window function w Use Hamming window; The expression for fast Fourier transform of the processed sound signal is: (2) In formula (2), represents the frequency index; Indicates frequency k The first i Frame waterwheel room sound signal; The amplitude is , the expression of phase is: (3) In formula (3), express The imaginary part of express The real part of .

[0008] In step 2, the expression for calculating the noise power spectrum estimation value of the waterwheel room sound signal after FFT of the current frame is: (4) In formula (4), Indicates the frequency of the waterwheel room sound signal in the i-th frame k The estimated value of the noise power spectrum at ; express i -1 frame waterwheel room sound signal at frequency k The estimated value of the noise power spectrum at ; Indicates the i Frame waterwheel room sound signal at frequency k The signal power spectrum at ; is the recursive smoothing factor, which is a value between 0 and 1 that controls i -1 frame (previous frame) waterwheel room sound signal at frequency k The noise power spectrum estimate at i Frame waterwheel room sound signal at frequency k The weighted ratio of the power spectrum at , The probability of the target sound existing Dynamic adjustment; Among them, the recursive smoothing factor It is obtained through two iterations. The specific process is: First iteration: Step S1: Sound signal of waterwheel room Perform the average processing between frequency points to obtain the i Frame waterwheel room sound signal at frequency k The average processed power spectrum estimate at ; The expression is: (5) In formula (5), Represents the weighted window function, using the Hamming window; M represents the half-width of the window function; 2M+1 represents the window length; Indicates the i Frame waterwheel room sound signal at frequency k + m The signal power spectrum at ; Step S2, through the first-order recursion Perform smoothing to obtain i Frame waterwheel room sound signal at frequency k The smoothed estimate of the signal power spectrum at ; The expression is: (6) In formula (6), is the spectral smoothing factor; Step S3, update i Frame waterwheel sound signal at frequency k The minimum smoothed estimate of ; The expression is: (7) In formula (7), L Indicates the search window length, which is equal to the window length of the frame windowing process; S i-j ( k ) indicates the ij A smoothed estimate of the frame, j =0,1,…, L ; Step S4, by Establish rules to obtain sound presence judgment results ; The expression is: (8) In formula (8), Indicates the frequency kThe target sound may be contained in the Indicates the frequency k There is only noise; Indicates the i Frame waterwheel room sound signal at frequency k The decision threshold of the smoothed estimate of the signal power spectrum at ; Indicates the Frame waterwheel room sound signal at frequency The decision threshold of the signal power spectrum at ; Second iteration: Step S5, the step S4 is judged to be The waterwheel room sound signal is smoothed for the second time; The expression is: (9) In formula (9), Indicates the i Frame waterwheel room sound signal at frequency k The second smoothed estimate of the signal power spectrum at ; Step S6, update the secondary i The minimum smoothed estimate of the frame waterwheel sound signal at frequency k ; The expression is: (10) In formula (10), L’ Indicates the length of the second search window, which is equal to the length of the first search window; S i-f ( k ) indicates the if A smoothed estimate of the frame, f =0,1,…, L’ ; Step S7: Calculate the probability of the target sound existing ; The expression is: (11) Step S8, by Calculate the recursive smoothing factor ; The expression is: (12) In formula (12), represents the minimum recursive smoothing factor; By Substitute into formula (4) to achieve dynamic update .

[0009] In step 3, the expression for calculating the signal-to-noise ratio is: (13) In formula (13), Indicates the i Frame rate k The signal-to-noise ratio of the sound signal in the waterwheel room.

[0010] In step 4, the expression for calculating the over-reduction factor is: (14) In formula (14), represents the initial oversubtraction factor; s Indicates the slope.

[0011] In step 5, the expression for calculating the power spectrum of the denoised signal is: (15) In formula (15), represents the gain compensation factor; In order to get a smaller ,use Replace the formula (15) , get the final denoised signal power spectrum; The expression is: (16).

[0012] In step 6, the pre-emphasis processing uses a first-order high-pass filter to perform a high-pass filtering operation on the restored signal, and weights the difference between each restored signal and the previous restored signal.

[0013] The specific process of step 7 is: For the clean signal after denoising Add different Gaussian white noise ,repeat I After reconstructing, we get a noisy signal ; The expression is: (17); In formula (17), Represents different weighting factors of Gaussian white noise; For each Perform EMD decomposition to obtain IMF components , retain the first component , calculate all first components The average value is used as the clean signal after denoising The final first IMF component IMF1; The expression is: (18) Subtract IMF1 from the denoised clean signal to obtain the residual component ; The expression is: (19) Will As a new signal, iterate; Similarly, we get the final p IMF components p , when the p The residual component obtained by decomposition is monotonic, stop computing.

[0014] The beneficial effects of the present invention are: (1) The present invention provides a method for diagnosing rotating machinery fault sound signals, which uses an improved minimum controlled recursive averaging algorithm (IMCRA) to estimate the noise power spectrum. The recursive estimation method of the IMCRA algorithm can more accurately estimate the noise power spectrum, effectively improve the accuracy of noise estimation, and avoid over- or under-estimation of noise. (2) The diagnostic method for rotating machinery fault sound signals of the present invention uses adaptive spectral subtraction to denoise the signal. By dynamically adjusting the over-subtraction factor, the noise removal effect is made more accurate, effectively improving the quality of the voice signal while ensuring the fidelity of the signal. (3) The diagnostic method for rotating machinery fault sound signals of the present invention retains the maximum value of the noise by calculating the maximum noise residual during the spectrum subtraction process, thereby reducing the noise residual as much as possible during the spectrum subtraction process, improving the clarity of the sound signal, and avoiding the pseudo-noise that occurs in traditional spectrum subtraction; (4) The diagnostic method for rotating machinery fault sound signals of the present invention uses the CEEMDAN algorithm to deeply analyze the sound signals and extract their multi-scale time-frequency features. Through multi-level deconstruction, the CEEMDAN algorithm not only improves the readability and feature recognition capabilities of the data, but also provides a solid foundation for constructing a feature dataset. (5) The diagnostic method of rotating machinery fault sound signals of the present invention, combined with the spatial feature extraction of convolutional neural networks and the temporal modeling of temporal convolutional networks, can accurately identify faults of the processed signals, thereby improving the accuracy and robustness of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1Schematic diagram of the prediction of the CNN-TCN neural network model in the method for diagnosing sound signals of rotating machinery faults of the present invention; Figure 2 It is a curve diagram of the change of the over-subtraction factor under different signal-to-noise ratios in the method of the present invention. DETAILED DESCRIPTION

[0016] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Example 1 The present invention provides a method for diagnosing a sound signal of a rotating machinery fault, and specifically comprises the following steps: collecting a sound signal from a waterwheel chamber, sequentially performing frame segmentation and windowing, and FFT processing to obtain the amplitude and phase of the signal; calculating a noise power spectrum estimate, a signal-to-noise ratio, an over-subtraction factor, and a power spectrum of a denoised signal of the sound signal from the waterwheel chamber after FFT of the current frame, restoring the signal through IFFT, performing pre-emphasis processing on the restored signal to obtain a denoised clean signal, decomposing the denoised clean signal into IMF components through a CEEMDAN method, calculating residual components, creating a data set, and inputting the data set into a trained CNN-TCN neural network model for fault identification.

[0018] Example 2 The method for diagnosing a fault sound signal of a rotating machinery according to the present invention is specifically implemented according to the following steps: Step 1: An acoustic sensor is set in the waterwheel chamber, and a sound signal of the waterwheel chamber is collected by the acoustic sensor. The collected sound signal of the waterwheel chamber is subjected to frame and window processing, and the processed sound signal is subjected to FFT, and the amplitude and phase of the sound signal of the waterwheel chamber after FFT are calculated; Step 2, calculating the noise power spectrum estimation value of the waterwheel chamber sound signal after FFT of the current frame; Step 3, calculating the signal-to-noise ratio based on the noise power spectrum estimate; Step 4, calculating the over-subtraction factor by the signal-to-noise ratio; Step 5, calculating the power spectrum of the denoised signal according to the over-subtraction factor; Step 6: Perform IFFT on the power spectrum of the final denoised signal obtained in step 5 and the phase to recover the signal, and perform pre-emphasis processing on the recovered signal to obtain a clean denoised signal; Among them, the pre-emphasis processing uses a first-order high-pass filter to perform high-pass filtering on the restored signal. The commonly used factors are , that is, weighting the difference between each restored signal and the previous one, so that the high-frequency portion of the signal occupies a more prominent position in the spectrum. This step is particularly important for speech signals, because high-frequency components often contain more detailed information, which can effectively improve the accuracy of subsequent feature extraction and fault diagnosis. Through this pre-emphasis processing, the time-frequency characteristics of the signal can be enhanced, providing clearer and richer information for subsequent signal decomposition and feature extraction; in, It represents the value of the signal autocorrelation function when the lag is 0, which represents the energy of the signal itself; It represents the value of the signal autocorrelation function when the lag is 1, which indicates the correlation between adjacent sampling points of the signal; Represents the first-order autocorrelation coefficient of the signal; represents unit delay; The purpose of pre-emphasis is to enhance the high-frequency component information of the signal and prevent the loss of high-frequency information during signal transmission. Step 7: Use the CEEMDAN method to remove the noise from the clean signal Decompose into IMF components and calculate the residual components; CEEMDAN decomposition not only improves data readability and feature recognition capabilities, but also provides a good data foundation for subsequent model performance; Step 8: Divide the IMF component IMFp obtained in step 7 into a training set, a test set, and a validation set in a ratio of 7:2:1; Step 9, such as Figure 1 As shown, the training set is input into the CNN-TCN neural network model for training to obtain the trained CNN-TCN neural network model, the validation set is input into the trained CNN-TCN neural network model for verification and optimization to obtain the trained CNN-TCN neural network model, and the test set is input into the trained CNN-TCN neural network model for fault identification.

[0019] Example 3 Based on Example 2, in step 1, the expression for the frame windowing process is: (1) In formula (1), Indicates the processed i Frame waterwheel room sound signal; Indicates the collected sound signal of the waterwheel room; N Indicates the frame length; R represents frame shift, R=N / 2; Indicates that the time index n Relative to the iStarting point of the sound signal in the frame waterwheel room iR Offset to get the window function w Relative index of Among them, the window function w Use Hamming window; The window length is 10ms~40ms, the overlap is 50%, and the frame length is equal to the window length; The purpose of windowing is to reduce the adverse effects of truncation effect and suppress the subjective auditory discontinuity of the frame edge; since the Hamming window has a better effect on the main lobe width, minimum stopband attenuation, and sidelobe peak, The expression for fast Fourier transform of the processed sound signal is: (2) In formula (2), represents the frequency index; Indicates frequency k The first i Frame waterwheel room sound signal; The amplitude is , the expression of phase is: (3) In formula (3), express The imaginary part of express The real part of .

[0020] Example 4 Based on Example 3, the specific process of step 2 is as follows: The expression for the noise power spectrum estimation value of the waterwheel room sound signal after the current frame FFT is calculated using the improved minimum controlled recursive averaging algorithm (IMCRA) is: (4) In formula (4), Indicates the Frame waterwheel room sound signal at frequency The estimated value of the noise power spectrum at ; express i -1 frame (previous frame) waterwheel room sound signal at frequency The estimated value of the noise power spectrum at ; Indicates the i Frame waterwheel room sound signal at frequency The signal power spectrum at ; is the recursive smoothing factor, which is a value between 0 and 1 that controls i -1 frame (previous frame) waterwheel room sound signal at frequency The noise power spectrum estimate at Frame waterwheel room sound signal at frequency The weighted ratio of the power spectrum at , The probability of the target sound existing Dynamic adjustment; Among them, the recursive smoothing factor It is obtained through two iterations. The specific process is: First iteration: Step S1: Sound signal of waterwheel room Perform the average processing between frequency points to obtain the i Frame waterwheel room sound signal at frequency k The average processed power spectrum estimate at ; The expression is: (5) In formula (5), Represents the weighted window function, using the Hamming window; M represents the half-width of the window function; 2M+1 represents the window length; Indicates the i Frame waterwheel room sound signal at frequency k + m The signal power spectrum at ; Step S2, through the first-order recursion Perform smoothing to obtain i Frame waterwheel room sound signal at frequency k The smoothed estimate of the signal power spectrum at ; The expression is: (6) In formula (6), is the spectral smoothing factor; Step S3, update i Frame waterwheel sound signal at frequency k The minimum smoothed estimate of ; The expression is: (7) In formula (7), L Indicates the search window length, which is equal to the window length of the frame windowing process; S i-j ( k ) indicates the ij A smoothed estimate of the frame, j =0,1,…, L ; Step S4, by Establish rules to obtain sound presence judgment results ; The expression is: (8) In formula (8), Indicates the frequency k The target sound may be contained in the Indicates the frequency k There is only noise; Indicates the i Frame waterwheel room sound signal at frequency k The decision threshold of the smoothed estimate of the signal power spectrum at ; Indicates the Frame waterwheel room sound signal at frequency The decision threshold of the signal power spectrum at ; Second iteration: Step S5, the step S4 is judged to be The waterwheel room sound signal is smoothed for the second time; The expression is: (9) In formula (9), Indicates the i Frame waterwheel room sound signal at frequency k The second smoothed estimate of the signal power spectrum at ; Step S6: Update the minimum smoothed estimated value of the waterwheel sound signal at frequency k for the second i-th frame ; The expression is: (10) In formula (10), L’ Indicates the length of the second search window, which is equal to the length of the first search window; S i-f ( k ) indicates the if A smoothed estimate of the frame, f =0,1,…, L’ ; Step S7: Calculate the probability of the target sound existing ; The expression is: (11) Step S8, by Calculate the recursive smoothing factor ; The expression is: (12) In formula (12), represents the minimum recursive smoothing factor; By Substitute into formula (4) to achieve dynamic update .

[0021] Example 5 Based on Example 4, in step 3, the expression for calculating the signal-to-noise ratio is: (13) In formula (13), Indicates the i Frame rate k The signal-to-noise ratio of the sound signal in the waterwheel room.

[0022] The signal-to-noise ratio (SNR) measures the ratio of useful components to noise in a signal. The SNR is calculated based on the power spectrum of each frequency point and can accurately reflect the signal quality of each frequency point. The signal-to-noise ratio (SNR) value determines whether the current frame falls within a low-noise or high-noise region. A low SNR indicates strong noise components, requiring stronger noise suppression. A high SNR indicates a clear signal, requiring a lower noise reduction level. Dynamic SNR calculation allows for more precise control of noise reduction in complex environments.

[0023] Example 6 Based on Example 5, in step 4, the expression for calculating the over-reduction factor is: (14) In formula (14), represents the initial oversubtraction factor, which is usually an empirically set constant; s Indicates the slope. The larger s is, the easier it is for the signal to be distorted. The optimal value of s is 3 / 20. like Figure 2 As shown, With signal-to-noise ratio Change, signal-to-noise ratio When low, Larger signal-to-noise ratio High, Smaller.

[0024] Example 7 Based on Example 6, in step 5, the over-subtraction factor is substituted into the improved spectrum subtraction method to calculate the power spectrum of the denoised signal. , then the expression is: (15) In formula (15), represents the gain compensation factor; In order to get a smaller ,use Replace the formula (15) , get the final denoised signal power spectrum; The expression is: (16).

[0025] Example 8 Based on Example 7, the specific process of step 7 is as follows: For the clean signal after denoising Add different Gaussian white noise ,repeat I After reconstructing, we get a noisy signal ; The expression is: (17); In formula (17), Represents different weighting factors of Gaussian white noise; For each Perform EMD decomposition to obtain IMF components , retain the first component , calculate all first components The average value is used as the clean signal after denoising The final first IMF component IMF1; The expression is: (18) Subtract IMF1 from the denoised clean signal to obtain the residual component ; The expression is: (19) Will As a new signal, iterate; Similarly, we get the final p IMF components p , when the p The residual component obtained by decomposition is monotonic, stop calculating; The final expression of the denoised clean signal is as follows: (20).

[0026] Example 9 Based on Example 8, the CNN-TCN neural network model consists of a convolutional neural network (CNN) and a temporal convolutional network (TCN). The CNN (convolutional neural network) extracts multi-scale features through convolutional layers and pooling layers, capturing information from different frequency bands. The TCN (temporal convolutional network) processes the long-term dependencies of time series signals through temporal convolutional layers, effectively enhancing the learning of temporal features. The convolutional neural network (CNN) consists of the first convolution module, the second convolution module, and the third convolution module; The first convolutional module consists of a convolutional layer (Conv1d), a ReLU activation layer, and a maximum pooling layer (MaxPool1d). The convolutional layer (Conv1d) takes an input shape of [32, 7, 1024] (batch size 32, input channels 7, sequence length 1024) and outputs a shape of [32, 32, 1024], with 32 output channels and 704 parameters. The ReLU activation layer performs nonlinear activation on the convolutional output, maintaining the shape of [32, 32, 1024]. The maximum pooling layer (MaxPool1d) reduces the sequence length from 1024 to 512, with an output shape of [32, 32, 512]. The second convolutional module consists of a convolutional layer (Conv1d), a ReLU activation layer, and a maximum pooling layer (MaxPool1d); the convolutional layer (Conv1d): the input shape is [32, 32, 512], the output shape is [32, 64, 512], contains 64 output channels, and has 6208 parameters; the ReLU activation layer: maintains the shape [32, 64, 512]; the maximum pooling layer (MaxPool1d): reduces the sequence length from 512 to 256, and the output shape is [32, 64, 256]; The third convolutional module consists of a convolutional layer (Conv1d), a ReLU activation layer, and a maximum pooling layer (MaxPool1d); the convolutional layer (Conv1d): the input shape is [32, 64, 256], the output shape is [32, 128, 256], contains 128 output channels, and has 24704 parameters; the ReLU activation layer: maintains the shape [32, 128, 256]; the maximum pooling layer (MaxPool1d): reduces the sequence length from 256 to 128, and the output shape is [32, 128, 128]; The temporal convolutional network (TCN) consists of the first TCN module (TemporalBlock), the second TCN module (TemporalBlock), the adaptive average pooling layer (AdaptiveAvgPool1d), and the fully connected layer (Linear); Among them, the first TCN module (TemporalBlock): the input shape is [32, 128, 128], the output shape is [32, 128, 128], including convolution operations (98816 parameters) and ReLU activation to handle long-term dependencies in time series; The second TCN module (TemporalBlock): has an input shape of [32, 128, 128] and an output shape of [32, 64, 128], including a convolution operation (37120 parameters) and an additional Conv1d layer (from 128 channels to 64 channels, 8256 parameters), as well as ReLU activation; AdaptiveAvgPool1d: The input shape is [32, 64, 128] and the output shape is [32, 64, 1], compressing the sequence length to 1; Fully connected layer (Linear): The input shape is [32, 64] (after flattening), the output shape is [32, 5], mapped to 5 categories, and the number of parameters is 325.

Claims

1. A method for diagnosing sound signals of rotating machinery faults, characterized in that: The specific steps are as follows: collect the sound signal of the waterwheel room, perform frame division and windowing, and FFT processing in sequence to obtain the amplitude and phase of the signal; calculate the noise power spectrum estimation value, signal-to-noise ratio, over-subtraction factor, and denoised signal power spectrum of the waterwheel room sound signal after FFT of the current frame, restore the signal through IFFT, and pre-emphasize the restored signal to obtain a denoised clean signal; decompose the denoised clean signal into IMF components through the CEEMDAN method, and calculate the residual component to create a data set, which is input into the trained CNN-TCN neural network model for fault identification.

2. The method for diagnosing a fault sound signal of a rotating machine according to claim 1, characterized in that: Please follow the steps below to implement it: Step 1: An acoustic sensor is set in the waterwheel chamber, and a sound signal of the waterwheel chamber is collected by the acoustic sensor. The collected sound signal of the waterwheel chamber is subjected to frame and window processing, and the processed sound signal is subjected to FFT, and the amplitude and phase of the sound signal of the waterwheel chamber after FFT are calculated; Step 2, calculating the noise power spectrum estimation value of the waterwheel chamber sound signal after FFT of the current frame; Step 3, calculating the signal-to-noise ratio based on the noise power spectrum estimate; Step 4, calculating the over-subtraction factor by the signal-to-noise ratio; Step 5, calculating the power spectrum of the denoised signal according to the over-subtraction factor; Step 6: Perform IFFT on the power spectrum of the final denoised signal obtained in step 5 and the phase to recover the signal, and perform pre-emphasis processing on the recovered signal to obtain a clean denoised signal; Step 7: Use the CEEMDAN method to remove the noise from the clean signal Decompose into IMF components and calculate the residual components; Step 8: Divide the IMF components obtained in step 7 into a training set, a test set, and a validation set; Step 9: Input the training set into the CNN-TCN neural network model for training to obtain the trained CNN-TCN neural network model; input the validation set into the trained CNN-TCN neural network model for verification and optimization to obtain the trained CNN-TCN neural network model; input the test set into the trained CNN-TCN neural network model for fault identification.

3. The method for diagnosing a fault sound signal of a rotating machine according to claim 2, characterized in that: In step 1, the expression for frame splitting and windowing is: （1） In formula (1), Indicates the processed i Frame waterwheel room sound signal; Indicates the collected sound signal of the waterwheel room; N Indicates the frame length; R represents frame shift, R=N / 2; Indicates that the time index n Relative to the i Starting point of the sound signal in the frame waterwheel room iR Offset to get the window function w The relative index of , where the window function w Use Hamming window; The expression for fast Fourier transform of the processed sound signal is: （2） In formula (2), represents the frequency index; Indicates frequency k The first i Frame waterwheel room sound signal; The amplitude is , the expression of phase is: （3） In formula (3), express The imaginary part of express The real part of .

4. The method for diagnosing a fault sound signal of a rotating machine according to claim 2, characterized in that: In step 2, the expression for calculating the noise power spectrum estimation value of the waterwheel room sound signal after FFT of the current frame is: （4） In formula (4), represents the noise power spectrum estimation value of the waterwheel room sound signal at frequency k in the i-th frame; represents the noise power spectrum estimation value of the waterwheel room sound signal at frequency k in frame i-1; represents the signal power spectrum of the waterwheel room sound signal at frequency k in the i-th frame; is the recursive smoothing factor, which is a value between 0 and 1 and is used to control the weighted ratio of the noise power spectrum estimate of the waterwheel room sound signal at frequency k in the i-1 frame (previous frame) to the power spectrum of the waterwheel room sound signal at frequency k in the i-th frame. The probability of the target sound existing Dynamic adjustment; Among them, the recursive smoothing factor It is obtained through two iterations. The specific process is: First iteration: Step S1: Sound signal of waterwheel room Perform the average processing between frequency points to obtain the i Frame waterwheel room sound signal at frequency k The average processed power spectrum estimate at ; The expression is: （5） In formula (5), Represents the weighted window function, using the Hamming window; M represents the half-width of the window function; 2M+1 represents the window length; Indicates the i Frame waterwheel room sound signal at frequency k + m The signal power spectrum at ; Step S2, through the first-order recursion Perform smoothing to obtain i Frame waterwheel room sound signal at frequency k The smoothed estimate of the signal power spectrum at ; The expression is: （6） In formula (6), is the spectral smoothing factor; Step S3, update i Frame waterwheel sound signal at frequency k The minimum smoothed estimate of ; The expression is: （7） In formula (7), L Indicates the search window length, which is equal to the window length of the frame windowing process; S i-j ( k ) indicates the ij A smoothed estimate of the frame, j =0,1,…, L ; Step S4, by Establish rules to obtain sound presence judgment results ; The expression is: （8） In formula (8), Indicates the frequency k The target sound may be contained in the Indicates the frequency k There is only noise; Indicates the i Frame waterwheel room sound signal at frequency k The decision threshold of the smoothed estimate of the signal power spectrum at ; Indicates the Frame waterwheel room sound signal at frequency The decision threshold of the signal power spectrum at ; Second iteration: Step S5, the step S4 is judged to be The waterwheel room sound signal is smoothed for the second time; The expression is: （9） In formula (9), Indicates the i Frame waterwheel room sound signal at frequency k The second smoothed estimate of the signal power spectrum at ; Step S6: Update the minimum smoothed estimated value of the waterwheel sound signal at frequency k in the second i-th frame ; The expression is: （10） In formula (10), L’ Indicates the length of the second search window, which is equal to the length of the first search window; S i-f ( k ) indicates the if A smoothed estimate of the frame, f =0,1,…, L’ ; Step S7: Calculate the probability of the target sound existing ; The expression is: （11） Step S8, by Calculate the recursive smoothing factor ; The expression is: （12） In formula (12), represents the minimum recursive smoothing factor; By Substitute into formula (4) to achieve dynamic update .

5. The method for diagnosing a fault sound signal of a rotating machine according to claim 2, characterized in that: In step 3, the expression for calculating the signal-to-noise ratio is: （13） In formula (13), Indicates the i Frame rate k The signal-to-noise ratio of the sound signal in the waterwheel room.

6. The method for diagnosing a fault sound signal of a rotating machine according to claim 2, characterized in that: In step 4, the expression for calculating the over-reduction factor is: （14） In formula (14), represents the initial oversubtraction factor; s Indicates the slope.

7. The method for diagnosing a fault sound signal of a rotating machine according to claim 2, characterized in that: In step 5, the expression for calculating the power spectrum of the denoised signal is: （15） In formula (15), represents the gain compensation factor; In order to get a smaller ,use Replace the formula (15) , get the final denoised signal power spectrum; The expression is: （16）。 8. The method for diagnosing a rotating machinery fault sound signal according to claim 2, characterized in that: In step 6, the pre-emphasis processing uses a first-order high-pass filter to perform a high-pass filtering operation on the restored signal, and weights the difference between each restored signal and the previous restored signal.

9. The method for diagnosing a fault sound signal of a rotating machine according to claim 2, characterized in that: The specific process of step 7 is: For the clean signal after denoising Add different Gaussian white noise ,repeat I After reconstructing, we get a noisy signal ; The expression is: （17）； In formula (17), Represents different weighting factors of Gaussian white noise; For each Perform EMD decomposition to obtain IMF components , retain the first component , calculate all first components The average value is used as the clean signal after denoising The final first IMF component IMF1; The expression is: （18） Subtract IMF1 from the denoised clean signal to obtain the residual component ; The expression is: （19） Will As a new signal, iterate; Similarly, we get the final p IMF components p , when the p The residual component obtained by decomposition is monotonic, stop computing.

Images