A method, medium and system for visual detection of sound signals of a dry-type reactor

Abstract

The application provides a kind of dry reactor sound signal visual detection method, medium and system, belong to dry reactor sound signal detection technical field, include: first, the sound signal of dry reactor is collected, and pretreatment is carried out to eliminate environmental noise.Then the time-frequency analysis is carried out to the sound signal after pretreatment, and the time-frequency spectrum is obtained.Next, adopt the way of bayesian probability inference and adaptive threshold increase, and highlight the small change in time-frequency spectrum, and obtain the increased time-frequency spectrum.Subsequently, energy distribution, peak frequency and harmonic structure are extracted from the increased time-frequency spectrum, and combined into a multi-dimensional feature vector.Apply dimension reduction algorithm, map high-dimensional feature vector to two-dimensional or three-dimensional space, and obtain the second feature vector.Finally, use unsupervised learning algorithm to carry out cluster analysis on the second feature vector, and assign color or label according to the clustering result for different categories, generate visual classification image output.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of dry-type reactor sound signal detection, and in particular relates to a dry-type reactor sound signal visual detection method, medium and system. BACKGROUND

[0002] Dry-type reactors are important passive power compensation devices in power systems, widely used in power transmission and distribution. They absorb or release reactive power to regulate the reactive power balance of the power grid, thereby improving voltage quality and increasing system power factor and transmission efficiency. Compared with traditional oil-immersed reactors, dry-type reactors have the advantages of small size, light weight, no oil, and maintenance-free, so they are the first choice in many cases.

[0003] However, dry-type reactors produce certain sound noise during operation, and these noise signals contain rich information about the operating state of the device, which can be used for fault diagnosis and condition monitoring. Existing diagnostic methods mainly rely on the subjective judgment of professionals on sound signals, which is low in efficiency and reliability. Therefore, there is an urgent need to develop an automatic and quantitative dry-type reactor sound signal analysis method to improve the accuracy and reliability of fault diagnosis.

[0004] In recent years, with the continuous progress of signal processing and pattern recognition technology, some sound signal analysis methods based on time-frequency analysis and machine learning have been widely applied. However, these methods still have limitations in dealing with small signal changes, and cannot fully capture the subtle fault characteristics in dry-type reactor sound signals. SUMMARY

[0005] Therefore, the present application provides a dry-type reactor sound signal visual detection method, medium and system, which can solve the technical problem that existing methods still have limitations in dealing with small signal changes and cannot fully capture the subtle fault characteristics in dry-type reactor sound signals.

[0006] The present application is implemented as follows:

[0007] The first aspect of the present application provides a dry-type reactor sound signal visual detection method, which includes the following steps:

[0008] S10, collect the sound signal of the dry-type reactor and pre-process it to eliminate environmental noise;

[0009] S20, perform time-frequency analysis on the pre-processed sound signal to obtain a time-frequency spectrum;

[0010] S30, use Bayesian probability reasoning and adaptive threshold expansion to highlight the small changes in the time-frequency spectrum to obtain an expanded time-frequency spectrum;

[0011] S40, extracting features of the enhanced time-frequency spectrogram, including energy distribution, peak frequency, harmonic structure;

[0012] S50, combining the extracted features into a multi-dimensional vector, denoted as a first feature vector;

[0013] S60, applying a dimension reduction algorithm to map the first feature vector from a high-dimensional feature vector to a two-dimensional or three-dimensional space to obtain a second feature vector;

[0014] S70, clustering the second feature vector;

[0015] S80, assigning colors or labels to different categories according to the clustering results, generating a visual classification image based on the second feature vector and outputting.

[0016] Wherein, the adaptive threshold enhancement is calculated by using multiple equations, including a time-frequency spectrogram normalization equation, a nonlinear contrast enhancement equation, an adaptive threshold calculation equation, a dynamic enhancement coefficient calculation equation, and a frequency-dependent adjustment equation.

[0017] Further, the time-frequency spectrogram normalization equation is specifically represented as:

[0018]

[0019] In the formula, S norm (t,f) is the normalized time-frequency spectrogram, S(t,f) is the original time-frequency spectrogram; t is time, f is frequency.

[0020] Wherein, the nonlinear contrast enhancement equation is specifically represented as:

[0021]

[0022] In the formula, C(t,f) is the nonlinear contrast enhanced spectrogram, tanh is the hyperbolic tangent function, γ is the contrast enhancement intensity parameter, μ local (t,f) is the weighted average value in the local window centered at (t,f), σ local (t,f) is the weighted standard deviation in the local window centered at (t,f), ∈ is a small positive number to prevent the denominator from being zero, δ is the gradient enhancement parameter, is the gradient of the normalized spectrogram at (t,f).

[0023] Wherein, the adaptive threshold calculation equation is specifically represented as:

[0024] T(t,f)=μ global +k·σ global ·(1-e -α·|C(t,f)|)·(1+lambda*entropy(t,f));

[0025] Wherein, T(t,f) is an adaptive threshold;mu global is a global average of the entire spectrogram;sigma global is a global standard deviation of the entire spectrogram;k is an adjustment parameter, controlling the sensitivity of the threshold;alpha is a non-linear mapping parameter, controlling the rate of change of the threshold with contrast;lambda is an entropy weight parameter;entropy(t,f) is the normalized Shannon entropy of the local window centered at (t,f).

[0026] Wherein, the dynamic expansion coefficient calculation equation is specifically represented as:

[0027]

[0028] Wherein, E(t,f) is a dynamic expansion coefficient;beta is an expansion intensity parameter, controlling the degree of expansion;eta is a sparsity weight parameter;sparsity(t,f) is a sparsity measure of the local window centered at (t,f).

[0029] Wherein, the frequency-dependent adjustment equation is specifically represented as:

[0030]

[0031] Wherein, S enhanced (t,f) is the final expanded time-frequency spectrogram;f max is the maximum frequency;phi is a frequency adjustment parameter.

[0032] Further, the method for clustering the second feature vector is k-means clustering or Gaussian mixture model clustering.

[0033] The second aspect of the application provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions, and the program instructions are used to execute the above-mentioned visual detection method of the dry-type reactor sound signal when running.

[0034] The third aspect of the application provides a visual detection system of the dry-type reactor sound signal, which comprises the above-mentioned computer readable storage medium.

[0035] Compared with the prior art, the visual detection method, medium and system of the dry-type reactor sound signal provided by the application have the following beneficial effects:

[0036] 1. The detection ability of micro-signal changes is enhanced. Through Bayesian probability reasoning and adaptive threshold expansion algorithm, the micro-change features in the time-frequency spectrogram can be effectively highlighted, the fine fault signal capture ability is improved, and the foundation for subsequent feature extraction and fault diagnosis is laid.

[0037] 2. Realize the visualization analysis of multi-dimensional features. The method maps the high-dimensional feature space to two-dimensional or three-dimensional space by using the dimension reduction algorithm, and automatically identifies different signal patterns by using unsupervised clustering, and generates intuitive color classification images. This kind of visual expression helps to more intuitively understand the internal law of the sound signal of the dry-type reactor, and provides convenient fault diagnosis support for the operation and maintenance personnel.

[0038] 3. Good adaptability and scalability. The method adopts modular design, and each processing step has strong flexibility and configurability. Different preprocessing, time-frequency analysis, feature extraction, dimension reduction and clustering algorithms can be selected and optimized according to the actual signal characteristics, so as to meet the diagnosis requirements in different scenes.

[0039] 4. Improve the accuracy and reliability of fault diagnosis. Through effective capture and intuitive visual expression of slight signal changes, the method can more accurately identify the running state of the dry-type reactor, provide reliable basis for subsequent fault prediction and decision support, and greatly improve the accuracy and reliability of diagnosis.

[0040] In summary, the present application solves the technical problem that the existing method still has limitations in processing slight signal changes and cannot fully capture the subtle fault characteristics in the sound signal of the dry-type reactor. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The flowchart of the method provided by the present application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application.

[0043] As shown in the flowchart of the sound signal visualization detection method of the dry-type reactor provided by the present application, the method includes the following steps: Figure 1 S10, collect the sound signal of the dry-type reactor and perform preprocessing to eliminate environmental noise;

[0044] S20, perform time-frequency analysis on the preprocessed sound signal to obtain a time-frequency spectrum;

[0045] S30, use Bayesian probability reasoning and adaptive threshold expansion to highlight the slight changes in the time-frequency spectrum to obtain an expanded time-frequency spectrum;

[0046]

[0047] ​S40, extract features of the enhanced time-frequency spectrogram, including energy distribution, peak frequency, and harmonic structure;

[0048] S50, combine the extracted features into a multi-dimensional vector, denoted as the first feature vector;

[0049] S60, apply a dimension reduction algorithm to map the first feature vector from a high-dimensional feature vector to a two-dimensional or three-dimensional space to obtain a second feature vector;

[0050] S70, cluster the second feature vector;

[0051] S80, assign colors or labels to different categories based on the clustering results, generate a visual classification image based on the second feature vector, and output.

[0052] The specific implementation of the above steps is described in detail as follows:

[0053] The specific implementation of step S10 is as follows: first, collect the sound signal generated by the dry reactor using a high-quality microphone device. In order to eliminate the influence of environmental noise, a pre-processing method combining frequency domain and time domain can be used. In the frequency domain, a band-pass filter can be applied to set an appropriate passband frequency range to remove low-frequency and high-frequency noise components. In the time domain, an adaptive filter can be used to adaptively suppress noise using a reference signal (such as an environmental noise signal), thereby effectively eliminating the pollution of environmental noise on the sound signal. This comprehensive preprocessing method can significantly improve the signal quality and lay a good foundation for subsequent time-frequency analysis.

[0054] The specific implementation of step S20 is as follows: perform time-frequency analysis on the pre-processed sound signal. Common time-frequency analysis methods include short-time Fourier transform (STFT) and wavelet transform. STFT divides the entire sound signal into multiple short time windows, and performs Fourier transform on the signal in each window to obtain a time-frequency spectrogram. Wavelet transform uses wavelet functions of different scales and positions to decompose and reconstruct the signal, which can obtain more detailed time-frequency distribution features. In practical applications, appropriate time-frequency analysis methods and parameter settings can be selected according to the characteristics of the dry reactor sound signal to obtain a clear time-frequency spectrogram.

[0055] The specific implementation of step S30 is as follows: use Bayesian probability reasoning and adaptive threshold enhancement to highlight the small changes in the time-frequency spectrogram. First, use a Bayesian probability model to analyze the time-frequency spectrogram, and infer the possible small signal changes in the time-frequency spectrogram based on prior knowledge and observation data. Then, use an adaptive threshold enhancement algorithm to process the time-frequency spectrogram. The algorithm includes the following sub-steps:

[0056] 1. Time-frequency spectrogram normalization: use The original time spectrum S(t,f) is normalized to obtain the normalized time spectrum S. norm (t,f).

[0057] 2. Non-linear contrast enhancement: Utilizing... Nonlinear contrast enhancement is performed on the normalized time-frequency spectrum, where γ is the contrast enhancement intensity parameter, ranging from 1 to 5, and μ... local (t,f) and σ local (t,f) are the weighted average and weighted standard deviation within the local window, respectively, and ∈ is a small positive number to prevent the denominator from being zero, which is taken as 10. -5 δ is the gradient enhancement parameter, which ranges from 0.1 to 1.

[0058] 3. Adaptive threshold calculation: Using T(t,f)=μ global +k·σ global ·(1-e -α·|C(t,f)| The adaptive threshold T(t,f) is calculated using (1+λ·entropy(t,f)), where μ global and σ global , respectively, are the global average and global standard deviation of the entire time spectrum, k is the adjustment parameter, ranging from 1 to 3, α is the nonlinear mapping parameter, ranging from 0.1 to 1, λ is the entropy weight parameter, ranging from 0.1 to 1, and entropy(t,f) is the normalized entropy of the local window.

[0059] 4. Dynamic expansion: adopts...

[0060]

[0061] E(t,f) is the dynamic expansion coefficient.

[0062] β is an amplification intensity parameter that controls the degree of amplification (typically ranging from 0.5 to 2).

[0063] η is the sparsity weighting parameter (typically ranging from 0.1 to 1).

[0064] sparsity(t,f) is a measure of the sparsity of a local window centered at (t,f).

[0065] 5. Frequency-dependent adjustment: using The final expanded time-frequency spectrum S is obtained. enhanced (t,f), where f max The maximum frequency is φ, which is a frequency adjustment parameter with a value ranging from 0.1 to 0.5.

[0066] It should be noted that:

[0067] 1. S(t,f) is obtained from the time-frequency analysis in step S20.

[0068] 2. μ local (t,f) and σ local (t,f) are calculated by sliding a fixed-size Gaussian weighting window (e.g., 11x11) over the time-frequency spectrogram.

[0069] 3. μ global and σ global are obtained by calculating the mean and standard deviation of the entire time-frequency spectrogram.

[0070] 4. is calculated using the Sobel operator.

[0071] 5. entropy(t,f) is obtained by calculating the probability distribution of pixel values within the local window and then applying the Shannon entropy formula.

[0072] 6. sparsity(t,f) can be measured using the proportion of non-zero elements within the local window.

[0073] 7. γ, δ, k, α, λ, β, η, and φ are hyperparameters that can be optimized using methods such as cross-validation, grid search, or Bayesian optimization.

[0074] These equations introduce advanced techniques such as nonlinear transformations, local features (e.g., entropy and sparsity), gradient information, and frequency-dependent adjustments, which can capture and enhance subtle changes in the time-frequency spectrogram more precisely while considering both local and global characteristics of the signal. This approach can better adapt to different types of dry-type reactor sound signals, improving the sensitivity and reliability of detection.

[0075] This adaptive threshold expansion algorithm can effectively highlight subtle signal changes in the time-frequency spectrogram, laying the foundation for subsequent feature extraction and classification analysis.

[0076] The specific implementation of step S40 is to extract key features from the expanded time-frequency spectrogram S enhanced (t,f), including:

[0077] 1. Energy distribution features: Calculate the energy distribution of the time-frequency spectrogram in the time and frequency dimensions, such as total energy, frequency band energy, instantaneous power, etc. These features can reflect the overall energy characteristics of the signal.

[0078] 2. Peak frequency features: Detect significant peaks in the time-frequency spectrogram and extract their peak frequencies, peak amplitudes, etc. These features can reflect the harmonic characteristics of the signal.

[0079] 3. Harmonic structure features: Analyze the frequency multiplication relationship between each peak frequency in the time-frequency spectrogram, extract relevant indicators of harmonic structure, such as harmonic richness, harmonic distortion, etc. These features can reflect the periodicity characteristics of the signal.

[0080] Through the comprehensive analysis of the above three types of features, the time-frequency characteristics of the dry-type reactor sound signal can be fully described, providing valuable information for subsequent pattern recognition and fault diagnosis.

[0081] The specific implementation of step S50 is: combining the features extracted in step S40, including energy distribution, peak frequency, harmonic structure, etc., into a multi-dimensional feature vector, denoted as the first feature vector This vector can describe the comprehensive time-frequency characteristics of the dry-type reactor sound signal.

[0082] The specific implementation of step S60 is: applying a dimension reduction algorithm to map the high-dimensional first feature vector from the original high-dimensional space to a two-dimensional or three-dimensional space, obtaining the second feature vector after dimension reduction. Common dimension reduction algorithms include principal component analysis (PCA), linear discriminant analysis (LDA), t-SNE, etc. Through dimension reduction processing, the original high-dimensional feature space can be projected into a low-dimensional space while preserving the main relationships between the original features as much as possible, providing suitable input data for subsequent clustering analysis.

[0083] The specific implementation of step S70 is: performing clustering analysis on the second feature vector after dimension reduction. Common clustering algorithms include k-means clustering, Gaussian mixture model clustering, spectral clustering, etc. Through clustering analysis, different patterns or states contained in the dry-type reactor sound signal can be automatically identified, providing a basis for subsequent visual classification. The specific parameter settings of the clustering algorithm, such as the number of cluster centers K, distance metric, convergence condition, etc., need to be adjusted and optimized according to the characteristics of the actual problem.

[0084] The specific implementation of step S80 is: according to the clustering results of step S70, assigning distinct colors or labels to different categories, and generating a visual classification image based on the second feature vector . The specific approach is as follows:

[0085] 1. Distance calculation: for each point (x, y) in the second feature vector space, calculate its Euclidean distance to each cluster center where D is the dimension (2 or 3) of the second feature vector; d i (x, y) is the Euclidean distance from the point (x, y) in the second feature vector space to the i-th cluster center; v j is the value of the point (x, y) in the j-th dimension; cij value of the i-th cluster center in the j-th dimension.

[0086] 2. Soft classification weight calculation: adopt to calculate the soft classification weight w i (x,y) of point (x,y) belonging to the i-th cluster. i (x,y) is the soft classification weight of point (x,y) belonging to the i-th cluster; K is the total number of clusters; σ is the bandwidth parameter of the Gaussian kernel, which controls the smoothing degree of soft classification.

[0087] 3. Color assignment: adopt to calculate the final color value C(x,y) of point (x,y). where is the representative color of the i-th cluster (predefined or dynamically generated according to the cluster features), and C(x,y) is the final color value (RGB or HSV format) of point (x,y).

[0088] 4. Contrast enhancement: adopt to perform contrast enhancement on the color value, where α is the contrast enhancement factor (usually in the range of 1 to 5), β is the contrast adjustment threshold (usually 0.5), and C enhanced (x,y) is the color value after enhancing the contrast.

[0089] 5. Label generation: adopt to assign the final cluster label L(x,y) to point (x,y), i.e., the index of the cluster with the largest weight.

[0090] It should be noted that

[0091] 1. v j and c ij come from the dimension reduction result of S60 step and the clustering result of S70 step.

[0092] 2. σ can be selected by cross-validation, or set to a certain proportion (such as 1 / 3) of the average distance between clusters.

[0093] 3. K is the number of clusters determined in S70 step.

[0094] 4. C i A set of contrasting colors can be predefined, or dynamically generated according to the features of each cluster.

[0095] 5. α and β can be adjusted by experiment to obtain the best visualization effect.

[0096] The steps of generating the final visualization classification image can be simply described as follows:

[0097] 1. For each point (x, y) in the second feature vector space, calculate its distance d to each cluster center i (x, y).

[0098] 2. Calculate the weight w of each point belonging to each cluster using the soft classification weight formula i (x, y).

[0099] 3. Calculate the color value C(x, y) of each point according to the weight and the predefined cluster color.

[0100] 4. Apply the contrast enhancement formula to get C enhanced (x, y).

[0101] 5. Assign a final cluster label L(x, y) to each point using the label generation formula.

[0102] 6. Map the calculated color value and label to a two-dimensional or three-dimensional image to generate the final visual classification image.

[0103] This method can generate a smooth transition color visualization image while retaining clear class boundaries, helping to more intuitively display the classification results of the dry-type reactor sound signal.

[0104] Through the above steps, a smooth transition color visualization classification image can be generated, while retaining clear class boundaries, helping to intuitively display the classification results of the dry-type reactor sound signal.

[0105] The second aspect of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions, and the program instructions are used to execute the above-mentioned dry-type reactor sound signal visualization detection method when running.

[0106] The third aspect of the present application provides a dry-type reactor sound signal visualization detection system, which comprises the above-mentioned computer readable storage medium.

[0107] Specifically, the principle of the present application is:

[0108] 1. Time-frequency analysis and micro change enhancement

[0109] Dry-type reactors produce relatively complex sound signals during operation, which contain rich information about the state of the device. By performing time-frequency analysis on the sound signal, such as short-time Fourier transform (STFT) or wavelet transform, the time-frequency characteristic distribution can be obtained. However, since device faults often cause subtle sound changes, it is difficult to accurately capture these changes by directly observing the time-frequency spectrum.

[0110] To this end, the application adopts Bayesian probability reasoning and adaptive threshold expansion technology to effectively highlight the subtle change characteristics in the time-frequency spectrogram. First, the Bayesian probability model is used to analyze the implicit change patterns that may exist in the time-frequency spectrogram based on prior knowledge and observation data. Then, through steps such as normalization, nonlinear contrast enhancement, adaptive threshold calculation, dynamic expansion, and frequency-dependent adjustment, the weak signal changes in the time-frequency spectrogram are fully amplified to facilitate subsequent feature extraction and fault diagnosis.

[0111] 2. Dimensionality reduction and clustering analysis of multi-dimensional features

[0112] From the enhanced time-frequency spectrogram, key features such as energy distribution, peak frequency, and harmonic structure are extracted to form a high-dimensional feature vector. Direct fault pattern recognition in high-dimensional space is difficult, so dimensionality reduction algorithms such as principal component analysis (PCA) and linear discriminant analysis (LDA) are used to map the high-dimensional feature vector to a two-dimensional or three-dimensional space, preserving the main relationships between the original features.

[0113] Based on the low-dimensional feature space after dimensionality reduction, the application uses unsupervised clustering algorithms such as k-means clustering and Gaussian mixture model clustering to automatically identify different patterns or states contained in the dry-type reactor sound signal. This unsupervised analysis method can effectively reveal the internal structural characteristics of the signal data, providing valuable clues for subsequent fault diagnosis.

[0114] 3. Visualization classification based on feature clustering

[0115] The low-dimensional feature space after clustering is mapped to a two-dimensional or three-dimensional visualization classification image, providing intuitive fault diagnosis support for operating and maintenance personnel. Specifically, the Euclidean distance of each feature point to the cluster center is calculated, and a Gaussian kernel function is used to convert it to a soft classification weight. Then, different clusters are assigned distinct colors according to the weight distribution, and a contrast enhancement algorithm is applied to generate the final visualization classification image.

[0116] This visualization method based on feature clustering can fully reveal the distribution characteristics of different fault patterns in the low-dimensional feature space, allowing maintenance personnel to intuitively identify and diagnose the operating state of the equipment. Compared with traditional single signal analysis, this method provides more comprehensive and interpretable fault diagnosis basis, helping to improve the accuracy and reliability of diagnosis.

[0117] In summary, the application integrates advanced digital signal processing and pattern recognition technologies such as time-frequency analysis, feature engineering, dimensionality reduction, and clustering, and skillfully handles the detection of subtle changes in dry-type reactor sound signals, and presents the results of fault diagnosis through visualization.

[0118] For a better understanding and implementation of the present application, an embodiment of a specific application scenario of the present application is provided below: A certain power company owns a batch of important dry-type reactor equipment, which is used to cooperate with the transformer to adjust the reactive power balance of the power grid. In order to ensure the stable operation of these devices, the company hopes to establish an automated sound signal detection system to timely discover potential equipment failure risks. After investigation, the company decides to use the dry-type reactor sound signal visualization detection method proposed by the present application to analyze and diagnose the collected sound data on site.

[0119] Firstly, the company installs a high-quality microphone acquisition device beside a key dry-type reactor, which continuously records the sound signals generated during the operation of the device. The collected raw sound data is preprocessed to remove the influence of environmental noise. The preprocessed sound signal is shown in Table 1.

[0120] Time Sound signal 0.0s 0.01 0.1s 0.02 0.2s 0.03 ... ... 10.0s 0.15

[0121] Table 1 Waveform of the preprocessed sound signal of the dry-type reactor

[0122] Next, the company performs time-frequency analysis on the preprocessed sound signal and uses the short-time Fourier transform (STFT) method to obtain the time-frequency spectrogram. As shown in Table 2, the time-frequency spectrogram shows the energy distribution characteristics of the dry-type reactor at different times and frequency ranges.

[0123] Frequency / Hz 0s 1s 2s ... 10s 0 1 2 3 ... 4 100 2 3 4 ... 5 200 3 4 5 ... 6 ... .. .. .. ... ... 1000 4 5 6 ... 7

[0124] Table 2 Time-frequency spectrogram of the sound signal of the dry-type reactor

[0125] By carefully observing the time-frequency spectrogram in Table 2, it can be found that there are some small change areas in it, which may be related to potential equipment failure risks. In order to highlight these small change characteristics, the company uses the Bayesian probability reasoning and adaptive threshold expansion algorithm proposed by the present application for processing.

[0126] Firstly, the Bayesian probability model is used to analyze the time-frequency spectrogram, and according to prior knowledge and observation data, the possible small signal change area is inferred. Then, the adaptive threshold expansion algorithm is applied to process the time-frequency spectrogram, and the specific steps are as follows:

[0127] 1. Time-frequency spectrogram normalization: adopt The original time-frequency spectrogram S(t,f) is normalized to obtain the normalized time-frequency spectrogram S norm (t,f). Wherein, the company measures min(log(1+S))=0.1, max(log(1+S))=2.5.

[0128] 2. Nonlinear contrast enhancement: adopt The normalized time-frequency spectrum is subjected to nonlinear contrast enhancement to obtain C(t,f). Through testing, the company selects γ = 3 and δ = 0.5.

[0129] 3. Adaptive threshold calculation: The adaptive threshold T(t,f) is calculated using T(t,f) = 0.8 + 2 · 0.3 · (1 - e -0.5·|C(t,f)| ) · (1 + 0.5 · entropy(t,f)), where μ global = 0.8, σ global = 0.3, k = 2, α = 0.5, and λ = 0.5. The normalized entropy entropy(t,f) of the local window is measured to be between 0.3 and 0.7.

[0130] 4. Dynamic expansion: The dynamic expansion coefficient E(t,f) is calculated using

[0131] E(t,f) = 1 + 0.5 · (1 - e ) · (1 + 0.5 · entropy(t,f))

[0132] β is the expansion intensity parameter, which controls the degree of expansion (usually the value range is 0.5 to 2)

[0133] η is the sparsity weight parameter (usually the value range is 0.1 to 1)

[0134] sparsity(t,f) is the sparsity measure of the local window centered at (t,f)

[0135] Measured to be between 0.2 and 0.8, β = 1.2, and η = 0.5.

[0136] 5. Frequency-dependent adjustment: The frequency-dependent adjustment is performed using to obtain the final expanded time-frequency spectrum S enhanced (t,f), where f max = 1000 Hz and φ = 0.3.

[0137] Through the adaptive threshold expansion algorithm of the above 5 steps, the company successfully highlights the small signal change characteristics in the time-frequency spectrum, as shown in Table 3. Compared with the original time-frequency spectrum, some changes that were not obvious in the enhanced time-frequency spectrum become more prominent.

[0138] Frequency / Hz 0s 1s 2s ... 10s 0 1 2 3 ... 4 100 2 3 4 ... 5 200 3 4 5 ... 6 ... .. .. .. ... ... 1000 4 5 6 ... 7

[0139] Table 3 Enhanced time-frequency spectrum of dry reactor sound signal

[0140] Next, the company extracts the key features of the enhanced time-frequency spectrum, including:

[0141] 1. Energy distribution feature: Calculate the energy distribution of the spectrogram in time and frequency dimensions, such as total energy E total =∑ t,f S enhanced (t,f)=3450, band energy, instantaneous power P(t)=∑ f S enhanced (t,f)。

[0142] 2. Peak frequency feature: Detect the significant peak in the spectrogram, extract its peak frequency f peak =320Hz, peak amplitude A peak =6.2.

[0143] 3. Harmonic structure feature: Analyze the frequency multiplication relationship between each peak frequency in the spectrogram, extract the harmonic richness H R =0.78, harmonic distortion H D =0.22.

[0144] The above 3 types of features are combined into a 12-dimensional first feature vector

[0145] After that, the company uses principal component analysis (PCA) algorithm to map the high-dimensional first feature vector from the original 12-dimensional space to a two-dimensional feature space as shown in Table 4. This dimensionality reduction operation can effectively remove the redundancy between features and highlight the main pattern differences.

[0146] u1 u2 2.1 1.5 2.3 1.7 2.1 1.6 ... ... 2.5 1.9

[0147] Table 4 Two-dimensional feature space distribution of dry reactor sound signal

[0148] In the two-dimensional feature space, the company uses k-means clustering algorithm to perform unsupervised classification on the sample points, and finds that they can be divided into 3 different patterns. The cluster center coordinates are respectively

[0149] According to the clustering results, the company assigns red, green and blue representative colors to the 3 patterns respectively. Then, the soft classification weight calculation formula is used where assigns the probability weight of belonging to 3 clusters to each feature point (u1, u2).

[0150] Finally, the company applies the contrast enhancement formula and maps the calculated color values to the two-dimensional plane to generate a visual classification image.

[0151] The red area corresponds to a normal operating state, the green area may have some minor fault hidden dangers, and the blue area indicates that the device has a relatively serious fault state.

[0152] By continuously monitoring this visual classification image, the company can timely discover the possible fault signs of the dry-type reactor, and provide an important basis for subsequent maintenance. Compared with the traditional single signal analysis method, the visual diagnosis result provided by the application is more intuitive and comprehensive, which helps to improve the accuracy and reliability of fault diagnosis, and contributes new technical support for intelligent management of power equipment.

[0153] Overall, this specific embodiment fully demonstrates the application value of the application in the sound signal analysis of the dry-type reactor. Through a series of innovative technical means such as time-frequency analysis, feature extraction, dimension reduction and clustering visualization, the company successfully extracts and presents the minor fault features in the sound signal of the dry-type reactor, which provides effective support for the state monitoring and preventive maintenance of the equipment.

[0154] The above is only a specific embodiment of the application, but the protection scope of the application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A method for visual detection of sound signals of dry-type reactors, characterized in that, The method comprises the following steps: S10, collecting a sound signal of a dry reactor and pre-processing the sound signal to eliminate environmental noise; S20, performing time-frequency analysis on the pre-processed sound signal to obtain a time-frequency spectrogram; S30, highlighting a slight change in the time-frequency spectrogram by using Bayesian probability reasoning and an adaptive threshold enhancement method to obtain an enhanced time-frequency spectrogram; S40, extracting features of the enhanced time-frequency spectrogram, including energy distribution, peak frequency, and harmonic structure; S50, combining the extracted features into a multi-dimensional vector, denoted as a first feature vector; S60, applying a dimension reduction algorithm to map the first feature vector from a high-dimensional feature vector to a two-dimensional or three-dimensional space to obtain a second feature vector; S70, clustering the second feature vector; S80, assigning colors or labels to different categories according to the clustering results, generating a visual classification image based on the second feature vector, and outputting the classification image.

2. A method of visualizing detection of sound signals of a dry reactor according to claim 1, characterized in that, The adaptive threshold enhancement method uses multiple equations for calculation, including a time-frequency spectrogram normalization equation, a nonlinear contrast enhancement equation, an adaptive threshold calculation equation, a dynamic enhancement coefficient calculation equation, and a frequency-dependent adjustment equation.

3. A method of visualizing detection of sound signals of a dry reactor according to claim 2, characterized in that, The time-frequency spectrogram normalization equation is specifically represented as: In the formula, S norm (t,f) is the normalized time-frequency spectrogram, S(t,f) is the original time-frequency spectrogram; t is time, and f is frequency.

4. A method of visualizing detection of sound signals of a dry reactor according to claim 3, characterized in that, The nonlinear contrast enhancement equation is specifically represented as: where C(t,f) is the nonlinear contrast enhanced spectrogram, tanh is the hyperbolic tangent function, γ is the contrast enhancement intensity parameter, μ local (t,f) is the weighted average value within the local window centered at (t,f), σ local (t,f) is the weighted standard deviation within the local window centered at (t,f), ∈ is a small positive number to prevent the denominator from being zero, δ is the gradient enhancement parameter, is the gradient of the normalized spectrogram at (t,f).

5. A method of visualizing detection of sound signals of a dry reactor according to claim 4, characterized in that, The adaptive threshold calculation equation is specifically represented as: T(t,f) = μ global + k · σ global · (1 - e -α·|C(t,f)| ) · (1 + λ · entropy(t,f)) ; where T(t,f) is the adaptive threshold; μ global is the global mean of the whole spectrum; σ global is the global standard deviation of the whole spectrum; k is a tuning parameter that controls the sensitivity of the threshold; a is a non-linear mapping parameter that controls the rate at which the threshold varies with contrast; l is an entropy weight parameter; and entropy(t,f) is the normalized Shannon entropy of the local window centered at (t,f).

6. A method of visualizing detection of sound signals of a dry reactor according to claim 5, characterized in that, The dynamic enhancement coefficient calculation equation is specifically represented as: In the equation, E(t,f) is the dynamic enhancement coefficient; β is the enhancement intensity parameter, which controls the degree of enhancement; η is the sparsity weight parameter; and sparsity(t,f) is the sparsity measure of the local window centered at (t,f).

7. A method of visualizing detection of sound signals of a dry reactor according to claim 6, characterized in that, The frequency-dependent adjustment equation is specifically represented as: In the formula, S enhanced (t,f) is the final expanded time-frequency spectrum; f max is the maximum frequency; and φ is a frequency adjustment parameter.

8. A method of visualizing detection of sound signals of a dry reactor according to claim 7, characterized in that, The method for clustering the second feature vector is k-means clustering or Gaussian mixture model clustering.

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

10. A visual detection system of sound signals of a dry-type reactor, characterized by, The computer-readable storage medium comprises the computer-readable storage medium according to claim 9.

Images