Machine learning based hplc channel noise identification method and system

Abstract

The present application relates to the technical field of low-voltage power line high-speed carrier communication, and particularly relates to a HPLC channel noise identification method and system based on machine learning, which comprises the following steps: collecting noise signals and preprocessing to obtain zero-mean noise sequences when the HPLC communication unit channel is idle; extracting local relative fluctuation features and adaptive discrete enhancement features based on a sliding window; constructing a clustering model containing background and pulse noise clusters; using an optimized distance measurement rule; iteratively updating the clustering center to convergence; and accurately distinguishing background noise and pulse noise according to the final clustering result. Through the local relative fluctuation features and adaptive discrete enhancement features, the present application eliminates channel interference and pulse self-masking problems, accurately distinguishes noise types, optimizes the distance measurement rule, introduces a gravitational enhancement coefficient to adaptively adjust the weight, effectively avoids clustering boundary deviation, improves recognition accuracy and accelerates convergence speed, and enhances the overall performance and stability of noise identification under complex channels.

Description

Technical Field

[0001] This invention relates to the field of high-speed carrier communication technology for low-voltage power lines. In particular, it relates to a machine learning-based HPLC channel noise identification method and system. Background Technology

[0002] High-speed low-voltage power line carrier communication (HPLC) is the core communication technology of smart grid power consumption information acquisition system. In practical applications, the low-voltage power line channel is not a dedicated communication channel. The random access, switching and operation of various household appliances, frequency converters, switching power supplies, industrial loads and other equipment in the distribution area will cause the line impedance to change drastically in real time with the load status. The channel exhibits complex characteristics of strong interference, high attenuation, time-varying and non-stationary. There are a lot of high-intensity, multi-type and randomly distributed noise interference in the channel.

[0003] HPLC channel noise mainly includes two categories: steady background noise and burst impulse noise. To improve communication reliability and data transmission accuracy, it is necessary to accurately identify the noise type and adopt targeted anti-interference and noise reduction strategies. Traditional noise identification methods mainly use the energy threshold decision method, which distinguishes between background noise and impulse noise by setting a fixed energy threshold. However, this method has obvious drawbacks: when the threshold is set too high, it is easy to miss small-amplitude, dense periodic impulse noise; when the threshold is set too low, large-amplitude fluctuating background noise is easily misidentified as impulse noise, resulting in low identification accuracy and poor applicability.

[0004] Existing technologies employ the K-means clustering algorithm for automatic noise pattern classification, improving the automation level to some extent. However, this algorithm defaults to using Euclidean distance as the sample similarity measure, making it only suitable for scenarios where the feature space is spherically distributed and the density of each category is uniform. In the actual environment of low-voltage power line HPLC, impulse noise is characterized by strong bursts, uneven energy amplitudes, and irregular temporal distribution, exhibiting a long-tailed distribution that is severely incompatible with the applicable conditions of Euclidean distance. Directly using Euclidean distance leads to cluster boundary shifts and cluster center drift, making it impossible to accurately distinguish between stable background noise and burst impulse noise. This results in low noise recognition accuracy and poor robustness, failing to meet the practical application requirements of high reliability, low bit error rate, and strong anti-interference in smart grid HPLC communication. Summary of the Invention

[0005] To address the problems of existing technologies being unable to adapt to the characteristics of sudden and uneven energy of impulse noise, which easily leads to cluster boundary shifts, inability to accurately distinguish between background noise and impulse noise, and poor classification results, this invention provides solutions in the following aspects.

[0006] In the first aspect, the HPLC channel noise identification method based on machine learning includes: acquiring noise signals in the idle state of the HPLC communication unit's channel; preprocessing the noise signals to obtain a zero-mean noise sequence; performing feature quantization on each sample point in the zero-mean noise sequence based on a sliding window mechanism to extract local relative fluctuation features and adaptive discrete enhancement features; wherein, the local relative fluctuation features are used to characterize the deviation of the instantaneous intensity of the sample point from the local background amplitude, and the adaptive discrete enhancement features are used to characterize the dispersion and non-uniformity of the data distribution within the window where the sample point is located; constructing a clustering model containing background noise clusters and impulse noise clusters, and constructing optimized distance metric rules based on the local relative fluctuation features and the adaptive discrete enhancement features; calculating the feature distance between each sample point and the center of each cluster based on the optimized distance metric rules, iteratively updating the cluster centers until convergence, and identifying background noise and impulse noise in the channel based on the converged clustering results.

[0007] Preferably, the step of extracting the local relative fluctuation features includes: Select any sample point at the current moment as the target sample point, calculate the ratio between the instantaneous absolute value of the target sample point and the average of the absolute values ​​of all sample points within the sliding window centered on the target sample point, and obtain the local relative fluctuation characteristics of the target sample point; If the local relative fluctuation feature falls within a preset reference fluctuation range, the target sample point is determined to belong to stationary background noise; if the local relative fluctuation feature exceeds the upper limit of the reference fluctuation range, the target sample point is determined to have the morphological characteristics of impulse noise.

[0008] Preferably, the steps for extracting adaptive discrete enhancement features include: Calculate the average absolute deviation of the local relative fluctuation characteristics of each sample point within the sliding window. The average absolute deviation is used to quantify the severity of signal fluctuations within the window. The average absolute deviation is weighted and fused with the local relative fluctuation features of the current sample point. This results in a significant increase in the value of the generated adaptive discrete enhancement feature when the data distribution dispersion within the window is large and the current sample point is the pulse center, thereby eliminating the self-masking effect of strong impulse noise in the feature extraction process.

[0009] Preferably, the optimized distance metric rule is as follows: When calculating the distance between a sample point and the center of an impulse noise cluster, a gravity enhancement coefficient determined by an adaptive discrete enhancement feature is introduced to perform a weighted correction on the distance from the sample point to the center of the impulse noise cluster. Among them, adaptive discrete enhancement features refer to statistical features that quantify the discreteness and fluctuation inhomogeneity of data distribution within the window where the noise sample is located and highlight the characteristics of impulse mutation. The number of clusters in the clustering model is set to two, corresponding to the background noise cluster and the impulse noise cluster respectively; the initial cluster center of the background noise cluster is set to 1, and the initial cluster center of the impulse noise cluster is set to the maximum value in the local relative fluctuation feature sequence. Calculate the distance from each sample point to the center of the background noise cluster and the distance from each sample point to the center of the impulse noise cluster, and use the distance to determine the category of the sample point.

[0010] Preferably, the gravity enhancement coefficient satisfies the following condition: The gravity enhancement coefficient is negatively correlated with the adaptive discrete enhancement feature, and is normalized based on the mean of all adaptive discrete enhancement features in the noise sequence. The gravity enhancement coefficient is used to adaptively adjust the distance weight according to the dispersion of the sample points when calculating the distance from the sample point to the center of the impulse noise cluster, so that the cluster center is iteratively updated in the direction of the high dispersion impulse sample.

[0011] Preferably, the distance between each sample point and the cluster center is calculated based on the optimized distance metric rule, and the sample points are divided into corresponding clusters based on minimizing the intra-cluster variance. The cluster centers of the background noise cluster and the impulse noise cluster are iteratively updated according to the divided sample points until the cluster centers no longer change. The clustering is then determined to be converged and the noise recognition result is output.

[0012] Preferably, the preprocessing of the noise signal includes: When the channel energy is detected to be lower than the preset signal decision threshold, the channel is determined to enter the idle state. After the channel enters the idle state, the gain amplifier is switched to high gain mode and the noise signal is continuously sampled at the preset sampling rate to obtain the original noise sequence. The original noise sequence is subjected to DC removal processing to eliminate circuit zero-point drift and obtain a zero-mean noise sequence.

[0013] Secondly, a machine learning-based HPLC channel noise identification system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned machine learning-based HPLC channel noise identification method is implemented.

[0014] The present invention has the following effects: 1. This invention extracts dual features of local relative fluctuation features and adaptive discrete enhancement features to eliminate the strong impulse self-masking effect and time-varying impedance interference, accurately characterize the differences in noise morphology, avoid cluster boundary shift, significantly improve the distinction accuracy between background noise and impulse noise, and solve the problem of poor classification effect in existing technologies.

[0015] 2. This invention optimizes the clustering distance metric rules and introduces a gravity enhancement coefficient associated with adaptive discrete enhancement features. It can dynamically adjust the distance weight according to the data dispersion, guide the cluster centers to move closer to high-dispersion pulse samples, and shorten the clustering iteration convergence time by combining reasonable initial cluster center settings, thereby improving the efficiency and robustness of noise recognition. Attached Figure Description

[0016] Figure 1 This is a flowchart of steps S1-S4 in the HPLC channel noise identification method based on machine learning according to an embodiment of the present invention.

[0017] Figure 2 This is a structural block diagram of the HPLC channel noise identification system based on machine learning, according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0019] Reference Figure 1 The HPLC channel noise identification method based on machine learning includes steps S1-S4, as detailed below: S1: Acquire noise signals in the idle state of the HPLC communication unit, preprocess the noise signals, and obtain a zero-mean noise sequence.

[0020] To avoid the communication signal masking and interference with noise acquisition and to ensure the purity and validity of the acquired noise data, the channel status is continuously monitored during the operation of the HPLC communication unit. When the channel is determined to be idle, the noise signal acquisition process is initiated to obtain the raw noise signal under the current channel environment. The acquired raw noise signal is preprocessed by removing the DC component from the signal and eliminating the zero-point drift error introduced by the ADC acquisition circuit and the front-end conditioning circuit, so that the noise signal fluctuates around zero level, thereby obtaining a stable, unbiased zero-mean noise sequence.

[0021] The steps for preprocessing noise signals include: The real-time energy level of the power line carrier communication channel is monitored, and the detected channel energy value is compared with a pre-set signal decision threshold. If the real-time channel energy is lower than the decision threshold, it is determined that there is no effective service signal transmission on the current communication channel, and the channel is in an idle state. After confirming that the channel has entered an idle state, in order to improve the sensitivity of weak noise signal acquisition, the gain amplifier in the front-end signal conditioning link is switched to high-gain operating mode to amplify low-amplitude noise signals. At the same time, the ambient noise in the channel is continuously sampled according to the system's preset fixed sampling rate, for example, the preset fixed sampling rate is set to 50MHz, to complete the time-domain signal acquisition and obtain the original noise sequence of the channel.

[0022] Because the signal acquisition circuit and the front-end analog conditioning circuit are prone to inherent zero-point drift deviation during operation, the original noise sequence will carry a fixed DC bias component, which will interfere with the accuracy of subsequent noise feature extraction. Therefore, the acquired original noise sequence is subjected to DC removal filtering to remove the fixed DC component superimposed in the sequence, completely eliminate the signal offset problem caused by the zero-point drift of the hardware circuit, and finally generate a zero-mean noise sequence with a stable waveform and no bias component.

[0023] Because the impedance of low-voltage power line carrier communication channels changes in real time with the load connection status, directly using the absolute amplitude of the noise signal cannot accurately reflect the true nature of the noise, easily leading to difficulty in distinguishing impulse noise from background noise. Simultaneously, strong impulse noise can inflate the window average during local statistical processes, masking its own characteristics and producing a self-masking effect, thus affecting noise identification accuracy. Therefore, it is necessary to perform dual quantization of zero-mean noise sequences using both relative and discrete features. The specific steps are as follows: S2: Based on the sliding window mechanism, feature quantization is performed on each sample point in the zero-mean noise sequence to extract local relative fluctuation features and adaptive discrete enhancement features. Among them, the local relative fluctuation features are used to characterize the degree of deviation of the instantaneous intensity of the sample point from the local background amplitude, and the adaptive discrete enhancement features are used to characterize the dispersion and non-uniformity of the data distribution within the window where the sample point is located.

[0024] Each sample point in the noise sequence is taken as the current target sample point, and a sliding window of fixed length is constructed with the target sample point as the center. The absolute value of the signal amplitude of the target sample point is calculated as the ratio of the absolute value of the amplitude of all sample points in the sliding window, and the ratio is used as the local relative fluctuation feature of the current target sample point.

[0025] By utilizing local relative fluctuation characteristics, noise signals can be converted from the absolute amplitude domain to the relative statistical characteristic domain, thereby eliminating the influence of time-varying impedance of the power line channel on noise amplitude and achieving decoupling between noise pattern and line impedance.

[0026] When the local relative fluctuation feature is in a stable range close to 1, it indicates that the amplitude of the current sample point is basically consistent with the local background amplitude, corresponding to stable background noise; when the local relative fluctuation feature is significantly greater than 1, it indicates that the amplitude of the current sample point is much higher than the local average level, exhibiting typical characteristics of impulse noise.

[0027] Specifically, the local relative fluctuation characteristics satisfy the following relationship: ; In the formula, Indicates the first Local relative fluctuation characteristics at each sample point Indicates the current time. The instantaneous intensity of the noise signal at each sample point. Represents a minimal constant. To prevent the denominator from being zero; Indicates noise signal at Local average absolute amplitude at that location Indicates the sliding window radius; sets the sliding window radius. ; Indicates the total length of the window; This represents the instantaneous intensity of each sample point within the sliding window; if If a point is located at the edge of a sequence, the local average absolute magnitude is calculated by padding with zeros.

[0028] To address the issues of time-varying channel impedance and the susceptibility of noise amplitude to load fluctuations in low-voltage power line carrier communication (HPLC), directly using the absolute amplitude of the noise signal for feature extraction would result in different amplitudes for the same noise source due to changes in line impedance, failing to accurately characterize the inherent form of the noise. Therefore, by constructing local relative fluctuation features using relative features instead of absolute amplitudes, the noise signal is converted from the absolute amplitude domain to the relative statistical domain, thereby decoupling the noise morphology characteristics from the line impedance and eliminating the interference of time-varying impedance on noise identification.

[0029] Each sample point with local relative fluctuation characteristics is used as the current target sample point. A sliding window of fixed length is constructed with the target sample point as the center. The mean absolute deviation of the local relative fluctuation characteristics of each sample point within the window is calculated, and the mean absolute deviation is weighted and fused with the local relative fluctuation characteristics of the current sample point. The fusion result is used as the adaptive discrete enhancement feature of the current target sample point.

[0030] Adaptive discrete enhancement features can effectively amplify the non-uniformity and dispersion of data distribution within a window, thereby eliminating the self-masking effect of strong impulse noise in local statistical processes and highlighting the abrupt change characteristics of impulse noise.

[0031] When the window contains only stable background noise, the data distribution is uniform and the dispersion is low, and the adaptive discrete enhancement feature value approaches 0; when the window contains impulse noise, the data distribution fluctuates violently and the dispersion increases significantly, and the feature value at the center of the impulse is greatly enhanced, exhibiting obvious impulse noise discrimination characteristics.

[0032] Specifically, the adaptive discrete enhancement features satisfy the following relationship: ; In the formula, Indicates the first Adaptive discrete enhancement features at each sample point Indicates the radius of the sliding window. Indicates the total length of the window. This represents the local relative fluctuation characteristic value of each sample point within the sliding window. Indicates the first The average value of the local relative fluctuation characteristics within a sliding window centered on a sample point. Indicates the first Local relative fluctuation characteristic values ​​at each sample point To calculate the average absolute deviation, the current window's range is quantified. The degree of fluctuation.

[0033] Adaptive discrete enhancement features are used to address the problem of strong impulse noise features being masked by local average values. By calculating the mean absolute deviation of local relative fluctuation features through a sliding window, the discreteness and fluctuation intensity of the data within the window are characterized. Combined with the weighting coefficient of the current sample point, the impulse position feature is enhanced, which can effectively highlight the abrupt change characteristics of impulse noise. The formula adopts the same window structure as the local relative fluctuation features to ensure uniform feature scale. The mean absolute deviation is more resistant to interference and more robust in value. The overall design fits the characteristics of HPLC noise distribution and has good rationality and practicality.

[0034] Local relative fluctuation features and adaptive discrete enhancement features are used together as input features for the K-means clustering algorithm to construct optimized distance metric rules, thereby achieving accurate clustering and differentiation between background noise and impulse noise.

[0035] Traditional clustering distance calculation methods rely solely on fixed formulas to classify samples, failing to adapt to the inherent distribution characteristics of noise and thus struggling to accommodate the random bursts and significant variations in the dispersion of power line channel noise. This can lead to biased impulse noise segmentation and low clustering convergence efficiency. Therefore, it is necessary to construct a clustering model that includes both background noise clusters and impulse noise clusters. The specific steps are as follows: S3: Construct a clustering model that includes background noise clusters and impulse noise clusters, and build an optimized distance metric rule based on local relative fluctuation features and adaptive discrete enhancement features.

[0036] To achieve adaptive adjustment of distance weights according to the discrete state of the data, the gravity enhancement coefficient is set to have a negative correlation with the adaptive discrete enhancement features. At the same time, the statistical mean of all adaptive discrete enhancement features in the entire noise sequence is used as a reference benchmark to complete the normalization process, unify the scale and range of coefficient values, and ensure that the coefficients have a unified comparison standard across the entire sequence.

[0037] When calculating the distance between sample points and the center of impulse noise clusters using optimized distance metric rules, the gravity enhancement coefficient can autonomously adjust the distance calculation weight according to the data dispersion of the corresponding region of the sample point. This effectively reduces the equivalent distance between high-dispersion impulse samples and the impulse cluster center, guiding the cluster center to gradually shift and converge towards the location of the high-dispersion impulse samples during the iteration process, thereby further improving the accuracy of impulse noise clustering.

[0038] Based on the inherent distribution characteristics of power line carrier channel noise, it can be seen that the noise in the channel is mainly divided into two types: stable background noise and transient impulse noise. Therefore, by setting the number of clusters in the clustering model to two types, the actual noise category can be accurately matched, the clustering operation process can be simplified, and the classification redundancy and discrimination error caused by redundant clustering categories can be avoided.

[0039] From the perspective of the numerical distribution of local relative fluctuation characteristics, the fluctuation amplitude of the stable background noise sample points is basically consistent with the local environment, and its characteristic value is stable and close to 1. Therefore, setting the initial cluster center of the background noise cluster to 1 can fit the overall distribution benchmark of the background noise. Impulse noise has the characteristics of instantaneous amplitude change and extremely strong relative fluctuation. Its local relative fluctuation characteristics are at the highest level in the whole sequence. Therefore, selecting the maximum value in the complete local relative fluctuation characteristic sequence as the initial cluster center of the impulse noise cluster can make the initial cluster center as close as possible to the typical characteristic distribution position of impulse noise, effectively shorten the convergence time of subsequent clustering iterations, and improve the rationality of cluster initialization.

[0040] S4: Based on the optimized distance metric rule, calculate the feature distance between each sample point and the center of each cluster, iteratively update the cluster center until convergence, and identify background noise and impulse noise in the channel based on the converged clustering results.

[0041] After completing the clustering model construction and initial cluster center configuration, the feature distance from each sample point to the center of the two clusters is calculated based on the optimized distance metric rule. The principle of minimizing intra-cluster variance is used as the division principle, and all sample points are divided into corresponding clusters according to the optimal distance matching principle, so as to ensure that the similarity of the same type of sample is high and the distinction of the different types of sample is obvious.

[0042] Specifically, the feature distance satisfies the following relationship: ; ; In the formula, Indicates the first The feature distance from each sample point to the center of the background noise cluster. Indicates the first Local relative fluctuation characteristics of individual sample points Represents the cluster centers of background noise clusters. Indicates the first The characteristic distance from each sample point to the center of the impulse noise cluster. This represents the mean of the adaptive discrete enhancement features for all sample points in the entire noise sequence. Indicates the first Adaptive discrete enhancement features for each sample point This represents the cluster center of the impulse noise cluster.

[0043] After completing a single round of sample classification, the cluster centers corresponding to the two types of noise are recalculated and updated using the classified sample data. The process of distance calculation, sample classification and cluster center update is continuously repeated until the positions of the cluster centers obtained in the two iterations no longer change. The clustering model is then considered to have converged. Finally, the background noise and impulse noise in the channel are distinguished and identified based on the converged clustering results, and the final noise identification result is output.

[0044] This invention also provides a machine learning-based HPLC channel noise identification system. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the machine learning-based HPLC channel noise identification method according to the first aspect of the present invention. The system also includes other components well-known to those skilled in the art, such as a communication bus and a communication interface, the setup and functions of which are known in the art and will not be described further here.

[0045] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A machine learning-based HPLC channel noise identification method, characterized in that, include: Noise signals were acquired during the idle state of the HPLC communication unit, and the noise signals were preprocessed to obtain a zero-mean noise sequence. Based on the sliding window mechanism, feature quantization is performed on each sample point in the zero-mean noise sequence to extract local relative fluctuation features and adaptive discrete enhancement features. The local relative fluctuation features are used to characterize the degree of deviation of the instantaneous intensity of the sample point from the local background amplitude, and the adaptive discrete enhancement features are used to characterize the dispersion and non-uniformity of the data distribution within the window where the sample point is located. A clustering model containing background noise clusters and impulse noise clusters is constructed, and an optimized distance metric rule is constructed based on the local relative fluctuation features and the adaptive discrete enhancement features; Based on the optimized distance metric rule, the characteristic distance between each sample point and the center of each cluster is calculated, and the cluster center is iteratively updated until convergence. The background noise and impulse noise in the channel are identified based on the converged clustering results.

2. The HPLC channel noise identification method based on machine learning according to claim 1, characterized in that, The steps for extracting the local relative fluctuation features include: Select any sample point at the current moment as the target sample point, calculate the ratio between the instantaneous absolute value of the target sample point and the average of the absolute values ​​of all sample points within the sliding window centered on the target sample point, and obtain the local relative fluctuation characteristics of the target sample point; If the local relative fluctuation feature falls within a preset reference fluctuation range, the target sample point is determined to belong to stationary background noise; if the local relative fluctuation feature exceeds the upper limit of the reference fluctuation range, the target sample point is determined to have the morphological characteristics of impulse noise.

3. The HPLC channel noise identification method based on machine learning according to claim 1, characterized in that, The steps for extracting adaptive discrete enhancement features include: Calculate the average absolute deviation of the local relative fluctuation characteristics of each sample point within the sliding window. The average absolute deviation is used to quantify the severity of signal fluctuations within the window. The average absolute deviation is weighted and fused with the local relative fluctuation features of the current sample point. This results in a significant increase in the value of the generated adaptive discrete enhancement feature when the data distribution dispersion within the window is large and the current sample point is the pulse center, thereby eliminating the self-masking effect of strong impulse noise in the feature extraction process.

4. The HPLC channel noise identification method based on machine learning according to claim 1, characterized in that, The optimized distance metric rule is specifically as follows: When calculating the distance between a sample point and the center of an impulse noise cluster, a gravity enhancement coefficient determined by an adaptive discrete enhancement feature is introduced to perform a weighted correction on the distance from the sample point to the center of the impulse noise cluster. Among them, adaptive discrete enhancement features refer to statistical features that quantify the discreteness and fluctuation inhomogeneity of data distribution within the window where the noise sample is located and highlight the characteristics of impulse mutation. The number of clusters in the clustering model is set to two, corresponding to the background noise cluster and the impulse noise cluster respectively; the initial cluster center of the background noise cluster is set to 1, and the initial cluster center of the impulse noise cluster is set to the maximum value in the local relative fluctuation feature sequence. Calculate the distance from each sample point to the center of the background noise cluster and the distance from each sample point to the center of the impulse noise cluster, and use the distance to determine the category of the sample point.

5. The HPLC channel noise identification method based on machine learning according to claim 4, characterized in that, The gravity enhancement coefficient satisfies the following condition: The gravity enhancement coefficient is negatively correlated with the adaptive discrete enhancement feature, and is normalized based on the mean of all adaptive discrete enhancement features in the zero-mean noise sequence. The gravity enhancement coefficient is used to adaptively adjust the distance weight according to the dispersion of the sample points when calculating the distance from the sample point to the center of the impulse noise cluster, so that the cluster center is iteratively updated in the direction of the high dispersion impulse sample.

6. The HPLC channel noise identification method based on machine learning according to claim 1, characterized in that, The distance between each sample point and the cluster center is calculated based on the optimized distance metric rule, and the sample points are assigned to the corresponding clusters based on minimizing the intra-cluster variance. Based on the divided sample points, the cluster centers of the background noise cluster and the impulse noise cluster are iteratively updated again until the cluster centers no longer change. The clustering is then determined to be converged and the noise recognition result is output.

7. The HPLC channel noise identification method based on machine learning according to claim 1, characterized in that, The preprocessing of the noise signal includes: When the channel energy is detected to be lower than the preset signal decision threshold, the channel is determined to enter the idle state. After the channel enters the idle state, the gain amplifier is switched to high gain mode and the noise signal is continuously sampled at the preset sampling rate to obtain the original noise sequence. The original noise sequence is subjected to DC removal processing to eliminate circuit zero-point drift and obtain a zero-mean noise sequence.

8. A machine learning-based HPLC channel noise identification system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the machine learning-based HPLC channel noise identification method according to any one of claims 1-7.

Images