Converter station noise map construction method, medium and system

Abstract

The invention provides a converter station noise map construction method, medium and system, and belongs to the technical field of converter station noise monitoring, and the method comprises the steps: dividing a converter station plan into a series of triangular regions; acquiring a noise signal set acquired by three noise sensors arranged at the three top ends of each triangular area, and preprocessing to obtain a preprocessed noise signal set; calculating noise features of the first region and noise source parameters corresponding to the noise features; clustering the first region noise features at each detection moment; performing noise source parameter marking on each typical noise feature to obtain a most matched noise source parameter; training a noise recognition model by adopting machine learning; acquiring all noises detected at continuous detection moments of the converter station, and inputting the noises into the model for identification to obtain corresponding noise source parameters; and mapping noise source parameters into RGB colors, and marking on the converter station plane graph to obtain a converter station noise distribution map changing along with time.

Description

Technical Field

[0001] The present invention belongs to the technical field of converter station noise monitoring. Specifically, it relates to a method, medium and system for constructing a noise map of a converter station. Background Art

[0002] With the continuous expansion of the power system capacity, as an important power system equipment, the operation and maintenance management difficulty of the converter station is also increasing day by day. The traditional manual inspection method is inefficient and cannot meet the requirements of the status monitoring of large-scale equipment. Therefore, the use of automatic monitoring technology to realize the equipment status evaluation has become an important research topic. At present, the use of acoustic monitoring technology to realize the power equipment status evaluation has been widely applied. This technology monitors the working state by analyzing the sound or vibration generated during the operation of the equipment, and has the advantages of simple installation and low cost.

[0003] In the application of converter station acoustic monitoring, a key problem is how to accurately obtain the acoustic information of the equipment. The traditional method mainly uses a single-point noise sensor to obtain the sound signal. However, this method cannot obtain the sound propagation information in the spatial dimension, and the accuracy of equipment fault location is poor. To solve this problem, some methods use multiple sensors to obtain the sound signal, but the large number of equipment and noise sensors makes it difficult to build the acquisition system. That is, in the prior art, in the process of realizing the establishment of a noise map by using multiple sensors to obtain the converter station noise signal, there are technical problems that the deployment position of the noise sensor depends on manual experience judgment, and high-precision noise source positioning and noise source characteristics cannot be obtained, resulting in an inaccurate noise map. Summary of the Invention

[0004] In view of this, the present invention provides a method, medium and system for constructing a noise map of a converter station, which can solve the technical problems in the prior art that in the process of realizing the establishment of a noise map by using multiple sensors to obtain the converter station noise signal, the deployment position of the noise sensor depends on manual experience judgment, and high-precision noise source positioning and noise source characteristics cannot be obtained, resulting in an inaccurate noise map.

[0005] The present invention is implemented as follows:

[0006] The first aspect of the present invention provides a method for constructing a noise map of a converter station, which includes the following steps:

[0007] S10. Divide the converter station floor plan into a series of triangular regions; obtain the noise signal sets collected by three noise sensors arranged at the three vertices of each triangular region, and after preprocessing the noise, obtain a preprocessed noise signal set; here, the noise sensor can be a pick-up or a microphone;

[0008] S20. For any triangular region, based on the preprocessed noise signal set corresponding to the triangular region, use the sound source localization method to determine the sound source position of the noise, and calculate the actual decibel level of the noise based on the sound source position and the detected decibel level, so as to form the first region noise characteristics and their corresponding noise source parameters. The first region noise characteristics include the noise decibel level and the time-frequency characteristics of the noise, and the noise source parameters include the sound source position of the noise, the sound source device, and the operating state of the sound source device;

[0009] S30. Cluster the first region noise characteristics at each detection moment of all triangular regions of the entire converter station to obtain multiple cluster centers, representing multiple typical noise characteristics;

[0010] S40. Mark the noise source parameters for each typical noise characteristic to obtain the most matching noise source parameters;

[0011] S50. Use the multiple typical noise characteristics and their most matching noise source parameters as training data to train a machine learning model to obtain a noise recognition model for determining the noise source parameters according to the noise characteristics;

[0012] S60. Obtain all the noises detected at consecutive detection moments of the converter station, input them into the noise recognition model for recognition, and obtain the corresponding noise source parameters;

[0013] S70. Map the first region noise characteristics corresponding to all the noises detected at consecutive detection moments of the converter station into RGB colors, and mark the noise source parameters on the floor plan of the converter station to obtain a converter station noise distribution map that changes with time.

[0014] Based on the above technical solutions, a method for constructing a noise map of a converter station according to the present invention can be further improved as follows:

[0015] Specifically, it further includes step S25, which is executed after step S20 is completed. Specifically: optimize the triangular region division of the floor plan of the converter station for the obtained preprocessed noise signal set or the first region noise characteristics and their corresponding noise source parameters, and after repeating the execution of steps S10 and S20, continue to execute the remaining steps.

[0016] The beneficial effect of adopting the above improvement scheme is that step S25 is used to ensure that the shape and size of the triangular region can better reflect the actual noise distribution while detecting the noise.

[0017] Among them, the step of dividing the converter station plan into a series of triangular regions is specifically as follows: First, obtain the outline of the converter station according to the blueprint or plan of the converter station; then use the Delaunay triangulation algorithm to triangulate the outline of the converter station, and divide the converter station into multiple triangular regions.

[0018] Among them, the step of determining the sound source position of the noise by the sound source localization method is specifically as follows: For each of the triangular regions, extract the preprocessed noise signals of the corresponding three sensors to form the preprocessed noise signal set of the triangular region, and then use the sound source localization algorithm of the advanced time difference to locate the spatial coordinates of the noise according to the signal differences of the three sensors, that is, the sound source position of the noise.

[0019] Among them, in the step of clustering the first region noise characteristics at each detection moment of all triangular regions of the entire converter station, the algorithm used for clustering is the K-Means algorithm. The specific clustering steps are as follows: First, randomly select multiple clustering centers, and then iteratively assign each first region noise characteristic to the clustering center closest to it and update the clustering center until the center no longer changes; finally, obtain multiple clustering centers to represent multiple typical noise characteristics.

[0020] Among them, the machine learning model uses LSTM as the framework of the noise recognition model.

[0021] Further, the step S25 specifically includes:

[0022] First, set the segmentation threshold according to the standard deviation of the noise intensity of each triangular region. The larger the standard deviation, the more uneven the noise distribution, and further segmentation is required;

[0023] Then, for the triangular regions with a standard deviation greater than the threshold, find the two sides with the largest noise intensity difference, and divide the triangular region along the midpoint connection line of the two sides with the largest difference;

[0024] Repeat the above segmentation process until the standard deviation of the noise intensity of all triangular regions is less than the threshold, indicating that the noise in the region is uniform.

[0025] Further, the segmentation threshold is set to 0.618 times the standard deviation of the noise intensity of each triangular region.

[0026] The second aspect of the present invention provides a computer-readable storage medium, wherein program instructions are stored in the computer-readable storage medium, and when the program instructions run, they are used to execute the above-mentioned method for constructing a converter station noise map.

[0027] The third aspect of the present invention provides a converter station noise map construction system, which includes the above-mentioned computer-readable storage medium.

[0028] Compared with the prior art, the beneficial effects of a method, medium and system for constructing a noise map of a converter station provided by the present invention are as follows: By using the method of triangulation and multi-sensor sound collection in the converter station area, the present invention realizes high-precision sound source localization. Triangulation ensures uniform sound distribution in each sub-region, avoiding monitoring blind spots. The deployment positions of multiple noise sensors are determined according to triangulation, improving the quality of obtaining spatial information of the sound source. In addition, the triangular region can be optimized and adjusted according to the obtained noise signals, so that the shape and size of the triangular region better conform to the actual noise distribution. In addition, by constructing a noise source feature database and using intelligent machine learning algorithms for noise pattern matching and classification, the present invention greatly reduces the dependence on expert experience and realizes automatic recognition and analysis of acoustic signals. Compared with the traditional manual analysis method, the automatic recognition process of the present invention is faster and more efficient, and the recognition performance will continuously improve with the accumulation of sample data, greatly improving the intelligence level of the entire acoustic monitoring system.

[0029] In summary, the technical solution of the present invention solves the technical problem that in the process of using multiple sensors to obtain noise signals of a converter station and establish a noise map in the prior art, the deployment positions of noise sensors rely on manual experience judgment, and high-precision noise source localization and noise source characteristics cannot be obtained, resulting in an inaccurate noise map. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments of the present invention. Obviously, the drawings in the following description are only the second embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0031] Figure 1 is a flowchart of the method provided by the present invention;

[0032] Figure 2 is a schematic diagram of cutting a triangular region on the plan view of the converter station provided by the present invention;

[0033] Figure 3 is a schematic flowchart of using the DBSCAN clustering algorithm in the method of the present invention;

[0034] Figure 4 is a schematic diagram of using the elbow principle to determine the optimal number of clusters K in the method of the present invention;

[0035] Figure 5 is a schematic diagram of the implementation of Few-Shot Learning in the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0037] As Figure 1 shown, it is a flowchart of a method for constructing a noise map of a converter station provided by the first aspect of the present invention. The method includes the following steps:

[0038] S10. Divide the floor plan of the converter station into a series of triangular regions; obtain the noise signal sets collected by three noise sensors arranged at the three vertices of each triangular region, and after preprocessing the noise, obtain a preprocessed noise signal set;

[0039] S20. For any triangular region, based on the preprocessed noise signal set corresponding to the triangular region, use the sound source localization method to determine the sound source position of the noise, and calculate the actual decibel number of the noise based on the sound source position and the detected decibel number, forming the first region noise characteristics and their corresponding noise source parameters. The first region noise characteristics include the noise decibel number and the time-frequency characteristics of the noise, and the noise source parameters include the sound source position, sound source equipment, and operating state of the sound source equipment of the noise;

[0040] S30. Cluster the first region noise characteristics at each detection moment of all triangular regions of the entire converter station to obtain multiple cluster centers, representing multiple typical noise characteristics; where the detection moment is defined as detecting once every certain time interval. Generally, the time interval is set to 1 - 10 minutes;

[0041] S40. Mark the noise source parameters for each typical noise characteristic to obtain the most matching noise source parameters;

[0042] S50. Use multiple typical noise characteristics and their most matching noise source parameters as training data to train a machine learning model to obtain a noise recognition model for determining noise source parameters based on noise characteristics;

[0043] S60. Obtain all the noises detected at consecutive detection moments of the converter station, input them into the noise recognition model for recognition, and obtain the corresponding noise source parameters;

[0044] S70. Map the first region noise characteristics corresponding to all the noises detected at consecutive detection moments of the converter station to RGB colors, and mark the noise source parameters on the floor plan of the converter station to obtain a converter station noise distribution map that changes with time.

[0045] Among them, in the above technical solution, it further includes step S25, which is executed after step S20 is completed. Specifically: optimize the triangular region division of the converter station floor plan based on the obtained preprocessed noise signal set or the noise characteristics of the first region and their corresponding noise source parameters, and after repeating the execution of steps S10 and S20, continue to execute the remaining steps.

[0046] The following is a description of the specific implementation of the first embodiment of the present invention:

[0047] Specific implementation of step S10:

[0048] 1) First, obtain the blueprint or floor plan of the converter station, preprocess the blueprint, including image denoising, edge extraction, etc., to obtain the outline of the converter station. Then, use the Delaunay triangulation algorithm to triangulate the outline of the converter station, dividing the converter station into multiple triangular regions. The principle of the Delaunay triangulation algorithm is to ensure that no other points are contained inside the region, thereby generating triangles as close to equilateral shapes as possible. Compared with other triangulation methods, Delaunay triangulation can produce a more natural and reasonable triangular division result.

[0049] 2) After triangulation, calculate the centroid of each triangular region, and virtually set noise sensors at the three vertices of each triangular region for subsequent noise signal acquisition. The purpose of setting three sensors is to collect the spatial noise information of each triangular region. Compared with setting only one sensor, setting three sensors can more accurately reflect the noise distribution of each triangular region. The specific division is as Figure 2 shown.

[0050] In summary, step S10 uses image processing and the Delaunay triangulation algorithm to divide the converter station floor plan into multiple triangular regions, and sets three virtual noise sensors in each triangular region, laying a foundation for subsequent noise signal acquisition. The purpose is to accurately capture the noise information of each region of the converter station.

[0051] Specific implementation of step S20:

[0052] 1) After completing the triangular region division and setting virtual noise sensors, first collect the noise signals detected by each sensor at different times and perform preprocessing. The preprocessing includes noise reduction, filtering, etc. to remove the random errors in the noise signals and obtain clean preprocessed noise signals.

[0053] 2) For each triangular region, extract the preprocessed noise signals of the corresponding three sensors to form the preprocessed noise signal set of this triangular region. Then, use the sound source localization algorithm based on the advanced time difference to locate the spatial coordinates of the noise according to the signal differences of the three sensors, that is, the sound source position of the noise. The principle of this algorithm is to calculate the position of the sound source by using the trilateral intersection method according to the sound propagation speed and the time difference of arrival at each sensor.

[0054] 3) Next, according to the located sound source position and the noise intensities collected by the three sensors, the actual noise intensity at this position can be calculated, that is, the decibel number of the noise. The time-frequency characteristics of the noise can also be extracted from the frequency domain analysis.

[0055] 4) Finally, according to the located sound source position and the equipment distribution, the noise source equipment and its working state can be determined to form the noise source parameters.

[0056] Thus, the first region noise characteristics and the corresponding noise source parameters can be generated for each triangular region. Step S20 uses the sound source localization technology to automatically analyze the noise information of each region, laying a foundation for map construction.

[0057] In the specific implementation manner of step S20, the relevant formulas involved are described as follows:

[0058] Let the vertex positions of the triangular region be A, B, and C, and the signals received by the corresponding sensors be s A , s B , s C .

[0059] First, preprocess the signal to remove noise:

[0060] s′ A = f(s A )

[0061] s′ B = f(s B )

[0062] s′ C = f(s c )

[0063] Among them, f is the preprocessing function, and noise reduction can be achieved by methods such as wavelet transform. s′ A , s′ B , s′ C represents the signal after preprocessing and noise reduction of the signal.

[0064] Then, select one signal s′ A as the reference signal and calculate the correlation coefficients of the other two signals with the reference signal:

[0065] R AB = corr(s′ A , s′ B )

[0066] R AC = corr(s′ A , s′ C )

[0067] Determine the signal s′ with the greatest correlation B . Based on the sound propagation speed v and the time difference Δt AB , Δt AC between each signal, the sound source position can be calculated as follows:

[0068]

[0069]

[0070] By solving the above system of equations, the coordinates of the sound source can be determined

[0071] Then, according to the signal intensity P A , P B , P C received by the sensor and the distance d A , d B , d C from the sound source, the sound source intensity is calculated using the propagation loss model:

[0072]

[0073] where k and α are model parameters, with the default k = 20 and α = 2, generally speaking. Similarly, P S can be calculated

[0074] Among them, in the sound propagation loss model, the parameters k and α need to be determined according to the actual situation

[0075] For outdoor sound propagation, the following empirical values are generally used:

[0076] The value of k is 20. This is a commonly used propagation loss coefficient in the empirical rule

[0077] The value of α is 2. This conforms to the distance - squared attenuation law in theory

[0078] That is, in the outdoor environment, the empirical parameters commonly used in the propagation loss model are:

[0079]

[0080] Of course, to further improve the model accuracy, problem data can also be established through multi-point acquisition to determine the parameters k and α. For example, the least squares method can be used for fitting to find the parameter combination that makes the model calculation results best match the measured values.

[0081] For complex environments such as indoors, sound will have multiple reflections and scatterings, and the propagation loss pattern is relatively complex. At this time, the parameters k and α need to be determined through a large number of experiments. Generally, the value of α will be taken as a value less than 2.

[0082] To sum up, for the empirical parameter values of the propagation loss model in the outdoor environment: k = 20, α = 2. For complex environments, the parameters of the model need to be determined through experiments.

[0083] Finally, the spectral characteristics of the noise are extracted through methods such as Fourier transform to form the first area noise characteristics of this area.

[0084] Specific implementation method of step S25:

[0085] After the execution of step S20, the purpose of this step is to optimize and adjust the triangular area according to the obtained noise signal, so that the shape and size of the triangular area better conform to the actual noise distribution.

[0086] First, the segmentation threshold can be set according to the standard deviation of the noise intensity of each triangular area. The larger the standard deviation, the more uneven the noise distribution, and further segmentation is required.

[0087] Then, for the triangular areas with a standard deviation greater than the threshold, find the two sides with the largest difference in noise intensity, and connect the midpoints along this side to divide the triangular area.

[0088] Repeat this segmentation process until the standard deviation of the noise intensity of all triangular areas is less than the threshold, indicating that the noise within the area is uniform.

[0089] In addition, the area size can also be optimized according to the time-frequency characteristics of the noise. For example, a larger triangular area can be set for low-frequency stable noise.

[0090] By adaptively adjusting the shape and size of the triangular area through noise information, the subsequent noise acquisition and map construction can be made more accurate and efficient.

[0091] The method for setting the segmentation threshold in step S25 can be as follows:

[0092] Suppose the noise intensities detected by n sensors in the triangular area are ω1, ω2,..., ω n , then the calculation of the standard deviation of the noise intensity in this area is:

[0093]

[0094] wherein, is the mean value of the noise intensity.

[0095] For the standard deviation σ, a reference standard deviation σ0 can be set. When σ is greater than σ0, it is considered that the noise distribution is uneven and needs to be segmented. Here, the reference standard deviation is defaulted to 0.618σ, and it can also be set according to empirical values or optimized through multiple tests. The specific setting idea is as follows:

[0096] 1. Collect the noise distribution samples of multiple triangular regions, manually judge whether the noise distribution is uniform, and label "uniform" or "non-uniform".

[0097] 2. Calculate the standard deviation σ of these sample regions.

[0098] 3. Try different values of σ0 and judge the partitioning effect. If the current σ0 partitions the region originally labeled as "uniform", or fails to partition the "non-uniform" region, then σ0 needs to be adjusted.

[0099] 4. Repeat the above process to find the best value of σ0 that makes the partitioning result consistent with the manual annotation.

[0100] Specific implementation of step S30:

[0101] After obtaining the first-region noise characteristics of all triangular regions at each moment, in order to summarize the typical noise patterns of the converter station, this step uses the clustering analysis method to cluster the noise characteristics of all triangular regions.

[0102] First, define a suitable distance metric, which can adopt the Euclidean distance or the Mahalanobis distance, to measure the difference between two noise characteristics.

[0103] Then, select a clustering algorithm, such as the K-Means, hierarchical clustering and other algorithms to cluster the noise characteristics. Among them, the K-Means algorithm first randomly selects K clustering centers, and then iteratively assigns each noise characteristic to the clustering center closest to it and updates the clustering center until the center no longer changes.

[0104] Repeat the clustering process and use evaluation indicators such as the silhouette coefficient and the DBI index to select the optimal number of clusters K.

[0105] Finally, K clustering centers can be obtained, and each clustering center represents a typical noise characteristic.

[0106] Through clustering analysis in step S30, the main noise characteristics in the dataset can be effectively summarized, providing data support for the construction of the subsequent noise recognition model.

[0107] The specific implementation of step S30 is described by the following formula:

[0108] Collect the first area noise features of all N triangular areas at T moments to form a sample set:

[0109] X = {x1, x2,..., x N}

[0110] where xi i represents the feature vector of the i-th area.

[0111] Define a distance function d(x, y) to calculate the distance between two samples, and the Euclidean distance can be used:

[0112]

[0113] where D is the dimension of the feature vector.

[0114] Then use the K-Means algorithm to cluster the samples:

[0115] 1. Randomly select K samples as the initial cluster centers μ1, μ2,..., μ K

[0116] 2. For each sample x, calculate its distance from each cluster center and classify it into the nearest class:

[0117]

[0118] 3. For the i-th class, update its cluster center:

[0119]

[0120] 4. Repeat steps 2-3 until the cluster centers no longer change.

[0121] The above ci i represents the cluster to which the sample xi i belongs, and Ci i represents all samples in the i-th class;

[0122] Finally, obtain K clustering results and the corresponding cluster centers as typical noise patterns.

[0123] The specific implementation of step S40 is as follows:

[0124] 1. Collect multiple typical noise features output by step S30.

[0125] 2. For each typical noise feature, extract its time-frequency features, waveform features and other information to form a feature vector.

[0126] 3. Collect all possible noise sources in the converter station, including equipment such as transformers, switching devices, reactors, etc., to form a noise source data set.

[0127] 4. For each noise source device, pre-record its operating noise and extract feature vectors to construct a feature database.

[0128] 5. For each typical noise feature, find the most matching noise source feature vector in the feature database to determine the most likely noise source. The matching method can use the Euclidean distance to compare vector similarities.

[0129] 6. Take the most matching noise source as the noise source parameter corresponding to this typical noise feature.

[0130] 7. For noise features that cannot be accurately matched, experts conduct manual annotation to determine the corresponding noise source.

[0131] 8. Finally, obtain each typical noise feature and its most matching noise source parameter.

[0132] Step S40 labels the corresponding noise source information for each typical noise feature through the method of feature matching, laying a data foundation for training the noise recognition model.

[0133] The specific implementation of Step S40 is described by the following formula:

[0134] For the i-th typical noise pattern, extract its feature vector f i .

[0135] Collect all M possible noise source devices to obtain a set of feature vectors:

[0136] F = {f1 f2…, f M}

[0137] Define the distance function between vectors:

[0138]

[0139] where d is the Euclidean distance.

[0140] Then the most matching noise source device is:

[0141]

[0142] That is, the source device with the most similar feature vector to f i , and j represents the subscript of the most matching noise source.

[0143] The specific implementation of Step S50 is as follows:

[0144] 1. Collect the labeled typical noise feature data set output by Step S40.

[0145] 2. Select a machine learning model suitable for processing time series data, such as RNN, LSTM, etc., as the model framework for noise recognition.

[0146] 3. Define the model structure, with the input layer being the time-frequency features of the noise and the output layer being the noise source classification.

[0147] 4. Configure the training set, validation set, and test set of the dataset. Use the training set to train the model, the validation set for hyperparameter tuning, and the test set to evaluate the model performance.

[0148] 5. In the training process, adopt the training algorithm corresponding to the model, such as backpropagation, gradient descent method, etc., and optimize the model parameters through iteration.

[0149] 6. After the training is completed, evaluate the recognition accuracy, recall rate and other indicators of the model on the test set.

[0150] 7. If the performance indicators do not meet the requirements, adjust the model structure and training algorithm, and retrain the model.

[0151] 8. Obtain the finally trained noise recognition model, which can predict the corresponding noise source according to the noise characteristics.

[0152] Step S50 obtains a model that can identify noise sources according to noise characteristics through machine learning methods, laying a foundation for constructing a noise map.

[0153] The specific implementation manner of step S60 is as follows:

[0154] 1. Arrange multiple noise sensors in the converter station to obtain noise data at multiple times and positions within the station.

[0155] 2. Preprocess the collected original noise signals, including filtering, removing outliers, etc., to obtain clean noise data.

[0156] 3. Extract features from the preprocessed noise data to obtain vector information such as its time-frequency features and energy features.

[0157] 4. Organize the extracted noise features into a dataset in chronological order as the input to the model trained in step S50.

[0158] 5. Input the dataset into the noise recognition model in sequence for inference.

[0159] 6. The noise recognition model outputs the predicted noise source classification results for each noise data.

[0160] 7. Post-process the recognition results output by the model, including smoothing, denoising, etc., to obtain the final noise source parameters.

[0161] 8. Finally, the noise source parameters corresponding to the noise data at all times of the converter station are obtained.

[0162] Step S60 uses the trained model to intelligently identify the in-station noise, providing data support for constructing a dynamic noise map.

[0163] The specific implementation manner of step S70 is as follows:

[0164] 1. Collect the noise at each time output by step S60 and the corresponding noise source parameters.

[0165] 2. According to the noise source parameters, search for device metadata to obtain the coordinate position information of the device.

[0166] 3. Take out the noise source coordinates and parameters at each time in turn and perform visual drawing on the plane map.

[0167] 4. Referring to the method of coloring drone images, represent the noise characteristics of different first regions with different RGB colors. The darker the color, the greater the noise; and mark the noise source parameters on the floor plan of the converter station.

[0168] 5. Use image special effects to generate the diffusion effect of the noise, presenting the propagation process of the noise.

[0169] 6. Stitch the visualization images at multiple times to generate a dynamic noise change video.

[0170] 7. Add a time watermark to the video to show the change process of the noise over time.

[0171] 8. Finally, an intuitive and dynamic noise distribution map of the converter station is obtained.

[0172] S70 vividly displays the recognition results through data visualization means, intuitively presenting the spatial distribution and temporal variation of the noise on the map, and achieving the goal of the entire method.

[0173] The specific implementation manner of step S70 is described by the following formula:

[0174] For the noise set at time t:

[0175]

[0176] where, represents the spatial coordinate of the i-th noise, which has been obtained through step S60; si represents the normalized value of the noise characteristics of the corresponding first region, obtained by using the cosine normalization method.

[0177] Define a mapping function to map the intensity to the RGB color space:

[0178] c = map(s)

[0179] Among them, map can be a linear or non-linear mapping function.

[0180] Then the noise distribution image at time t is as follows:

[0181] For each sample

[0182] At the coordinate Plot the point with color c i = map(s i )

[0183] Repeat the visualization drawing and presentation for all noise samples at time t, and finally output the noise dynamic distribution map.

[0184] The following is the second embodiment of the present invention. The specific variables used in this embodiment are independent of other embodiments:

[0185] In addition, in the second embodiment, the sound source localization method specifically includes the following steps:

[0186] Calculate the time differences detected by three noise sensors for the same noise, including the time difference Δt1 detected by the first noise sensor and the second noise sensor, and the time difference Δt2 detected by the first noise sensor and the third sensor;

[0187] Multiply the corresponding time differences by the speed of sound to obtain the corresponding distance differences Δd1 and Δd2;

[0188] Take Δd1 as the hyperbolic distance difference, and take the first noise sensor and the second noise sensor as the hyperbolic foci to draw a hyperbola;

[0189] Take Δd2 as the hyperbolic distance difference, and take the first noise sensor and the third noise sensor as the hyperbolic foci to draw a hyperbola;

[0190] Take the intersection point of the two hyperbolas as the position corresponding to the noise.

[0191] In the second embodiment, the actual decibel level is defined as the decibel level measured at a reference distance dS from the sound source.

[0192] In the second embodiment, the reference distance is set to 1 meter.

[0193] In the second embodiment, the calculation formula for the actual decibel level LA is:

[0194]

[0195] Among them,

[0196] LT is the decibel value detected by the noise sensor;

[0197] dS is the reference distance;

[0198] dA is the actual distance from the noise source to the noise sensor;

[0199] log represents logarithm.

[0200] In the second embodiment, the actual decibel value of the noise occurring within a triangular region is obtained by averaging the three actual decibel values calculated correspondingly by three noise sensors at the vertices of the triangular region.

[0201] In the second embodiment, the process of clustering the noise data includes:

[0202] In the second embodiment, the process of clustering the noise data includes:

[0203] Extracting features from each noise sample; the features include time-domain features, frequency-domain features, and Mel-frequency cepstral coefficients;

[0204] Performing standardization and normalization operations on all features;

[0205] Selecting a clustering algorithm; K-means, hierarchical clustering, DBSCAN, etc. are all acceptable;

[0206] Selecting the number of clusters; this can be determined by methods such as the elbow method, silhouette coefficient, etc.;

[0207] Clustering the extracted features using the selected clustering algorithm and parameters.

[0208] Figure 3 is a schematic flowchart of using the DBSCAN clustering algorithm in the method of the present invention. The present invention preferably uses DBSCAN as the clustering algorithm, including the following steps:

[0209] A1: As Figure 3 shown, querying the neighborhood. For each point in the dataset, the algorithm calculates the number of points within its eps-neighborhood. Each point needs to consider the matching attributes within a certain radius around it to determine the attribute of this point.

[0210] A2: Selecting core points. If the number of points within the eps-neighborhood of a point exceeds minPts, mark it as a core point.

[0211] A3: Constructing the neighborhood chain. For each core point, connect all the points (including other core points) within its eps-neighborhood to form a cluster.

[0212] A4: Attribution of border points. Assign the border points to the cluster of the core point they are connected to.

[0213] A5 marks noise. Finally, the points that are not assigned to any cluster are marked as noise.

[0214] Figure 4 It is a schematic diagram for determining the optimal number of clusters K using the elbow method in the method of the present invention. The present invention preferably uses the elbow method and the silhouette coefficient as the search algorithms for the optimal number of clusters K, including the following steps:

[0215] B1: Solve the sum of the distortion degrees of each class. The distortion degree of each class is equal to the sum of the squares of the distances between the centroid of the class and the positions of its internal members. Divide n samples into K classes, denote the k-th class as Ck (k = 1, 2,... K), and denote the central position of this class as uk. Define the total distortion degree of all classes:

[0216]

[0217] Among them, J is called the aggregation coefficient.

[0218] B2: The distortion degree will decrease as the number of classes increases. However, for data with a certain degree of discrimination, at a certain critical point, the distortion degree will be greatly improved and then decrease slowly. This critical point can be considered as a point with better clustering performance. Take the aggregation coefficient J obtained in B1 as the ordinate and the number of clustering classes K as the abscissa to generate a line graph, and find the critical point K.

[0219] B3: For a clustering task, in an ideal class cluster, the samples within the cluster are as close as possible, and the clusters are as far away from each other as possible. The silhouette coefficient is an evaluation index for the density and dispersion degree of a class, and the formula is expressed as follows

[0220] s = (b - a) / max(a, b)

[0221] a is the average of the distances between the samples in the cluster

[0222] b is the average of the distances from the representative sample to the samples in the nearest cluster other than its own cluster

[0223] The value of s ranges from [-1, 1].

[0224] If s is close to 1, it means that the cluster where the representative sample is located is reasonable. If s is close to -1, it means that s should be assigned to other clusters. Similarly, using the above indicators, train multiple models and compare to select the appropriate clustering class K.

[0225] B4: Compare the clustering classes K1 and K2, and select the one with the best actual classification effect as the final number of clusters K.

[0226] Figure 5It is a schematic diagram of the Few-Shot Learning implementation of the method of the present invention. In practical applications, there are situations where the number of samples is insufficient and the degree of sample variation is too large. The present invention constructs a converter station noise map using clustering and Few-Shot Learning machine learning, including the following steps:

[0227] C1: Increase the training data through prior knowledge and use data augmentation methods. High-order data augmentation methods use the information of other data sets to generate more samples of the target class, learn the within-sample and out-of-sample change distributions through encoder-decoder and generate new samples; obtain more samples of the target class based on weakly labeled samples or unlabeled samples, mainly including semi-supervised learning and active learning; train a GAN based on a small sample of the target class to directly generate more samples of the target class. It is applicable to scenarios where only small samples and unlabeled samples are difficult to collect.

[0228] C2: Narrow the space that the model needs to search. Use a neural network to learn multiple tasks simultaneously, so that a part of the hidden layer focuses on the general information across tasks and a part of the hidden layer focuses on the information of specific tasks. During the learning process, the model parameters of one task are constrained by other tasks, that is, there is a regularization-like effect between tasks. It is divided into parameter sharing methods that directly share the parameters of part of the hidden layer between different tasks and parameter typing methods that penalize the parameter differences of different tasks; project the samples into a low-dimensional space that is easier to distinguish different classes. A lower dimension and a feature representation that are easier to distinguish different classes mean that the space that the model needs to search is smaller.

[0229] The second aspect of the present invention provides a computer-readable storage medium, wherein program instructions are stored in the computer-readable storage medium, and when the program instructions run, they are used to execute the above-mentioned method for constructing a converter station noise map.

[0230] The third aspect of the present invention provides a system for constructing a converter station noise map, which includes the above-mentioned computer-readable storage medium.

Claims

1. A method for constructing a noise map of a converter station, characterized in that It includes the following steps: S10. Divide the converter station floor plan into a series of triangular regions; Obtain the noise signal sets collected by three noise sensors set at the three vertices of each triangular region, and after preprocessing the noise, obtain a preprocessed noise signal set; S20. For any triangular region, according to the preprocessed noise signal set corresponding to the triangular region, use the sound source localization method to determine the sound source position of the noise, and calculate the actual decibel level of the noise based on the sound source position and the detected decibel level, forming the first region noise characteristics and their corresponding noise source parameters. The first region noise characteristics include the noise decibel level and the time-frequency characteristics of the noise, and the noise source parameters include the sound source position, sound source equipment, and operating state of the sound source equipment of the noise; S30. Cluster the first region noise characteristics at each detection moment of all triangular regions of the entire converter station to obtain multiple cluster centers, representing multiple typical noise characteristics; S40. Mark the noise source parameters for each typical noise characteristic to obtain the most matching noise source parameters; S50. Use the multiple typical noise characteristics and their most matching noise source parameters as training data to train a machine learning model to obtain a noise recognition model for determining the noise source parameters according to the noise characteristics; S60. Obtain all the noises detected at consecutive detection moments of the converter station, input them into the noise recognition model for recognition, and obtain the corresponding noise source parameters; S70. Map the first region noise characteristics corresponding to all the noises detected at consecutive detection moments of the converter station to RGB colors, and mark the noise source parameters on the converter station floor plan to obtain a converter station noise distribution map that changes with time.

2. The method for constructing a noise map of a converter station according to claim 1, wherein It also includes step S25, which is used to be executed after step S20 is completed. Specifically: optimize the triangular region division of the converter station floor plan based on the obtained preprocessed noise signal set or the first region noise characteristics and their corresponding noise source parameters, and after repeating steps S10 and S20, continue to execute the remaining steps.

3. The method for constructing a noise map of a converter station according to claim 1, characterized in that, The step of dividing the converter station floor plan into a series of triangular regions is specifically: first, obtain the outline of the converter station according to the blueprint or floor plan of the converter station; then use the Delaunay triangulation algorithm to triangulate the outline of the converter station to divide the converter station into multiple triangular regions.

4. A method for constructing a noise map of a converter station according to claim 1, characterized in that The step of using the sound source localization method to determine the sound source position of the noise is specifically: for each triangular region, extract the preprocessed noise signals of the corresponding three sensors to form the preprocessed noise signal set of the triangular region, and then use the advanced time difference sound source localization algorithm to locate the spatial coordinates of the noise, that is, the sound source position of the noise, according to the signal differences of the three sensors.

5. A method for constructing a noise map of a converter station according to claim 1, characterized in that In the step of clustering the first regional noise features at each detection moment of all triangular regions of the entire converter station, the algorithm used for clustering is the K-Means algorithm. The specific clustering steps are as follows: First, randomly select multiple cluster centers, and then iteratively assign each first regional noise feature to the cluster center that is closest to it and update the cluster centers until the centers no longer change; finally, obtain multiple cluster centers to represent multiple typical noise features.

6. The method for constructing a noise map of a converter station according to claim 1, wherein The machine learning model uses LSTM as the framework of the noise recognition model.

7. A method for constructing a noise map of a converter station according to claim 2, characterized in that The step S25 specifically includes: First, set a segmentation threshold according to the standard deviation of the noise intensity of each triangular region. The larger the standard deviation, the more uneven the noise distribution, and further segmentation is required; Then, for the triangular regions with a standard deviation greater than the threshold, find the two sides with the largest difference in noise intensity, and divide the triangular region along the midpoint connection line of the two sides with the largest difference; Repeat the above segmentation process until the standard deviation of the noise intensity of all triangular regions is less than the threshold, indicating that the noise in the region is uniform.

8. A method for constructing a noise map of a converter station according to claim 7, characterized in that, The segmentation threshold is set to 0.618 times the standard deviation of the noise intensity of each triangular region.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which are used to execute a method for constructing a noise map of a converter station according to any one of claims 1-8 when the program instructions run.

10. A converter station noise map construction system, characterized in that, It includes the computer-readable storage medium described in claim 9.

Images