Array sound signal processing method, medium and system for dry type reactor detection

Through the multi-sensor array and CUDA parallel processing technology, combined with multi-task learning model and time-frequency analysis, the problem of information omission in dry reactor detection technology is solved, and the full range of acquisition and efficient processing of dry reactor sound signals is realized. , improves the accuracy and reliability of fault detection.

CN119007750APending Publication Date: 2024-11-22ELECTRIC POWER RESEARCH INSTITUTE OF STATE GRID NINGXIA ELECTRIC POWER COMPANY
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411021569.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing dry reactor detection technology is difficult to fully capture the time-frequency characteristics of the equipment's working sound, resulting in information missed and fault detection is not timely.

Method used

Using multi-sensor array combined with CUDA parallel processing technology, the sound characteristics of the reactor are extracted through a pre-trained multi-task learning model, and combined with time-frequency analysis and feature-time correlation method to achieve all-round and multi-angle sound signals of the dry reactor Acquisition and efficient processing.

Benefits of technology

It significantly improves the integrity and representativeness of information collection, improves signal processing speed, enhances the accuracy and reliability of fault detection, and can dynamically monitor the reactor status and predict the development trend of potential faults.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007750A_ABST
    Figure CN119007750A_ABST
Patent Text Reader

Abstract

The invention provides an array sound signal processing method, medium and system for dry-type reactor detection, and belongs to the technical field of array dry-type reactor sound signal detection, and the method comprises the steps: building CUDA flows with the same number as sound sensors, and obtaining the original sound signals of each sensor in real time; and the signal of each CUDA flow is preprocessed, such as noise reduction and filtering, so that the signal quality is improved. And extracting the working sound of the reactor from the preprocessed signal by using a pre-training model. And time alignment is carried out on the extracted working sound to ensure synchronization of different sensor signals. And fusing the aligned working sound on the GPU, and obtaining an enhanced sound signal by adopting beam forming or principal component analysis. And performing time-frequency analysis on the fused signal, extracting sound features, associating the sound features with time information, and generating output data. The technical problem that the time-frequency characteristic of the working sound of the dry-type reactor cannot be comprehensively captured due to the fact that information is easy to miss in the prior art is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of array dry-type reactor detection acoustic signals, and in particular, relates to an array acoustic signal processing method, medium and system for dry-type reactor detection. Background Art

[0002] As a key device in the power system, dry-type reactors play an important role in improving power quality, harmonic suppression and reactive power compensation. With the continuous expansion of the power system and the increasing complexity of power loads, higher requirements are placed on the reliability and stability of dry-type reactors. Traditional dry-type reactor monitoring methods mainly rely on regular manual inspections and offline testing, which is not only time-consuming and labor-intensive, but also difficult to detect potential fault hazards in a timely manner.

[0003] In recent years, with the rapid development of sensor technology and signal processing technology, equipment status monitoring methods based on sound signal analysis have gradually attracted attention. In the field of dry-type reactors, sound signals contain rich information about the equipment's operating status, which can reflect a variety of problems such as loose windings, abnormal core vibration, and cooling system failures. However, the existing dry-type reactor monitoring technology based on sound signals still has some limitations: when the existing technology collects the sound signals of the operation of the dry-type reactor, it often uses filtering and noise reduction processing. Due to the complex structure of the dry-type reactor, there may be signals with short duration or extremely large frequency fluctuation ranges. These signals may be cleared as abnormal signals, resulting in the inability to fully capture the time-frequency characteristics of the working sound of the dry-type reactor. Summary of the invention

[0004] In view of this, the present invention provides an array acoustic signal processing method, medium and system for dry-type inductor detection, which can solve the technical problem that the existing dry-type inductor detection technology is prone to missing information, resulting in the inability to fully capture the time-frequency characteristics of the dry-type inductor working sound.

[0005] The present invention is achieved in that:

[0006] A first aspect of the present invention provides an array acoustic signal processing method for dry-type reactor detection, which comprises the following steps:

[0007] S10, establishing CUDA streams with the same number as the sound sensor array, for processing signals of multiple sound sensors in parallel;

[0008] S20, acquiring the original sound signal collected by the corresponding sound sensor in real time through each CUDA stream;

[0009] S30, preprocessing the original sound signal in each CUDA stream, including noise reduction and filtering, to improve signal quality;

[0010] S40, in each CUDA stream, using a pre-trained reactor working sound extraction model to extract the reactor working sound from the preprocessed signal;

[0011] S50, performing time alignment on the reactor working sounds extracted from each CUDA stream to ensure that the signals collected by different sensors are synchronized in time;

[0012] S60, fusing the aligned reactor working sounds in the GPU, and obtaining an enhanced sound signal by using a beam forming or principal component analysis method;

[0013] S70, performing time-frequency analysis on the fused sound signal to extract sound features, including spectrum, Mel-frequency cepstral coefficients or wavelet coefficients;

[0014] S80: Associating the extracted sound features with the time information, generating sound feature-time data and outputting the data.

[0015] Among them, the reactor working sound extraction model adopts a multi-task learning method to form a self-supervised learning framework, including a continuous feature sub-network, a discrete feature sub-network, a time-frequency feature sub-network, an attention sub-network and a fusion sub-network.

[0016] Furthermore, the continuous feature subnetwork is used to extract time-continuous sound signal features. The specific structure is a recurrent neural network based on a long short-term memory network, which is used to capture dependencies in long time series and output vectors representing continuous sound features.

[0017] Furthermore, the discrete feature subnetwork is used to extract features of multiple sound signal segments that are discontinuous in time. The specific structure is a multi-layer one-dimensional convolutional neural network, each layer is followed by batch normalization and ReLU activation function, and finally global average pooling is used to output discrete feature vectors.

[0018] Furthermore, the time-frequency feature subnetwork is used to extract time-frequency joint features from the time-frequency diagram of the sound signal. The specific structure is a deep convolutional network based on ResNet. Void convolution is used in the last few layers to expand the receptive field, and a feature pyramid structure is used to fuse features of different depth layers, and finally output a time-frequency feature vector, where the last few layers default to the last 2 layers.

[0019] Furthermore, the attention subnetwork is used to adaptively focus on the most important features of the input signal. The specific structure is a dual attention mechanism including two branches: time attention and frequency attention. The time attention is implemented using a self-attention mechanism, and the frequency attention is implemented using frequency selective convolution.

[0020] Furthermore, the fusion subnetwork is used to integrate the features extracted by each subnetwork. The specific structure is to first concatenate the feature vectors, then perform feature fusion through a multi-layer fully connected network, and finally output a comprehensive feature representation, which is used for subsequent reactor working sound recognition and fault diagnosis.

[0021] Furthermore, the sound sensor array is composed of 8 high-sensitivity condenser microphones.

[0022] Specifically, the step S10 includes:

[0023] Step 101, determining the number of sensors in the sound sensor array;

[0024] Step 102, using a computing unified device architecture programming model to create an independent computing unified device architecture flow having the same number of sensors as the number of sensors in the sound sensor array;

[0025] Step 103, associating each computing unified device architecture flow with a specific sound sensor in the sound sensor array;

[0026] Step 104, create a stream object using the compute unified device architecture runtime application program interface function;

[0027] Step 105: setting the priority of each computing unified device architecture flow, the priority range is from negative one to zero, where negative one represents the highest priority;

[0028] Step 106: Allocate GPU memory to store the original sound signal collected by each sound sensor;

[0029] Step 107: Prepare for processing the signals of multiple sound sensors in parallel to improve the overall signal processing efficiency.

[0030] Wherein, the step S20 specifically includes:

[0031] Step 201, allocating a graphics processor memory to each created computing unified device architecture stream for storing the original sound signal collected by the corresponding sound sensor;

[0032] Step 202: using an asynchronous memory copy function to transfer the sound data in the CPU memory to the GPU memory;

[0033] Step 203: using double buffering technology, while one buffer is being processed, the other buffer receives new data;

[0034] Step 204, setting the sampling rate of the sound signal to 44,100 Hz or 48,000 Hz;

[0035] Step 205, setting the data block size, the data block size is one thousand and twenty-four or two thousand and forty-eight sampling points;

[0036] Step 206: ensuring that the original signals of each sound sensor can be transmitted to the graphics processor in a timely and efficient manner;

[0037] Step 207: Prepare for subsequent processing to balance delay and processing efficiency.

[0038] Wherein, the step S30 specifically includes:

[0039] Step 301: In each computing unified device architecture flow, perform noise reduction processing on the original sound signal;

[0040] Step 302: Use spectral subtraction to perform noise reduction, estimate the noise power spectrum in the frequency domain, and then subtract the noise power spectrum from the original signal power spectrum;

[0041] Step 303: Perform short-time Fourier transform on the signal to obtain a time-frequency representation of the signal;

[0042] Step 304, estimating the noise power spectrum, using the non-speech segment or the silent segment of the signal for estimation;

[0043] Step 305: Calculate the enhanced signal power spectrum, and set the over-reduction factor and spectrum lower limit factor;

[0044] Step 306: perform inverse short-time Fourier transform based on the original phase information to obtain a de-noised time domain signal;

[0045] Step 307: bandpass filter the denoised signal to remove irrelevant frequency components.

[0046] Wherein, the step S40 specifically includes:

[0047] Step 401: In each computing unified device architecture flow, use a pre-trained reactor working sound extraction model to process the pre-processed signal;

[0048] Step 402, dividing the preprocessed signal into frames of fixed length, each frame containing 2,048 sampling points, and adjacent frames overlapping by 50%;

[0049] Step 403: extracting time-continuous features using a continuous feature subnetwork;

[0050] Step 404: extracting time discontinuous features using a discrete feature subnetwork;

[0051] Step 405: extract joint features from the time-frequency graph using the time-frequency feature subnetwork;

[0052] Step 406: Use the attention sub-network to adaptively focus on important features through time attention and frequency attention mechanisms;

[0053] Step 407: Use the fusion sub-network to integrate the features extracted by each sub-network and output a comprehensive feature representation.

[0054] Wherein, the step S50 specifically includes:

[0055] Step 501: Select a reference sensor, usually a sensor with the highest signal-to-noise ratio is selected as a reference;

[0056] Step 502: Calculate the cross-correlation function between other sensor signals and the reference signal;

[0057] Step 503: Find the time delay value that maximizes the cross-correlation function as the optimal alignment delay;

[0058] Step 504: Use interpolation technology to accurately align the signal at a sub-sampling level;

[0059] Step 505: Use a dynamic time warping algorithm to perform further fine-grained alignment to handle possible sampling rate drift;

[0060] Step 506: setting an alignment window, and performing local alignment within the window;

[0061] Step 507: gradually slide the window to complete the alignment of the entire signal, ensuring that the signals from different sensors are accurately aligned in time.

[0062] Wherein, the step S60 specifically includes:

[0063] Step 601, using a beamforming method to fuse the aligned reactor working sounds;

[0064] Step 602, calculating the time delay from each sensor to the sound source;

[0065] Step 603: Perform corresponding time delay compensation on each sensor signal;

[0066] Step 604: performing weighted summation on the compensated signal;

[0067] Step 605: dynamically adjust the weight of each sensor according to the signal quality;

[0068] Step 606: construct a signal covariance matrix using a principal component analysis method;

[0069] Step 607: Calculate the eigenvalues ​​and eigenvectors of the covariance matrix, select the eigenvector corresponding to the maximum eigenvalue as the principal component, and project it to obtain an enhanced signal.

[0070] Wherein, the step S70 specifically includes:

[0071] Step 701, using short-time Fourier transform to calculate the time-frequency representation of the signal;

[0072] Step 702: Calculate power spectrum density;

[0073] Step 703, calculate the Mel frequency cepstrum coefficients, first convert the linear frequency spectrum into a Mel frequency spectrum, then take the logarithm, and finally perform discrete cosine transform;

[0074] Step 704, usually taking the first thirteen Mel-frequency cepstral coefficients;

[0075] Step 705: Calculate wavelet coefficients using continuous wavelet transform or discrete wavelet transform;

[0076] Step 706, selecting a suitable scale range, usually covering a frequency range of 20 Hz to 5 kHz;

[0077] Step 707: Describe the time-frequency characteristics of the sound signal from multiple perspectives to provide a comprehensive feature representation for subsequent feature time association and fault diagnosis.

[0078] Wherein, the step S80 specifically includes:

[0079] Step 801: construct a feature vector for each time frame, including the spectrum, Mel-frequency cepstral coefficients and wavelet coefficients at that time point;

[0080] Step 802: Associating the feature vector with the corresponding timestamp to form a feature time pair;

[0081] Step 803: Calculate the first and second order time derivatives of the feature to capture the time evolution of the feature;

[0082] Step 804: construct an augmented feature vector, including the original feature, the first-order derivative and the second-order derivative;

[0083] Step 805: Use principal component analysis or autoencoder to perform feature dimension reduction to reduce the amount of data and highlight important features;

[0084] Step 806: Save the reduced-dimensional feature vector together with the time information in a file format suitable for large-scale scientific data storage;

[0085] Step 807: Set an appropriate compression level during the saving process to balance storage space and reading speed.

[0086] Optionally, the training step of the reactor working sound extraction model includes:

[0087] Step 1001, collecting sound data of multiple dry-type reactors of different models and different working states;

[0088] Step 1002: Use a high-quality microphone array to record the working sound of the reactor under different environmental conditions;

[0089] Step 1003, dividing the recorded sound data into segments of fixed length;

[0090] Step 1004: preprocess each sound clip, including noise reduction and normalization;

[0091] Step 1005: Calculate the harmonic distortion, sound pressure level, spectrum center and working status index of each sound clip as training labels;

[0092] Step 1006: Initialize model parameters and randomly select sound clips for each training batch for training;

[0093] Step 1007: Calculate the loss function between the predicted output and the true label, calculate the gradient using the back propagation algorithm, and use the optimizer to update the model parameters.

[0094] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned array acoustic signal processing method for dry-type inductor detection.

[0095] A third aspect of the present invention provides an array acoustic signal processing system for dry-type inductor detection, which includes the above-mentioned computer-readable storage medium.

[0096] Compared with the prior art, the array acoustic signal processing method, medium and system for dry-type reactor detection provided by the present invention have the following beneficial effects:

[0097] First, the present invention adopts multi-sensor array acquisition technology, combined with the CUDA parallel processing architecture, to achieve all-round, multi-angle sound signal acquisition and efficient processing of dry-type reactors. Compared with the traditional single-point acquisition method, the present invention can more comprehensively capture the sound characteristics of each part of the reactor, greatly improving the integrity and representativeness of information collection. Through the parallel processing of CUDA streams, this method increases the signal processing speed by 3-5 times, meeting the needs of real-time monitoring.

[0098] Secondly, the present invention designs an innovative multi-stage noise reduction and signal enhancement algorithm. By combining spectral subtraction and adaptive filtering, the influence of environmental noise is effectively suppressed, and the signal-to-noise ratio is improved by 5-8dB. At the same time, beamforming technology is used to perform spatial filtering and signal fusion on multi-channel signals, further enhancing the target sound signal.

[0099] Furthermore, the present invention proposes a multi-task learning model for extracting the working sound of a reactor, including a continuous feature sub-network, a discrete feature sub-network, a time-frequency feature sub-network, an attention sub-network, and a fusion sub-network. This innovative network structure can simultaneously capture the temporal continuity characteristics and discrete event characteristics of the sound signal, and adaptively focus on important features through the attention mechanism. Compared with the traditional single network structure, this method has significantly improved the comprehensiveness of feature extraction and the generalization ability of the model.

[0100] In addition, the present invention adopts advanced time-frequency analysis methods, combined with short-time Fourier transform, Mel-frequency cepstral coefficients and wavelet transform and other technologies, to comprehensively extract the time-frequency characteristics of sound signals. Compared with the traditional single feature analysis, this multi-dimensional feature extraction method can more accurately characterize the working state of the dry-type reactor.

[0101] Finally, the present invention designs an innovative feature-time correlation method, which not only extracts rich sound features, but also closely associates these features with time information. This method enables the system to track the dynamic changes of the reactor state and predict the development trend of potential faults. Compared with the traditional static analysis method, the dynamic monitoring capability of the present invention is significantly enhanced.

[0102] In summary, the present invention comprehensively improves the accuracy, real-time and reliability of dry-type inductor sound signal detection through innovative technologies such as multi-sensor array acquisition, CUDA parallel processing, multi-task deep learning model, comprehensive time-frequency analysis and dynamic feature-time correlation, and solves the technical problem that the existing dry-type inductor detection technology is prone to missing information, resulting in the inability to comprehensively capture the time-frequency characteristics of the working sound of the dry-type inductor. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0104] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0105] like Figure 1 FIG. 1 is a flow chart of an array acoustic signal processing method for dry-type reactor detection provided by the present invention. The method comprises the following steps:

[0106] S10, establishing CUDA streams with the same number as the sound sensor array, for processing signals of multiple sound sensors in parallel;

[0107] S20, acquiring the original sound signal collected by the corresponding sound sensor in real time through each CUDA stream;

[0108] S30, preprocessing the original sound signal in each CUDA stream, including noise reduction and filtering, to improve signal quality;

[0109] S40, in each CUDA stream, using a pre-trained reactor working sound extraction model to extract the reactor working sound from the preprocessed signal;

[0110] S50, performing time alignment on the reactor working sounds extracted from each CUDA stream to ensure that the signals collected by different sensors are synchronized in time;

[0111] S60, fusing the aligned reactor working sounds in the GPU, and obtaining an enhanced sound signal by using a beam forming or principal component analysis method;

[0112] S70, performing time-frequency analysis on the fused sound signal to extract sound features, including spectrum, Mel-frequency cepstral coefficients or wavelet coefficients;

[0113] S80: Associating the extracted sound features with the time information, generating sound feature-time data and outputting the data.

[0114] The specific implementation methods of the above steps are described in detail below:

[0115] The specific implementation method of step S10 is: first, determine the number N of sensors in the sound sensor array. Then, use the CUDA (CUDA is a parallel computing platform and programming model, which includes the CUDA instruction set architecture and the parallel computing engine inside the GPU.) programming model to create N independent CUDA streams. Each CUDA stream is associated with a specific sound sensor and is used to process the signal collected by the sensor. In specific implementation, the CUDA runtime API function CUDAStreamCreate() can be used to create a stream object. In order to optimize performance, the priority of the stream can be set. Use the CUDAStreamCreateWithPriority() function to set the priority in the range of [-1,0], where -1 represents the highest priority. The purpose of this step is to prepare for the parallel processing of signals from multiple sound sensors, make full use of the parallel computing capabilities of the GPU, and improve the overall signal processing efficiency.

[0116] The specific implementation of step S20 is: for each created CUDA stream, GPU memory is allocated to store the original sound signal collected by the corresponding sound sensor. The asynchronous memory copy function CUDAMemcpyAsync() is used to transfer the sound data in the host (CPU) memory to the GPU memory. In order to achieve real-time processing, double buffering technology can be used: set one buffer to be processed and the other buffer to receive new data at the same time. Set an appropriate sampling rate, usually 44.1kHz or 48kHz, to ensure high fidelity of the sound signal. At the same time, set a suitable data block size, such as 1024 or 2048 sampling points, to balance delay and processing efficiency. The purpose of this step is to ensure that the original signals of each sound sensor can be transmitted to the GPU in a timely and efficient manner for subsequent processing.

[0117] The specific implementation of step S30 is: in each CUDA stream, the original sound signal is preprocessed, mainly including two sub-steps of noise reduction and filtering. The noise reduction process adopts spectral subtraction, the principle of which is to estimate the noise power spectrum in the frequency domain and then subtract the noise power spectrum from the original signal power spectrum. In specific implementation, the signal is first subjected to a short-time Fourier transform (STFT) to obtain the time-frequency representation X(k,m) of the signal, where k represents the frequency index and m represents the time frame index. Then the noise power spectrum |N(k)| is estimated. 2 , can be estimated using the non-speech or silence segments of the signal. Next, calculate the enhanced signal power spectrum:

[0118]

[0119] Among them, α is the over-subtraction factor, which usually ranges from [1, 2], and β is the spectrum lower limit factor, which ranges from [0.001, 0.1]. Finally, the inverse short-time Fourier transform (ISTFT) is performed on the original phase information to obtain the denoised time domain signal.

[0120] The filtering process uses a bandpass filter to remove irrelevant frequency components. Considering the working sound characteristics of the dry-type reactor, a bandpass filter with a frequency range of [100Hz, 5000Hz] can be designed. It is implemented using a finite impulse response (FIR) filter, and its transfer function is:

[0121]

[0122] Where h(n) is the impulse response of the filter and N is the filter order. The filter order can be set to an odd number, such as 201 or 301, to ensure a linear phase response. The filter coefficients are designed using a Hamming window function to reduce passband ripple and stopband attenuation. The purpose of this step is to improve the quality of the signal, remove environmental noise and irrelevant frequency components, and prepare for subsequent feature extraction.

[0123] The specific implementation of step S40 is: in each CUDA stream, a pre-trained reactor working sound extraction model is used to extract the reactor working sound from the preprocessed signal. The model adopts a multi-task learning framework, including a continuous feature sub-network, a discrete feature sub-network, a time-frequency feature sub-network, an attention sub-network, and a fusion sub-network. First, the preprocessed signal is divided into frames of fixed length, each frame usually contains 2048 sampling points, and adjacent frames overlap by 50%. Feature extraction is performed on each frame signal: the continuous feature sub-network uses a long short-term memory network (LSTM) to extract time-continuous features; the discrete feature sub-network uses a one-dimensional convolutional neural network to extract time-discontinuous features; the time-frequency feature sub-network uses a deep convolutional network based on ResNet to extract joint features from the time-frequency graph. The attention sub-network adaptively focuses on important features through time attention and frequency attention mechanisms. Finally, the fusion sub-network integrates the features extracted by each sub-network and outputs a comprehensive feature representation. The purpose of this step is to accurately extract the working sound of the reactor from a complex sound signal and provide a reliable feature representation for subsequent analysis and diagnosis.

[0124] The specific implementation of step S50 is: time align the reactor working sounds extracted from each CUDA stream to ensure that the signals collected by different sensors are synchronized in time. First, select a reference sensor, usually the sensor with the highest signal-to-noise ratio as the reference. Then, calculate the cross-correlation function between other sensor signals and the reference signal:

[0125]

[0126] Where x(n) is the reference signal, y(n) is the signal to be aligned, and τ is the time delay. Find the value that makes R xy (τ) The largest τ value is the optimal alignment delay. Use interpolation techniques (such as cubic spline interpolation) to accurately align the signals at the sub-sampling level. In order to deal with possible sampling rate drift, the dynamic time warping (DTW) algorithm can be used for further fine-grained alignment. Set an alignment window, such as 1 second, perform local alignment within the window, and then gradually slide the window to complete the alignment of the entire signal. The purpose of this step is to ensure that the signals from different sensors are accurately aligned in time, providing a basis for subsequent signal fusion and feature extraction.

[0127] The specific implementation of step S60 is: the aligned reactor working sound is fused in the GPU, and the enhanced sound signal is obtained by using the beamforming or principal component analysis method. The beamforming method is based on the delay-and-sum principle, and first calculates the time delay from each sensor to the sound source:

[0128]

[0129] Among them, r i is the position vector of the ith sensor, r0 is the reference point position vector, is the unit vector of the sound source direction, and c is the speed of sound. Then, the corresponding time delay compensation is performed on each sensor signal, and a weighted sum is performed:

[0130]

[0131] Among them, w i is the weight of the i-th sensor, which can be adjusted dynamically according to the signal quality. The principal component analysis (PCA) method first constructs the signal covariance matrix:

[0132]

[0133] Among them, X(t) is a vector containing all sensor signals, and T is the signal length. Then the eigenvalues ​​and eigenvectors of C are calculated, and the eigenvector corresponding to the maximum eigenvalue is selected as the principal component, and the enhanced signal is projected. The purpose of this step is to use the information of multiple sensors to improve the signal-to-noise ratio and spatial resolution of the signal, and obtain a clearer and more reliable working sound of the reactor.

[0134] The specific implementation of step S70 is: perform time-frequency analysis on the fused sound signal to extract sound features. First, use short-time Fourier transform (STFT) to calculate the time-frequency representation of the signal:

[0135]

[0136] Where w(n) is the window function (such as the Hanning window), R is the frame shift, and N is the FFT length. Then, the spectrum characteristics are calculated, including the power spectral density (PSD):

[0137]

[0138] Calculate the Mel Frequency Cepstral Coefficients (MFCC): First convert the linear frequency spectrum to a Mel frequency spectrum, then take the logarithm, and finally perform a discrete cosine transform (DCT). Usually take the first 13 MFCC coefficients.

[0139]

[0140] Among them, S k is the Mel filter bank output, and K is the number of filters, typically 26.

[0141] Calculate wavelet coefficients: Use continuous wavelet transform (CWT) or discrete wavelet transform (DWT). Take CWT as an example:

[0142]

[0143] Among them, ψ(t) is the mother wavelet function (such as Morlet wavelet), a is the scale parameter, and b is the translation parameter. Select an appropriate scale range, usually covering the frequency range of 20Hz to 5000Hz. The purpose of this step is to describe the time-frequency characteristics of the sound signal from multiple angles and provide a comprehensive feature representation for subsequent feature-time association and fault diagnosis.

[0144] The specific implementation of step S80 is: associate the extracted sound features with time information, generate sound feature-time data and output it. First, for each time frame t, construct a feature vector F(t), which contains the spectrum, MFCC and wavelet coefficients at that time point. Then, associate the feature vector with the corresponding timestamp t to form a feature-time pair (F(t), t). In order to capture the time evolution of the feature, the first and second order time derivatives of the feature can be calculated:

[0145]

[0146] Δ 2 F(t)=F(t+1)-2F(t)+F(t-1)

[0147] Constructing augmented eigenvectors In order to reduce the amount of data and highlight important features, principal component analysis (PCA) or autoencoders can be used for feature dimensionality reduction. Finally, the reduced feature vector is saved together with the time information, and a file format suitable for large-scale scientific data storage such as HDF5 or NetCDF can be used. During the saving process, an appropriate compression level (such as GZIP compression level 6) is set to balance storage space and reading speed. The purpose of this step is to generate a feature-time series dataset containing rich time-frequency information to provide comprehensive data support for subsequent reactor condition monitoring and fault diagnosis.

[0148] During the entire processing, some important parameters and threshold reference values ​​involved are as follows:

[0149] 1.CUDA stream priority: [-1,0], -1 means the highest priority.

[0150] 2. Sound signal sampling rate: 44.1kHz or 48kHz.

[0151] 3. Data block size: 1024 or 2048 sampling points.

[0152] 4. Oversubtraction factor α in spectral subtraction: [1,2].

[0153] 5. Spectral lower limit factor β in spectral subtraction: [0.001, 0.1].

[0154] 6.Bandpass filter frequency range: [100Hz,5000Hz].

[0155] 7.FIR filter order: 201 or 301.

[0156] 8. Signal frame length: 2048 sampling points, adjacent frames overlap 50%.

[0157] 9. Time alignment window: 1 second.

[0158] 10. Number of Mel filter banks: 26.

[0159] 11. Number of MFCC coefficients: 13.

[0160] 12. Wavelet transform frequency range: 20Hz to 5000Hz.

[0161] 13. Data compression level: GZIP compression level 6.

[0162] These parameters and thresholds can be fine-tuned according to specific application scenarios and hardware conditions to obtain the best processing effect. In practical applications, the optimal parameter combination can be determined through methods such as cross-validation or grid search.

[0163] In the process of training the reactor working sound extraction model, a multi-task learning strategy is adopted to optimize the continuous feature prediction, discrete feature classification, and time-frequency feature reconstruction tasks at the same time. By sharing the underlying feature representation and task-specific output layers, the model learns more robust and universal reactor working sound features. The following are the training steps for the reactor working sound extraction model, including the establishment of a training data set, data label processing, and training process:

[0164] 1. Create a training data set:

[0165] a) Collect sound data of multiple dry-type reactors of different models and working conditions.

[0166] b) Use a high-quality microphone array to record the working sound of the reactor under different environmental conditions (such as different temperatures, humidity, background noise, etc.).

[0167] c) Divide the recorded sound data into segments of fixed length (eg, 1 to 10 seconds).

[0168] d) Preprocess each sound clip, including noise reduction, normalization, etc.

[0169] 2. Data label processing:

[0170] For each sound clip, the following features are calculated as training labels:

[0171] a) Harmonic distortion (THD):

[0172]

[0173] Among them, A i represents the amplitude of the i-th harmonic, A1 represents the fundamental amplitude, and N is the number of harmonics considered.

[0174] b) Sound Pressure Level (SPL):

[0175]

[0176] Among them, p rms is the root mean square pressure of the sound signal, p ref is the reference sound pressure (usually 20μPa).

[0177] c) Spectrum Centre (SC):

[0178]

[0179] Among them, f k is the kth frequency component, P(f k ) is the power spectral density of the frequency component, and K is the total number of frequency components.

[0180] d) Working Status Index (WSI):

[0181] WSI=w1·THD+w2·SPL+w3·SC

[0182] Among them, w1, w2, w3 are weight coefficients, which need to be adjusted according to the specific application.

[0183] 3. Model training process:

[0184] a) Initialize model parameters.

[0185] b) For each training batch:

[0186] A batch of sound clips are randomly selected from the training dataset.

[0187] Input the sound clip into the model and get the predicted output.

[0188] Calculate the loss function between the predicted output and the true label:

[0189] L=α·L THD +β·L SPL +γ·L SC +δ·L WSI

[0190] Among them, L THD ,L SPL ,L SC ,L WSI They are the losses of various indicators (such as mean square error), and α, β, γ, and δ are weight coefficients.

[0191] The gradient is calculated using the back-propagation algorithm.

[0192] Use an optimizer such as Adam to update the model parameters.

[0193] c) Evaluate the model performance on the validation set and save the model checkpoint if the performance improves.

[0194] d) Repeat steps b and c until the preset number of training rounds is reached or the model converges.

[0195] 4. Model Evaluation:

[0196] The model performance was evaluated on the test set, and the prediction accuracy of each indicator and the accuracy of the overall working status classification were calculated.

[0197] Variables and constants explained:

[0198] THD: Total harmonic distortion, which indicates the strength of the harmonic components in the signal relative to the fundamental wave.

[0199] A i : The amplitude of the i-th harmonic.

[0200] N: The number of harmonics considered, usually 5-7.

[0201] SPL: Sound Pressure Level, a logarithmic measure of sound intensity.

[0202] p rms : RMS pressure of the sound signal.

[0203] p ref : Reference sound pressure, usually 20μPa (the audible threshold of the human ear).

[0204] SC: Spectral Center, which represents the "center of gravity" of the signal spectrum.

[0205] f k : kth frequency component.

[0206] P(f k ): frequency f k The power spectral density at .

[0207] K: total number of frequency components.

[0208] WSI: Working status index, which comprehensively reflects the working status of the reactor.

[0209] w1,w2,w3: weight coefficients in WSI calculation.

[0210] L: total loss function.

[0211] L THD ,L SPL ,L SC ,L WSI : Loss function of each indicator.

[0212] α, β, γ, δ: weight coefficients of each part in the total loss function.

[0213] Specifically, the principle of the present invention is:

[0214] 1. Multi-sensor array acquisition and CUDA parallel processing principle:

[0215] The present invention adopts multi-sensor array acquisition technology, based on the theory of sound field spatial sampling, and realizes the all-round acquisition of the sound field of the dry reactor by rationally arranging multiple sound sensors. According to the Nyquist sampling theorem, the sensor spacing is no more than half the wavelength of the sound, which can effectively avoid spatial aliasing. CUDA parallel processing technology is based on the multi-core architecture of GPU, and uses CUDA streams to realize parallel processing of multiple sensor signals. Each CUDA stream corresponds to an independent hardware queue, which can be executed concurrently, making full use of the computing resources of the GPU and significantly improving the efficiency of signal processing.

[0216] 2. Principle of multi-stage noise reduction and signal enhancement:

[0217] The present invention adopts a noise reduction strategy combining spectral subtraction and adaptive filtering. Spectral subtraction is based on the short-term stationary assumption. It effectively suppresses additive noise by estimating the noise power spectrum and subtracting it from the original signal power spectrum. Adaptive filtering uses the minimum mean square error criterion to dynamically adjust the filter coefficients to adaptively suppress non-stationary noise. Beamforming technology is based on the principle of spatial coherence of acoustic waves. By adjusting the phase and amplitude of multiple sensor signals, spatial filtering and signal enhancement are achieved to improve the signal-to-noise ratio of the target signal.

[0218] 3. Principle of sound extraction model of multi-task learning:

[0219] The multi-task learning model proposed in the present invention is based on the representation learning and transfer learning theories in deep learning. The continuous feature sub-network adopts the LSTM structure and uses its gating mechanism to capture the dependencies of long time series. The discrete feature sub-network uses a one-dimensional CNN to extract temporal local features through local receptive fields and weight sharing. The time-frequency feature sub-network is based on the ResNet structure and uses residual connections to alleviate the gradient vanishing problem of deep networks. The attention sub-network implements adaptive feature weighting and enhances the interpretability of the model. The multi-task learning framework realizes knowledge transfer between different feature extraction tasks through shared representations and task-specific layers, thereby improving the generalization ability of the model.

[0220] 4. Principle of comprehensive time-frequency analysis:

[0221] The present invention combines a variety of time-frequency analysis methods to fully capture the time-frequency characteristics of sound signals. Short-time Fourier transform is based on the principle of time-frequency localization and realizes time-varying spectrum analysis of signals through sliding windows. Mel-frequency cepstrum coefficients use the auditory characteristics of the human ear to perform nonlinear transformations on the spectrum and extract features that are more in line with auditory perception. Wavelet transform is based on multi-resolution analysis theory and realizes localized representation of signals in the time and frequency domain through scale and translation operations, which is particularly suitable for analyzing non-stationary signals.

[0222] 5. Feature-time correlation principle:

[0223] The feature-time association method proposed in this invention is based on time series analysis and dynamic system theory. By calculating the time derivative of the feature, the dynamic change trend of the feature is captured. The construction of the augmented feature vector is based on the state space representation theory, combining static features with dynamic information to fully describe the system state. The feature dimension reduction adopts principal component analysis or autoencoder, based on the information compression theory, to retain the main variation of the data and reduce redundancy.

[0224] 6. Principle of self-supervised learning framework:

[0225] The present invention adopts a self-supervised learning framework based on representation learning and contrastive learning theory. By designing appropriate pre-training tasks such as reconstruction, prediction, and contrastive learning, the model can learn meaningful feature representations from a large amount of unlabeled data. This approach makes full use of the intrinsic structural information in the data and improves the performance of the model in downstream tasks.

[0226] In summary, the technical solution of the present invention is based on the theoretical foundations of multiple disciplines such as acoustics, signal processing, deep learning and dynamic system analysis, and the various modules are logically rigorous and mutually supportive. Multi-sensor array acquisition provides comprehensive data for subsequent processing, CUDA parallel processing ensures real-time performance, multi-stage noise reduction and signal enhancement improve signal quality, multi-task learning models achieve comprehensive feature extraction, comprehensive time-frequency analysis captures the multi-dimensional characteristics of signals, and feature-time correlation methods achieve dynamic monitoring. The organic combination of these technical principles enables the present invention to effectively solve the key problems in the detection of sound signals in dry-type reactors and achieve high-precision and high-reliability state monitoring and fault diagnosis.

[0227] The following is an example of a specific application scenario of the present invention: This example takes a 330kV dry-type reactor in a large substation as the object, and describes in detail the specific implementation process of the array acoustic signal processing method for dry-type reactor detection of the present invention.

[0228] 1. System Configuration

[0229] The hardware system used in this embodiment includes:

[0230] 8 high-sensitivity condenser microphones (frequency response range: 20Hz-20kHz, sensitivity: -38dB±3dB);

[0231] 1 multi-channel data acquisition card (sampling rate: 48kHz, resolution: 24bit);

[0232] 1 workstation equipped with NVIDIA RTX 3090 GPU (CUDA cores: 10496, video memory: 24GB);

[0233] The software system is developed based on Python 3.8, and the main libraries used include NumPy, SciPy, PyTorch and PyCUDA.

[0234] 2. Sensor layout

[0235] According to the structural characteristics of the dry-type reactor, the 8 microphones are arranged in the following manner:

[0236] The four microphones are evenly distributed around the reactor housing, 20 cm away from the housing surface;

[0237] Two microphones were placed on top of the reactor, 1m apart;

[0238] Two microphones were placed at the bottom of the reactor, 1m apart;

[0239] This arrangement ensures all-round sound collection of the reactor.

[0240] 3. Data Collection and Preprocessing

[0241] The system was run continuously for 30 days, collecting 1 minute of sound data every hour. The sampling rate was set to 48kHz, and each CUDA stream processed a data block of 1024 sampling points.

[0242] In CUDA parallel processing, eight CUDA streams are created, each corresponding to a microphone. The asynchronous memory copy function CUDAMemcpyAsync() is used to transfer the sound data in the host memory to the GPU memory. In order to achieve real-time processing, double buffering technology is used: when one buffer is processed, the other buffer receives new data.

[0243] 4. Noise reduction

[0244] The original sound signal in each CUDA stream is denoised. First, spectral subtraction is applied:

[0245] Use Hanning window for framing, frame length 1024 points, frame shift 512 points;

[0246] Perform FFT on each frame of signal to obtain the time-frequency representation of the signal;

[0247] Estimate the noise power spectrum, using the first 100ms (non-working state) of the signal for estimation;

[0248] Set the oversubtraction factor α = 1.2 and the spectrum lower limit factor β = 0.002;

[0249] Apply the spectrum subtraction formula to calculate the power spectrum of the enhanced signal;

[0250] Combine the original phase information with IFFT to obtain the noise-reduced time domain signal;

[0251] Then apply a bandpass filter:

[0252] Design a FIR bandpass filter with a passband of 100Hz-5000Hz;

[0253] The filter order is 201, and it is designed using a Hamming window;

[0254] Filter the denoised signal to remove irrelevant frequency components;

[0255] 5. Reactor working sound extraction

[0256] Use the pre-trained reactor working sound extraction model to process the denoised signal. The model structure is as follows:

[0257] Continuous feature subnetwork:

[0258] 2-layer bidirectional LSTM, 128 hidden units per layer;

[0259] Each layer of LSTM is followed by BatchNorm and ReLU activation functions;

[0260] The last layer is followed by a fully connected layer, which outputs a 64-dimensional feature vector;

[0261] Discrete feature subnetwork:

[0262] 4-layer 1D CNN, with convolution kernel sizes of 3, 5, 7, and 9 respectively;

[0263] Each convolution layer is followed by BatchNorm and ReLU activation functions;

[0264] Finally, global average pooling is used to output a 64-dimensional feature vector;

[0265] Time-frequency feature sub-network:

[0266] Based on the ResNet18 structure, the input is a spectrogram (Mel spectrum);

[0267] The last two layers use dilated convolutions with dilation rates of 2 and 4;

[0268] Use feature pyramid structure to fuse features from different layers;

[0269] Output 128-dimensional feature vector;

[0270] Attention subnetwork:

[0271] Temporal attention: self-attention mechanism, 8 attention heads;

[0272] Frequency attention: 3-layer 1D CNN with kernel sizes of 3, 5, and 7;

[0273] Using gating mechanism to fuse time and frequency attention;

[0274] Fusion sub-network:

[0275] Concatenate the feature vectors of each sub-network;

[0276] 3-layer fully connected network with hidden layer sizes of 256, 128, and 64;

[0277] Each fully connected layer is followed by BatchNorm and ReLU activation functions;

[0278] Finally, a 64-dimensional comprehensive feature representation is output;

[0279] 6. Time Alignment

[0280] Select the microphone signal with the highest signal-to-noise ratio as the reference and time-align the other 7 microphone signals:

[0281] Calculate the cross-correlation function and find the time delay corresponding to the maximum cross-correlation;

[0282] Use cubic spline interpolation for accurate sub-sampling level alignment;

[0283] Set the alignment window to 1 second and slide gradually to complete the alignment of the entire signal;

[0284] 7. Signal Fusion

[0285] The 8 microphone signals are aligned using the beamforming method and then fused:

[0286] Calculate the time delay from each microphone to the hypothetical sound source based on the microphone array geometry configuration;

[0287] Perform corresponding time delay compensation for each microphone signal;

[0288] Dynamically adjust the weight of each microphone based on signal quality;

[0289] Perform weighted summation to obtain an enhanced sound signal;

[0290] 8. Time-frequency analysis

[0291] Perform a comprehensive time-frequency analysis on the fused signal:

[0292] Short-time Fourier transform: frame length 2048 points, frame shift 1024 points, using Hanning window;

[0293] Calculate Mel frequency cepstral coefficients: 26 Mel filters, take the first 13 MFCC coefficients;

[0294] Continuous wavelet transform: using Morlet wavelet, the scale range corresponds to 20Hz-5000Hz;

[0295] 9. Feature-Time Correlation

[0296] Construct a feature vector containing spectrum, MFCC and wavelet coefficients, and calculate the first and second order time derivatives:

[0297] Original feature vector dimension: 257 (FFT amplitude spectrum) + 13 (MFCC) + 50 (wavelet coefficients) = 320 dimensions;

[0298] Augmented feature vector: 320 (original features) + 320 (first-order derivatives) + 320 (second-order derivatives) = 960 dimensions;

[0299] PCA is used for dimensionality reduction, retaining 95% of the variance information, and finally a 128-dimensional feature vector is obtained;

[0300] 10. Status Diagnosis

[0301] Use pre-trained hidden Markov models for state diagnosis:

[0302] Define 5 hidden states: normal, slightly abnormal, moderately abnormal, severely abnormal, and emergency state;

[0303] Use the forward-backward algorithm to calculate the likelihood and evaluate the current state;

[0304] Decode the most likely state sequence using the Viterbi algorithm;

[0305] 11. Results Analysis

[0306] During the 30-day monitoring period, the system processed a total of 720 hours of sound data. The following results were obtained through analysis:

[0307] 1) Detection performance:

[0308] Normal working state recognition accuracy: 99.2%;

[0309] Abnormal state detection rate: 95.7%;

[0310] False positive rate: 1.3%

[0311] 2) Detected abnormal events:

[0312] Slight abnormality: 27 times;

[0313] Moderate abnormality: 5 times;

[0314] Severe abnormality: 1 time;

[0315] Emergency: 0 times;

[0316] 3) Main abnormal types and proportions:

[0317] Winding looseness: 42%;

[0318] Abnormal core vibration: 31%;

[0319] Cooling system failure: 18%;

[0320] Others: 9%;

[0321] 4) System performance:

[0322] Average signal processing delay: 23.6ms;

[0323] GPU utilization: 78%

[0324] CPU utilization: 32%

[0325] Table 1 shows the statistics of abnormal events detected within a week:

[0326] Table 1 Abnormal statistics

[0327]

[0328]

[0329] 12. Case Analysis

[0330] During the monitoring period, the system successfully warned of a potential serious fault. On July 5, the system detected several minor anomalies in succession, mainly manifested as abnormal increase in energy in a specific frequency band (500-800Hz). By analyzing the time-frequency characteristics and state transition probability, the system predicted that the equipment might have a loose winding fault.

[0331] Through this embodiment, the effectiveness and practicality of the array acoustic signal processing method for dry-type reactor detection of the present invention are fully verified. This method not only significantly improves the operating reliability and maintenance efficiency of dry-type reactors, but also provides strong support for the intelligent and digital transformation of power systems.

[0332] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for processing array acoustic signals for dry-type reactor detection, characterized in that: The following steps are involved: S10, establishing CUDA streams with the same number as the sound sensor array, for processing signals of multiple sound sensors in parallel; S20, acquiring the original sound signal collected by the corresponding sound sensor in real time through each CUDA stream; S30, preprocessing the original sound signal in each CUDA stream, including noise reduction and filtering; S40, in each CUDA stream, using a pre-trained reactor working sound extraction model to extract the reactor working sound from the preprocessed signal; S50, performing time alignment on the reactor working sounds extracted from each CUDA stream to ensure that the signals collected by different sound sensors are synchronized in time; S60, fusing the aligned reactor working sounds in the GPU, and obtaining an enhanced sound signal by using a beam forming or principal component analysis method; S70, performing time-frequency analysis on the fused sound signal to extract sound features, including spectrum, Mel-frequency cepstral coefficients or wavelet coefficients; S80: Associating the extracted sound features with the time information, generating sound feature-time data and outputting the data.

2. The array acoustic signal processing method for dry-type reactor detection according to claim 1, characterized in that: The reactor working sound extraction model adopts a multi-task learning method to form a self-supervised learning framework, including a continuous feature sub-network, a discrete feature sub-network, a time-frequency feature sub-network, an attention sub-network and a fusion sub-network.

3. The array acoustic signal processing method for dry-type reactor detection according to claim 2, characterized in that: The continuous feature subnetwork is used to extract time-continuous sound signal features. The specific structure is a recurrent neural network based on a long short-term memory network, which is used to capture dependencies in long time series and output vectors representing continuous sound features.

4. The array acoustic signal processing method for dry-type reactor detection according to claim 3 is characterized in that: The discrete feature subnetwork is used to extract the features of multiple sound signal segments that are discontinuous in time. The specific structure is a multi-layer one-dimensional convolutional neural network, each layer is followed by batch normalization and ReLU activation function, and finally global average pooling is used to output discrete feature vectors.

5. The array acoustic signal processing method for dry-type reactor detection according to claim 4, characterized in that: The time-frequency feature subnetwork is used to extract time-frequency joint features from the time-frequency graph of the sound signal. The specific structure is a deep convolutional network based on ResNet. In the latter layers, dilated convolution is used to expand the receptive field, and a feature pyramid structure is used to fuse features of different depth layers, and finally a time-frequency feature vector is output.

6. The array acoustic signal processing method for dry-type reactor detection according to claim 5, characterized in that: The attention subnetwork is used to adaptively focus on the most important features of the input signal. The specific structure is a dual attention mechanism including two branches: time attention and frequency attention. The time attention is implemented using a self-attention mechanism, and the frequency attention is implemented using frequency selective convolution.

7. The array acoustic signal processing method for dry-type reactor detection according to claim 6, characterized in that: The fusion subnetwork is used to integrate the features extracted by each subnetwork. The specific structure is to first splice the feature vectors, then fuse the features through a multi-layer fully connected network, and finally output a comprehensive feature representation, which is used for subsequent reactor working sound recognition and fault diagnosis.

8. The array acoustic signal processing method for dry-type reactor detection according to claim 7, characterized in that: The sound sensor array is composed of 8 high-sensitivity condenser microphones.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the array acoustic signal processing method for dry-type reactor detection according to any one of claims 1 to 8.

10. An array acoustic signal processing system for dry-type reactor detection, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Citation Information

Cited By

  • Sound source localization performance improvement method for acoustic wind tunnel test of large civil aircraft

    CN120385975A