A channel noise identification method and system

Abstract

The present invention discloses a channel noise identification method and system, and relates to the field of noise detection technology. The method comprises: collecting typical noise, and randomly combining the typical noise to construct a typical noise time domain image library; obtaining a typical noise time domain image from the typical noise time domain image library, and performing preprocessing to obtain a noise time domain image with interference signals removed; extracting edge features based on the noise time domain image with interference signals removed; using an improved HOG to extract feature descriptors for the edge features to train a classifier; obtaining the field noise of the power line channel, extracting a time domain image for the field noise, and using the trained classifier to identify the noise type. The second-order Laplacian edge detection operator can accurately locate the edge contour, and combined with the improved HOG to extract rich feature vectors, thereby improving the accuracy of noise classification.

Description

Technical Field

[0001] The present invention relates to the technical field of noise detection, and in particular to a channel noise identification method and system. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Due to electromagnetic radiation from devices on the power line, external radio signal interference, and emergencies within the power system, there are noises such as colored background noise, narrowband noise, and impulse noise in the power line carrier channel. Existing technologies usually use machine learning to convert complex noise signals into intuitive image forms and perform noise recognition based on images. However, the current method of combining machine learning with noise classification has certain limitations.

[0004] Existing technologies such as the invention patent with application number 2022112731267, entitled A Power Line Carrier Channel Noise Detection Method and System, propose a noise recognition method using machine learning. It uses the Sobel operator to locate the edge of the noisy time domain image, but the response of the Sobel operator in the edge direction is not optimal, especially when the edge direction does not completely match the operator direction, which may lead to a decrease in the accuracy of edge detection. At the same time, it uses traditional HOG to ensure the effective extraction of texture features. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a channel noise recognition method and system, which can accurately locate the edge contour by adopting a second-order Laplacian edge detection operator, and extract rich feature vectors in combination with an improved HOG, thereby improving the accuracy of noise recognition.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a channel noise identification method, comprising:

[0008] Collect typical noises, randomly combine them, and build a typical noise time domain image library;

[0009] Acquire a typical noise time domain image from a typical noise time domain image library, perform preprocessing, and obtain a noise time domain image with interference signals removed;

[0010] Extract edge features based on noisy time domain images with interference signals removed;

[0011] Using an improved HOG extraction feature descriptor for the edge features to train a classifier;

[0012] The on-site noise of the power line channel is obtained, and the time domain image of the on-site noise is extracted, and the trained classifier is used to identify the noise type.

[0013] Preferably, the typical noise includes colored background noise, narrowband noise and impulse noise.

[0014] Preferably, the obtaining of a typical noise time domain image from a typical noise time domain image library and preprocessing to obtain a noise time domain image with interference signals removed specifically includes:

[0015] After graying the typical noise time domain image, wavelet decomposition is performed to obtain high-frequency coefficients and low-frequency coefficients;

[0016] Threshold processing is performed on the low-frequency coefficients to filter out the noise component;

[0017] The high-frequency coefficients and the low-frequency coefficients after threshold processing are reconstructed by wavelet to obtain the noisy time domain image with the interference signal removed.

[0018] Preferably, a second-order Laplacian edge detection operator is used to extract image edge features to define edge contours.

[0019] Preferably, the edge features are extracted using an improved HOG feature descriptor, specifically comprising:

[0020] Divide the edge features into cell units, and combine every 2×2 cell units into a block unit; collect the gradient of each pixel point in the cell unit or the direction histogram of the edge;

[0021] Performing multi-scale scaling on the histogram to obtain multiple gradient direction histograms of different scales, each gradient direction histogram generating 9 integral images;

[0022] The sliding detection window performs sliding detection on the integral image; specifically, the block unit is slid in the detection window, the gradient amplitude of each pixel is used as the weight, and the gradient direction histogram of the cell unit is counted by retrieving the integral image to form a set of 9-dimensional feature vectors;

[0023] The eigenvectors corresponding to all block units are normalized, and the normalized eigenvectors are cascaded to form a HOG feature vector, which constitutes a feature descriptor.

[0024] Preferably, the method further comprises adopting PCA to reduce the dimension of the feature descriptor.

[0025] Preferably, the classifier comprises constructing an SVM binary classifier by using a one-to-one method between any two types of samples to perform classification with multiple noises.

[0026] In a second aspect, the present invention provides a channel noise identification system, comprising:

[0027] Typical noise modeling module, used to collect typical noises, randomly combine typical noises, and build a typical noise time domain image library;

[0028] A typical noise processing module is used to obtain a typical noise time domain image from a typical noise time domain image library, perform preprocessing, and obtain a noise time domain image with interference signals removed;

[0029] A feature extraction module, used for extracting edge features based on the noisy time domain image with interference signals removed;

[0030] A training module, used for extracting feature descriptors using improved HOG for the edge features to train the classifier;

[0031] The noise detection module is used to obtain the on-site noise of the power line channel, extract the time domain image of the on-site noise, and use the trained classifier to identify the noise type.

[0032] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the channel noise identification method described in the first aspect.

[0033] In a fourth aspect, the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the channel noise identification method described in the first aspect are implemented.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention proposes a channel noise identification method and system, models typical noise, constructs a machine learning network model based on an improved HOG algorithm, uses the improved HOG algorithm to extract time domain images, obtains multi-scale feature information, forms feature descriptors, and then uses the SVM algorithm to perform noise classification prediction. Among them, the present invention uses a second-order Laplacian edge detection operator to identify edge contours. Compared with first-order derivative operators (such as Sobel and Prewitt), the Laplacian operator is more sensitive to details and textures, and can highlight edge areas more clearly, laying the foundation for the later use of the improved HOG algorithm to extract multi-scale features. The combination of the second-order Laplacian edge detection operator and the improved HOG algorithm effectively improves the accuracy of edge feature extraction of noisy time domain images.

[0036] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The exemplary embodiments of the present disclosure and the description thereof are used to explain the present disclosure but do not constitute a limitation of the present disclosure.

[0038] Figure 1 A schematic diagram of a channel noise identification method provided by an embodiment of the present invention;

[0039] Figure 2 A schematic diagram of colored background noise provided by an embodiment of the present invention;

[0040] Figure 3 A schematic diagram of narrowband noise provided by an embodiment of the present invention;

[0041] Figure 4 A schematic diagram of impulse noise provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0043] Embodiment 1

[0044] like Figure 1 As shown, this embodiment discloses a channel noise identification method, including:

[0045] S1: Collect typical noises, randomly combine them, and build a typical noise time domain image library;

[0046] S2: Obtain a typical noise time domain image from a typical noise time domain image library, perform preprocessing, and obtain a noise time domain image with interference signals removed;

[0047] S3: Extract edge features based on the noisy time domain image with interference signals removed;

[0048] S4: extracting feature descriptors using improved HOG for the edge features to train a classifier;

[0049] S5: Acquire the on-site noise of the power line channel, extract the time domain image of the on-site noise, and use the trained classifier to identify the noise type.

[0050] In this embodiment, firstly, a typical noise is modeled, and a single noise and a mixed noise after the single noise is superimposed are obtained by randomly combining the typical noise, and the noise time domain is extracted to establish a typical noise time domain image library;

[0051] The typical noises include colored background noise, narrowband noise and impulse noise; Figure 2-Figure 4 As shown;

[0052] The colored background noise is generally generated by low-power noise sources in the power line, such as thermal noise and harmonics generated by small motors. The background noise does not change significantly over time in the time domain, but its spectrum is very wide, occupies the entire signal bandwidth, and persists.

[0053] The power spectral density of colored background noise is low, distributed over the entire frequency band, slightly decreases with increasing frequency, and has a larger value near low frequencies. In general, this type of noise is closest to Gaussian white noise, especially when a large number of statistically independent colored background noises are superimposed on the power line, according to the central limit theorem, its overall characteristics will be more Gaussian.

[0054] After Gaussian white noise is processed by a shaping filter, colored background noise can be obtained. The transfer function of the shaping filter is shown in formula (1):

[0055] (1)

[0056] in, is the filter coefficient, n1 is the order, and Z is a complex variable with real and imaginary parts, often expressed in polar coordinates, i.e. Z= , r is the amplitude and Ω is the phase angle.

[0057] The narrowband noise has a narrow bandwidth. Generally, this type of noise is caused by crosstalk in the power line channel caused by radio signals in shortwave and medium wave broadcasting. This interference is ubiquitous and lasts for a long time. Narrowband noise is similar to a sinusoidal signal and can be regarded as the superposition of several independent sinusoidal signals, similar to an AM broadcast signal. The amplitude of narrowband noise changes little over time. This type of noise band is very narrow, and generally there will be multiple such narrowbands within the signal bandwidth.

[0058] Narrowband noise can be regarded as the superposition of several independent sinusoidal signals. The theoretical model is shown in equation (2):

[0059] (2)

[0060] in, represents the amplitude of a sinusoidal signal, is the frequency, represents the phase, t is the time information in the sinusoidal signal time series, and n2 is the number of medium and short waves.

[0061] The random pulse noise appears randomly and has no periodicity. It is usually generated by the electrical equipment of the power line network at the moment of switching, or by the instantaneous interference of lightning, so the duration is relatively short, at the microsecond level. The pulse noise is in the form of an attenuated sine wave or a superposition of attenuated sine waves. The form of a single pulse is shown in formula (3):

[0062] (3)

[0063] in, represents the amplitude of a pulse, is the frequency, Indicates the phase, is the time constant, n3 is the number of pulses, and t is the time information in the sinusoidal signal time series.

[0064] In this embodiment, the theoretical formulas of the above-mentioned various typical noises are used for modeling, and a single typical noise is randomly generated by uniform distribution. The generated typical noise is controlled within 0.02 seconds, wherein the colored background noise is generated by Gaussian white noise through a shaping filter with a coefficient of (1 -0.3889 -0.2413 0.0296 0.0276); the impulse noise is the superposition of multiple attenuated sine waves, and the impulse noise is generated from the switching circuit. The range of the on pulse is 1.9 to 11.2 microseconds, and the range of the off pulse is 15 to 150 microseconds. The number of attenuated sine waves is controlled by setting the number of switches to simulate the number of switching devices; the narrowband noise is mainly the interference generated by the radio signal in the broadcast. For the on-site environment, 100 shortwave signals are uniformly selected from 3MHz to 23MHz, and 5 medium wave signals are selected from 0.531MHz to 1.173MHz. The narrowband noise is generated by the superposition of attenuated sine waves.

[0065] In this embodiment, because the noise in the power line is mostly additive noise, a single typical noise is randomly combined, and the single typical noise generated by MATLAB is used to perform corresponding superposition of time domain points;

[0066] As an optional implementation method, the random combination of typical noises can have seven combinations: pulse noise, background noise, narrowband noise, background noise superimposed on pulse noise, background noise superimposed on narrowband noise, pulse noise superimposed on narrowband noise, and simultaneous superposition of three typical noises, thereby obtaining a combination of multiple typical noises and establishing a typical noise time domain image library.

[0067] In this embodiment, classification detection of power line channel noise is performed based on machine learning, and the machine learning mainly uses a method of combining HOG (Histogram of Oriented Gradients) + SVM (Support Vector Machines).

[0068] Specifically:

[0069] (1) The noise image in the typical noise database contains three color channels, RGB, while the features that truly reflect the noise information have nothing to do with color. If the original RGB image is used for processing, the redundant information will be greatly increased, and the complexity of the calculation will also increase. Therefore, this embodiment first performs a grayscale operation on the typical noise time domain image, converting the RGB image into a grayscale image operation, which is beneficial to further processing, making the image simpler and reducing the amount of data.

[0070] (2) The main difference between different noisy time domain images is the edge of the noise signal. The edge of an image is the most basic feature of an image. The edge refers to the discontinuity of the local characteristics of an image. The general method of extracting the edge of an image is to use the edge operator of the first-order derivative as the kernel to perform convolution and operation with each pixel of the image, and then select a suitable threshold to extract the edge of the image. Currently, the Sobel operator is often used, but the Sobel operator is a first-order edge detection operator. Its detection results have problems such as weak edge detection directionality and wide extracted edges, which will affect the accuracy of edge extraction.

[0071] This embodiment uses a second-order Laplacian edge detection operator to extract image edges to define edge contours.

[0072] The second-order Laplacian edge detection operator is an edge detection operator based on the second-order derivative, which is rotationally invariant. It uses the zero-crossing point between the positive and negative peaks of the second-order differential to determine the details of steep and slowly changing edges, which is very suitable for accurately locating edges in images.

[0073] The principle of edge detection by the second-order Laplacian edge detection operator is that when the grayscale of the central pixel in the neighborhood is higher than that of other pixels, the grayscale of the central pixel is increased; when the grayscale of the central pixel in the neighborhood is lower than that of other pixels, the grayscale of the central pixel is reduced.

[0074] The result of the convolution operation is processed with an appropriate attenuation factor and added to the original center pixel to achieve image sharpening.

[0075] Since the second-order Laplacian edge detection operator is very sensitive to noise, the image needs to be processed for noise before edge detection.

[0076] This embodiment uses wavelet decomposition to process noise, specifically:

[0077] Get the grayscale image, perform wavelet decomposition, and obtain high-frequency coefficients and low-frequency coefficients;

[0078] Threshold processing is performed on the low-frequency coefficients to filter out the noise component;

[0079] The high-frequency coefficients and the low-frequency coefficients after threshold processing are reconstructed by wavelet to obtain a denoised grayscale image.

[0080] In order to further refine the edge containing useful information, this embodiment adopts NMS processing to retain the point with the largest local amplitude change.

[0081] The edges of the grayscale image after NMS processing will be clearer, and then the adaptive threshold is used to remove most of the invalid edges to obtain a binary image. Finally, the edge features in the image are located and extracted.

[0082] Specifically, this embodiment adopts an adaptive threshold method based on the median filter concept. For an image of size m×n, a 3x3 window is selected to slide across the entire image. After sorting the window pixels by row, the minimum value, the middle value, and the maximum value are obtained in turn, and the average value of the above three data is used as the threshold. Assume For pixels in the window, the adaptive threshold t(x, y) at the target point (x, y) can be expressed as:

[0083] t(x,y)=[min( )+med( )+max( ) / 3]

[0084] Among them, min( )、med( ) and max( ) represent the minimum, middle and maximum values ​​of the window pixels respectively. Finally, the edge binary image B(x,y) corresponding to the original image is obtained as follows:

[0085] B(x,y)=

[0086] After the above-mentioned edge detection, non-maximum suppression refinement and adaptive threshold processing, the grayscale image can accurately locate the edge position.

[0087] (3) When the HOG algorithm is used to extract feature descriptors, the features of the entire image are extracted. However, when modeling typical noise, the edges of the generated typical noise time domain image will have identification information such as coordinate axes that interfere with the accuracy of recognition and classification.

[0088] Therefore, this embodiment performs truncation correction on the delineated edge contour, retains the useful pixels that need to be identified, reduces the interference of identification information such as the coordinate axis generated by modeling, and improves the recognition accuracy;

[0089] Specifically, the cropping position is set according to the image size, the edge contour is cropped and corrected according to the cropping position, and the coordinate axis of the edge portion is removed to obtain more effective information.

[0090] (4) The HOG algorithm is used on the processed typical noise time domain image to calculate and count the directional gradient histogram of the local area to form a feature descriptor;

[0091] Specifically, the processed typical noise time domain image is divided into small connected areas, which are called cell units, and each 2×2 cell units are combined into a block unit; the gradient or edge direction histogram of each pixel point in the cell unit is collected;

[0092] Performing multi-scale scaling on the histogram to obtain multiple gradient direction histograms of different scales, each gradient direction histogram generating 9 integral images;

[0093] The sliding detection window performs sliding detection on the integral image; specifically, the block unit is slid in the detection window, the gradient amplitude of each pixel is used as the weight, and the gradient direction histogram of the cell unit is counted by retrieving the integral image to form a set of 9-dimensional feature vectors;

[0094] The eigenvectors corresponding to all block units are normalized, and the normalized eigenvectors are cascaded to form a HOG feature vector, which constitutes a feature descriptor.

[0095] This embodiment performs multi-scale scaling on the gradient direction histogram and generates multiple gradient direction histograms of different scales, so that the algorithm can capture the feature information of the image at different scales. This multi-scale feature extraction method helps to improve the algorithm's ability to detect features of different sizes in the image, and enhances the algorithm's adaptability and generalization ability.

[0096] (5) Since the amount of information data extracted by the HOG algorithm is relatively large and contains a lot of redundant information, and the HOG feature matrix is ​​a typical sparse matrix, this embodiment uses PCA to reduce the dimension, thereby reducing the data dimension.

[0097] In this embodiment, SVM is used for classification; the SVM algorithm is a supervised learning model and related learning algorithm for analyzing data in classification and regression analysis. SVM is a binary classification model that maps the feature vector of an instance to some points in space. The purpose of SVM is to draw a line to "best" distinguish the two types of points, so that if there are new points in the future, this line can also make a good classification.

[0098] This embodiment designs a multi-class SVM classifier, and constructs an SVM binary classifier by using a one-vs-one method between any two types of samples, thereby achieving classification of multiple types of noise.

[0099] This embodiment integrates the improved HOG and SVM algorithms to construct a power line channel noise detection system and method. The second-order Laplacian operator is used to accurately identify edge contours, enhance details and texture responses, and lay the foundation for improving the HOG algorithm to extract multi-scale features. The algorithm combines SVM classification to achieve efficient analysis of noise time domain images and improve noise classification accuracy. Combining power line communication and AI technology, the measured and theoretical noise are matched to optimize system performance, ensuring more accurate on-site noise judgment and communication system optimization.

[0100] Embodiment 2

[0101] This embodiment provides a channel noise identification system, including:

[0102] Typical noise modeling module, used to collect typical noises, randomly combine typical noises, and build a typical noise time domain image library;

[0103] A typical noise processing module is used to obtain a typical noise time domain image from a typical noise time domain image library, perform preprocessing, and obtain a noise time domain image with interference signals removed;

[0104] A feature extraction module, used for extracting edge features based on the noisy time domain image with interference signals removed;

[0105] A training module, used for extracting feature descriptors using improved HOG for the edge features to train the classifier;

[0106] The noise detection module is used to obtain the on-site noise of the power line channel, extract the time domain image of the on-site noise, and use the trained classifier to identify the noise type.

[0107] Embodiment 3

[0108] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the channel noise identification method described in the first embodiment are implemented.

[0109] Embodiment 4

[0110] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the channel noise identification method described in the first embodiment are implemented.

[0111] The steps or modules involved in the above embodiments 2 to 4 correspond to those in embodiment 1. For the specific implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A channel noise identification method, characterized in that: include: Collect typical noises, randomly combine them, and build a typical noise time domain image library; A typical noise time domain image is obtained from a typical noise time domain image library, and preprocessed to obtain a noise time domain image with interference signals removed, specifically including: graying the typical noise time domain image, and then performing wavelet decomposition to obtain high-frequency coefficients and low-frequency coefficients; thresholding the low-frequency coefficients to filter out noise components; and wavelet reconstructing the high-frequency coefficients and the low-frequency coefficients after thresholding to obtain a noise time domain image with interference signals removed; The edge features are extracted based on the noisy time domain image with the interference signal removed; specifically, the second-order Laplacian edge detection operator is used to extract the edge features of the image to delineate the edge contour; in order to further refine the edges containing useful information, NMS processing is used to retain the points with the largest local amplitude changes; the binary image is obtained by further removing most of the invalid edges through adaptive thresholding, so as to locate and extract the edge features in the image; specifically, an adaptive thresholding method based on the median filtering idea is used to select a 3x3 window for an image of size m×n and slide it over the entire image; after sorting the window pixels by row, the minimum value, the middle value and the maximum value are obtained in turn, and then the average value of the above three data is used as the threshold; assuming For pixels in the window, the adaptive threshold t(x, y) at the target point (x, y) can be expressed as: t(x,y)=[min( )+with( )+max( ) / 3] Among them, min( )、med( ) and max( ) represent the minimum, middle and maximum values ​​of the window pixels respectively; the edge binary image B(x, y) corresponding to the original image is finally obtained as follows: B(x,y)= After the grayscale image has been processed by the above-mentioned edge detection, non-maximum suppression refinement and adaptive thresholding, the edge position can be accurately located; The improved HOG feature descriptor is used to extract the edge feature to train the classifier; the improved HOG feature descriptor is used to extract the edge feature, specifically including: dividing the edge feature into cell units, and combining every 2×2 cell units into a block unit; collecting the gradient or edge direction histogram of each pixel in the cell unit; performing multi-scale scaling on the histogram to obtain multiple gradient direction histograms of different scales, and each gradient direction histogram generates 9 integral images; sliding the detection window to perform sliding detection on the integral image; specifically, sliding the block unit in the detection window, taking the gradient amplitude of each pixel as the weight, counting the gradient direction histogram of the cell unit by retrieving the integral image, and forming a set of 9-dimensional feature vectors; normalizing the eigenvectors corresponding to all block units, and cascading the normalized feature vectors to form a HOG feature vector to constitute a feature descriptor; The on-site noise of the power line channel is obtained, and the time domain image of the on-site noise is extracted, and the trained classifier is used to identify the noise type.

2. A channel noise identification method as claimed in claim 1, characterized in that: The typical noise includes colored background noise, narrowband noise and impulse noise.

3. A channel noise identification method as claimed in claim 1, characterized in that: It also includes using PCA to reduce the dimension of the feature descriptor.

4. A channel noise identification method as claimed in claim 1, characterized in that: The classifier includes constructing a SVM binary classifier by using a one-to-one method between any two types of samples to perform classification with multiple noises.

5. A channel noise identification system, executing a channel noise identification method as claimed in claim 1, characterized in that: include: Typical noise modeling module, used to collect typical noises, randomly combine typical noises, and build a typical noise time domain image library; A typical noise processing module is used to obtain a typical noise time domain image from a typical noise time domain image library, perform preprocessing, and obtain a noise time domain image with interference signals removed; A feature extraction module, used for extracting edge features based on the noisy time domain image with interference signals removed; A training module, used for extracting feature descriptors using improved HOG for the edge features to train the classifier; The noise detection module is used to obtain the on-site noise of the power line channel, extract the time domain image of the on-site noise, and use the trained classifier to identify the noise type.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of a channel noise identification method according to any one of claims 1 to 4 are implemented.

7. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the channel noise identification method according to any one of claims 1 to 4 are implemented.

Images